text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
address:
- 'Department of Physics & Astronomy, University of Southampton, Highfield, Southampton SO17 1BJ, UK'
- 'INAF, Osservatorio Astronomico di Brera Merate, via E. Bianchi 46, I-23807Merate, Italy'
- 'INAF, Osservatorio Astronomico di Roma, Via Frascati 33, I-00078 Monteporzio Catone, Italy'
author:
- 'Federico M. Vincentelli'
- Piergiorgio Casella
bibliography:
- 'bib.bib'
title: 'Sub-second variability in black-hole X-ray binary jets'
---
Introduction {#sec1}
============
Black-hole low mass X-ray binaries (BH LMXRBs), are transient sources which are known to present emission along almost all the electromagnetic spectrum, from radio to hard X-rays. Systematic studies in X-rays showed that during their outburst these sources can display two main spectral state: a soft and a hard state. In the first case, the X-ray spectrum can be described by the black-body emission from an optically thick, geometrically thin accretion disk: in the second, the spectrum is dominated by a power-law up to $\approx$ 100 keV which has been explained in terms of Comptonization of the disk photons from a hot optically thin, geometrically thick inflow [@dgk07]. It has also been shown that during this state, BH LMXRBs display radio emission with an approximately flat spectrum which can extend up to optical-infrared (O-IR) frequencies [@corbel2002; @corbel2013; @russell_d2013]. Such feature has been successfully modelled with the superposition of the syncrothron emission profiles generated by the acceleration of the electrons in a relativistic jet . In order to reproduce the observed spectral energy distribution, such model assumes that the energy of the electrons is continuously replenished, so that the radiative and adiabatic energy losses are balanced. However, the origin of the process which provides such energy is still a matter of debate.
To solve this issue, in analogy to what is generally believed to happen in Gamma-Ray Bursts [@kobayashi1997] and in blazars [@spada2001], different models tried to apply the internal shocks scenario also to the jets in these sources. According to such scenario, shells of matter are launched into the jet with variable velocities. Due to differences in velocity they collide and merge. Application to XRBs showed that through the conversion of kinetic to internal energy in the shocks, this process can balance the energy losses and reproduce the observed flat spectrum at lower energies[@jamil2010]. Of course, the key prediction of such scenario is the presence of variable emission from the jet.
First evidences of strong variability from BHXRB jets were found in GRS 1915+105. Multiwavelength observations revealed the presence of spectacular infrared and radio flares on timescales of $\approx$ hours [@mirabel1998; @eikenberry1998; @fender1998; @fender2000; @rothstein2005; @lasso-cabrera2013]. These variations, however, were found to be due to relativistic ejections launched from the system, in correspondence to strong accretion episodes, and not to the presence of shocks in a compact jet. Nevertheless the observation of this behaviour increased the scientific attention towards the variability at lower energies, initiating the modern approach to this kind of studies.
In the early 2000s, new fast read-out detectors allowed to probe with sub-second time resolution the optical emission of these sources. The first simultaneous sub-second optical/X-ray observations revealed a complex connection between the emission in the two bands [@kanbach2001; @hynes2003; @hynes2006; @durant2008]. The optical response to the X-ray fluctuations could not be explained just with the emission from a compact jet [@malzac2004]. Such complexity derives from the presence of different processes which can take place in the optical band (e.g. synchrotron radiation from the hot inflow [@veledina2011; @veledina2013], or reprocessing from the outer disk [@obrien]).
A crucial step in the study of jets is represented by the detection of a 0.1 s lag optical-infrared (O-IR) emission in respect to the X-rays during the hard state of GX 339-4. Such measurement showed how the fluctuations from the inflow can be transferred into the compact jet in fractions of a second [@casella2010; @gandhi2008; @kalamkar2016]. This was then confirmed with the recent detection of a similar lag in V404 Cygni, which also allowed to put new constrains on the size of the jet base [@gandhi2017]. Moreover such discovery opened the possibility to study in detail the physical processes which take place inside the jet, through the analysis of the fast multiwavelength variability [@gandhi2010; @malzac2018; @vincentelli2018]. The aim of this paper is to review the latest developments in the study of the jet fast variability. Therefore, even though fast optical observations have been key to define the phenomenological scenario, we choose to focus only on the results found in the IR band, which are less affected by other components and allowed to reach a deeper insight on the jet’s internal processes.
Fast IR variability {#sec2}
===================
First results
-------------
The first unambiguous detection of fast variability from a compact jet was found in the IR in GX 339-4 during the so called low-hard state [@casella2010]. The fast infrared observation presented variability down to $\approx 200$ ms timescale, which, with brightness temperature arguments, excluded any possibility of emission from the inflow. Moreover also the measured 0.1 s lag in the cross-correlation function (CCF), presented a symmetric structure, inconsistent with being reprocessing from the outer disk.
Such result, together with the discovery in the same years of a similar lag also between the optical and the X-ray variability in GX 339-4 [@gandhi2008; @gandhi2010], had great relevance from the theoretical point of view. Evidence of fast variability from the base of the jet strongly suggested that internal shocks could effectively play a crucial role also in BH XRBs. This was confirmed by the model developed by [@malzac2013; @malzac2014]: by linking the velocities of the injected shells to the fluctuations measured in the X-ray luminosity, such model managed to reproduce not only the sub-second O-IR variability, but also the observed 0.1 s lag.
Further observations were then performed to investigate how the inflow/outflow connection could evolve during the outburst. In particular, this was done by [@kalamkar2016] who observed the same source during a higher luminosity state, near to the transition to the soft state (also known as high-intermediate state). This brought to the discovery of the first IR Quasi-periodic oscillation (QPO), which resulted also in harmonic relation with a simultaneous QPO in the X-rays. The CCF together with the usual 0.1 s lag, showed also evidence of some oscillations. While the fast variability associated to the 0.1 s lag was easily associated to fluctuations in the jet, the QPO could not be explained in terms only of a hot inflow precessing, and therefore was interpreted as the effect of a jet precessing together with the hot inflow.
Deeper analysis
---------------
Given the physical interpretation of these results, the described observations could also be used to investigate how the jet modifies the fluctuations which are injected from the inflow. In particular, this was done by [@vincentelli2018] and [@malzac2018] by applying Fourier domain techniques (i.e. cross-spectral analysis) to simultaneous X-ray/IR light curves. The two main quantities which can be measured are the the coherence and phase/time lags (For a review see [@uttley2014]). While the first gives a measure of the degree of linear correlation between two curves, the seconds estimate a possible lag as a function of the Fourier frequency. Fig. 1 shows these quantities for the two datasets analysed in GX 339-4.
{width="\textwidth"}
As expected from the shape of the CCF, in the low-hard state a simple constant time lag at 0.1 s was measured [@vincentelli2018]. The coherence, instead, showed a rather particular trend, decreasing gradually as a function of frequency. Such result indicated that the way fluctuations are transferred into the jet cannot be seen as a simple linear impulse response function (see [@nowak1999] and [@uttley2014]), suggesting the presence of some kind of non-linearity as the origin of the IR emission. However, given that a classical non-linear process would just give a very low coherence, the authors suggested that the smooth decrease of the coherence could be due to a non-stationarity of the system, which leads to a broad IR-emitting region in the jet. Such scenario would also be in agreement with the internal shock models, as the region where shock takes place, and therefore variability is emitted, is not stably localized (see also Fig. 4 in [@malzac2018]).
The cross-spectral analysis measurements done in the higher luminosity state (performed by [@malzac2018]) were found to be significantly different. In particular, the lags showed a complex and non-intuitive trend. While at high frequency the “usual” 0.1 s lag was clearly detected, at lower frequency the lags changed sign tending at $\approx$ -$\pi/2$. Moreover, a distinct component was also seen in correspondence of the two QPOs. Such complex lags were found to be associated to very low coherence, indicating the presence of a highly non-linear process. Despite the strong differences, the authors found that such peculiar trend could be still naturally explained by internal shocks: the origin of the negative lag was identified as a consequence of the Doppler boosting modulation. Such relativistic effect leads to an anti-correlation between X-ray and IR on longer timescales, generating a constant negative phase lag. Such effect is partially distorted, due the presence of two QPOs at 0.08 and 0.16 Hz, but still visible.
Discussion {#sec1}
==========
Recent application of cross-spectral analysis to simultaneous sub-second IR/X-ray observations of GX 339-4 permitted to shed light on the physical process which take place inside the jet. In particular, the observations seem to be in good agreement with the expectations from internal-shock models. Moreover thanks to the application of Fourier techniques it has been possible for the first time to appreciate in the IR variability effects due to the relativistic motion of matter accelerated in the jet.
The significant differences observed in the two states can also give indications on the behaviour of the jet. For example the appearance at higher luminosities of an anti-correlation on longer timescales, suggests some changes in the physical/geometrical conditions of the jet, which enhance the Doppler boosting effect. Given that the inclination of the source cannot change, a possible explanation for this behaviour could be the change in the average $\Gamma$ Lorentz factor which leads to a smaller viewing angle of the jet itself. This would then introduce an anti-correlation as the jet points away from our line of sight. It is interesting to notice that a recent application of this same model to multi-epoch observations of MAXI J1836-194 (Peault et al. in prep.) found similar results, requiring an increase of the average jet velocity during the evolution of the outburst.
Conclusions {#sec2}
===========
We gave a synthetic review of the latest results on fast variability from the jet in BH LMXRBs. While the first fast IR observations managed mainly to put strong constrains on the geometry of the jet, recent application of cross-spectral analysis techniques are permitting to probe the internal processes of the jet. Two observations performed at different stages of the outburst suggest that the jet is evolving together with the inflow properties. The main differences seem to be related to relativistic effects, which would then indicate an increase in the jet average velocity. Multiple observations along the same outburst are needed to put much stronger constrains on the physical scenario. The great effort put in the coordination between observing facilities is now facilitating the building of multi-wavelength campaigns of transient sources. New campaigns are indeed now managing to follow with several facilities the evolution of the multi-wavelength variability of these source, and will give in the near future a further significant contribution to the understanding of jets.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the energetics of vortices and vortex lattices produced by rotation in the cyclic phase of $F=2$ spinor Bose condensates. In addition to the familiar triangular lattice predicted by Tkachenko for $^4$He, many more complex lattices appear in this system as a result of the spin degree of freedom. In particular, we predict a magnetic-field-driven transition from a triangular lattice to a honeycomb lattice. Other transitions and lattice geometries are driven at constant field by changes in the temperature-dependent ratio of charge and spin stiffnesses, including a transition through an aperiodic vortex structure.'
author:
- 'Ryan Barnett,$^1$ Subroto Mukerjee,$^{2,3}$ and Joel E. Moore$^{2,3}$'
title: Vortex lattice transitions in cyclic spinor condensates
---
One of the many remarkable properties of superfluids is the appearance of vortex lattices in rotated systems [@tkachenko66; @yarmchuk79]. These lattices are periodic arrangements of vortices that allow the superflow outside the vortex cores to remain irrotational and are analogous to the mixed state of type-II superconductors in a magnetic field. Bose condensates of atoms with nonzero integer spin [@stenger98; @ho98; @ohmi98], referred to as “spinor Bose condensates”, combine spin and superfluid ordering in different ways depending on the spin and the interatomic interaction. These condensates, and the vortices and other topological defects that they allow, have been actively studied in recent years.
Since the physics of individual defects in spinor Bose condensates is now understood for the most experimentally relevant cases with total spin $F\leq3$, [@zhou01; @makela06a; @mukerjee06; @semenoff07; @yip07; @barnett07], a natural next step is to understand physical situations controlled by the collective physics of many defects. Two examples are the vortex lattice in a rotated condensate and the superfluid transition in a two-dimensional condensate. In general, the lowest-energy vortex defects of spinor condensates have both superfluid and spin character. Although external rotation of the condensate couples only to the superfluid part, the mixed nature of the vortices means that the interaction between the spin parts of different vortices is also important in determining the vortex lattice.
This letter uses a general approach to vortex lattice phases in spinor condensates, including the quadratic Zeeman anisotropy normally present in experimental systems, to show that the cyclic phase of an $F=2$ spinor condensate undergoes an unusual vortex lattice transition in a weak applied magnetic field. This transition allows collective physics resulting from the nontrivial spin configuration of vortices to be imaged using spin-insensitive measurements. The comparison of energies of different lattices uses an Ewald summation trick that exactly reproduces previous results obtained for simpler lattices using elliptic functions [@tkachenko66]. More generally, the methods of this letter allow the energy of any periodic arrangement of vortices in a spin-anisotropic spinor condensate to be rapidly calculated. We also show that under some conditions there is a strictly aperiodic vortex structure rather than a true lattice.
Dilute $F=2$ bosons interact via the potential $V(|{\bf r}_1-{\bf
r}_2|)=\delta({\bf r}_1 - {\bf r}_2) (g_0 P_0 + g_2 P_2 + g_4
P_4)$, where $P_F$ projects into the total-spin $F$ state and $g_F
= 4 \pi \hbar^2 a_F / M$ determines $g_F$ given $a_F$, the scattering length in the spin-$F$ channel. This two-body potential gives the interaction Hamiltonian [@ciobanu00; @ueda00] $${\cal H}_{\rm int} = \int d {\bf r} :
{\alpha \over 2} (\psi^\dagger \psi)^2 + {\beta \over 2}
|\psi^\dagger {\bf F} \psi|^2 + {\tau \over 2} |\psi^\dagger
\psi_t|^2 :,$$ with $\psi$ a five-component vector field whose component $\psi_m({\bf r})$ destroys a boson at point ${\bf r}$ with $F_z =
m$, $m =-2, \ldots, +2$, and ${\bf F}$ denoting the spin-2 representation of the $SU(2)$ generators. $\psi_t$ is the time-reversal conjugate of $\psi$: $\psi_{tm} = (-1)^m
\psi^\dagger_m$. The parameters in this Hamiltonian are determined by $g_0, g_2, g_4$ via $\alpha = (3 g_4 + 4 g_2)/7$, $\beta = -(g_2 - g_4)/7$, $\tau = {1 \over 5} (g_0 - g_4) - {2
\over 7} (g_2 - g_4).$ To $H_{\rm int}$ must be added the one-body Hamiltonian for an isotropic and spatially uniform condensate $${\cal H}_0 = \int d
{\bf r} \; \frac{\hbar^2}{2M} \nabla \psi^{\dagger} \cdot \nabla \psi
- \mu \psi^\dagger \psi,$$ where $\mu$, the chemical potential. Minimizing this Hamiltonian over single-particle condensates leads to three phases: ferromagnetic, antiferromagnetic, and cyclic. The cyclic phase that will be the focus of our work occurs when $\beta, \tau
>0$ and is expected to be realized in a condensate of $^{85}$Rb atoms [@ciobanu00]. The spinor structure of this state, having the symmetry of the tetrahedron, results in a nonabelian homotopy group which has been pursued in the liquid physics community for several years.
In all existing experiments, an important effect even at the single-particle level is the existence of anisotropy in spin space resulting from the magnetic fields used as part of the trapping process. Including the hyperfine interaction, the bosons we consider interact with the external magnetic field as ${\cal H}_z
= \Gamma {\bf I} \cdot {\bf S} - 2 \mu_B B_z$ where $\Gamma$ is the magnitude of the hyperfine interaction, ${\bf I}$ is the nuclear spin, ${\bf S}$ is the electronic spin, $\mu_B$ is the Bohr magneton, and $B$ is the magnitude of the magnetic field taken to point in the $z$-direction. Within the manifold of spin-two states, a Hamiltonian which reproduces the correct energies up to a constant is given by ${\cal H}_z=\sqrt{\Gamma^2
+(\mu_B B)^2 + \Gamma \mu_B B F_z}$ [@briet31]. This Hamiltonian can be expanded in powers of $F_z$. Since the relaxation time of the total magnetization is typically longer than the condensate lifetime, the linear term can be neglected. Particular attention has been paid to the next term which gives rise to the quadratic Zeeman effect [@stenger98]. However, due to the high symmetry of the cyclic state, this quadratic term alone is not enough to select its orientation. For this case, one therefore must consider the cubic term which is at next order.
![Orientations of the cyclic state in an external magnetic field which breaks the spin rotational symmetry. Upon increasing the magnetic field, the spinor will undergo a transition from state (a) to state (b).[]{data-label="Fig:shapes"}](shapes){width="3.5in"}
To determine the spin states in the presence of such a magnetic field, one must also consider the spin exchange interaction energy of the condensate per particle which is $E_s=\frac{1}{2} n \beta
{\langle {\bf F} \rangle} \cdot {\langle {\bf F} \rangle} +\frac{1}{2} n \tau
|{\langle {\chi} |}{ {\chi_t} \rangle}|^2$, where $n$ is the condensate density. Since the total spin is assumed to be conserved in experiments, we can neglect the first term in this expression. Minimizing $E_s+ E_z$ (where $E_z={\langle {{\cal H}_z} \rangle}$) over possible spinor states we find the following: At small magnetic fields, the spinor $\chi_{t1}=\left(\sqrt{1/3},0,0,\sqrt{2/3},0\right)^T$ (up to any rotation about the $z$-axis) is selected. In the classification scheme described in [@barnett06], this state is represented by a tetrahedron with one of its faces parallel to the $xy$ plane. Upon increasing the magnetic field there is a transition at $\mu_B B_c = n \tau/16$ to the spin orientation $\chi_{t2}=\left(\sin(\theta)/\sqrt{2},0,\cos(\theta),0,
-\sin(\theta)/\sqrt{2}\right)^T$, where $\theta$ changes continuously with increasing magnetization. This spinor has the symmetries of a distorted tetrahedron with one of its edges parallel to the $xy$ plane (when $\theta=\pi/4$ it has the symmetries of the regular tetrahedron). These two types of orientations are summarized in Fig. \[Fig:shapes\]. The magnitude of the critical field is of the order of ambient fields in current experiments [@sadler06] but smaller fields can in principle be simulated by optical means [@gerbier06]. Having identified the different types of tetrahedral states that are stabilized in an applied magnetic field, we now discuss the effects of rotation in addition to the applied field. The rotation couples to the phase of the condensate and has the effect of producing point vortices in two dimensions or lines of vortices in three dimensions. A vortex is a special type of configuration in an ordered phase that breaks a continuous symmetry: sufficiently far from a “core” region (linear in 3D or point-like in 2D) in which the order is destroyed. The configuration is locally in an ordered state, but cannot be smoothly deformed to the uniform configuration. We find in general that vortices form a two-dimensional lattice whose properties depend on the nature of the constituent vortices and the interactions between them. For simplicity, it is assumed that the magnetic field and the axis of rotation are in the same direction. Owing to the $\pi/3$ spin rotation symmetry of the state (a) that is stabilized at fields $B < B_c$, its vortices are of three types: $(n,m)$, $(n-1/3,m+1/3)$ and $(n+1/3,m-1/3)$, where $n$ and $m$ are integers and the first argument inside the parentheses is the winding number of the phase while the second is that of the spin. The vortex lattice that is formed has a net nonzero winding number for the phase and zero winding number for the spin.
The energetics, in addition to the above constraints on the winding numbers, will depend on the stiffnesses $K_c$ and $K_s$ of the condensate corresponding respectively to the charge (phase) and the spin. (The expected behavior of these stiffnesses will be discussed in closing.) The interaction energy of two vortices $(x_1,y_1)$ and $(x_2,y_2)$ separated by a distance $r$ in the state (a) is given by $$E = 2\pi K_cx_1x_2\log(\xi/r)+\pi K_sy_1y_2\log(\xi/r),
\label{Eq:vorten}$$ where $\xi$ is the typical radius of a vortex. For $K_s > K_c$, it is energetically favorable to produce only vortices of the type $(1,0)$. However, for $K_s/K_c < 1$, the $(1,0)$ vortex breaks up into $(2/3,1/3)$ and $(1/3,-1/3)$ vortices as can be seen by putting the appropriate values of the winding numbers into Eqn. \[Eq:vorten\]. Once again, the energetics are subject to the constraints of rotation mentioned in the previous paragraph. For values of $K_s/K_c < 1/4$, each $(2/3,1/3)$ vortex breaks up into $(1/3,2/3)$ and $(1/3,-1/3)$ vortices. Thus, in this regime, there are only $(1/3,2/3)$ and $(1/3,-1/3)$ vortices with twice as many of the latter as the former. Since, the nature of the vortices is determined by the ratio $K_s/K_c$, so too is the lattice they form as will be described later. For $B > B_c$, the state (b) is stabilized. This state can be shown to have vortices only of the type $(n,m)$. When subjected to a rotation only $(1,0)$ vortices will be produced like at low fields with $K_s >
K_c$.
Thus, to summarize, the following kinds of vortices are produced by rotation: 1) $(1,0)$ vortices for $K_s
> K_c$ or $B > B_c$, 2) an equal number of $(2/3,1/3)$ and $(1/3,-1/3)$ vortices for $1/4 <K_s/K_c < 1$ and $B < B_c$ and 3) twice as many $(1/3,-1/3)$ vortices as $(1/3,2/3)$ for $0 <
K_s/K_c < 1/4$ and $B < B_c$. In each case, the density of vortices is determined by the angular velocity of the rotation.
Having determined the types of vortices that are produced at the various values of magnetic field and stiffnesses, we now evaluate the energies of vortex lattices. Due to the long-ranged nature of the logarithmic interactions, the energy of a vortex lattice is difficult to evaluate directly. Thus, we develop a method that is similar to the Ewald summation technique for the cohesive energy of three-dimensional ionic crystals [@ewald21; @tosi64]. For simplicity, we use a scalar condensate to demonstrate the technique; the generalization to spinor condensates is straightforward and will be given presently. The energy (in units of the stiffness) of a single vortex taken to be at the origin is given by $$\phi(0)= \sum_{{\bf R} \ne 0} \log\left(\frac{\xi}{\bf R}\right) -
\int d^2 r \rho_0 \log\left(\frac{\xi}{r}\right),
\label{Eq:latsum}$$ where ${\bf R}$ are the lattice vectors and $\xi$, the size of the vortex. The second term is due to a uniform negative background charge of density $\rho_0$ which arises from the fact we are working in a rotating frame of reference. Note that each of these terms diverges individually but their difference does not. The Ewald trick is to add and subtract a normalized Gaussian function $\frac{1}{\pi \sigma^2} e^{-r^2/\sigma^2}$ from each point charge, where $\sigma$ is a screening length. For instance, the potential of a point charge at the origin screened by such a Gaussian function is $
\phi(r) = \frac{1}{2} {\rm Ei}(r^2/\sigma^2),
$ where ${\rm Ei}(x) = \int_x^{\infty} dt e^{-t}/t$ is the exponential integral. Note that this will decay exponentially fast at large distances. In three dimensions the corresponding potential is $\phi(r) = (1-{\rm erf}(r))/r$ where erf is the error function.
Proceeding along these lines, the resulting potential corresponding to Eq. \[Eq:latsum\] is $$\begin{aligned}
\phi(0) =& \frac{1}{2} \sum_{{\bf R} \ne 0}
{\rm Ei}\left(\frac{R^2}{\sigma^2}\right)
+ \sum_{{\bf G} \ne 0} \rho_0 \frac{2\pi}{G^2} e^{-\frac{\sigma^2}{4}G^2}
\notag
\\
&- \log\left(\frac{\xi}{\sigma}\right)-\frac{\gamma}{2} - \rho_0
\frac{\pi}{2}\sigma^2,
\label{Eq:ewald1}\end{aligned}$$ where ${\bf G}$ are the reciprocal lattice vectors and $\gamma$ is the Euler-Mascheroni constant. The first term comes from the density of point charges screened by the Gaussian function while second term comes from difference of the charge density of the Gaussian functions and the uniform charge density. The term $-
\log\left(\frac{\xi}{\sigma}\right)-\frac{\gamma}{2} $ is obtained after subtracting off the additional Gaussian function at the origin where we omit the charge. Finally, the last term is to make the average of the screened potential zero [@tosi64]. On the other hand, the potential of a test charge away by ${\bf d}$ from the origin is $$\begin{aligned}
\phi({\bf d}) = \frac{1}{2} \sum_{{\bf R} }
{\rm Ei}\left(\frac{|{\bf R}-{\bf d}|^2}{\sigma^2}\right)
&+ \sum_{{\bf G} \ne 0} \rho_0 \frac{2\pi}{G^2} e^{-\frac{\sigma^2}{4}G^2}
e^{i{\bf G} \cdot {\bf d}}
\notag
\\ &- \rho_0 \frac{\pi}{2}\sigma^2
\label{Eq:ewald2}\end{aligned}$$ The best check of this procedure is to see if the sum is independent of the parameter $\sigma$. The two sums in real and reciprocal space in Eqns. \[Eq:ewald1\] and \[Eq:ewald2\] both converge exponentially fast. In this way, the energy per vortex of the square lattice is found to be $\phi_s(0) = -\log\left(
\frac{\xi}{a}\right)-1.3105$ (where $a$ is the lattice constant) while that of the triangular lattice at the same density is found to be $\phi_t(0) = -\log\left( \frac{\xi}{a}\right)-1.3211$. Both are in precise agreement with the results obtained by integrating over the full spatial flow pattern [@tkachenko66]. For vortices in spinor condensates, which contain windings of phase and spin, the above procedure is applied individually to each sector with the Ewald sums being weighted by the corresponding stiffnesses. The main advantage of the Ewald technique is that it can be generalized to treat complicated unit cells with an arbitrary number of vortices in them in a straightforward and numerically efficient manner.
Let us first consider the case $B>B_c$. As noted earlier, the vortices produced by the rotation are of the type $(1,0)$. These form the usual triangular lattice for all values of $K_s/K_c$. For $B<B_c$, the fractional winding numbers of the fundamental vortices give rise to more interesting possibilities. For $1/4<K_s/K_c<1$, the lattice is bipartite with equal numbers of $(2/3,1/3)$ and $(1/3,-1/3)$ vortices. We use the Ewald summation technique to numerically evaluate the energy of the lattice assuming the same parallelogram unit cell for both sublattices and an arbitrary displacement between them. We then perform a minimization of the energy over these parameters to identify the lattice that is produced at different values of $K_s/K_c$. The sequence of lattices is described in Fig. \[Fig:phasediag1\]. At exactly $K_s/K_c=1$, the two sublattices do not interact with each other and each is a triangular lattice. As soon as $K_s/K_c$ is lowered and the two begin to interact, the honeycomb lattice is stabilized and remains so till $K_s/K_c=0.76$. Below this value, the vortices of one type move to the centers of the rhombic unit cells formed by the other type which we term an interpenetrating rhombic lattice. The internal angle of the rhombus changes continuously with $K_s/K_c$ from $\pi/3$ at $K_s/K_c=0.76$ to $\pi/2$ at $K_s/K_c=0.64$. The interpenetrating square lattice thus obtained at $K_s/K_c=0.64$ is stable down to $K_s/K_c=1/4$. This sequence of lattices is the same as obtained for rotating two-component condensates in the quantum Hall regime [@muellerho02], or equivalently the $F=1$ polar condensate, but the values where the transitions occur are different for $F=2$.
For $K_s/K_c<1/4$, a lattice with $(1/3,2/3)$ and $(1/3,-1/3)$ vortices is obtained with twice as many of the latter as the former. Exactly at $K_s/K_c=1/4$, the two sublattices do not interact and each is a triangular lattice. The sublattice of the $(1/3,2/3)$ vortices has a unit cell of length $\sqrt{2}$ times that of the $(1/3,-1/3)$ vortices. These two lattices are [*incommensurate*]{} for any rotation angle between them, which follows from showing that the nonzero squared lengths of lattice vectors in one lattice are disjoint from those in the other lattice. This incommensurability implies that the energy of interaction between the two lattices can be calculated using the Ewald technique by averaging over all displacement vectors instead of specific lattice points, and the result is zero. In the other limit, $K_s/K_c \rightarrow 0$, the interaction between all pairs of vortices is identical and a triangular lattice is obtained. While there are several way to distribute the two kinds of vortices in such a lattice, the lattice where the $(1/3,-1/3)$ vortices form a honeycomb lattice while the $(1/3,2/3)$ vortices are at the centers of each hexagon is the most symmetric one with three vortices per unit cell. The behavior between the incommensurate structure at $K_s/K_c = 1/4$ and this specific triangular structure as $K_s/K_c \rightarrow 0$ is difficult to determine reliably by our technique, since given the existence of the incommensurate structure, there is no justification for a numerical search over unit cells with a finite number of basis vectors.
![A schematic depiction of the different types of vortex lattices obtained at different values of $K_s/K_c$ and magnetic field. The color code for the vortices is: (1,0) yellow, (2/3,1/3) cyan, (1/3,-1/3) magenta and (1/3, -2/3) white.[]{data-label="Fig:phasediag1"}](phasediag1){width="3.5in"}
As demonstrated above, transitions between different vortex lattices can be tuned by a magnetic field $B$ or the ratio $K_s/K_c$. While the field $B$ can be applied directly or its effect simulated through optical techniques in experiments [@gerbier06], the ratio $K_s/K_c$ is more difficult to manipulate directly. In spinor condensates at low temperatures, this ratio is typically close to 1 but is renormalized by quantum and thermal fluctuations. Increasing temperature acts to reduce $K_s$ more rapidly than $K_c$, because the soft spin modes that are excited at finite temperature have a larger phase space than the phase modes (assuming that the quadratic Zeeman term can be neglected). Both a nonlinear-sigma-model analysis and a study of Bogoliubov-like excitations suggest that under normal experimental conditions at nonzero temperature, $K_s$ is slightly less than $K_c$ so that the magnetic transition will be observable. The best possibility to observe evolving vortex structure as $K_s/K_c$ is further reduced is to raise the temperature very close to $T_c$ of the condensate: if the magnetic order is lost before the superfluid order, as allowed by Landau theory, this ratio will rapidly decrease to zero in a narrow temperature range.
To conclude, we have shown using the Ewald summation technique that different types of vortex lattices can be produced in cyclic condensates as functions of magnetic field and the ratio of the charge and the spin stiffnesses. In particular, there is a magnetic-field-driven transition from a triangular to a honeycomb lattice. In the low-field limit, there are both abrupt transitions and continuous families of lattices as functions of the ratio of the stiffnesses, including the appearance of an incommensurate structure at one value.
The authors would like to thank D. A. Huse, M. Lucianovic, O. Motrunich, D. Podolsky, D. Stamper-Kurn, M. Vengalattore, and A. Vishwanath for useful discussions. RB was supported by the Sherman Fairchild Foundation, SM by DOE/LBNL, and JEM by NSF DMR-0238760.
[20]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , ****, ().
, , , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, **** ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Transiting exoplanetary systems are surpassingly important among the planetary systems since they provide the widest spectrum of information for both the planet and the host star. If a transiting planet is on an eccentric orbit, the duration of transits ${T_{\rm D}}$ is sensitive to the orientation of the orbital ellipse relative to the line of sight. The precession of the orbit results in a systematic variation in both the duration of individual transit events and the observed period between successive transits, $P_{{\mathrm{\rm obs}}}$. The periastron of the ellipse slowly precesses due to general relativity and possibly the presence of other planets in the system. This secular precession can be detected through the long-term change in $P_{{\mathrm{\rm obs}}}$ (transit timing variations, TTV) or in ${T_{\rm D}}$ (transit duration variations, TDV). We estimate the corresponding precession measurement precision for repeated future observations of the known eccentric transiting exoplanetary systems (XO-3b, HD 147506b, GJ 436b and HD 17156b) using existing or planned space-borne instruments. The TDV measurement improves the precession detection sensitivity by orders of magnitude over the TTV measurement. We find that TDV measurements over a $\sim4$year period can typically detect the precession rate to a precision well exceeding the level predicted by general relativity.'
author:
- |
András Pál$^{1,2}$[^1] and Bence Kocsis$^{1,3}$[^2]\
$^{1}$Harvard-Smithsonian Center for Astrophysics, 60 Garden street, Cambridge, MA, 02138, USA\
$^{2}$Department of Astronomy, Loránd Eötvös University, Pázmány P. st. 1/A, Budapest H-1117, Hungary\
$^{3}$Department of Atomic Physics, Institute of Physics, Loránd Eötvös University, Pázmány P. st. 1/A, Budapest H-1117, Hungary
date: 'Accepted ..., Received ... ; in original form ...'
title: Periastron Precession Measurements in Transiting Extrasolar Planetary Systems at the Level of General Relativity
---
\[firstpage\]
binaries: eclipsing – planetary systems – relativity – methods: observational – techniques: photometric
Introduction
============
Since the discovery of the first transiting extrasolar planet [@charbonneau2000; @brown2001], the number of such systems has increased to more than 30[^3]. These transiting extrasolar planets (TEPs) provide unique information on the properties of the system. Based on the geometry provided by the transit light curve(s), the inclination, the physical radius and mass, therefore the density and the surface gravity can be derived, in addition to the mass of the planet. Moreover, the time between successive transits can be measured with an exceedingly high accuracy ($\sim 10^{-6}$ – $10^{-7}$, relative to the period). The detection of long–term transit timing variations can be used to learn more beyond the properties of the parent-star system [@miralda2002; @steffen2007]. They can be indicative of the presence of other planetary companions [see e.g. @holman2005; @agol2005; @millerricci2008], co-orbital companions [Trojans, see @ford2007], or satellites [@simon2007] in the system, could provide information on the oblateness of the host star, or can be used to detect the additional prograde periastron precession predicted by general relativity (GR) [@miralda2002; @heyl2007]. Secular variations in the semimajor axis (and therefore in the transit timing) are also predicted on the time scale of stellar life due to the anisotropic light redistribution [a.k.a. Yarkovski-effect, see @fabrycky2008]. Furthermore, @iorio2006 has shown that TEP observations can in principle also test the gravitoelectric correction of GR by measuring the radial velocity amplitude and transiting periodicity simultaneously, in order to verify that the third Kepler’s law requires a semimajor axis dependent correction.
In a pioneer study, @miralda2002 derived the modification of the observed time period between successive transits $P_{{\mathrm{\rm obs}}}$, called transit timing variations (TTVs), caused by the standard periastron precession due to GR [e.g. @misner1973] and the perturbations of other planets if present. Recently, @heyl2007 have extended these studies and estimated the precision of precession rate measurements for long–term mock observations of eccentric transiting extrasolar planets (ETEPs). Both studies restricted to small eccentricities. At that time, the existence of close eccentric planets was known only through radial velocity measurements, and no ETEPs had been observed. Since their publication, four transiting extrasolar planets have been discovered with significant eccentricity: XO-3 [@johnskrull2007], HD 147506 [a.k.a. HAT-P-2, @bakos2007], GJ 436 [@gillon2007; @butler2004], and HD 17156 [@fischer2007]. Therefore it is now possible, for the first time, to make specific predictions for future, long–term measurements of periastron precession effects for real exoplanetary systems.
In this paper we determine the precision by which repeated long–term future ETEP observations will be able to detect the periastron precession rate for existing systems. In addition to TTVs, i.e. the slow modulation of $P_{{\mathrm{\rm obs}}}$ considered previously [@miralda2002; @heyl2007], the periastron precession also changes the time durations ${T_{\rm D}}$ of individual transits. We examine whether these transit duration variations (TDVs) can be used to improve the sensitivity of periastron precession measurements. We estimate the precession rate measurement precision for long term repeated observations of $P_{{\mathrm{\rm obs}}}$ and ${T_{\rm D}}$ for the known ETEPs. Since several of the observed ETEPs have large eccentricities, we derive expressions for both TTVs and TDVs which are applicable for arbitrary eccentricities. We estimate whether future observations of currently known ETEPs will be able to reach the sensitivity necessary to test the prediction of GR, using existing or planned space-borne instruments. We refer the reader to a recent independent study by @jordan2008, of precession rates in eccentric transiting extrasolar planets.
The next section of this paper introduces the geometrical description which is the basis of our calculations, and derives the expected transit timings and durations for planets orbiting a star with an arbitrarily large eccentricity. In § \[s:realsystems\], we utilize our results for the confirmed four ETEP systems, and give predictions for future observations of periastron precession with space-borne observations. Our conclusions are discussed in § \[s:summary\].
Transit timings and durations for eccentric orbits {#s:transit}
==================================================
The reference frame used for the description of exoplanetary systems as well as for binary/multiple stellar systems is fixed to the sky: the plane of the sky is defined by the $(X+,Y+)$ while $Y+$ points towards to north. For planetary transit observations, the line-of-sight lies close to the orbital plane, i.e. perpendicular to the plane of the sky. The orbit is given by Cartesian coordinates $(\xi,\eta)$, where $\xi+$ is parallel to $X-$ and $\eta+$ oriented toward the observer (see also Fig. \[fig:geometry\]). The Lagrange vector or eccentricity vector is given in these coordinates as $(k,h)=(e\cos\omega,e\sin\omega)$. Let us define the angle $\varphi_0$ as the angle relative to $\xi+$ in the orbital plane at the instance[^4] of the transit. From the definitions of $(\xi,\eta)$, observing from Earth is equivalent to setting $\varphi_0=\pi/2$.
Now let us denote the mean longitude of the planet at the instance of the transit by $\lambda$. For a circular orbit, $\lambda=\varphi_0$. From its standard definition in celestial mechanics, it is straightforward to derive the mean longitude for arbitrary eccentricities (see ). The result is $$\begin{aligned}
\lambda & \equiv & \lambda(\varphi_0,k,h)\equiv \lambda(\varphi_0,e\cos\omega,e\sin\omega)= \nonumber \\
& = & \arg\left(k+\cos\varphi_0+\frac{he_{\perp}}{2-\ell},
h+\sin\varphi_0-\frac{ke_{\perp}}{2-\ell}\right) - \nonumber \\
& & -
\frac{e_{\perp}(1-\ell)}{1+e_{\parallel}}, \label{lambdaattransit}\end{aligned}$$ where $e_{\parallel}=k\cos\varphi_0+h\sin\varphi_0$, $e_{\perp}=k\sin\varphi_0-h\cos\varphi_0$ are the components of the eccentricity vector relative to the line-of-sight, $\ell=1-\sqrt{1-e^2}$ is the oblateness of the orbit, and $\arg(x,y)=\arctan(y/x)$ if $x\geq 0$ and $\pi+\arctan(y/x)$ otherwise. Plugging in the observed values of $e$ and $\omega$, provides a simple way of calculating the mean longitude of the orbit for an arbitrary transit observation. Note that this formalism omits the direct usage of the mean anomaly, true anomaly and eccentric anomaly which have no meanings for $e\to0$. All of our derived formulae are based on the well–behaved parameters $\lambda$, $k$, and $h$, and therefore are valid for arbitrary eccentricities.
In the following subsections, we derive the expressions describing TTVs and TDVs, discuss the corresponding observational implications and estimate the precision of periastron precession observations.
.
System $M_\star/M_{\odot}$ $P$ (d) $e$ $\omega$ (degrees) $H$ (d) $b$ $P_{\rm sec}$ (years) $\frac{m_2/M_{\oplus}}{(a_2/3a)^3}$
------------ --------------------- ------------- ------------------- -------------------- --------- --------------- ----------------------- -------------------------------------
HD 147506b $1.298\pm0.07$ $ 5.63341 $ $0.520\pm0.010$ $179.3\pm3.6$ $0.083$ $0$ $19790\pm740$ $12.2$
XO-3b $1.41\pm0.08$ $ 3.19154$ $0.260\pm0.017$ $344.6\pm6.6$ $0.050$ $0.8$ $9280\pm410$ $15.9$
GJ 436b $0.452\pm0.013$ $ 2.64385 $ $0.150\pm0.012$ $351.0\pm1.2$ $0.065$ $0.85\pm0.02$ $15180\pm400$ $2.6$
HD 17156b $1.2\pm0.1$ $21.21725 $ $0.6717\pm0.0027$ $121.23\pm0.40$ $0.098$ $0.50\pm0.12$ $143000\pm7900$ $5.9$
Modulation of the transit period
--------------------------------
The period between successive transits $P_{{\mathrm{\rm obs}}}$ is modified by a possible slow precession of the orbital elements. These modulations are referred to as transit timing variations (TTVs). We derive the modulation of $P_{\rm obs}$ due to periastron precession in two steps. First we demonstrate that the change in the period between successive transits is simply related to the change in the mean longitude $\Delta \lambda$. Then relating the mean longitude shift to the periastron precession rate we derive the expected TTV rate.
Let $P_0$ be the orbital period, $n=2\pi/P_0$ be the mean angular velocity. The mean longitude of the planet increases steadily in time, $\lambda=nt+\lambda_0$. The variation in the mean longitude at the transit center would result in a variation in the transit cadence. During a transit at time $t_1$, the mean longitude is $\lambda_{\rm tr}(t_1)=nt_1+\lambda_0$, and after an observed revolution, at the instance $t_2=t_1+P_{{\mathrm{\rm obs}}}$ it is $\lambda_{\rm tr}(t_2)+2\pi=nt_2+\lambda_0$. Therefore the observed period between transits is $$\label{e:Pobsdef}
P_{{\mathrm{\rm obs}}}=\frac{2\pi+\Delta \lambda}{n}=P_0\left(1+\frac{\Delta\lambda}{2\pi}\right)$$ where $\Delta \lambda=\lambda_{\rm tr}(t_2)-\lambda_{\rm tr}(t_1)$.
In the following we assume that the shift in the mean longitude is caused by the perihelion shift $\Delta \omega = P_0 \dot\omega$ per period, (i.e. we assume a constant eccentricity), then from the chain rule $$\label{e:DLambda}
\Delta \lambda= \frac{\partial\lambda}{\partial \omega}\Delta\omega=
\frac{\partial\lambda}{\partial k} \frac{\partial k}{\partial \omega}\Delta\omega + \frac{\partial\lambda}{\partial h} \frac{\partial h}{\partial \omega}\Delta \omega.$$ Here $\partial k/\partial \omega=-e\sin\omega$ and $\partial h/\partial \omega=e\cos\omega$, from definition (see above), and the partial derivatives $\partial\lambda/\partial k$ and $\partial\lambda/\partial h$ can be found from and are given explicitly in the Appendix (\[e:lambda\_k\]–\[e:lambda\_h\]). At transit, we get $$\label{e:lambda/omega}
\frac{\partial \lambda}{\partial \omega}=
\frac{e^2}{1+\sqrt{1-e^2}}+\sqrt{1-e^2}\frac{e^2+(2+e\sin\omega)e\sin\omega}{(1+e\sin\omega)^2}.$$ Combining and defining the secular period of the periastron precession as $P_{\rm sec}=(2\pi)/\dot\omega$, we get $$P_{{\mathrm{\rm obs}}}=P_0+\frac{P_0^2}{2\pi} \frac{\partial\lambda}{\partial\omega} \dot\omega
=P_0+ \frac{\partial\lambda}{\partial\omega}\frac{P_0^2}{P_{\rm sec}}.\label{pobs1}$$
Since $\partial\lambda/\partial\omega$ itself is not constant due to periastron precession, the observed period between transits slowly changes. Differentiating with respect to time, we get $$\label{pobs1b}
\dot P_{\rm obs} = 2\pi\frac{\partial^2\lambda}{\partial\omega^2} \frac{P_0^2}{P_{\rm sec}^2}
=\frac{4\pi(1-e^2)^{3/2}e\cos\omega}{(1+e\sin\omega)^3}\frac{P_0^2}{P_{\rm sec}^2},$$ since the partial derivative of with respect to $\omega$ is $$\frac{\partial^2\lambda}{\partial\omega^2}=\frac{2(1-e^2)^{3/2}e\cos\omega}{(1+e\sin\omega)^3}. \label{d2ldo2}$$ Note that this equation clearly shows that the small eccentricity approximation $\partial^2\lambda/\partial\omega^2 \approx e\cos \omega$ [e.g. used by @miralda2002; @heyl2007] is very imprecise for moderate to large eccentricites. Depending on the actual value of $\omega$, can result even $6-8$ times smaller or larger values for $\partial^2\lambda/\partial\omega^2$ as its first order approximation for eccentricities $0.5-0.7$.
We have to mention here that the observed period and therefore the individual transit timings are also affected by the light time effect (LTE). Since the precession of an elliptical orbit causes the distance between the host star and the planet at the transit instances to vary, the light time delay will change for transit event to transit event. The magnitude of this effect can be derived as follows. The distance between the star and the planet at transit time (i.e. when $\lambda=\varphi_0=\pi/2$) is $$r=\frac{a(1-e^2)}{1+e\cos v}=\frac{a(1-e^2)}{1+e\cos(\varphi_0-\omega)}=
\frac{a(1-e^2)}{1+e\sin\omega}. \label{rattransit}$$ This difference in the distance implies an additional $-r(1-\mu)/c$ time shift, respective to the barycentric reference frame of the star–planet system. Here, $\mu=M_{\rm p}/(M_{\rm p}+M_\star) \ll 1$ is the mass parameter, $a$ is the semimajor axis of the orbit and $c$ is the speed of light. Therefore, the correction in the observed period is $$\begin{aligned}
P^{\rm +LTE}_{{\mathrm{\rm obs}}} & = & P_{{\mathrm{\rm obs}}}-\frac{a}{c}(1-e^2)\frac{\partial (1+e\sin\omega)^{-1}}{\partial\omega}P_0\dot\omega= \nonumber \\
& = & P_{{\mathrm{\rm obs}}}+2\pi \frac{a}{c}\frac{(1-e^2)e\cos\omega}{(1+e\sin\omega)^2}\frac{P_0}{P_{\rm sec}},\end{aligned}$$ where we neglected the barycentric correction. Thus, the variation in this period (corrected for LTE) due to the variation in $\omega$ is $$\label{e:LC}
\dot P^{\rm +LTE}_{{\mathrm{\rm obs}}} = \dot P_{{\mathrm{\rm obs}}} - 4\pi^2(1-e^2)\frac{e(e+e\cos^2\omega+\sin\omega)}{(1+e\sin\omega)^3}\frac{a}{c}\frac{P_0}{P_{\rm sec}^2}.$$ Comparing and (\[e:LC\]), and assuming that the motion of the planet is non-relativistic, $a/c\ll P_0$, we find that $|\dot P^{\rm +LTE}_{{\mathrm{\rm obs}}}-\dot P_{{\mathrm{\rm obs}}}| \ll |\dot P_{{\mathrm{\rm obs}}}|$. We conclude that the period variation due to the LTE is negligible compared to the period variation caused by the changing geometry.
In summary, , along with , give the modulation of the actual observable period between transit events. These results are valid for arbitrary eccentricities and are independent of the physical mechanism causing the secular precession of the periastron. We calculate the secular precession period caused by general relativity [see e.g., @wald1984] using $$P_{\rm sec} = \frac{c^2 (1-e^2)}{3(2\pi G M_\star)^{2/3}}
P_0^{5/3}, \label{pprecmmotion}$$ where $M_\star$ is the mass of the parent star, $G$ is Newton’s gravitational constant. This secular period is of order $10^4$–$10^5\,$years for the known ETEP systems (see Table \[table:basicdata\] and for more details). We note that if other planets are also present in these systems, they might also cause additional periastron precession of a larger magnitude. The Yarkovski-effect [@fabrycky2008] and the tidal circularization [see @johnskrull2007 and the references therein] lead to negligible modifications for our purposes, as these effects are relevant on timescales exceeding $0.1\,$Gyr for these systems. In the following, we compare the precession measurement sensitivities with the general relativistic rate $P_{\rm sec}$ given by .
Modulation of the transit duration
----------------------------------
Here we investigate how the periastron precession affects the duration of a transit. Let us denote the half duration of the transit by $H={T_{\rm D}}/2$, which we define as half the time between the instances when the center of the planetary disk intersects the limb of the star, i.e. between the center of the ingress and egress. Note that this is not the time between the first and last contact. This is important because the instances of the center of the ingress and egress can be measured more accurately than their beginning or end. Since we are interested in estimating the *variations* of the duration of the transit to leading order, we perform a first-order calculation, i.e. assuming a constant apparent tangential velocity for the planet and neglecting the changes in the impact parameter due to the elliptical orbit and/or the curvature of the projection of the orbit due to the inclination. From Kepler’s Second Law and , one can calculate the tangential distance $\Delta x$ traveled by the planet during time $H$, $$\Delta x=v_{\rm tan} H = an\frac{1+e\sin\omega}{\sqrt{1-e^2}} H$$ (see for the derivation of $v_{\rm tan}$). This can be related to the impact parameter $b$ and the radius of the star $R_\star$ for the geometry of the transit as $$\frac{\Delta x}{R_\star}= \sqrt{1-b^2}.$$ Thus, to leading order, $$\label{e:H}
H = \frac{P_0}{2\pi}\left\{\frac{R_\star}{a}\frac{\sqrt{1-e^2}}{1+e\sin\omega}\sqrt{1-b^2} +
\mathcal{O}\left[\left(\frac{R_\star}{a}\right)^{3}\right]\right\}.$$ Note that depends on $\omega$ through the $(1+e\sin\omega)^{-1}$ term and also implicitly through $b$, $$\label{e:b}
b=\frac{r\cos i}{R_\star} = \frac{a}{R_\star}\frac{1-e^2}{1+e\sin\omega}\cos i.$$ The variation in $H$ caused by the variation in the periastron can be found from and (\[e:b\]), $$\frac{\partial H}{\partial\omega}=
\frac{e\cos\omega}{1+e\sin\omega}H\frac{1-2b^2}{1-b^2}. \label{dhdomega}$$ Note that equation reflects the qualitative expections implied by Kepler’s Second Law. Namely, if an eccentric orbit advances, the distance between the planet and the star will change. If this distance decreases, the impact parameter will also decrease (resulting a longer transit duration) but due to Kepler’s Second Law, the apparent tangential velocity of the transiting object will increase (resulting a shorter transit duration). Therefore at a certain value of the inclination and/or the impact parameter, the two effects cancel each other yielding no TDV effect. Equation \[dhdomega\] clearly shows that it occurs when the impact parameter is $b=1/\sqrt{2}\approx0.707$. The long-term variation in the duration of the transit is then $$\label{doth}
\dot H =\frac{\partial H}{\partial\omega}\dot\omega=\frac{\sqrt{1-e^2} e \cos \omega}{(1+e\sin \omega)^2}\frac{1-2b^2}{\sqrt{1-b^2}}\frac{R_{\star}}{a}\frac{P_0}{P_{\rm sec}}$$ Comparing and the TDV effect relates to the TTV effect as $$\label{e:TDV/TTV}
\frac{\dot H}{\dot P_{\rm obs}} =
\frac{1+e\sin \omega}{6\pi}\frac{1-2b^2}{\sqrt{1-b^2}} \frac{R_{\star}}{R_{\rm Sch}}$$ where $R_{\rm Sch}=2 G M/c^2$ is the Schwarzschild radius of the star. As an example, note that $R_{\star}/R_{\rm Sch}=2.5\times 10^{5}$ for the Sun. Therefore, shows that the TDV effect is always much larger than the TTV effect. In particular, the ratio is larger for increasing $b$. In the limit $b\to 1$ breaks down because the periastron precession shifts the orbit out of the transiting region.
Observational implications
--------------------------
Now, using the results for the TDV and TTV effects, and (\[doth\]), we can estimate how these timing and transit duration variations might be observed on long timescales. In the following we discuss these observational implications.
### Transit Timing Variations
Equation (\[pobs1b\]) shows that the observed period between successive transits increases or decreases at a practically constant rate during the observations, $\dot P_{{\mathrm{\rm obs}}}$. The time of the $m$th transit from an arbitrary epoch $T_0$ can be found from adding up the contributions of the observed $m$ number of periods $$\begin{aligned}
T_m&=& T_0 + P_{{\mathrm{\rm obs}}}m+Dm^2, \label{ttiming}\end{aligned}$$ where $P_{{\mathrm{\rm obs}}}\approx P_0$ denotes the time of the first observed orbit, $D$ is the transit timing variation factor, $$D=\frac{P_{0}\dot P_{{\mathrm{\rm obs}}}}{2}=2\pi G_{\rm TV}(e,\omega)P_0\left(\frac{P_0}{P_{\rm sec}}\right)^2, \label{dtiming}$$ where we have introduced the geometrical factor $$\label{e:G_TV}
G_{\rm TV}(e,\omega)=(1-e^2)^{3/2}\frac{e\cos\omega}{(1+e\sin\omega)^3}.$$
The chance of detecting the periastron precession increases with the geometrical factor $G_{\rm TV}(e,\omega)$ which is related to the alignment of the orbital ellipse with the line of sight. The optimal value for detecting the precession is $\omega=0$ or $\pi$ for small eccentricities, i.e. the semimajor axis should be perpendicular to the line of sight. For arbitrary eccentricities, the optimal value for $\omega$ for the TTV observation is $$\omega^{\rm best}_{\rm TV}=\frac{3}{2}\pi\pm\arccos\left(\frac{6e}{1+\sqrt{1+24e^2}}\right).$$ which approaches $0^{\circ}$ and $180^{\circ}$ for small eccentricities as expected, and $270^{\circ}$ for large eccentricities. In case of $e=0.5$, $\omega_{\rm extr}=\{235.4^\circ,304.6^\circ\}$. The least favorable value for $\omega$ occurs when $G_{\rm TV}(e,\omega)=0$, i.e. at $\omega^{\rm worst}_{\rm TV}=\{90^{\circ},270^{\circ}\}$,.
### Transit Duration Variations
Now let us turn to the TDV effect. The observed duration of the $m$th transit can be calculated in the same way, namely $$H_m=H_0+Fm,$$ where $F$ is the shift in the transit duration per orbit. This factor is $$\label{e:F}
F= P_0 \dot H = 2\pi G_{\rm DV}(e,\omega,b)H\frac{P_0}{P_{\rm sec}},$$ and $$\label{e:G_DV}
G_{\rm DV}(e,\omega,b)=\frac{1-2b^2}{1-b^2}\frac{e\cos\omega}{1+e\sin\omega}.$$ The optimal orientation $\omega^{\rm best}_{\rm DV}$ for detecting the TDV effect for fixed $e$, $b$, and $P_0$ can be found by maximizing $ H G_{\rm DV}(e,\omega,b)$. The result is $$\omega^{\rm best}_{\rm DV}=\frac{3}{2}\pi\pm\arccos\left(\frac{4e}{1+\sqrt{1+8e^2}}\right).$$ and the worst orientation is at $\omega^{\rm worst}_{\rm DV}=\{90^{\circ},270^{\circ}\}$, just like for the TTV case. Comparing $\omega^{\rm best}_{\rm TV}$ and $\omega^{\rm best}_{\rm DV}$ it is clear that the most favorable orientation in terms of the two effects are similar, hence the chance of detecting the periastron motion through transit timing variations or transit duration variations is correlated. Both effects go away if the eccentricity is oriented parallel to the line of sight. We also note that for moderate values of $e$ and small impact parameters, $|G_{\rm TV}|\approx|G_{\rm DV}|$ which also implies that the most favorable geometry for detecting either TTVs or TDVs is similar.
Error analysis
--------------
Next we estimate the parameter measurement precision of the TTV and TDV effects for future observations. We consider the repeated observation of a particular transiting system over a total timespan $T_{\rm tot}$, measuring the transit timing $T_m$ and duration $H_m$ for each transit with respective errors $\sigma(T)$ and $\sigma(H)$. (We discuss the specific values of $\sigma(T)$ and $\sigma(H)$ for transit observations in Section \[s:photometric\]). For simplicity, let us assume that these measurements are equidistant and in total $N$ independent transits are observed, i.e. the $m$th transit is observed if $m=0,d,\dots,(N-1)d$, where $d=T_{\rm tot}/(NP_0)$.
### Transit Timing Variations
Using , we can fit a second-order polynomial to these observations with unknown coefficients $T_0$, $P_{\rm obs}$ and $D$ by minimizing the merit function $$\chi^2_{\rm TV} = \sum\limits_{m=0,d,\dots,(N-1)d}
\left[\frac{T_m - (T_0 + P_{\rm obs}m+Dm^2)}{\sigma(T)}\right]^2.$$ The minimization of the above function results in a linear set of equations in the parameters $p_{i}=\{T_0, P_{\rm obs}, D\}$. Assuming Gaussian errors, the parameter estimation covariance matrix can be found from the Fisher matrix method [@finn1992]: $$\left< \delta p_i \delta p_j \right> = (\mathcal{F}^{-1})_{ij}$$ Here $\mathcal{F}$ is the Fisher matrix defined as $$\mathcal{F}_{ij} = \sum\limits_{m=0,d,\dots,(N-1)d} \frac{1}{\sigma^2(T)}\frac{\partial T^{\rm fid}_m}{\partial p_i} \frac{\partial T^{\rm fid}_m}{\partial p_j}$$ where $T^{\rm fid}_m$ is the fiducial value of $T_m$ given by . The marginalized expected squared parameter estimation error is given by the diagonal elements of the covariance error matrix $\sigma^2(p_i)=(\mathcal{F}^{-1})_{ii}$. In particular, the resulting uncertainty of $D$ becomes $$\begin{aligned}
\sigma(D) &=& \frac{\sqrt{180}\,\sigma(T)}{d^2 \sqrt{N(N^2-1)(N^2-4)}} \nonumber\\
&\approx& \sqrt{180} \left(\frac{P_0}{T_{\rm tot}}\right)^2 \left(1+\frac{5}{2N^2} \right) \frac{\sigma(T)}{\sqrt{N}}. \label{duncert}\end{aligned}$$ Here, the first equality is valid for arbitrary $N$, while the second is its first order approximation for large $N$. The leading order approximation is verified against @press1992.
### Transit Duration Variations
We can repeat the same calculations as above for the observation of the half transit duration $H_m$ to measure the variation factor $F$. The merit function in this case $$\chi^2_{\rm DV} = \sum\limits_{m=0,d,\dots,(N-1)d}
\left[\frac{H_m - (H_0 + Fm)}{\sigma(H)}\right]^2,$$ has to be minimized for the same set of observations. This minimization again leads to a linear set of equations in the parameters $H_0$ and $F$. The Fisher matrix method in this case gives the uncertainty in $F$ as $$\begin{aligned}
\sigma(F) &=& \frac{\sqrt{12}\sigma(H)}{d\sqrt{N(N^2-1)}}\approx \nonumber\\
&\approx& \sqrt{12} \frac{P_0}{T_{\rm tot}} \left(1+\frac{1}{2N^2} \right)
\frac{\sigma(H)}{\sqrt{N}}. \label{funcert}\end{aligned}$$
Figure \[fig:tdvksig\] shows the detection significance of the TDV measurement $|F|/\sigma(F)$, if the precession rate in $F$ is given by the general relativistic formula, . Here each transit is assumed to be measured (i.e. $d=1$) with a precision $\sigma(H)=5$sec for a total observation time of $T_{\rm tot}=4$years. These assumptions are realistic for the future Kepler mission (see § \[s:photometric\] below). Other parameters are $M_{\star}=M_{\odot}$, $R_{\star}=R_{\odot}$, $b=0$, and we averaged over the possible orientations of $\omega$. For other parameters, $$\begin{aligned}
\frac{|F|/\sigma(F)}{|F_{0}|/\sigma(F_{0})} & = & \frac{1}{d}\frac{1}{\sqrt{1-b^2}}
\left(\frac{R_{\star}}{R_{\odot}}\right) \left(\frac{M_{\star}}{M_{\odot}}\right)^{1/3}\cdot \nonumber \\
& & \cdot
\left(\frac{\sigma(H)}{5\,{\rm sec}}\right)^{-1}
\left(\frac{T_{\rm tot}}{4\,{\rm yr}}\right)^{3/2},\end{aligned}$$ implying that the detection significance can be even better.
Figure \[fig:tdvksig\] clearly shows that the chances of detecting the precession effects through the TDV effect is encouraging. The detection significance of the general relativistic precession of a transiting exoplanet with eccentricity $e\gtrsim 0.2$ and period $P\lesssim 5$ days, is typically over the level. Generally, and can be used directly to check what kind of observations are required to detect the precession of the periastron through the TTV or TDV methods, respectively.
The case of XO-3, HD 147506, GJ436 and HD 17156 {#s:realsystems}
===============================================
As of this writing, four TEPs are known with a non-zero eccentricity within , namely XO-3 [@johnskrull2007], HD 147506 [@bakos2007], GJ 436b [@butler2004; @gillon2007], and HD 17156b [@fischer2007; @barbieri2007]. The planet TrES-1 [@alonso2004] has also been reported as an object with non-zero eccentricity, however, it is zero within thus we omit from our analysis. We note here that recently both GJ 436 and HD 17156 have been suggested to have another planetary companions [see @ribas2008; @short2008]. The secular period of the periastron motion are determined by the mass of the star $M_\star$, the orbital period $P_0$, and the eccentricity $e$, while the timing variation constant $D$ is also affected by the actual argument of pericenter, $\omega$. The transit duration variation factor $F$ is affected indirectly by the geometrical ratio $a/R_\star$ and directly by the impact parameter $b$. These parameters are summarized in the first seven columns of Table \[table:basicdata\] for these four ETEP systems. The derived GR periastron precession period, $P_{\rm sec}$ can be found in the 8th column of the table.
In addition to the inevitable periastron precession caused by GR, there might be other sources of perturbations causing periastron precession. The last column gives the minimum mass to semimajor axis ratio of a hypothetical exterior perturber (e.g. a planet or an asteroid belt), in Earth mass units, which causes the same periastron presession rate as that caused by the general relativity. This estimate based on @price1979, and is valid for $a_2\gtrsim 3a$ and for non-resonant cases. Note that the minimum mass of the perturber scales with the third power of the semimajor axis ratio to cause a comparable precession rate as GR. The numbers show that the precession caused by additional planets in the system, if present, can easily cause a larger precession rate than GR. In case of orbital resonances with exterior planets, the precession rate can be even larger [@holman2005; @agol2005]. To be conservative, we examine whether the precession rate can be measured to a precision better than that corresponding to GR.
To obtain a high significance, detection of the periastron precession using transit timing variations, we need $k\sigma(D)\approx|D|$. Using , the total number of transits necessary to measure $D$ with this precision is $$N_{\rm TV}=180\left[\frac{\sigma(T)}{k|D|}\right]^2\left(\frac{P_0}{T_{\rm tot}}\right)^4.$$ Thus, the number of such required transit observations is extremely sensitive to the orbital period $P_0$ and the observation timespan. Table \[table:ttvdata\] gives the corresponding values of the transit timing variation factors for the known ETEP systems and this number of observations, assuming a $T_{\rm tot}=20$ year long observational timespan, timing precision of $\sigma(T)=2~{\rm sec}$ and sensitivity of the GR periastron precession level. The table shows the recently discovered XO-3b system is a promising candidate to detect the GR periastron precession through the TTV effect, while the other ETEP systems require unrealistically many observations. We note that if other perturbing planets are present in these systems and lead to a precession rate that is larger by a factor of 10 than the GR precession rate, then the number of detections (during the same $T_{\rm tot}$ timespan) is lower by a factor of 100. It is interesting that the best candidate (by far) is XO-3b even though its eccentricity is not as large as that of HD 147506b or HD 17156b.
System $D$ (seconds) $N_{20\,{\rm y},~\sigmalevel{3},~2\,{\rm sec}}$
------------ ---------------------------- -------------------------------------------------
HD 147506b $-(5.9\pm0.7)\cdot10^{-7}$ $6600$
XO-3b $+(4.3\pm0.6)\cdot10^{-7}$ $1280$
GJ 436b $+(5.0\pm0.5)\cdot10^{-8}$ $44160$
HD 17156b $-(6.9\pm0.8)\cdot10^{-8}$ $97\cdot10^{6}$
: Transit timing variation factor ($D$, in seconds) and the number of transits ($N_{20\,{\rm y},\sigmalevel{3},2\,{\rm sec}}$) what should be detected almost uniformly in a 20 year long timespan, each with an error of $2$ second to confirm the precession within . []{data-label="table:ttvdata"}
System $F$ (seconds) $N_{4\,{\rm y},~\sigmalevel{3},~2\,{\rm sec}}$
------------ ---------------------------- ------------------------------------------------
HD 147506b $-(1.8\pm0.3)\cdot10^{-2}$ $20$
XO-3b $-(5.4\pm0.6)\cdot10^{-3}$ $70$
GJ 436b $-(4.1\pm0.5)\cdot10^{-3}$ $85$
HD 17156b $-(3.2\pm0.2)\cdot10^{-3}$ $8970$
: Transit duration variation factor ($F$, in seconds) and the number of transits ($N_{2\,{\rm y},~\sigmalevel{3},~2\,{\rm sec}}$) what should be detected almost uniformly in a 4 year long timespan, each with an error of $2$sec to confirm the precession within . []{data-label="table:tdvdata"}
Let us now turn to the observational constraints for the detection of transit duration variations. Using , the total number of required observations within $T_{\rm tot}$ is $$N_{\rm DV}=12\left[\frac{\sigma(H)}{k|F|}\right]^2
\left(\frac{P_0}{T_{\rm tot}}\right)^2.$$ Note that $N_{\rm DV}$ is not as sensitive to the $P_0/T_{\rm tot}$ ratio as $N_{\rm TV}$, and implies that a smaller number of observations is typically necessary.
In Table \[table:tdvdata\] we present the values of the transit duration variation factor, $F$, and the number of required observations to reach the same confidence for detecting the GR periastron precession[^5]. Here we assumed a shorter observation timespan, $T_{\rm tot}=4$ year (i.e. shorter compared to the 20 year long timespan necessary for the detection of the TTV effect), the same timing precision of $\sigma(H)=2~{\rm sec}$ and the same level of detection, . The best known ETEP system for TDV detection is therefore HD 147506b, but the number of necessary observations is feasible for XO-3b and GJ 436b as well.
Photometric detection {#s:photometric}
---------------------
The precision for measuring the transit timing and transit duration for a photometric observation can be estimated as follows. Since the time of the ingress ($T_{\rm I}$) and the time of the egress ($T_{\rm E}$) – i.e. when the center of the planet crosses the limb of the star inwards or outwards, respectively – defines both the time of the transit center and the half duration like $$\begin{aligned}
T & = & \frac12(T_{\rm E}+T_{\rm I}), \\
H & = & \frac12(T_{\rm E}-T_{\rm I}),\end{aligned}$$ moreover $T_{\rm I}$ and $T_{\rm E}$ can be treated as uncorrelated variables, therefore the uncertainties of the transit time and half duration would be nearly the same, i.e. $\sigma(T)\approx\sigma(H)$. We have estimated these uncertainties for the four distinct planets using Monte-Carlo simulations by fitting transit light curves on mock data sets. We have used the observed planetary parameters as an input for these artificial light curves. The fit was performed assuming quadratic limb darkening [see @mandel2002] in the Sloan $z'$ band. The mock light curves were sampled with $\Delta\tau_1=1\,{\rm sec}$ cadence and an additional Gaussian noise of $\sigma_1(m)=1\,{\rm mmag}$ was added. The resulting uncertainties, $\sigma_{1,1}(T)$ and $\sigma_{1,1}(H)$ for the four planets are presented in Table \[table:ttvtdvuncert\]. Since the depth of the four transits are nearly the same (see the appropriate normalized radii, $p=R_p/R_\star$, all between $0.068\lesssim p\lesssim0.085$), the uncertainties $\sigma_{1,1}(T)$ and $\sigma_{1,1}(H)$ are almost the same for the four cases. Using these normalized values, one can easily estimate the uncertainties for arbitrary sampling cadence $\Delta \tau$ and photometric precision $\sigma(m)$ using $$\begin{aligned}
\sigma(T) & \approx & \sigma_{1,1}(T)\frac{\sigma(m)}{1\,\rm mmag}
\sqrt{\frac{\Delta\tau}{1\,\rm sec}}, \\
\sigma(H) & \approx & \sigma_{1,1}(H)\frac{\sigma(m)}{1\,\rm mmag}
\sqrt{\frac{\Delta\tau}{1\,\rm sec}}.\end{aligned}$$ For comparison, note that the expected photometric precision of the Kepler space telescope [see @borucki2007] is $1\,{\rm mmag}$ for observing a light curve of a bright, $M_v=8.8$ star with a $1\,{\rm sec}$ sampling cadence. Since the star XO-3 has almost the same apparent magnitude, it is clear, the transit durations would be detected with an accuracy of $\sigma(H)\approx 5\,{\rm sec}$ if this star was in the field of Kepler. Therefore, and Table \[table:tdvdata\] shows that the transit duration variations would be detectable for HD 147506(b) or XO-3(b)–like systems due to the GR periastron precession, within confidence with a Kepler–type mission within approximately 3 or 4 years, respectively.
System $\sigma_{1,1}(T)$ (sec) $\sigma_{1,1}(H)$ (sec)
------------ ------------------------- -------------------------
HD 147506b 5.3 4.8
XO-3b 6.9 4.7
GJ 436b 6.7 8.4
HD 17156b 5.4 6.1
: Uncertainties of the transit time and transit duration measurements for the four known ETEPs, assuming Sloan $z'$-band photometric data taken with a 1sec cadence and 1mmag photometric precision. []{data-label="table:ttvtdvuncert"}
Summary {#s:summary}
=======
The first four eccentric transiting exoplanetary systems have been discovered during 2007. The precession of an eccentric orbit causes variations both in the transit timings and transit durations. We estimated the significance of measuring the corresponding observable effects compared to the inevitable precession rate of general relativity. We applied these calculations to predict the significance of measuring the effect for the four known eccentric transiting planetary systems. Our calculations show that a space-borne telescope is adequate to detect the change in the transit durations to a high significance better than the GR periastron precession rate within a 3 – 4 year timespan (in a continuous observing mode). The same kind of instruments would need more than a decade to detect the corresponding transit time variations to this sensitivity even for the most optimistic known system.
The CoRoT mission has already found two transiting planets [see @barge2008; @alonso2008] and there are two known planets in the planned field-of-view of the Kepler mission [see @odonovan2006; @pal2008]. Our results suggest that if an *eccentric* transiting planet is found in the Kepler or CoRoT field, these missions will be able to measure the periastron precession rate to a very high significance within their mission lifetime or with the support of ground-based observations on a longer time scale. This will provide an independent test of the theory of general relativity and will also be useful for testing for the presence of other planets in these systems.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank the hospitality and support of the Harvard-Smithsonian Center for Astrophysics where this work was partially carried out. We thank Andres Jordan for useful comments on the manuscript. BK acknowledges support from OTKA grant No. 68228.
[99]{}
Agol, E., Steffen, J., Sari, R., & Clarkson, W., 2005, MNRAS, 359, 567
Alonso, R. et al., 2004, ApJ, 613, 153
Alonso, R. et al., 2008, astro-ph:0803.3207
Bakos, G. Á. et al., 2007, ApJ, 670, 826
Barbieri, M. et al., 2007, A&A, 476, 13
Barge, P. et al., 2008, astro-ph:0803.3202
Borucki, W. J. et al., 2007, ASP Conf. Ser., 366, 309
Brown, T. M. et al., 2001, ApJ, 552, 699
Butler et al., 2004, ApJ, 617, 580
Charbonneau, D., Brown, T. M., Latham, D. W. & Major, M., 2000, ApJ, 529, 45
Fischer, D. A. et al., 2007, ApJ, 669, 1336
Fabrycky, D., 2008, astro-ph:0803.1839
Finn, L. S., 1992, Phys. Rev. D, 46, 5236
Ford, E. B. & Holman, M. J., 2007, ApJ, 664, 51
Gillon, M. et al., 2007, A&A, 472, 13
Heyl, J. S. & Gladman, B. J., 2007, MNRAS, 377, 1511
Holman, M. J, & Murray, N. W., 2005, Science, 307, 1288
Iorio, L., 2006, NewA, 11, 490
Jordan,A. & Bakos, G, 2008, ApJ, in press
Johns-Krull, C .M. et al., 2007, astro-ph:0712.4283
Mandel, K., Agol, E., 2002, ApJ, 580, 171
Miller-Ricci, E. et al., 2008, astro-ph:0802.0718
Miralda-Escude, J., 2002, ApJ, 564, 1019
Misner, C. W., Thorne, K., & Wheeler, J. A., 1973, Gravitation, Second Edition, W. H. Freeman and Company, San Francisco
O’Donovan, F. T. et al., 2006, ApJ, 651, 61
Pál, A. et al., 2008, astro-ph:0803.0746
Press, W. H., Teukolsky, S. A., Vetterling, W. T. & Flannery, B. P., 1992, Numerical Recipes in C: the art of scientific computing, Second Edition, Cambridge University Press
Price, M. P., & Rush, W. E., 1979, Am. J. Phys., 47, 531
Ribas, I., Font-Ribera, A. & Beaulieu, J., 2008, ApJ, 677, 59
Short, D., Welsh, W. F., Orosz, J. A. & Windmiller G., 2008, astro-ph:0803.2935
Simon, A., Szatmáry, K. & Szabó, Gy. M., 2007, A&A, 470, 727
Steffen, J. H. & Agol, E., 2007, ASP Conf. Ser, 366, 158
Wald, R. M., 1984 General Relativity, The University of Chicago Press
Mean longitude at the transit instances {#appendixlambdaderiv}
=======================================
The derivation of goes as follows. According to Kepler’s equation, $E-e\sin E=M=\lambda-\omega$, one can write $\lambda=\omega+E-e\sin E$. The only thing what is to be done is to calculate the eccentric anomaly $E$ for the instance when the orbiting body intersect the semi-line with the argument angle $\varphi_0$. The latter means that the true anomaly $v$ of the body is $v=\varphi_0-\omega$, by definition. The relation between the eccentric and true anomaly is $$\tan\frac{E}{2}=\sqrt{\frac{1-e}{1+e}}\tan{\frac{v}{2}},$$ which is equivalent with $$\begin{aligned}
\cos E & = & \frac{e+\cos v}{1+e\cos v}, \label{coseatt}\\
\sin E & = & \frac{\sqrt{1-e^2}\sin v}{1+e\cos v}. \label{sineatt}\end{aligned}$$ Using the addition theorem, the sine and cosine of the angle $\omega+E$ can be written as: $$\begin{aligned}
\cos(\omega+E) & = & \cos\omega\frac{e+\cos(\varphi_0-\omega)}{1+e\cos(\varphi_0-\omega)} - \nonumber \\
& & - \sin\omega\frac{\sqrt{1-e^2}\sin(\varphi_0-\omega)}{1+e\cos(\varphi_0-\omega)}, \\
\sin(\omega+E) & = & \sin\omega\frac{e+\cos(\varphi_0-\omega)}{1+e\cos(\varphi_0-\omega)} + \nonumber \\
& & + \cos\omega\frac{\sqrt{1-e^2}\sin(\varphi_0-\omega)}{1+e\cos(\varphi_0-\omega)}.\end{aligned}$$ Thus, the mean longitude itself is going to be $$\begin{aligned}
\lambda & = & \omega+E-e\sin E=\arg\left[\cos(\omega+E),\sin(\omega+E)\right]- \nonumber \\
& & - e\frac{\sqrt{1-e^2}\sin v}{1+e\cos v}.\end{aligned}$$ If both arguments of the above $\arg[\cdot,\cdot]$ function is multiplied by the always positive common denominator $1+e\cos(\varphi_0-\omega)$, one gets after some simplification: $$\begin{aligned}
\omega+E & = & \arg\left[k+\cos\varphi_0+\frac{h(k\sin\varphi_0-h\cos\varphi_0)}{1+\sqrt{1-e^2}}, \right. \nonumber \\
& & \left.h+\sin\varphi_0-\frac{k(k\sin\varphi_0-h\cos\varphi_0)}{1+\sqrt{1-e^2}}\right].\end{aligned}$$ Putting all terms together and replacing the appropriate terms by $e_\perp=k\sin\varphi_0-h\cos\varphi_0$, $e_\parallel=k\cos\varphi_0+h\sin\varphi_0$ and $\ell=1-\sqrt{1-e^2}$, we get . The partial derivatives of become $$\begin{aligned}
\frac{\partial\lambda}{\partial k} & = & -\frac{h}{2-\ell}-(1-\ell)\frac{h+(2+e_
\parallel)\sin\varphi_0}{(1+e_\parallel)^2}, \label{e:lambda_k}\\
\frac{\partial\lambda}{\partial h} & = & +\frac{k}{2-\ell}+(1-\ell)\frac{k+(2+e_
\parallel)\cos\varphi_0}{(1+e_\parallel)^2}. \label{e:lambda_h}\end{aligned}$$
Tangential velocity and position at the transit {#appendixtangentvelocity}
===============================================
It is known from the theory of the two-body problem that the angular momentum of a body orbiting around a mass of $GM=\mu$ and having an orbit with the semimajor axis of $a$ and eccentricity $e$ is $C=\sqrt{\mu a(1-e^2)}$. Since $C=rv_{\rm tan}$ for all points, the tangential velocity would be $$v_{\rm tan}=\frac{C}{r}=\sqrt{\mu a(1-e^2)}\frac{1+e\cos(\varphi_0-\omega)}{a(1-e^2)}\label{vtran1}$$ Using Kepler’s Third Law, i.e. $\mu=n^2a^3$, the above equation can be reordered to $$v_{\rm tan}=an\frac{1+e\cos(\varphi_0-\omega)}{\sqrt{1-e^2}}.\label{vtran2}$$ For $\varphi_0=\pi/2$, becomes $$v_{\rm tan}=an\frac{1+e\sin\omega}{\sqrt{1-e^2}}.$$
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
[^3]: See http://exoplanet.eu for up to date information
[^4]: i.e. at the center of a transit
[^5]: Note that since the measurement of the TDV effect relies on fitting 2 parameters, instead of 3 parameters for the TTV effect, the confidence corresponds to a higher confidence level for the TDV effect.
| {
"pile_set_name": "ArXiv"
} |
---
address:
- 'Dipartimento di Fisica, Università di Torino, 10125 Torino, Italy'
- 'INFN, Sezione di Torino, 10125 Torino, Italy'
- 'Departamento de Física Atómica, Molecular y Nuclear and Instituto de Física Teórica y Computacional Carlos I, Universidad de Granada, 18071 Granada, Spain'
- 'Departamento de Física Atómica, Molecular y Nuclear, Universidad de Sevilla, 41080 Sevilla, Spain'
- 'Center for Theoretical Physics, Laboratory for Nuclear Science and Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA'
author:
- 'M.B. Barbaro'
title: 'The role of meson exchange currents in charged current (anti)neutrino-nucleus scattering'
---
[1,2]{}, , , , , ,
[1]{}
[2]{}
[3]{}
[4]{}
[5]{}
We present our recent progress in the description of neutrino-nucleus interaction in the GeV region, of interest for ongoing and future oscillation experiments. In particular, we discuss the weak excitation of two-particle-two-hole states induced by meson exchange currents in a fully relativistic framework. We compare the results of our model with recent measurements of neutrino scattering cross sections, showing the crucial role played by two-nucleon knockout in the interpretation of the data.
Introduction
============
The accurate understanding of neutrino-nucleus scattering in the GeV region is a challenging problem of many-body nuclear physics and a necessary input for particle physics studies related to neutrinos. The modeling of this reactionis indeed the largest and most complicated source of uncertainty in present neutrino long-baseline experiments (T2K, NOvA, MicroBooNE) and it will be a crucial limitation for the sensitivity of next generation neutrino oscillation experiments (Hyper-Kamiokande, DUNE). These experiments, as other ones already completed (MiniBooNE, SciBooNE, NOMAD, ArgoNeuT), aim at precision measurements of the oscillation parameters; most importantly, they search for signals of leptonic CP violation, which may help explaining the matter-antimatter asymmetry of the Universe. Being the detectors made of large volumes of complex nuclei (typically carbon, oxygen and argon), good control of neutrino-nucleus interactions is required in order to extract significant information on the neutrino physics. On the other hand, the same experiments can provide interesting informations on the nuclear dynamics, complementary to the ones obtained using electromagnetic and hadronic probes. This has motivated a very intense activity of nuclear theorists on the subject over the last few years. Comprehensive reports on the state of the art of the field can be found in several review articles [@GGZReview; @FZReview; @AHNReview; @UMReview].
One of the main difficulties in the analysis of the above mentioned experiments is the lack of precise knowledge of the incoming neutrino energy $E_\nu$, which is broadly distributed around an average value. As a consequence, $E_\nu$ can only be reconstructed from the outgoing particle(s) kinematics; this procedure is strongly dependent upon the model adopted to describe the nucleus in its initial and final states, on the treatment of final-state interactions (FSI) between the outgoing nucleon(s) and the residual nucleus, as well as on the assumptions about the different mechanisms contributing to the observed cross section. The theoretical description of these effects is a complicated many-body problem. A wide variety of models has been proposed to describe CCQE (Charged-Current-Quasi-Elastic) neutrino and antineutrino cross sections, identified experimentally by the absence of pions in the final state. These models rely on quite different hypotheses and approximations and utilize diverse theoretical frameworks: Relativistic Mean Field Theory [@Gonzalez], Random-Phase-Approximation [@Martini; @Nieves; @Ghent], Relativistic Green Function [@Giusti], Spectral Function Formalism[@Benhar], Green Function Monte Carlo [@Schiavilla], Coherent Density Fluctuation Model [@Antonov], Transport Theory [@Mosel]. Despite these differences, there is now a general consensus on the important role played by two-particle-two-hole (2p2h) excitations in the interpretation of the CCQE data. Such excitations are induced by two-body currents, hence they go beyond the usual impulse approximation scheme, in which the probe interacts with one single nucleon. Nevertheless, due to the above mentioned definition of CCQE events in neutrino experiments (no pions in the final state), they also contribute to the observed cross section.
Though the present precision of neutrino experimental data is generally too low to discriminate between different models, valuable informations on the reliability of the latter - in a given kinematical regime - are provided by the large amount of high quality electron scattering data. The validation against these data represents a mandatory (albeit not sufficient) test any model should pass before being applied to neutrino scattering reactions, where other sources of uncertainties are present – [*e.g.*]{} the limited knowledge of the elementary neutrino-nucleon form factors. In order to fulfill this requirement, we have developed a nuclear model based on some general properties of electron scattering data and known as SuSA (Super-Scaling-Approach). The model has been first introduced in Ref. [@Amaro:2004bs] and subsequently refined by including effects from Relativistic Mean Field (RMF) theory and adding the contribution of meson-exchange-currents (MEC) in the 2p2h sector: the resulting model (called SuSAv2-MEC) has been recently validated versus electron scattering data in a wide range of kinematical conditions [@Megias:2016lke]. The model is capable of describing different kinematical regions: quasielastic, 2p2h and inelastic. For the neutrino case the latter is presently limited to the $\Delta$-resonance excitation, while for electron scattering the full inelastic spectrum is considered. Here we will focus mainly on the treatment of the 2p2h region, which occurs for transferred energies between the QE and $\Delta$ peaks.
The present contribution is organized in the following way: in Section \[sec:formalism\] we briefly define the formalism needed for studying CC neutrino and antineutrino scattering, with particular reference to the 2p2h contribution. In Section \[sec:results\] we present some selected results and compare our predictions with experimental data. Finally, in Section \[sec:conclusions\] we summarize our work and outline future developments.
Formalism {#sec:formalism}
=========
In this Section we summarize the essential formalism for $(\nu_l,l^-)$ and antineutrino $(\overline{\nu}_l,l^+)$ CC reactions in nuclei. Let $K^\mu=(\epsilon,{{\bf k}})$, $K'{}^\mu=(\epsilon',{{\bf k}}')$ be the incident and scattered lepton four-momenta, respectively, $Q^{\mu}=(\omega,{{\bf q}})=(K-K')^{\mu}$ the four-momentum transfer, $\omega$ and ${{\bf q}}$ the energy and momentum transfer. The $z$ direction is chosen along the vector ${{\bf q}}$. The double-differential cross section is $$\frac{d\sigma}{d\Omega'd\epsilon'}
= \sigma_0 {\cal S_{\pm}},
\label{sigma}$$ where $\Omega'$ is the scattering solid angle and $\sigma_0$ is a kinematical factor including the weak couplings (see Appendix). The nuclear structure function ${\cal S_{\pm}}$ is the linear combination of five response functions $${\cal S_{\pm}}
=
V_{CC} R^{CC}
+ 2 V_{CL} R^{CL}
+ V_{LL} R^{LL}
+ V_{T} R^{T}
\pm 2 V_{T'} R^{T'},
\label{Spm}
$$ where the sign of the last term is positive for neutrinos and negative for antineutrinos. The $V_K$’s, whose expressions are given in the Appendix, are kinematical factors depending on the lepton’s kinematics. The five response functions $$\begin{aligned}
R^{CC} &=& W^{00} \label{rcc} \\
R^{CL} &=& -\frac12\left(W^{03}+ W^{30}\right) \\
R^{LL} &=& W^{33} \\
R^{T} &=& W^{11}+ W^{22} \\
R^{T'} &=& -\frac{i}{2}\left(W^{12}- W^{21}\right) \label{rtp}\end{aligned}$$ embody the nuclear structure and dynamics and are given in terms of components of the hadronic tensor $$W^{\mu\nu} = \overline{\sum_i} \sum_f \langle f | {\hat J}^\mu(Q)| i \rangle^*
\langle f | {\hat J}^\nu(Q)| i \rangle\, \delta(E_i+\omega-E_f) .$$ In the above expression ${\hat J}^\mu(Q)$ represents the nuclear many-body current operator, the initial and final nuclear states $|i\rangle$ and $|f\rangle$ are exact eigenstates of the nuclear Hamiltonian, with energies $E_i$ and $E_f$, respectively, and the symbol $\overline{\sum}_i$ means average over initial states. This form is very general and includes all possible final states that can be reached through the action of the current operator ${\hat J}^\mu(Q)$ on the exact ground state. Thus the hadronic tensor can be expanded as the sum of one-particle one-hole (1p1h), two-particle two-hole (2p2h), plus additional channels: $$W^{\mu\nu} = W^{\mu\nu}_{1p1h} + W^{\mu\nu}_{2p2h} + \cdots$$ Here we focus on 2p2h excitations, within a model which will be shortly presented in the next subsection.
2p2h excitations {#sec:MEC}
----------------
In order to evaluate the 2p2h hadronic tensor $W^{\mu\nu}_{2p2h}$, we choose to describe the nuclear ground state as a relativistic Fermi gas (RFG), characterized by a Fermi momentum $k_F$. The main justification for this choice, which is undoubtedly too simple to encompass all aspects of nuclear dynamics, is that we are interested in a kinematical region where typical energies are of the order of or higher than the nucleon mass and therefore relativistic effects cannot be neglected. The RFG model is one of the few nuclear models in which Lorentz covariance can be maintained.
![ Feynman diagrams of the MEC considered in the present study, including the seagull (a,b), pion-in-flight (c), pion-pole (d,e), and $\Delta$ pole (f–i). []{data-label="diagrams"}](diagmec2.pdf)
In the RFG, the 2p2h channel corresponds to states with two nucleons of momenta ${{\bf p}}'_1$ and ${{\bf p}}'_2$ above the Fermi momentum, $p'_i>k_F$, and two hole states of momenta ${{\bf h}}_1$ and ${{\bf h}}_2$ below the Fermi momentum, $h_i<k_F$. The 2p2h hadronic tensor is given by $$\begin{aligned}
W^{\mu\nu}_{2p2h}
&=&
\frac{V}{(2\pi)^9}\int
d^3p'_1
d^3h_1
d^3h_2
\frac{M^4}{E_1E_2E'_1E'_2} \Theta(p'_1,p'_2,h_1,h_2)
\nonumber \\
&\times&
r^{\mu\nu}({{\bf p}}'_1,{{\bf p}}'_2,{{\bf h}}_1,{{\bf h}}_2)
\delta(E'_1+E'_2-E_1-E_2-\omega) ,
\label{hadronic}\end{aligned}$$ where $V$ is the volume of the system, $\bf p'_2= h_1+h_2+q-p'_1$ is fixed by momentum conservation, $M$ is the nucleon mass, the energies $E_i$ and $E'_i$ are the on-shell energies of the holes and particles, and $$\Theta(p'_1,p'_2,h_1,h_2)
\equiv
\theta(p'_2-k_F)
\theta(p'_1-k_F)
\theta(k_F-h_1)
\theta(k_F-h_2).
$$ By exploiting the energy $\delta$-function and rotational symmetry the above expression can be reduced to a 7-dimensional integral, to be computed numerically. The non-trivial part of the calculation is contained in the function $r^{\mu\nu}$, which represents the elementary hadronic tensor for the basic 2p2h transition, with given initial and final momenta, summed over spin ($s_i$, $s'_i$) and isospin ($t_i$, $t'_i$) projections: $$r^{\mu\nu}({{\bf p}}'_1,{{\bf p}}'_2,{{\bf h}}_1,{{\bf h}}_2)=
\frac{1}{4}\sum_{s_1s_2s'_1s'_2}
\sum_{t_1t_2t'_1t'_2}
j^{\mu}(1',2',1,2)^*_A
j^{\nu}(1',2',1,2)_A .
\label{elementary}$$ In the above, $j^{\mu}(1',2',1,2)_A$ is the antisymmetrized matrix element of the two-body weak MEC, containing vector and axial components. The present model is the extension to the axial sector of the calculation of Ref. [@Arturo]. We work at tree level and include only one-pion exchange. Then the MEC operator can be written as the sum of four contributions, denoted as seagull, pion-in-flight, pion-pole and Delta-pole and represented diagrammatically in Fig. \[diagrams\]. The explicit expression are too long to be reported here and can be found in Ref. [@Simo:2016ikv].
The sum over isospin in Eq. (\[elementary\]) combines all the possible charge channels in the final state, corresponding to emission of $pp$, $nn$ and $pn$ pairs. CC neutrino scattering can induce two possible 2p2h transitions, $np \rightarrow pp$ and $nn \rightarrow np$, while in the case of antineutrino the allowed channels $np\rightarrow nn$ and $pp\rightarrow np$. In our formalism, it is possible to separate the contributions of these charge states to the response functions, namely to know how many pairs of each kind ($pp$, $nn$ and $pn$) participate to this specific process.
Results {#sec:results}
=======
Let us first present the predictions of our MEC model for the different response functions. We consider $\nu_\mu$-$^{12}$C scattering and typical kinematics of interest for current neutrino oscillation experiments.
![Comparison between 1p1h and 2p2h response functions for CC neutrino scattering off $^{12}$C for two values of the momentum transfer [@Simo:2016ikv].[]{data-label="fig1"}](c12h9ob1.pdf)
In Fig. \[fig1\] we show the 1p1h and 2p2h responses as functions of the energy transfer $\omega$ for $q$ = 800 and 1200 MeV/c. The 1p1h responses are computed in the RFG and only contain the one-body current. For these values of $q$ the MEC effects in the $T$ and $T'$ channels are large, the 2p2h strength at the maximum of the peak being around 1/2 of the 1p1h response. Moreover, the MEC effects are similar in the $T$ and $T'$ responses. The $CC$ response appears to be extremely sensitive to MEC effects, which are even larger than the one-body response. However, the contribution of this response to the cross section and the ones associated to the $CL$ and $LL$ channels largely cancel out, so that the net charge/longitudinal cross section is generally smaller than the transverse ones. The balance of the different response functions of course depends on the kinematics.
![ Doubly differential 2p2h neutrino cross section per neutron of $^{12}$C, for fixed muon scattering angle and for three neutrino energies, as a function of the muon kinetic energy. The separate $np$ and $pp$ channels are shown [@Simo:2016ikv]. []{data-label="fig7"}](pap7.pdf)
The separate $pp$ and $np$ emission channels in the differential neutrino cross section are shown in fig. \[fig7\] for three different values of the neutrino energy. Proton-proton final states clearly dominate the 2p2h cross section. The $pp/np$ ratio is around 5-6 near the maximum, but its precise value depends on the kinematics. Note that the $np$ distribution is shifted towards higher muon energies respect to the $pp$ one.
![MiniBooNE flux-averaged CCQE $\nu_\mu$-$^{12}$C ($\bar\nu_\mu$-$^{12}$C) differential cross section per nucleon as a function of the muon scattering angle (top panels) and of the muon kinetic energy (bottom panels). The left panels correspond to neutrino cross sections and the right ones to antineutrino reactions. Data are from [@AguilarArevalo:2010zc; @AguilarArevalo:2013hm].[]{data-label="CS_single"}](dsdcosMB_newfitMEC.pdf "fig:")![MiniBooNE flux-averaged CCQE $\nu_\mu$-$^{12}$C ($\bar\nu_\mu$-$^{12}$C) differential cross section per nucleon as a function of the muon scattering angle (top panels) and of the muon kinetic energy (bottom panels). The left panels correspond to neutrino cross sections and the right ones to antineutrino reactions. Data are from [@AguilarArevalo:2010zc; @AguilarArevalo:2013hm].[]{data-label="CS_single"}](dsdcosMB_bar_newfitMEC.pdf "fig:")\
![MiniBooNE flux-averaged CCQE $\nu_\mu$-$^{12}$C ($\bar\nu_\mu$-$^{12}$C) differential cross section per nucleon as a function of the muon scattering angle (top panels) and of the muon kinetic energy (bottom panels). The left panels correspond to neutrino cross sections and the right ones to antineutrino reactions. Data are from [@AguilarArevalo:2010zc; @AguilarArevalo:2013hm].[]{data-label="CS_single"}](dsdTmuMB_newfitMEC.pdf "fig:")![MiniBooNE flux-averaged CCQE $\nu_\mu$-$^{12}$C ($\bar\nu_\mu$-$^{12}$C) differential cross section per nucleon as a function of the muon scattering angle (top panels) and of the muon kinetic energy (bottom panels). The left panels correspond to neutrino cross sections and the right ones to antineutrino reactions. Data are from [@AguilarArevalo:2010zc; @AguilarArevalo:2013hm].[]{data-label="CS_single"}](dsdTmuMB_bar_newfitMEC.pdf "fig:")
We now present some comparison with experimental data. In order to do that, we combine the present calculation of the 2p2h MEC contribution with the SuSAv2 model, used to describe the QE and $\Delta$-excitation regions. We refer to this as “SuSAv2-MEC” model. Due to the limited space we choose some representative examples. The interested reader can find more results in Ref. [@Megias:2016fjk].
In Fig. \[CS\_single\] the MiniBooNE flux-averaged CCQE $\nu_\mu (\overline{\nu}_\mu)-^{12}$C differential cross section per nucleon is shown as a function of the muon scattering angle (top panels) and the muon kinetic energy (bottom panels). Panels on the left (right) correspond to neutrinos (antineutrinos). The mean beam energy in the MiniBooNE experiment is 0.788 GeV in the neutrino mode and 0.665 GeV in the antineutrino mode. As shown, the present model provides an excellent representation of the experimental cross section. In all of the cases the MEC contribution is essential in order to reproduce the data and it amounts to about 20-25% of the total response for neutrino, 15-20% for antineutrino. Analogous results [@Megias:2016fjk], not shown here, are found for the CCQE cross section measured in the T2K experiment, where the neutrino energy is close to the MiniBooNE one, but the flux is less broadly distributed.
![Flux-folded $\nu_\mu-^{12}$C CCQE (upper panel) and $\bar\nu_\mu-$CH (lower panel) scattering cross section per target nucleon as a function of $Q^2_{QE}$ and evaluated in the SuSAv2 and SuSAv2+MEC models. MINERvA data are from [@MINERVA; @MINERVApriv].[]{data-label="Minerva_numu"}](MINERVA_nu.pdf "fig:")\
![Flux-folded $\nu_\mu-^{12}$C CCQE (upper panel) and $\bar\nu_\mu-$CH (lower panel) scattering cross section per target nucleon as a function of $Q^2_{QE}$ and evaluated in the SuSAv2 and SuSAv2+MEC models. MINERvA data are from [@MINERVA; @MINERVApriv].[]{data-label="Minerva_numu"}](MINERVA_nubar.pdf "fig:")
A similar trend is shown in Fig. \[Minerva\_numu\], where the MINERvA flux-averaged CCQE $\nu_\mu (\overline{\nu}_\mu)$ differential cross section per nucleon is displayed as a function of the reconstructed four-momentum $Q^2_{QE}$. The top panel refers to $\nu_\mu-^{12}$C whereas the bottom panel contains predictions and data for $\overline{\nu}_\mu-$CH [^1]. Note that the mean energy of the MINERvA flux is much higher than the MiniBooNE one, about 3.5 GeV for both $\nu_\mu$ and $\overline{\nu}_\mu$. Also in this case the contribution of the 2p2h MEC is needed in order to reproduce the experimental data.
![The CC-inclusive T2K flux-folded $\nu_e-^{12}$C differential cross section per nucleon evaluated in the SuSAv2+MEC model is displayed as a function of the electron momentum. The separate contributions of the QE, 1$\pi$ and 2p2h MEC are displayed [@Ivanov:2015aya]. The data are from [@T2Kinclelectron].[]{data-label="inclusive_T2K_e"}](T2K_electron_Pe.pdf)
In Fig. \[inclusive\_T2K\_e\] the inclusive electron-neutrino flux-averaged single differential cross section on $^{12}$C measured by the T2K experiment is shown as a function of the electron momentum. The neutrino energy is very similar to the of the MiniBooNE experiment. However, here the data are inclusive, namely only the final electron is observed and pions, or any other product of the reaction, can be present in the final state. We show the separate contributions corresponding to the QE response, the 2p2h MEC, pionic and the total response. The colored bands correspond to the estimated uncertainty in the scaling analysis extended to the $\Delta$ region (see [@Ivanov:2015aya]). Although the role associated with the $\Delta$ resonance is essential, the model is underestimating the data, especially for increasing values of the electron momentum. This indicates that other inelastic channels, not taken into account in the present description, may play a significant role in explaining these data. Work along this line is presently in progress.
Conclusions {#sec:conclusions}
===========
Two-particle-two-hole excitations induced by meson-exchange currents play a very important role in interpreting neutrino scattering data. We have illustrated this point by comparing the SuSAv2+MEC predictions with data on neutrino and antineutrino scattering off $^{12}$C corresponding to three different experiments: MiniBooNE, MINERvA and T2K.
We have also studied the separate charge channels contributing to this process and shown that for neutrino scattering $pp$ final states give a contribution five to six times larger than the $np$ ones. Having the separate isospin contributions will allow us to apply this formalism to asymmetric nuclei $N\ne Z$, of interest for neutrino experiments based on $^{40}$Ar.
Future developments will focus on the extension of the model to different nuclei and to the inelastic region, of paramount importance for future experiments.
Appendix {#appendix .unnumbered}
========
In this Appendix we provide the explicit expressions of some quantities used in the text.
The factor $\sigma_0$ appearing in Eq.(\[sigma\]) is $$\sigma_0=
\frac{G_F^2\cos^2\theta_c}{2\pi^2}
\left(k^\prime \cos\frac{\tilde\theta}{2}\right)^2 ,$$ where $G_F$ is the Fermi constant, $\theta_c$ the Cabibbo angle and $\tilde\theta$ the generalized scattering angle, defined as $$\tan^2 \frac{\tilde\theta}{2} = \frac{|Q^2|}{(\epsilon+\epsilon')^2-q^2} .$$
The kinematical factors in Eq.(\[Spm\]) are given by $$\begin{aligned}
V_{CC}
&=&
1-\delta^2\tan^2\frac{\tilde{\theta}}{2}
\label{vcc}\\
V_{CL}
&=&
\frac{\omega}{q}+\frac{\delta^2}{\rho'}\tan^2\frac{\tilde{\theta}}{2}
\\
V_{LL}
&=&
\frac{\omega^2}{q^2}+
\left(1+\frac{2\omega}{q\rho'}+\rho\delta^2\right)\delta^2
\tan^2\frac{\tilde{\theta}}{2}
\\
V_{T}
&=&
\tan^2\frac{\tilde{\theta}}{2}+\frac{\rho}{2}-
\frac{\delta^2}{\rho'}
\left(\frac{\omega}{q}+\frac12\rho\rho'\delta^2\right)
\tan^2\frac{\tilde{\theta}}{2}
\\
V_{T'}
&=&
\frac{1}{\rho'}
\left(1-\frac{\omega\rho'}{q}\delta^2\right)
\tan^2\frac{\tilde{\theta}}{2}\ .
\label{vtp}\end{aligned}$$ In Eqs. (\[vcc\]–\[vtp\]) we have defined $$\begin{aligned}
\delta &=& \frac{m_l}{\sqrt{|Q^2|}}\\
\rho &=& \frac{|Q^2|}{q^2}\\
\rho' &=& \frac{q}{\epsilon+\epsilon'}\ ,\end{aligned}$$ where $m_l$ is the final charged lepton mass.
[99]{}
H. Gallagher, G. Garvey, G. Zeller, *Ann. Rev. Nuc. Part. Sci.* **61** (2011), 355.
J. Formaggio, G. Zeller, *Rev. Mod. Phys.* **84** (2012), 1307.
L. Alvarez-Ruso, Y. Hayato and J. Nieves, *New J. Phys.* **16** (2014) 075015.
U. Mosel, *Ann. Rev. Nuc. Part. Sci.* **66** (2016), 1.
R. Gonzalez-Jimenez, G.D. Megias, M.B. Barbaro, J.A. Caballero and T.W. Donnelly, *Phys. Rev. C* **90** (2014), 035501.
M. Martini, M. Ericson, G. Chanfray and J. Marteau, *Phys. Rev. C* **81** (2010), 045502.
R. Gran, J. Nieves, F. Sanchez and M.J. Vicente Vacas, *Phys. Rev. D* **88** (2013), 113007.
N. Jachowicz, K. Heyde, J. Ryckebusch and S. Rombouts, *Phys. Rev. C* **65** (2002), 025501.
A. Meucci, C. Giusti and F.D. Pacati, *Nucl. Phys. A* **739** (2004), 277.
O. Benhar, N. Farina, H. Nakamura, M. Sakuda and R. Seki, *Phys. Rev. D* **72** (2005), 053005.
A. Lovato, S. Gandolfi, J. Carlson, S.C. Pieper and R. Schiavilla, *Phys. Rev. Lett.* **112** (2014), 182502.
M.V. Ivanov, A.N. Antonov, J.A. Caballero, G.D. Megias, M.B. Barbaro, E.Moya de Guerra and J.M. Udías, *Phys. Rev. C* **89** (2014), 014607.
T. Leitner, L. Alvarez-Ruso and U. Mosel, *Phys. Rev. C* **73** (2006), 065502.
J.E. Amaro, M.B. Barbaro, J.A. Caballero, T.W. Donnelly, A. Molinari and I. Sick, *Phys. Rev. C* **71** (2005), 015501.
A. De Pace, M. Nardi, W.M. Alberico, T.W. Donnelly and A. Molinari, *Nucl. Phys. A* **726** (2003), 303.
G.D. Megias, J.E. Amaro, M.B. Barbaro, J.A. Caballero and T.W. Donnelly, *Phys. Rev. D* **94** (2016), 013012.
I. Ruiz Simo, J.E. Amaro, M.B. Barbaro, A. De Pace, J.A. Caballero and T.W. Donnelly, arXiv:1604.08423 \[nucl-th\].
A.A. Aguilar-Arevalo [*et al.*]{}, MiniBooNE Collaboration, journal = “Phys. Rev. D”, *Phys. Rev. D* **81** (2010), 092005.
A.A. Aguilar-Arevalo [*et al.*]{}, MiniBooNE Collaboration, journal = “Phys. Rev. D”, *Phys. Rev. D* **88** (2013), 032001.
G.D. Megias, J.E. Amaro, M.B. Barbaro, J.A. Caballero, T.W. Donnelly and I. Ruiz Simo, arXiv:1607.08565 \[nucl-th\]. G.A. Fiorentini [*et al.*]{}, MINERvA Collaboration, *Phys. Rev. Lett.* **111** (2013), 022502; L. Fields [*et al.*]{}, MINERvA Collaboration, *Phys. Rev. Lett.* **111** (2013), 022501.
K. McFarland, private communication.
L. Aliaga [*et al.*]{} \[MINERvA Collaboration\], arXiv:1607.00704 \[hep-ex\]. M.V. Ivanov, G.D. Megias, R. Gonzalez-Jimanez, O. Moreno, M.B. Barbaro, J.A. Caballero and T.W. Donnelly, *J. Phys. G* **43** (2016), 045101.
K. Abe [*et al.*]{}, T2K Collaboration, *Phys. Rev. Lett.* **113** (2014), 241803.
[^1]: These data correspond to a new analysis of the MINERvA collaboration [@MINERVApriv], based on the updated NuMI flux prediction [@Aliaga:2016oaz], and exceed by $\sim 20\%$ the ones published in [@MINERVA].
| {
"pile_set_name": "ArXiv"
} |
=1
Introduction {#intro}
============
The Jacobi groups are semidirect products of appropriate semisimple real algebraic groups of Hermitian type with Heisenberg groups [@bs; @ez; @LEE03; @TA99]. As unimodular, nonreductive, algebraic groups of Harish-Chandra type [@SA80], the Jacobi groups are intensively studied in mathematics [@gem; @bb; @TA90; @TA92; @TA99; @Y02; @Y08; @Y10]. The Siegel–Jacobi (partially bounded [@Y08; @Y10]) domains are homogeneous manifolds associated to the Jacobi groups by the generalized Harish-Chandra embedding. It was underlined [@csg; @SB14; @berr] that the Siegel–Jacobi disk is a nonsymmetric, reductive [@nomizu], quantizable [@Cah], Q-K. Lu Kähler manifold [@lu66], not Einstein with respect to the balanced metric [@arr; @don], with constant negative scalar curvature. In the present paper we investigate similar geometric properties for the Siegel–Jacobi ball.
As was already emphasized [@jac1; @sbj; @nou], the Jacobi group is relevant in several branches of physics, as: quantum mechanics, geometric quantization, quantum optics, nuclear structure, signal processing [@BERC08B; @KRSAR82; @Q90; @SH03]. In particular, the Jacobi group is responsible [@BERC09] for the squeezed states in quantum optics [@mandel]. Recently, new significant applications of the Jacobi group have been underlined in various domains as: quantum tomography [@marmo], quantum teleportation [@chi], Vlasov kinetic equation [@gbt], general symmetry methods for analyzing solutions of differential equations using Lie’s prolongation method [@hunz; @sep], establishing the link between the standard Gaussian distribution and the Siegel–Jacobi space [@mol]. Berezin’s quantization and Berezin symbols related to the Jacobi group have been also investigated [@ca1; @ca2].
The real Jacobi group of index $n$ is defined as $G^J_n({{\amsbb{R}}})={{\mathrm{H}}}_n({{\amsbb{R}}})\rtimes\operatorname{Sp}(n,{{\amsbb{R}}}) $, where ${{\mathrm{H}}}_n({{\amsbb{R}}})$ is the real Heisenberg group of real dimension $(2n+1)$ [@nou; @Y02]. Let $g=(M,X,k)$, $g'=(M',X',k')\in G^J_n({{\amsbb{R}}})$, where $X=(\lambda,\mu)\in{{\amsbb{R}}}^{2n}$ and $(X,k)\in {{\mathrm{H}}}_n({{\amsbb{R}}})$. Then the composition law in $G^J_n({{\amsbb{R}}})$ is $$\begin{gathered}
gg'=\big(MM',XM'+ X', k+k'+XM'JX'^t\big),\end{gathered}$$ where $$\begin{gathered}
J=\left(\begin{matrix} {{\mathbb{0}}_n}& {{\mathbbm{1}}_n}\\-{{\mathbbm{1}}_n}&
{{\mathbb{0}}_n}\end{matrix}\right) .\end{gathered}$$ Also it is considered the restricted real Jacobi group $G^J_n({{\amsbb{R}}})_0$, consisting only of elements of the form above, but $g=(M,X)$.
To the Jacobi group $G^J_n({{\amsbb{R}}})$ it is associated the homogeneous manifold – the Siegel–Jacobi upper half-plane – ${{\mathcal{X}}}^J_n\approx{{\amsbb{C}}}^n \times
{{\mathcal{X}}}_n$, where the Siegel upper half-plane ${{\mathcal{X}}}_n= {{\mathrm{Sp}}}(n,{{\amsbb{R}}})/{{\mathrm{U}}}(n)$ is realized as $$\begin{gathered}
{{\mathcal{X}}}_n:=\{V\in M(n,{{\amsbb{C}}})\,| \,V=S+\ii R, \, S,
R\in M(n,{{\amsbb{R}}}),\, R>0, \,
S^t=S, \, R^t=R\} . \end{gathered}$$
The (complex version of the) Jacobi group of index $n$ is defined as $G^J_n={{\mathrm{H}}}_n\rtimes\operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$, where ${{\mathrm{H}}}_n$ denotes the $(2n+1)$-dimensional Heisenberg group [@sbj; @nou; @Y02]. The composition law is $$\begin{gathered}
(g_1,\alpha_1,t_1)(g_2,\alpha_2, t_2)= \big(g_1 g_2, g_2^{-1}\times \alpha_1+\alpha_2, t_1+ t_2 +\Im \big(g^{-1}_2\times\alpha_1\bar{\alpha}_2\big)\big),\end{gathered}$$ where $\alpha_i \in{{\amsbb{C}}}^n$, $t_i\in{{\amsbb{R}}}$ and $g_i\in\operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$, $i=1,2$ have the form $$\begin{gathered}
\label{dgM}
g = \left( \begin{matrix}p & q\\ \bar{q} & \bar{p}\end{matrix}\right), \qquad p,q\in M(n,{{\amsbb{C}}}),\end{gathered}$$ and $g\times\alpha = p \alpha + q \bar{\alpha} $, and $g^{-1}\times\alpha ={p}^*\alpha -q^t\bar{\alpha}$.
The Siegel–Jacobi ball, denoted ${{\mathcal{D}}}^J_n$ [@sbj], is the homogeneous manifold associated with the Jacobi group $G^J_n $, whose points are in ${{\amsbb{C}}}^n\times{{\mathcal{D}}}_n$, where ${{\mathcal{D}}}_n$ denotes the Siegel (open) ball. The non-compact hermitian symmetric space $ \operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}/\operatorname{U}(n)$ admits a matrix realization as a bounded homogeneous domain: $$\begin{gathered}
\label{dn}
{{\mathcal{D}}}_n:=\big\{W\in M (n, {{\amsbb{C}}})\colon W=W^t,\, N>0,\, N:={{\mathbbm{1}}_n}-W\bar{W} \big\}.\end{gathered}$$ ${{\mathcal{D}}}_n$ is a hermitian symmetric space of type CI (cf. [@helg Table V, p. 518]), identified with the symmetric bounded domain of type II, ${{\mathfrak{R}}}_{\text{II}}$ in Hua’s notation [@hua].
In [@JGSP; @sbj; @nou] we have attached coherent states (CS) [@perG] to the Jacobi group $G^J_n$ with support on the Siegel–Jacobi ball ${{\mathcal{D}}}^J_n$. The particular case of coherent states attached to the Jacobi group $G^J_1$ defined on the Siegel–Jacobi disk ${{\mathcal{D}}}^J_1$ has been investigated in [@jac1; @csg]. In the present paper we don’t use effectively the coherent states, but we use results obtained in geometry using the coherent state approach. The homogeneous Kähler two-form on ${{\mathcal{D}}}^J_n$, denoted $\omega_{{{\mathcal{D}}}^J_n}$, has been obtained in [@sbj; @nou] from the scalar product of two coherent state vectors, and this will be the starting point of the present investigation. The reproducing kernel function for the Siegel–Jacobi ball was obtained previously, see [@neeb p. 532] and references there. We recall that in [@jac1] we have shown that when expressed in variables on ${{\mathcal{X}}}_1$ obtained (via the partial Cayley transform) from variables on ${{\mathcal{D}}}^J_1$, the Kähler two-form $\omega_{{{\mathcal{X}}}^J_1}$ is the one determined by Berndt [@bern; @bs] and Kähler [@cal3; @cal]. Applying the partial Cayley transform, which we recall later in Proposition \[THETAS\], from $\omega_{{{\mathcal{X}}}^J_n}$ obtained in [@Y07], it was obtained the homogeneous Kähler two-form on the Siegel–Jacobi ball $\omega_{{{\mathcal{D}}}^J_n} $ – in fact, Yang considers a Jacobi group more general than $G^J_n$. In [@nou] we have shown that when expressed in appropriate variables, $\omega_{{{\mathcal{D}}}^J_n}$ is a particular case of that obtained by Yang in [@Y10]. In [@berr] we have underlined that the homogeneous metric corresponding to $\omega_{{{\mathcal{D}}}^J_1}$ is a balanced metric [@arr; @don]. In the present paper it is emphasized that the metric corresponding to $\omega_{{{\mathcal{D}}}^J_n}$ [@JGSP; @sbj; @nou] is the balanced metric. We also point out several geometric aspects related with Berezin quantization on the Siegel–Jacobi ball.
The paper is laid out as follows. In Section \[PRL1\] we recall several notions which are used in the paper. The notion of balanced metric on homogeneous Kähler manifolds is briefly recalled in Section \[PRL2\]. Also other notions which appear in connection with Berezin quantization, as quantizable manifold, Berezin kernel, CS-type group, Q.-K. Lu manifold and diastasis are briefly recalled. The homogenous metric on the Siegel upper half plane and Siegel ball, which we need later, are recalled in Section \[PRL3\]. We have considered useful to collect in Section \[PRL4\] several formulas referring to the differential geometry of the Siegel–Jacobi disk ${{\mathcal{D}}}^J_1$, a guide for the formulas deduced in the present paper for the Siegel–Jacobi ball ${{\mathcal{D}}}^J_n$. Section \[GMM\] contains the original results of the present paper. The balanced metric is obtained from the Kähler potential, calculated previously [@sbj] using the coherent states. Do to the symmetry of the variables $W\in M(n,{{\amsbb{C}}})$ describing points on the Siegel ball ${{\mathcal{D}}}_n$, we write down the matrix associated with the metric $h_{{{\mathcal{D}}}^J_n}(z,W)$ on the Siegel–Jacobi ball as a four-blocks matrix expressed in $(z,W)\in{{\amsbb{C}}}^n\times{{\mathcal{D}}}_n$, but also in a variable $\eta$ related with $(z,W)$ by $\eta=({{\mathbbm{1}}_n}-W\bar{W})^{-1}(z+W\bar{z})$. The significance of the change of coordinates ${\rm FC}\colon (\eta,W) \rightarrow (z,W)$ as a homogeneous Kähler diffeomorphism was underlined in [@nou] in the context of the [*fundamental conjecture*]{} [@DN] for homogeneous Kähler manifolds (Gindikin and Vinberg [@GV]). In order to determine the inverse of the metric matrix attached to the metric $h_{{{\mathcal{D}}}^J_n}$, we need the “inverse” of the metric matrix ${{{{\mathfrak{h}}}}}^k_{{{\mathcal{D}}}_n}$ of the Siegel ball, presented in Section \[PRL3\] – the complication comparatively with that on the noncompact Grassmann manifold [@gras] comes from the fact the matrices describing the points on Siegel ball are symmetric – see Lemma \[CUL\]. $(h^k_{{{\mathcal{D}}}_n})^{-1}$ is calculated in Proposition \[LLB\], where more details on the calculation of the Laplace–Beltrami operator on Siegel ball are given. Theorem \[mainTH\] contains the matrix of the metric of the Siegel–Jacobi ball and its “inverse” (in the sense of Lemma \[CUL\]). Proposition \[geoPL\] collects a description of the geometry of the Siegel–Jacobi ball from the point of view of Berezin’s quantization. In Section \[SCR\] it is shown that the scalar curvature of the Siegel–Jacobi ball is constant and negative. This is compatible with Theorem 4.1 in [@loi04] which asserts that a Kähler manifold that admits a regular quantization has constant scalar curvature. It is also shown that ${{\mathcal{D}}}^J_n$ is not an Einstein manifold with respect to the balanced metric, but it is one with respect to the Bergman metric. Section \[LBO\] contains the Laplace–Beltrami operator on ${{\mathcal{D}}}^J_n$. Remark \[equivv\] essentially asserts that if the metric on the homogeneous space in $M = G/H$ is invariant to the action of $G$ on $M$, then the same is true for the Laplace–Beltrami operator, a fact used in the proof of Theorem \[mainTH\]. In Appendix \[APP1\] we reproduce also a lemma which calculates the invariant volume of the Siegel–Jacobi ball used in [@sbj]. In Appendix \[APP2\] already mentioned are presented auxiliary calculations to the paper [@maa] of Maass and a remark concerning the results obtained in [@hua59] referring to the Laplace–Beltrami operator on the Siegel ball and Siegel upper half plane.
The main new results of the present paper are contained in Remark \[kra\], Lemma \[CUL\], Theorem \[mainTH\], Propositions \[geoPL\], \[mainPR\] and \[incaun\]. They generalize to the Siegel–Jacobi ball the results recalled in Section \[PRL4\] obtained for the Siegel–Jacobi disk.
**Notation.** We denote by $\amsbb{R}$, $\amsbb{C}$, $\amsbb{Z}$, and $\amsbb{N}$ the field of real numbers, the field of complex numbers, the ring of integers, and the set of non-negative integers, respectively. We denote the imaginary unit $\sqrt{-1}$ by $\ii$, and the real and imaginary part of a complex number by $\Re$ and respectively $\Im$, i.e., we have for $z\in{{\amsbb{C}}}$, $z=\Re z+\ii \Im z$, and $\bar{z}= cc(z)= \Re z-\ii \Im z$. denotes the set of all $m\times n $ matrices with entries in the field $\amsbb{F}$. $M(n\times 1,\amsbb{F})$ is identified with $\amsbb{F}^n$. Set $M(n,\amsbb{F}):=M(n\times n,\amsbb{F})$. For any $A\in M_{n}(\amsbb{F})$, $A^{t}$ denotes the transpose matrix of $A$. For $A\in M_{n}(\amsbb{C})$, $\bar{A}$ denotes the conjugate matrix of $A$ and $A^{*}=\bar{A}^{t}$. For $A\in M_n(\amsbb{C})$, the inequality $A>0$ means that $A$ is positive definite. The identity matrix of degree $n$ is denoted by ${{\mathbbm{1}}_n}$ and ${{\mathbb{0}}_n}$ denotes the $M(n,\amsbb{F})$-matrix with all entries $0$. We consider a complex separable Hilbert space ${{\mathfrak{H}}}$ endowed with a scalar product which is antilinear in the first argument, $(\lambda x,y)=\bar{\lambda}(x,y)$, $x,y\in{{\mathfrak{H}}}$, $\lambda\in{{\amsbb{C}}}\setminus 0$. If $A$ is a linear operator, we denote by $A^{\dagger}$ its adjoint. If $\pi$ is representation of a Lie group $G$ on the Hilbert space ${{{{\mathfrak{H}}}}}$ and ${{\mathfrak{g}}}$ is the Lie algebra of $G$, we denote ${{\boldsymbol{X}}}:={\operatorname{d}}\pi(X)$ for $X\in{{\mathfrak{g}}}$. A complex analytic manifold is [*Kählerian*]{} if it is endowed with a Hermitian metric whose imaginary part $\omega$ has ${\operatorname{d}}\omega = 0$ [@helg]. We denote the action of a Lie group $G$ on the space $M$ by $G\times M\rightarrow M$. A coset space $M=G/H$ is [*homogenous Kählerian*]{} if it caries a Kählerian structure invariant under the group $G$ [@bo]. By a [*Kähler homogeneous diffeomorphism*]{} we mean a diffeomorphism $\phi\colon M\rightarrow N$ of homogeneous Kähler manifolds such that $\phi^*\omega_N=\omega_M$. We use Einstein convention that repeated indices are implicitly summed over. If the Kähler-two form has the local expression $$\begin{gathered}
\label{kall}
\omega_M(z)=\ii\sum_{\alpha,\beta=1}^n h_{\alpha\bar{\beta}} (z) {\operatorname{d}}z_{\alpha}\wedge
{\operatorname{d}}\bar{z}_{\beta}, \qquad h_{\alpha\bar{\beta}}= \bar{h}_{\beta\bar{\alpha}}= h_{\bar{\beta}\alpha},\end{gathered}$$ we denote $$\begin{gathered}
\label{corectt}
h^{\alpha\bar{\beta}}:=(h_{\alpha\bar{\beta}})^{-1},\end{gathered}$$ i.e., we have $$\begin{gathered}
\label{sum2}
h_{\alpha\bar{\epsilon}}h^{\epsilon \bar{\beta}}=\delta_{\alpha\beta}.\end{gathered}$$ In this paper we use the following expression for the Laplace–Beltrami operator on Kähler manifolds $M$ with the Kähler two-form : $$\begin{gathered}
\label{LPB}
\Delta_M(z):=\sum_{\alpha,\beta=1}^n(h_{\alpha\bar{\beta}})^{-1}\frac{{\partial}^2}{{\partial}\bar{z}_{\alpha}{\partial}{{z}_{\beta}}},\end{gathered}$$ cf., e.g., Lemma 3 in the Appendix of [@ber74] or in [@jost equation (5.2.15), p. 253], where the Laplace–Beltrami operator is $-\Delta_M(z)$ and the author uses the convention $h^{i\bar{j}}h_{k\bar{j}}=\delta_{ij}$ (see [@jost equation (5.2.14), p. 252]) instead of our notation , . This means that $h^{l\bar{k}}$ in Jost’s notation corresponds to $h^{k\bar{l}}=(h_{k\bar{l}})^{-1}$ in our notation.
Preliminaries {#PRL1}
=============
Balanced metric and Berezin quantization via coherent states {#PRL2}
------------------------------------------------------------
We consider a $G$-invariant Kähler two-form $\omega_M$ on the $2n$-dimensional homogeneous manifold $M=G/H$ derived from the Kähler potential $f(z,\bar{z})$ [@chern] $$\begin{gathered}
\label{Ter}
h_{\alpha\bar{\beta}}= \frac{{\partial}^2 f}{{\partial}{z}_{\alpha}{\partial}\bar{z}_{{\beta}}} .\end{gathered}$$
As was underlined in [@berr] for ${{\mathcal{D}}}^J_1$, the homogeneous hermitian metric determined in [@jac1; @sbj; @nou] is in fact a balanced metric, because it corresponds to the Kähler potential calculated as the product of two CS-vectors $$\begin{gathered}
\label{FK}
f(z,\bar{z})=\ln K_M(z,\bar{z}), \qquad K_M(z,\bar{z})=(e_{\bar{z}},e_{\bar{z}}).\end{gathered}$$ In the approach of Perelomov [@perG] to CS, it is supposed that there exists a continuous, unitary, irreducible representation $\pi$ of a Lie group $G$ on the separable complex Hilbert space ${{{{\mathfrak{H}}}}}$. [*The coherent vector mapping*]{} $\varphi$ is defined locally, on a coordinate neighborhood ${{\mathcal{V}}}_0\subset M$ (cf. [@sb6; @last]): $$\begin{gathered}
\varphi \colon \ M\rightarrow \bar{{{{{\mathfrak{H}}}}}}, \qquad \varphi(z)=e_{\bar{z}},\end{gathered}$$ where $ \bar{{{{{\mathfrak{H}}}}}}$ denotes the Hilbert space conjugate to ${{{{\mathfrak{H}}}}}$. The vectors $e_{\bar{z}}\in\bar{{{{{\mathfrak{H}}}}}}$ indexed by the points $z \in M $ are called [*Perelomov’s CS vectors*]{}. Using Perelomov’s CS vectors, we consider Berezin’s approach to quantization on Kähler manifolds with the supercomplete set of vectors verifying the Parceval overcompletness identity [@ber73; @ber74; @berezin; @ber75] $$\begin{gathered}
\label{PAR}
(\psi_1,\psi_2)_{{{\mathcal{F}}}_{K}}=\int_M (\psi_1,e_{\bar{z}}) (e_{\bar{z}},\psi_2) {\operatorname{d}}\nu_M(z,\bar{z}), \qquad \psi_1,\psi_2\in{{{{\mathfrak{H}}}}},\\
{\operatorname{d}}{\nu}_{M}(z,\bar{z})=\frac{\Omega_M(z,\bar{z})}{(e_{\bar{z}},e_{\bar{z}})},
\qquad \Omega_M:=\frac {1}{n!} \underbrace{\omega\wedge\cdots\wedge\omega}_{\text{$n$ times}}.\nonumber\end{gathered}$$ The reproducing kernel for the Hilbert space of holomorphic, square integrable functions with respect to the scalar product of the type was calculated via CS as the scalar product $K_M(z,\bar{w})=(e_{\bar{z}},e_{\bar{w}})$ [@jac1; @sbj; @nou].
On the other side, let us consider the weighted Hilbert space ${{{{\mathfrak{H}}}}}_f$ of square integrable holomorphic functions on $M$, with weight ${{\mbox{\rm e}}}^{-f}$ [@eng] $$\begin{gathered}
\label{HIF}
{{{{\mathfrak{H}}}}}_f=\left\{\phi\in\operatorname{hol}(M) \,| \int_M{{\mbox{\rm e}}}^{-f}|\phi|^2 \Omega_M <\infty \right\}.\end{gathered}$$
In order to identify the Hilbert space ${{{{\mathfrak{H}}}}}_f$ defined by with the Hilbert ${{{\mathcal{F}}}_{K}}$ space with scalar product , we have to consider the $\epsilon$-function [@cah+; @Cah; @raw] $$\begin{gathered}
\epsilon(z) = {{\mbox{\rm e}}}^{-f(z)}K_M(z,\bar{z}).\end{gathered}$$ If the Kähler metric on the complex manifold $M$ is obtained from the Kähler potential via , and is such that $\epsilon(z)$ is a positive constant, then the metric is called [*balanced*]{}. This denomination was firstly used in [@don] for compact manifolds, then it was used in [@arr] for noncompact manifolds and also in [@alo] in the context of Berezin quantization on homogeneous bounded domains.
The [*balanced hermitian metric*]{} of $M$ in local coordinates is $$\begin{gathered}
\label{herm}
{\operatorname{d}}s^2_M(z,\bar{z}) =\sum_{\alpha,\beta=1}^n \frac{{\partial}^2}{{\partial}z_{\alpha} {\partial}\bar{z}_{\beta}} \ln (K_M(z,\bar{z})) {\operatorname{d}}z_{\alpha}\otimes {\operatorname{d}}\bar{z}_{\beta} ,
\qquad K_M(z,\bar{z})=(e_{\bar{z}},e_{\bar{z}}).\end{gathered}$$
We recall that in [@cah+; @Cah; @raw] Berezin’s quantization via coherent states was globalized and extended to non-homogeneous manifolds in the context of geometric (pre-)quantization [@Kos]. To the Kähler manifold $(M,\omega)$, it is also attached the triple $\sigma =({{{{\mathcal{L}}}}},h,\nabla)$, where ${{{{\mathcal{L}}}}}$ is a holomorphic (prequantum) line bundle on $M$, $h$ is the hermitian metric on ${{{{\mathcal{L}}}}}$ and $\nabla$ is a connection compatible with metric and the Kähler structure [@SBS]. The manifold is called [*quantizable*]{} if the curvature of the connection [@chern] $F(X,Y)=\nabla_X\nabla_Y-\nabla_Y\nabla_X-\nabla_{[X,Y]}$ has the property that $F=-\ii \omega_M $, or ${\partial}\bar{{\partial}} \log \hat{h} =\ii \omega_M$, where $\hat{h}$ is a local representative of $h$. Then $\omega_M$ is integral, i.e., $c_1[{{\mathcal{L}}}]=[\omega_M]$. The reproducing (weighted Bergman) kernel admits the series expansion $$\begin{gathered}
\label{funck}
K_M(z,\bar{w})\equiv (e_{\bar{z}},e_{\bar{w}}) = \sum_{i=0}^{\infty}\varphi_i(z)\bar{\varphi}_i(w), \end{gathered}$$ where $\Phi= (\varphi_0,\varphi_1,\dots )$ is an orthonormal base with respect to the scalar product : $$\begin{gathered}
\int_M\bar{\varphi}_i\varphi_j\frac{\Omega_M}{K_M}=\delta_{ij}, \qquad i,j\in {{\amsbb{N}}}.\end{gathered}$$ The base $\Phi$ is finite-dimensional for compact manifolds $M$.
Let $\xi\colon {{{{\mathfrak{H}}}}}\setminus 0\rightarrow{{\amsbb{P}}}({{{{\mathfrak{H}}}}})$ be the canonical projection $\xi({{\boldsymbol{z}}})=[{{\boldsymbol{z}}}]$. The [*Fubini–Study metric*]{} in the nonhomogeneous coordinates $[z]$ is the hermitian metric on ${{\amsbb{CP}}}^{\infty}$ (see [@koba] for details) $$\begin{gathered}
\label{FBST}
{\operatorname{d}}s^2|_{\rm FS}({{\boldsymbol{[}}}{z}])= \frac{({\operatorname{d}}{{\boldsymbol{z}}},{\operatorname{d}}{{\boldsymbol{z}}}) ({{\boldsymbol{z}}},{{\boldsymbol{z}}})-({\operatorname{d}}{{\boldsymbol{z}}},{{\boldsymbol{z}}}) ({{\boldsymbol{z}}},{\operatorname{d}}{{\boldsymbol{z}}})}{({{\boldsymbol{z}}},{{\boldsymbol{z}}})^2}.\end{gathered}$$
It was proved [@raw] (see also [@berr]) that:
\[asaofi\]If $\epsilon(z)$ is constant on $M$, then the balanced Hermitian metric on $M$ is the pullback $$\begin{gathered}
\label{KOL}
{\operatorname{d}}s^2_M(z)=\iota_M^*{\operatorname{d}}s^2_{\rm FS}(z)= {\operatorname{d}}s^2_{\rm FS}(\iota_M(z))\end{gathered}$$ of the Fubini–Study metric via the embedding $$\begin{gathered}
\label{invers}
\iota_M\colon \ M\hookrightarrow {{\amsbb{CP}}}^{\infty},\qquad \iota_M(z) = [\varphi_0(z):\varphi_1(z):\dots ] .\end{gathered}$$ If $M$ is a compact manifold, the embedding is the Kodaira embedding.
If the homogeneous Kähler manifold $M=G/H$ to which we associate the Hilbert space of functions ${{\mathcal{F}}}_K$ with respect to the scalar product admits a holomorphic embedding $\iota_M \colon M \hookrightarrow {{{{\amsbb{P}}}({{\mathcal{H}^{\infty}}})}}$, then $M$ is called a CS-[*orbit*]{}, and $G$ is called a CS-[*type group*]{} [@sb6; @lis; @neeb].
We denote the [*normalized Bergman kernel*]{} ([*the two-point function*]{} of $M$ [@ber97; @SBS]) by $$\begin{gathered}
\label{kmic}
\kappa_M(z,\bar{z}'):=\frac{K_M(z,\bar{z}')}{\sqrt{K_M(z)K_M(z')}}=
(\tilde{e}_{\bar{z}},\tilde{e}_{\bar{z}'})=\frac{(e_{\bar{z}},e_{\bar{z}'})}{\|e_{\bar{z}}\|\|e_{\bar{z}'}\|} .\end{gathered}$$ The set $\Sigma_z:=\{ z'\in M \,|\,\kappa_M(z,\bar{z}')=0\} $ was called [@ber97; @SBS] [*polar divisor*]{} relative to $z\in M$, while a manifold for which $\Sigma_z=\varnothing$, $\forall\, z\in M$ was called in [@berr] a [*Q.-K. Lu manifold*]{}, extending to manifolds a denomination introduced for domains in ${{\amsbb{C}}}^n$ [@lu66]. Note that for a particular class of compact homogeneous manifolds that includes the hermitian symmetric spaces, $\Sigma_z$ is equal with the [*cut locus*]{} relative to $z\in M$ (see the definition of the cut locus, e.g., [@koba2 p. 100]), and $\Sigma_z$ is a [*divisor*]{} in the sense of algebraic geometry [@ber97; @SBS].
By [*Berezin kernel*]{} $b_M\colon M\times M\rightarrow [0,1]\in {{\amsbb{R}}}$, in this paper we mean: $$\begin{gathered}
b_M(z,z'):=|\kappa_M(z,\bar{z}')|^2.\end{gathered}$$ The [*Calabi’s diastasis*]{} [@calabi1] expressed via the CS-vectors [@cah+] reads $$\begin{gathered}
\label{DIA}
D_M(z,z'):= -\ln b_M(z,z') = -2\ln \left\vert(\tilde{e}_{\bar{z}},\tilde{e}_{\bar{z}'})\right\vert .\end{gathered}$$
Generalizing a theorem proved for homogeneous bounded domains [@alo], in [@LM15] Loi and Mossa have proved:
\[LMN\] Let $(M,\omega)$ be a simply-connected homogeneous Kähler manifold such that the associated Kähler form $\omega$ is integral. Then there exists a constant $\mu_0>0$ such that $M$ equipped with $\mu_0\omega$ is projectively induced.
With Theorem \[LMN\] and the recalled definitions, we can formulate the
\[kra\] Let $M=G/H$ be a simply-connected homogeneous Kähler manifold. Then the following assertions are equivalent:
1. $M$ is a quantizable Kähler manifold,
2. $M$ admits a balanced metric,
3. $M$ is CS-type manifold and $G$ is a CS-type group,
4. $M$ is projectively induced and we have , .
Loi and Mossa have proved [@LM15] the following
\[LMM\] Let $(M,\omega)$ be a homogeneous Kähler manifold. Then the following are equivalent:
1. $M$ is contractible,
2. $(M,\omega)$ admits a global Kähler potential,
3. $(M,\omega)$ admits a global diastasis $D_M\colon M\times M\rightarrow {{\amsbb{R}}}$,
4. $(M,\omega)$ admits a Berezin quantization.
The proof of the Theorem \[LMM\] is based on the fundamental conjecture for homogeneous bounded domains [@DN] and the asymptotic of the Bergman kernel [@eng].
Theorems \[LMN\], \[LMM\] and Remark \[kra\] will be used in the proof of Proposition \[geoPL\].
Elements of geometry of the Siegel upper half-plane and Siegel ball {#PRL3}
-------------------------------------------------------------------
Firstly we recall some standard facts about the symplectic group and its algebra, see references and details in [@sbj; @nou]. Even if we will be concerned mainly with the Jacobi group $G^J_n$, we recall some basic facts about Cayley and partial Cayley transform which will appear in the paper.
### The Cayley transform
If $X \in{{\mathfrak{sp}}}(n,{{\amsbb{R}}})$, i.e., $X^tJ+JX=0$, then $$\begin{gathered}
X=\left( \begin{matrix} a & b \\c & -a^t\end{matrix}\right),
\qquad\text{where} \quad b=b^t,\quad c=c^t, \quad a,b,c\in M(n,{{\amsbb{R}}}).\end{gathered}$$
If $ g=\left(\begin{matrix}a & b\\c& d\end{matrix}\right) \in {{\operatorname{Sp}}}(n,{{\amsbb{R}}})$, i.e., $g^tJg=J$, then the matrices $a,b,c,d\in M(n,{{\amsbb{R}}})$ have the properties $$\begin{gathered}
a^tc = c^ta, \!\!\qquad b^t d = d^t b, \!\!\qquad a^td - c^t b ={{\mathbbm{1}}_n},\!\! \qquad ab^t = ba^t, \!\!\qquad cd^t=dc^t, \!\!\qquad ad^t-bc^t ={{\mathbbm{1}}_n}. $$ We have the correspondence $$\begin{gathered}
\label{pana}
g=\left(\begin{matrix}a & b\\c & d\end{matrix}\right) \in M(2n,{{\amsbb{R}}})\leftrightarrow g_{{{\amsbb{C}}}} = {{\mathcal{C}}}^{-1}g{{\mathcal{C}}} = \left(\begin{matrix} p & q \\ \bar{q}
&\bar{p} \end{matrix}\right), \qquad p, q\in M(n,{{\amsbb{C}}}),\end{gathered}$$ where $$\begin{gathered}
{{\mathcal{C}}}=\left(\begin{matrix} \ii {{\mathbbm{1}}_n}& \ii {{\mathbbm{1}}_n}\\
-{{\mathbbm{1}}_n}& {{\mathbbm{1}}_n}\end{matrix}\right), \qquad {{\mathcal{C}}}^{-1}= \frac{1}{2}
\left(\begin{array}{cc} - \ii {{\mathbbm{1}}_n}& -{{\mathbbm{1}}_n}\\ -\ii {{\mathbbm{1}}_n}&
{{\mathbbm{1}}_n}\end{array}\right), \nonumber\\
2 a = p+q +\bar{p}+\bar{q} ,\qquad
2 b = \ii (\bar{p}-\bar{q}-p+q), \qquad 2 c = \ii (p+q-\bar{p}-\bar{q}), \nonumber\\
2 d = p-q +\bar{p}-\bar{q}, \qquad 2p = a+d+\ii (b-c), \qquad 2q= a-d -\ii (b+c). \label{CUCURUCU}
$$ To every $g\in \operatorname{Sp}(n,{{\amsbb{R}}})$, we associate via $g\mapsto g_{{{\amsbb{C}}}}\in \operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}\equiv\operatorname{Sp}(n, {{\amsbb{C}}})\cap {\rm U}(n,n)$ as in , where the matrices $p,q\in M(n,{{\amsbb{C}}})$ have the properties $$\begin{gathered}
\label{simplectic}
pp^*- qq^* = {{\mathbbm{1}}_n},\qquad pq^t=qp^t, \qquad p^*p-q^t\bar{q} = {{\mathbbm{1}}_n},\qquad p^t\bar{q}=q^*p . \end{gathered}$$
Let us consider an element $h=(g,l)$ in $G^J_n({{\amsbb{R}}})_0$, i.e., $$\begin{gathered}
\label{mM}
g=\left(\begin{matrix} a & b\\ c & d\end{matrix}\right)\in{\operatorname{Sp}}(n,{{\amsbb{R}}}),
\qquad l=(n,m)\in{{\amsbb{R}}}^{2n}, \end{gathered}$$ and $V\in{{\mathcal{X}}}_n$, $u\in{{\amsbb{C}}}^n\equiv{{\amsbb{R}}}^{2n}$.
Let $g\in \operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$ be of the form , and $\alpha, z\in{{\amsbb{C}}}^n$. The (transitive) action $(g,\alpha)\times(W,z)=(W_1,z_1)$ of the Jacobi group $G^J_n$ on the Siegel–Jacobi ball ${{\mathcal{D}}}^J_n$ is given by the formulas [@sbj] $$\begin{gathered}
W_1 =(pW+q)(\bar{q}W+\bar{p})^{-1}=(Wq^*+p^*)^{-1}(q^t+Wp^t),\nonumber\\
z_1 = (Wq^*+ p^*)^{-1}(z+ \alpha -W\bar{\alpha}).
\label{TOIU}\end{gathered}$$
Now we consider the partial Cayley transform [@sbj; @nou] $\Phi\colon {{\mathcal{X}}}^J_n\rightarrow {{\mathcal{D}}}_n^J$, $\Phi(V,u)=(W,z)$ $$\begin{gathered}
\label{bigtransf}
W = (V-\ii{{\mathbbm{1}}_n})(V+\ii{{\mathbbm{1}}_n})^{-1}, \qquad z = 2\ii (V+\ii{{\mathbbm{1}}_n})^{-1} u,$$ with the inverse partial Cayley transform $\Phi^{-1}\colon {{\mathcal{D}}}^J_n\rightarrow {{\mathcal{X}}}^J_n$, $\Phi^{-1}(W,z)=(V,u)$ $$\begin{gathered}
V = \ii ({{\mathbbm{1}}_n}-W ) ^{-1} ({{\mathbbm{1}}_n}+W ),\qquad
u = ({{\mathbbm{1}}_n}-W)^{-1}z. \label{big11}\end{gathered}$$ Let us define $\Theta\colon G^J_n({{\amsbb{R}}})_0\rightarrow (G^J_n)_0$, $\Theta(h)=h_*$, $h=(g,n,m)$, $h_*=(g_{{{\amsbb{C}}}},\alpha)$.
We have proved in [@nou Proposition 2] (see also [@gem; @Y08])
\[THETAS\] $\Theta$ is an group isomorphism and the action of $(G^J_n)_0$ on ${{\mathcal{D}}}^J_n$ is compatible with the action of $G^J_n({{\amsbb{R}}})_0$ on ${{\mathcal{X}}}^J_n$ through the biholomorphic partial Cayley transform , i.e., if $\Theta(h)=h_*$, then $\Phi h=h_*\Phi$. More exactly, if the action of $(G^J_n)_0$ on ${{\mathcal{D}}}^J_n$ is given by , then the action of $G^J_n({{\amsbb{R}}})_0$ on ${{\mathcal{X}}}^J_n$ is given by $(g,l)\times(V,u)\rightarrow (V_1,u_1)\in{{\mathcal{X}}}^J_n$, where $$\begin{gathered}
V_1 = (aV+b)(cV+d)^{-1} =(Vc^t+d^t)^{-1}(Va^t+b^t), \nonumber\\
u_1 = (Vc^t+d^t)^{-1}(u+Vn+m).\label{conf}$$ The matrices $g$ in and $g_{{{\amsbb{C}}}}$ in are related by , , while $\alpha=m+\ii n$, $m,n\in{{\amsbb{R}}}^n$.
### The metric
Siegel has determined the metric on ${{\mathcal{X}}}_n$, $\operatorname{Sp}(n,{{\amsbb{R}}})$-invariant to the action (see [@sieg equation (2)] or [@hua44 Theorem 3, p. 644]): $$\begin{gathered}
\label{msig}
{\operatorname{d}}s^2_{{{\mathcal{X}}}_n}(R)={\operatorname{Tr}}\big(R^{-1}{\operatorname{d}}V R^{-1} {\operatorname{d}}\bar{V}\big).\end{gathered}$$ With the Cayley transform and the relations $$\begin{gathered}
\label{scg1}
\dot{V} = 2\ii U \dot{W} U, \qquad U=({{\mathbbm{1}}_n}-W)^{-1},\qquad
2 R =({{\mathbbm{1}}_n}+W)U+({{\mathbbm{1}}_n}+\bar{W})\bar{U} ,\end{gathered}$$ introduced in , it is obtained the metric on ${{\mathcal{D}}}_n$, $\operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$-invariant to the action : $$\begin{gathered}
\label{mtrball}
{\operatorname{d}}s^2_{{{\mathcal{D}}}_n}(W)= 4{\operatorname{Tr}}(M{\operatorname{d}}W\bar{M}{\operatorname{d}}\bar{W}),\qquad W\in{{\mathcal{D}}}_n, \qquad M=({{\mathbbm{1}}_n}-W\bar{W})^{-1},\end{gathered}$$ associated with the Kähler two-form , modulo the factor $\frac{2}{k}$.
The metric can be written as in , .
Equation can be written down as $$\begin{gathered}
\tfrac{1}{4}{\operatorname{d}}s^2_{{{\mathcal{D}}}_n}(W) = \sum_{p,q,m,n}H_{pq\bar{m}\bar{n}}{\operatorname{d}}w_{pq}{\operatorname{d}}\bar{w}_{mn},\\
H_{pq\bar{m}\bar{n}}:=M_{np}M_{mq}, \qquad M:=({{\mathbbm{1}}_n}- W\bar{W})^{-1}.\end{gathered}$$ We have $$\begin{gathered}
\tfrac{1}{4}{\operatorname{d}}s^2_{{{\mathcal{D}}}_n}(W) =\sum_{p<q;\, m<n}(H_{pq\bar{m}\bar{n}}+H_{pq\bar{n}\bar{m}}+H_{qp\bar{m}\bar{n}}+H_{qp\bar{n}\bar{m}}){\operatorname{d}}w_{pq}{\operatorname{d}}\bar{w}_{mn} \nonumber\\
\hphantom{\tfrac{1}{4}{\operatorname{d}}s^2_{{{\mathcal{D}}}_n}(W) =}{} + \sum_{p<q}(H_{pq\bar{m}\bar{m}}+H_{qp\bar{m}\bar{m}}) {\operatorname{d}}w_{pq}{\operatorname{d}}\bar{w}_{mm}\nonumber\\
\hphantom{\tfrac{1}{4}{\operatorname{d}}s^2_{{{\mathcal{D}}}_n}(W) =}{} +\sum_{m<n}( H_{pp\bar{m}\bar{n}}+H_{pp\bar{n}\bar{m}}){\operatorname{d}}w_{pp}{\operatorname{d}}\bar{w}_{mn}+ \sum H_{pp\bar{m}\bar{m}}{\operatorname{d}}w_{pp}{\operatorname{d}}\bar{w}_{mm}.\label{SPS}\end{gathered}$$ For reasons which will be clarified further (see formulas , ), we write as $$\begin{gathered}
\label{NEMM}
\tfrac{1}{4}{\operatorname{d}}s^2_{{{\mathcal{D}}}_n}(W)= \sum_{p\le q; \, m\le n}h^k_{pqmn}{\operatorname{d}}w_{pq}{\operatorname{d}}\bar{w}_{mn}, \end{gathered}$$ where $$\begin{gathered}
\label{neM}
h^k_{pq\bar{m}\bar{n}} = 2M_{mp}M_{nq}(1-\delta_{pq})+2M_{mq}M_{np}(1-\delta_{mn})+
M^2_{mp}\delta_{pq}\delta_{mn},\end{gathered}$$ which proves the statement of the remark.
In order to take into account the symmetry of the matrix $W=(w_{ij})_{i,j=1,\dots,n}$, we introduce the notation (see [@nou equation (4.5)]): $$\begin{gathered}
\label{mircea}
\Delta^{ij}_{pq}:=\frac{{\partial}w_{ij}}{{\partial}w_{pq}} = \delta_{ip}\delta_{jq}+\delta_{iq}\delta_{jp}-\delta_{ij}\delta_{pq}\delta_{ip},
\qquad w_{ij}=w_{ji} .\end{gathered}$$
For calculation of the “inverse” $h^{-1}$ of the metric matrix $h$ of ${{\mathcal{D}}}^J_n$ – see Theorem \[mainTH\] – we need the “inverse” of the matrix , which we calculate in Proposition \[LLB\], exploiting the formula given in [@maa] for the Laplace–Beltrami operator on ${{\mathcal{X}}}_n$ with the result (see ): $$\begin{gathered}
\label{kpqmn}
k_{mn\bar{u}\bar{v}}=\tfrac{1}{2}(N_{vn}\bar{N}_{mu}+N_{vm}\bar{N}_{nu}), \qquad N={{\mathbbm{1}}_n}-W\bar{W}.\end{gathered}$$
We calculate $$\begin{gathered}
\label{BIGM}
E_{pq}^{uv}:=\sum_{m\leq n}h^k_{pq\bar{m}\bar{n}}k_{mn\bar{u}\bar{v}}.\end{gathered}$$
We shall prove a relation which gives the sense of the “inverse” matrix of , important in Theorem \[mainTH\]:
\[CUL\] The “inverse” matrix of the metric matrix on the Siegel ball ${{\mathcal{D}}}_n$ is the matrix and we have $$\begin{gathered}
\label{culmea}
E^{uv}_{pq}= \Delta^{uv}_{pq}.\end{gathered}$$
a\) Firstly, we consider the case $p=q$ in . We get successively $$\begin{gathered}
E^{uv}_{pp} = \sum_{m\le n}(2-\delta_{mn})({{\mathbbm{1}}_n}- W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}- \bar{W}W)^{-1}_{pn}k_{mn\bar{u}\bar{v}}\nonumber\\
\hphantom{E^{uv}_{pp}}{} = ({{\mathbbm{1}}_n}- W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}- \bar{W}W)^{-1}_{pm}({{\mathbbm{1}}_n}-W\bar{W})_{vm}({{\mathbbm{1}}_n}- \bar{W}W)_{mu}\nonumber\\
\hphantom{E^{uv}_{pp}=}{}+ \sum_{m<n}({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}-\bar{W}W)^{-1}_{pn} \nonumber\\
\hphantom{E^{uv}_{pp}=}{} \times [({{\mathbbm{1}}_n}-W\bar{W})_{vn}({{\mathbbm{1}}_n}-\bar{W}W)_{mu}+({{\mathbbm{1}}_n}-W\bar{W})_{vm}({{\mathbbm{1}}_n}-\bar{W}W)_{nu}]\nonumber\\
\hphantom{E^{uv}_{pp}}{} =({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}-\bar{W}W)^{-1}_{pm}({{\mathbbm{1}}_n}-W\bar{W})_{vm}({{\mathbbm{1}}_n}-\bar{W}W)_{mu}\nonumber\\
\hphantom{E^{uv}_{pp}=}{} +
\sum_{m<n}({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{np}({{\mathbbm{1}}_n}-\bar{W}W)^{-1}_{pm}({{\mathbbm{1}}_n}-W\bar{W})_{vn}({{\mathbbm{1}}_n}-\bar{W}W)_{mu}\nonumber\\
\hphantom{E^{uv}_{pp}=}{} +
\sum_{n<m}({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}-\bar{W}W)^{-1}_{pm}({{\mathbbm{1}}_n}-W\bar{W})_{vn}({{\mathbbm{1}}_n}-\bar{W}W)_{mu}\nonumber\\
\hphantom{E^{uv}_{pp}}{} =\delta_{up}\delta_{vp}.\label{PSLIT1}\end{gathered}$$ b) Now we consider the case $p\not=q$ in . We get successively $$\begin{gathered}
E^{uv}_{pq} = 2\!\sum_{m\le n}\!\! \big[({{\mathbbm{1}}_n}- W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}- \bar{W}W)^{-1}_{qn}\! +
({{\mathbbm{1}}_n}- W\bar{W})^{-1}_{mq}({{\mathbbm{1}}_n}- \bar{W}W)^{-1}_{pn}(1 - \delta_{mn})\big]k_{mn\bar{u}\bar{v}}\!\nonumber\\
\hphantom{E^{uv}_{pq}}{}
=2({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}-\bar{W}W)^{-1}_{qm}({{\mathbbm{1}}_n}-W\bar{W})_{vm}({{\mathbbm{1}}_n}-\bar{W}W)_{mu}\nonumber\\
\hphantom{E^{uv}_{pq}=}{}+\sum_{m<n}\big[({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{mp}({{\mathbbm{1}}_n}-\bar{W}W)^{-1}_{qn}+({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{mq}({{\mathbbm{1}}_n}-\bar{W}W)^{-1}_{pn}\big] \nonumber\\
\hphantom{E^{uv}_{pq}=}{} \times\big[({{\mathbbm{1}}_n}- W\bar{W})_{vn}({{\mathbbm{1}}_n}- \bar{W}W)_{vn}({{\mathbbm{1}}_n}- \bar{W}W)_{mu} + ({{\mathbbm{1}}_n}- W\bar{W})_{vm}({{\mathbbm{1}}_n}- \bar{W}W)_{nu}\big]\nonumber\\
\hphantom{E^{uv}_{pq}}{}= \delta_{pu}\delta_{vq}+\delta_{pv}\delta_{qu}.\label{PSLIT2}\end{gathered}$$ Putting together and , with formula $$\begin{gathered}
E^{pq}_{uv}=E^{pq}_{uv}(1-\delta_{pq})+E^{pq}_{uv}\delta_{pq},\end{gathered}$$ we have proved .
Geometry of the Siegel–Jacobi disk {#PRL4}
----------------------------------
We recall some formulas describing the differential geometry of the Siegel–Jacobi disk ${{\mathcal{D}}}^J_1$ [@jac1; @csg; @berr].
The (transitive) action of the group $ G^J_1={{\mathrm{H}}}_1\rtimes\text{SU}(1,1)\ni (g,\alpha)\times(z,w)\rightarrow(z_1,w_1)\in {{\mathcal{D}}}^J_1$ on the Siegel disk is given by the formulas $$\begin{gathered}
\label{dg} w_1
=\frac{a w+ b}{\delta}, \qquad \delta=\bar{b}w+\bar{a},\qquad
\text{SU}(1,1) \ni g= \left( \begin{matrix}a & b\\ \bar{b} &
\bar{a}\end{matrix}\right),\end{gathered}$$ where $|a|^2-|b|^2=1$, $$\begin{gathered}
\label{xxx}
z_1=\frac{\gamma}{\delta}, \qquad \gamma = z +\alpha-\bar{\alpha}w.\end{gathered}$$ The balanced Kähler two-form is obtained with formulas , $$\begin{gathered}
-\ii \omega_{k\mu}(z,w) = h_{z\bar{z}}{\operatorname{d}}z\wedge {\operatorname{d}}\bar{z}+h_{z\bar{w}}{\operatorname{d}}z\wedge
{\operatorname{d}}\bar{w} -h_{\bar{z}w}{\operatorname{d}}\bar{z}\wedge {\operatorname{d}}w +h_{w\bar{w}}{\operatorname{d}}w\wedge {\operatorname{d}}\bar{w} ,\end{gathered}$$ where the Kähler potential is $$\begin{gathered}
\label{keler}
f(z,w) =\mu\frac{2z\bar{z}+z^2\bar{w}+\bar{z}^2w}{2P} -2k\ln (P), \qquad P=1-w\bar{w}.\end{gathered}$$ $k$ indexes the positive discrete series of $\text{SU}(1,1)$ ($2k\in{{\amsbb{N}}}$), while $\mu>0$ indexes the representations of the Heisenberg group.
\[prop2\] The balanced Kähler two-form $\omega_{k\mu}$ on ${{\mathcal{D}}}^J_1$, $G^J_1$-invariant to the action , , can be written as $$\begin{gathered}
-\ii\omega_{k\mu}(z,w) =2k\frac{{\operatorname{d}}w \wedge {\operatorname{d}}\bar{w}}{P^2} + \mu\frac{{{\mathcal{A}}}\wedge \bar{{{\mathcal{A}}}}}{P},
\qquad {{\mathcal{A}}}={\operatorname{d}}z+\bar{\eta}(z,w){\operatorname{d}}w,
\\ z=\eta-w\bar{\eta},\qquad\text{and} \qquad \eta = \eta(z,w):= \frac{z+\bar{z}w}{1-w\bar{w}}.\end{gathered}$$
The matrix of the balanced metric $h=h(\varsigma)$, $\varsigma:=(z,w)\in{{\amsbb{C}}}\times{{\mathcal{D}}}_1$, determined with the Kähler potential , reads $$\begin{gathered}
\label{metrica}
h(\varsigma):=\left(\begin{matrix}h_{z\bar{z}} & h_{z\bar{w}}\\
h_{w\bar{z}}=\bar{h}_{z\bar{w}} & h_{w\bar{w}}\end{matrix}\right) =\left(\begin{matrix} \frac{\mu}{P} & \mu \frac{\eta}{P} \\
\mu\frac{\bar{\eta}}{P} &
\frac{2k}{P^2}+\mu\frac{|\eta|^2}{P}\end{matrix}\right).\end{gathered}$$
The inverse of the matrix reads $$\begin{gathered}
\label{hinv}
h^{-1}(\varsigma)= \left(\begin{matrix}h^{z\bar{z}}&
h^{z\bar{w}}\\h^{w\bar{z}}&h^{w\bar{w}}\end{matrix}\right) = \left(\begin{matrix}
\frac{P}{\mu}+\frac{P^2|\eta|^2}{2k} & -\frac{P^2\eta}{2k} \\
-\frac{P^2\bar{\eta}}{2k} & \frac{P^2}{2k}\end{matrix}\right).\end{gathered}$$
If we introduce the notation $$\begin{gathered}
{{\mathcal{G}}}_M(z):=\det
(h_{\alpha\bar{\beta}})_{\alpha,\beta=1,\dots,n},\end{gathered}$$ then we find $$\begin{gathered}
\label{g311}
{{\mathcal{G}}}_{{{\mathcal{D}}}^J_1}(z,w)=\frac{2k\mu}{(1-w\bar{w})^3},\qquad z\in{{\amsbb{C}}}, \qquad |w|<1.\end{gathered}$$
Elements of geometry of the Siegel–Jacobi ball {#GMM}
==============================================
The balanced metric {#BLM}
-------------------
The Jacobi algebra is the semi-direct sum ${{\mathfrak{g}}}^J_n:= {{\mathfrak{h}}}_n\rtimes {{\mathfrak{sp}}}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$ [@JGSP; @sbj; @nou]. The Heisenberg algebra ${{\mathfrak{h}}}_n$ is generated by the boson creation (respectively, annihilation) operators ${a}_i^{\dagger}$ (${a}_i$), $i,j =1,\dots,n$, which verify the canonical commutation relations $$\begin{gathered}
\label{baza1M}
\big[a_i,a^{\dagger}_j\big]=\delta_{ij}, \qquad [a_i,a_j] = \big[a_i^{\dagger},a_j^{\dagger}\big]= 0 .\end{gathered}$$ ${{\mathfrak{h}}}_n$ is an ideal in ${{\mathfrak{g}}}^J_n$, i.e., $[{{\mathfrak{h}}}_n,{{\mathfrak{g}}}^J_n]={{\mathfrak{h}}}_n$, determined by the commutation relations:
\[baza3M\] $$\begin{gathered}
\label{baza31}\big[a^{\dagger}_k,K^+_{ij}\big] = [a_k,K^-_{ij}]=0, \\
[a_i,K^+_{kj}] = \tfrac{1}{2}\delta_{ik}a^{\dagger}_j+\tfrac{1}{2}\delta_{ij}a^{\dagger}_k ,\qquad
\big[K^-_{kj},a^{\dagger}_i\big] = \tfrac{1}{2}\delta_{ik}a_j+\tfrac{1}{2}\delta_{ij}a_k , \\
\big[K^0_{ij},a^{\dagger}_k\big] = \tfrac{1}{2}\delta_{jk}a^{\dagger}_i,\qquad
\big[a_k,K^0_{ij}\big]= \tfrac{1}{2}\delta_{ik}a_{j} .\end{gathered}$$
$K^{\pm,0}_{ij}$ are the generators of the ${{\mathfrak{sp}}}({{\amsbb{R}}})_{{{\amsbb{C}}}}$ algebra:
\[baza2M\] $$\begin{gathered}
[K_{ij}^-,K_{kl}^-] = [K_{ij}^+,K_{kl}^+]=0 , \qquad 2\big[K^-_{ij},K^0_{kl}\big] = K_{il}^-\delta_{kj}+K^-_{jl}\delta_{ki}\label{baza23}, \\
2[K_{ij}^-,K_{kl}^+] = K^0_{kj}\delta_{li}+
K^0_{lj}\delta_{ki}+K^0_{ki}\delta_{lj}+K^0_{li}\delta_{kj}
\label{baza22}, \\
2\big[K^+_{ij},K^0_{kl}\big] = -K^+_{ik}\delta_{jl}-K^+_{jk}\delta_{li},\qquad
2\big[K^0_{ji},K^0_{kl}\big] = K^0_{jl}\delta_{ki}-K^0_{ki}\delta_{lj} . \label{baza24}\end{gathered}$$
Applying the same arguments as in the proof of Remark 3 in [@csg] (see also the Appendix in [@csg], where the notion of homogeneous reductive space in the meaning of Nomizu [@nomizu] is recalled; also see [@SB14]) to the Jacobi algebra ${{\mathfrak{g}}}^J_n$ determined by , , , we make the
\[geom\] The Jacobi group $G^J_n$ is an unimodular, non-reductive, algebraic group of Harish-Chandra type. The Siegel–Jacobi domain ${{\mathcal{D}}}^J_n$ is a reductive, non-symmetric manifold associated to the Jacobi group $G^J_n$ by the generalized Harish-Chandra embedding.
We have attached to the Jacobi group $G^J_n:={{\mathrm{H}}}_n\rtimes\operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$ coherent states based on the Siegel–Jacobi ball ${{\mathcal{D}}}^J_n$ [@sbj]. Perelomov’s CS vectors [@perG] associated to the group $G^J_n$ with Lie algebra the Jacobi algebra ${{\mathfrak{g}}}^J_n$ based on the complex $d$-dimensional ($d= n(n+3)/2$) manifold – the Siegel–Jacobi ball – $ {{\mathcal{D}}}^J_n\approx{{\mathrm{H}}}_n/{{\amsbb{R}}}\times \operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}/\text{U}(n)= {{\amsbb{C}}}^n\times{{\mathcal{D}}}_n $ – have been defined as [@sbj] $$\begin{gathered}
\label{csuX}
e_{z,W}= \exp ({{{\boldsymbol{X}}}})e_0,
\qquad {{\boldsymbol{X}}} := \sqrt{\mu}\sum_{i=1}^n z_i {{\boldsymbol{a}}}^{\dagger}_i + \sum_{i,j =1}^n w_{ij}{{\boldsymbol{K}}}^+_{ij},\qquad
z\in {{\amsbb{C}}}^n, \qquad W\in{{\mathcal{D}}}_n.\end{gathered}$$
The vector $e_0$ appearing in (\[csuX\]) verifies the relations $$\begin{gathered}
{{\boldsymbol{a}}}_ie_0= 0,\qquad i=1, \dots, n, \nonumber\\
{{\boldsymbol{K}}}^+_{ij} e_0 \not= 0 ,\qquad
{{\boldsymbol{K}}}^-_{ij} e_0 = 0 ,\qquad
{{\boldsymbol{K}}}^0_{ij} e_0 = \frac{k}{4}\delta_{ij} e_0, \qquad i,j=1,\dots,n .\label{vacuma}\end{gathered}$$
In (\[vacuma\]) $e_0=e^H_0\otimes e^K_0$, where $e^H_0$ is the minimum weight vector (vacuum) for the Heisenberg group ${{\mathrm{H}}}_n$, while $e^K_0$ is the extremal weight vector for $\operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$ corresponding to the weight $k$ in (\[vacuma\]), and $\mu$ parametrizes the Heisenberg group [@sbj; @nou; @gem]. Holomorphic irreducible representations of the Jacobi group based on Siegel–Jacobi domains have been studied in mathematics [@bb; @bs; @TA90; @TA92; @TA99]. In [@jac1; @nou; @gem] we have compared our results on the geometry of ${{\mathcal{D}}}^J_n$ and holomorphic representations of $G^J_n$ obtained using CS with those of the mentioned mathematicians.
The scalar product $K:{{\mathcal{D}}}^J_n\times {{\mathcal{D}}}^J_n \rightarrow{{\amsbb{C}}}$, $K(\bar{x},\bar{V};y,W)=(e_{x,V},e_{y,W})_{k\mu} $ is [@sbj; @gem]: $$\begin{gathered}
(e_{x,V},e_{y,W})_{k\mu} = \det (U)^{k/2} \exp\mu
F(\bar{x},\bar{V};y,W), \qquad U= ({{\mathbbm{1}}_n}-W\bar{V})^{-1}, \nonumber \\
2 F(\bar{x},\bar{V};y,W) = 2(x, U y) +( V\bar{y},U y) +( x,U W\bar{x}) .\label{KHKX}\end{gathered}$$ If we denote $\zeta = (x,V)$, $\zeta' = (y,W)$, then we find for the normalized Bergman kernel of the Siegel–Jacobi ball the expression $$\begin{gathered}
\label{kapS}
\kappa_{{{\mathcal{D}}}^J_n}(\zeta,\zeta')=\kappa_{{{\mathcal{D}}}_n}(V,W)\exp\mu[
2F(\zeta,\zeta')- F(\zeta)-F(\zeta')],\end{gathered}$$ where $\kappa_{{{\mathcal{D}}}_n}(V,W)$ is the normalized Bergman kernel of the Siegel ball $$\begin{gathered}
\kappa_{{{\mathcal{D}}}_n}(V,W)= \det{}^{k/2}\left[\frac{({{\mathbbm{1}}_n}-V\bar{V})({{\mathbbm{1}}_n}-W\bar{W})}{({{\mathbbm{1}}_n}-W\bar{V})^2}\right]. \end{gathered}$$ We find for the Calabi diastasis of the Siegel–Jacobi ball the expression $$\begin{gathered}
\label{DiaS}
D_{{{\mathcal{D}}}^J_n}(\zeta,\zeta')=D_{{{\mathcal{D}}}_n}(V,W)+2\mu\big[\Re F(\zeta)+\Re F(\zeta')- 2\Re F(\zeta,\zeta')\big],\\
D_{{{\mathcal{D}}}_n}(V,W)=-2\ln |\kappa_{{{\mathcal{D}}}_n}(V,W)|.\nonumber\end{gathered}$$ In particular, the reproducing kernel $K(z,W)=(e_{z,W},e_{z,W})_{k\mu}$, $z\in{{\amsbb{C}}}^n$, $W\in{{\mathcal{D}}}_n$ is $$\begin{gathered}
\label{kul}
K(z,W) = \det(M)^{k/2}\exp\mu F, \qquad M=({{\mathbbm{1}}_n}-W\bar{W})^{-1},
\\
2F =2\bar{z}^tMz+z^t\bar{W}Mz+\bar{z}^tMW\bar{z},\qquad 2F = 2\bar{\eta}^t\eta -\eta^t\bar{W}\eta-\bar{\eta}^tW\bar{\eta},\nonumber\\
\label{etaZ}
\eta=M(z+W\bar{z}), \qquad z=\eta-W\bar{\eta}.\end{gathered}$$ The Hilbert space of holomorphic functions ${{\mathcal{F}}}_K$ associated to the kernel $K$ given by is endowed with the scalar product of the type $$\begin{gathered}
(\phi ,\psi ) _{{{\mathcal{F}}}_{K}}= \Lambda_n \int_{z\in{{\amsbb{C}}}^n;\,W=W^t;\,
1-W\bar{W}>0}\bar{f}_{\phi}(z,W)f_{\psi}(z,W)Q K^{-1} {\operatorname{d}}z {\operatorname{d}}W,\\
{\operatorname{d}}z = \prod_{i=1}^n {\operatorname{d}}\Re z_i {\operatorname{d}}\Im z_i, \qquad {\operatorname{d}}W = \prod_{1\le i\le j \le n} {\operatorname{d}}\Re w_{ij} {\operatorname{d}}\Im {w}_{ij},\end{gathered}$$ where the normalization constant $\Lambda_n$ is given by $$\begin{gathered}
\Lambda_n = \mu^n\frac{k-3}{2\pi^{n(n+3)/2}}\prod_{i=1}^{n-1} \frac{\big(\frac{k-3}{2}-n+i\big)\Gamma (k+i-2)}{\Gamma [k+2(i-n-1)]} ,\end{gathered}$$ and the density of volume is [@sbj] (see also Lemma \[lemmahua\]) $$\begin{gathered}
\label{QQQ}
Q (z,W)= \det ({{\mathbbm{1}}_n}-W\bar{W})^{-(n+2)}.\end{gathered}$$
The manifold ${{\mathcal{D}}}^J_n$ has the Kähler potential , $f=\log K$, with $K$ given by , $$\begin{gathered}
f = -\tfrac{k}{2}\log \det ({{\mathbbm{1}}_n}-W\bar{W})\nonumber\\
\hphantom{f=}{} +\mu \big\{ \bar{z}^t({{\mathbbm{1}}_n}-W\bar{W})^{-1}{z}
+ \tfrac{1}{2}z^t\big[\bar{W}({{\mathbbm{1}}_n}-W\bar{W})^{-1}\big]z +\tfrac{1}{2}\bar{z}^t \big[ ({{\mathbbm{1}}_n}-W\bar{W})^{-1}W\big]\bar{z} \big\}.\label{kelerX}\end{gathered}$$ Because the symmetry of the matrix $W\in{{\mathcal{D}}}_n$, we write down the Kähler two-form on the Siegel–Jacobi ball as $$\begin{gathered}
-\ii \omega _{{{\mathcal{D}}}^J_n}(z,W) = h_{i\bar{j}}{\operatorname{d}}z_i\wedge{\operatorname{d}}\bar{z}_j+ \sum_{p\leq q}(h_{i\bar{p}\bar{q}}{\operatorname{d}}z_i\wedge{\operatorname{d}}\bar{w}_{pq}-\bar{h}_{i\bar{p}\bar{q}}{\operatorname{d}}\bar{z}_i\wedge {\operatorname{d}}w_{pq}) \nonumber\\
\hphantom{-\ii \omega _{{{\mathcal{D}}}^J_n}(z,W) =}{} + \sum_{p\le q; m\leq n}h_{pq\bar{m}\bar{n}}{\operatorname{d}}w_{pq}\wedge {\operatorname{d}}\bar{w}_{mn}.\label{omww}\end{gathered}$$ We use the following notation for the matrix of the metric $$\begin{gathered}
\label{math}
h= \left(\begin{matrix} h_1 &h_2\\h_3& h_4\end{matrix}\right) =
\left(\begin{matrix} h_{i\bar{j}}& h_{i\bar{p}\bar{q}} \\
h_{pq\bar{i}} & h_{pq\bar{m}\bar{n}} \end{matrix}\right) \in
M(n(n+3)/2,{{\amsbb{C}}}),\\
p\leq q, \qquad m\leq n, \qquad h=h^{*},\nonumber\end{gathered}$$ where $h_1\in M(n,{{\amsbb{C}}})$, $h_2\in M(n\times n(n+1)/2,{{\amsbb{C}}})$, $h_3\in
M(n(n+1)/2\times n,{{\amsbb{C}}})$, $h_4\in M(n(n+1)/2,{{\amsbb{C}}})$.
We use the following convention: indices of $z\in M(n\times 1,{{\amsbb{C}}})$ are denoted with $i$, $j$, $k$, $l$; indices of $W=W^t$, $W\in M(n,{{\amsbb{C}}})$ are denoted with $m$, $n$, $p$, $q$, $r$, $s$, $t$, $u$, $v$.
Let as write the “inverse” $h^{-1}$ of the matrix $h$ as $$\begin{gathered}
\label{matrh1}
h^{-1}=\left(\begin{matrix}h^1&h^2\\h^3&h^4\end{matrix}\right) =
\left(\begin{matrix}(h^1)_{i\bar{j}}&(h^2)_{i\bar{p}\bar{q}}\\(h^3)_{pq\bar{i}}&(h^4)_{pq\bar{m}\bar{n}}\end{matrix}\right), \qquad
p\leq q, \qquad m\leq n,\end{gathered}$$ where the matrices $h^1$, $h^2$, $h^3$, $h^4$ have the same dimensions as the matrices $h_1$, $h_2$, $h_3$, respectively $h_4$.
In fact, we shall determine the “inverse” of such that $$\begin{gathered}
\label{ciudat}
\left(\begin{matrix} h_1 &h_2\\h_3& h_4\end{matrix}\right)
\left(\begin{matrix}h^1&h^2\\h^3&h^4\end{matrix}\right)=
\left(\begin{matrix} {{\mathbbm{1}}_n}& 0\\ 0 & \Delta \end{matrix}\right),\end{gathered}$$ where $\Delta$ is defined in .
It is useful to introduce the notation $$\begin{gathered}
f_{pq}:=1-\tfrac{1}{2}\delta_{pq}.\end{gathered}$$
We shall prove the following (partial results have been presented in [@JGSP Proposition 11], [@sbj Proposition 3.11] and [@nou Proposition 1]):
\[mainTH\] The Kähler two-form $\omega_{{{\mathcal{D}}}^J_n}$, associated with the balanced metric of the Siegel–Jacobi ball ${{\mathcal{D}}}^J_n$, $G^J_n$-invariant to the action is $$\begin{gathered}
- \ii\omega_{{{\mathcal{D}}}^J_n}(z,W) = \tfrac{k}{2}{\operatorname{Tr}}({{\mathcal{B}}}\wedge\bar{{{\mathcal{B}}}})
+\mu {\operatorname{Tr}}({{\mathcal{A}}}^t\bar{M}\wedge \bar{{{\mathcal{A}}}}), \qquad {{\mathcal{A}}} ={\operatorname{d}}z+ {\operatorname{d}}W\bar{\eta},\nonumber\\
{{\mathcal{B}}} = M{\operatorname{d}}W,\qquad M = ({{\mathbbm{1}}_n}-W\bar{W})^{-1}.\label{aabX}\end{gathered}$$ The matrix of the metric on ${{\mathcal{D}}}^J_n$ has the matrix elements :
\[hcomp\] $$\begin{gathered}
h_{i\bar{j}} =\mu \bar{M}_{ij},\label{hc11}\\
h_{i\bar{p}\bar{q}} = \mu (\eta_q\bar{M}_{ip}
+\eta_p\bar{M}_{iq})f_{pq},\label{hc12}\\
h_{pq\bar{i}} = \mu (\bar{\eta}_q\bar{M}_{pi} +\bar{\eta}_p\bar{M}_{qi})f_{pq},\label{hc21}\\
h_{pq\bar{m}\bar{n}} =\frac{k}{2}h^k_{pq\bar{m}\bar{n}}+\mu
h^{\mu}_{pq\bar{m}\bar{n}},\label{hc22}\\
h^k_{pq\bar{m}\bar{n}} =
2(M_{mp}M_{nq}+M_{mq}M_{np})-2M_{mp}(M_{np}\delta_{pq}+M_{mq}\delta_{mn})
+ M^2_{mp}\delta_{pq}\delta_{mn}\nonumber \\
\hphantom{h^k_{pq\bar{m}\bar{n}}}{}
= 2M_{mp}M_{nq}(1-\delta_{pq})+2M_{mq}M_{np}(1-\delta_{mn})+ M^2_{mp}\delta_{pq}\delta_{mn} ,\label{hc23}\\
h^{\mu}_{pq\bar{m}\bar{n}} = \bar{\eta}_p(\eta_n M_{mq}+\eta_m
M_{nq})+\bar{\eta}_q({\eta}_nM_{mp}+\eta_mM_{np})
-\bar{\eta}_p(\eta_n M_{mp}\nonumber\\
\hphantom{h^{\mu}_{pq\bar{m}\bar{n}}=}{} +\eta_m
M_{np})\delta_{pq}-{\eta}_m(\bar{\eta}_pM_{mq}+\bar{\eta}_qM_{mp})\delta_{mn}
+\bar{\eta}_p\eta_mM_{mp}\delta_{pq}\delta_{mn}\nonumber\\
\hphantom{h^{\mu}_{pq\bar{m}\bar{n}}}{}
= [\bar{\eta}_p(\eta_n\bar{M}_{qm}+\eta_m\bar{M}_{qn})+\bar{\eta}_q(\eta_n\bar{M}_{pm}+\eta_m\bar{M}_{pn})]f_{pq}f_{mn}.\label{hc24}\end{gathered}$$
The “inverse” $h^{-1}$ of the metric matrix $h$ which verifies , with components , obtained with the inverse of $h^k$ has the elements $h^1$–$h^4$ given by
\[nr11\] $$\begin{gathered}
\big(h^1\big)_{ij}
=\theta\bar{M}^{-1}_{ij},
\qquad \theta = \frac{1}{\mu}+\alpha\frac{2}{k},\qquad
\alpha=\eta^t\bar{M}^{-1}\bar{\eta}=\bar{\eta}_nS_n, \qquad S_n=\sum\eta_q\bar{M}^{-1}_{qn},\!\!\label{NR1}\\
\big(h^2\big)_{i\bar{m}\bar{n}} =-\frac{1}{k}\big(S_n\bar{M}^{-1}_{im}+S_m\bar{M}^{-1}_{in}\big),\label{NR2}\\
\big(h^3\big)_{mn\bar{i}} =-\frac{1}{k}\big(\bar{S}_n\bar{M}^{-1}_{mi}+\bar{S}_m\bar{M}^{-1}_{ni}\big),\label{NR3}\\
\big(h^4\big)_{pq\bar{m}\bar{n}} =
\frac{2}{k}(h^k)^{-1}_{pq\bar{m}\bar{n}}=\frac{1}{k}\big(\bar{M}^{-1}_{qn}\bar{M}^{-1}_{pm}+\bar{M}^{-1}_{pn}\bar{M}^{-1}_{qm}\big). \label{NR4}\end{gathered}$$
The determinant of the metric matrix $h$ is $$\begin{gathered}
\label{DTH}
{{\mathcal{G}}}_{{{\mathcal{D}}}^J_n}(z,W)= \det h _{{{\mathcal{D}}}^J_n}(z,W)= \left(\frac{k}{2}\right)^{\frac{n(n+1)}{2}}\mu^n\det
({{\mathbbm{1}}_n}-W\bar{W})^{-(n+2)}.\end{gathered}$$ In the case $n=1$ formulas , , became the formulas , , respectively , with $2k\leftrightarrow\frac{k}{2}$.
Firstly, we determine the matrix elements of the metric $h$ on the Siegel–Jacobi disk ${{\mathcal{D}}}^J_n$ applying formula to the Kähler potential as in .
We use the formulas (see [@nou equations (4.23)])
\[LIKK\] $$\begin{gathered}
\frac{{\partial}M_{ab} }{{\partial}w_{ik}} = M_{ai}X_{kb}+ M_{ak}X_{ib}- M_{ai}X_{ib}\delta_{ik},\qquad \text{where} \quad X=X^t=\bar{W}M=\bar{M}\bar{W},\label{LIK}\\
\frac{{\partial}X_{ab} }{{\partial}w_{ik}} = X_{ai}X_{bk}+X_{ak}X_{ib}-X_{ai}X_{ib}\delta_{ik} , \\
\frac{{\partial}\bar{ X}_{ab} }{{\partial}w_{ik}} =M_{ai}M_{bk}+M_{ak}M_{bi}-M_{ai}M_{bk}\delta_{ik} \label{LIK2}.\end{gathered}$$
From , we get
\[morD\] $$\begin{gathered}
\frac{{\partial}\eta_q}{{\partial}z_l} = M_{ql},\\
\frac{{\partial}\bar{\eta}_q}{{\partial}z_j} ={X}_{qj}, \\
\frac{{\partial}\eta_t}{{\partial}w_{pq}} = M_{tp}\bar{\eta}_q +M_{tq}\bar{\eta}_p- M_{tp}\bar{\eta}_p\delta_{pq}, \\
\frac{{\partial}\bar{\eta}_n}{{\partial}w_{pq}} =
\bar{\eta}_pX_{qn}+\bar{\eta}_qX_{pn}-\bar{\eta}_qX_{pn}\delta_{pq} .\end{gathered}$$
With , and , we find the matrix elements of the metric of the Siegel–Jacobi ball.
Now we prove . We calculate $$\begin{gathered}
\label{HKJ} h^k _{mn\bar{p}\bar{q}}=\frac{{\partial}}{{\partial}\bar{w}_{pq}} J_{mn}, \qquad J_{mn}= \frac{{\partial}}{{\partial}w_{mn}}\ln\det M. \end{gathered}$$ With formula $$\begin{gathered}
\frac{{\operatorname{d}}(\det A)}{{\operatorname{d}}t}=\det A {\operatorname{Tr}}\left(A^{-1}\frac{{\partial}A}{{\partial}t}\right),\end{gathered}$$ we have $$\begin{gathered}
J_{mn}= M^{-1}_{ij}\frac{{\partial}M_{ji}}{{\partial}w_{mn}}.\end{gathered}$$ We find $$\begin{gathered}
J_{mn}=X_{nm}+X_{mn}-X_{mn}\delta_{mn}=2X_{mn}f_{mn}.\end{gathered}$$ In order to obtain $ h^k _{mn\bar{p}\bar{q}}$, we calculate $$\begin{gathered}
\frac{{\partial}X_{nm}}{{\partial}\bar{w}_{pq}}= \frac{{\partial}\bar{w}_{nk}}{{\partial}\bar{w}_{pq}}M_{kn}+w_{nk}\frac{{\partial}M_{kb}}{{\partial}\bar{w}_{pq}}.\end{gathered}$$ But with $$\begin{gathered}
\frac{{\partial}M_{km}}{{\partial}\bar{w}_{pq}}=\overline{\left(\frac{M_{mk}}{{\partial}w_{pq}}\right)}, \qquad \text{and with} \qquad {{\mathbbm{1}}_n}+\bar{W}\bar{X}=\bar{M},\end{gathered}$$ we find $$\begin{gathered}
\frac{{\partial}X_{nm}}{{\partial}\bar{w}_{pq}}=2(M_{pm}\bar{M}_{nq}+M_{qm}\bar{M}_{np}-M_{pm}\bar{M}_{nq}\delta_{pq})
-(2M_{pm}\bar{M}_{mq}-M_{pm}\bar{M}_{mq}\delta_{pq})\delta_{mn},\end{gathered}$$ i.e., we reobtain formula .
With , we have $$\begin{gathered}
h^{k}_{pp\bar{m}\bar{m}} = M^2_{mp}, \nonumber\\
h^{k}_{pp\bar{m}\bar{n}} = 2M_{mp}M_{nq},\qquad m<n,\nonumber\\
h^{k}_{pq\bar{m}\bar{m}} = 2M_{mp}M_{mq},\qquad p<q,\nonumber\\
h^{k}_{pq\bar{m}\bar{n}} = 2(M_{mp}M_{nq}+M_{mq}M_{np}),\qquad p<q,\qquad m<n.\label{620Ec}\end{gathered}$$
In order to find $-\ii \frac{2}{k}{\operatorname{d}}\omega_{{{\mathcal{D}}}_n}(W) $, we make the summation $$\begin{gathered}
\sum h^k_{pp\bar{m}\bar{m}}{\operatorname{d}}w_{pp}\wedge{\operatorname{d}}\bar{w}_{mm} + \sum_{m<n} h^k_{pp\bar{m}\bar{n}}{\operatorname{d}}w_{pp}\wedge{\operatorname{d}}\bar{w}_{mn} \nonumber\\
\qquad\quad{} +\sum_{p<q} h^k_{pq\bar{m}\bar{m}}{\operatorname{d}}w_{pq}\wedge{\operatorname{d}}\bar{w}_{mm} + \sum_{p<q,\, m<n} h^k_{pq\bar{m}\bar{n}}{\operatorname{d}}w_{pq}\wedge{\operatorname{d}}\bar{w}_{mn}\nonumber\\
\qquad{} = \sum M_{mp}\bar{M}_{pm}{\operatorname{d}}w_{pp}\wedge{\operatorname{d}}\bar{w}_{mm} + 2\sum_{m<n}M_{mp}M_{pn}{\operatorname{d}}w_{pp}\wedge{\operatorname{d}}\bar{w}_{mn} \nonumber\\
\qquad\quad{} + 2\sum_{p<q}M_{mp}M_{mq} + 2\sum_{p<q, \, m<n}(M_{mp}M_{nq}+M_{mq}M_{np}) {\operatorname{d}}w_{pq}\wedge{\operatorname{d}}\bar{w}_{mn} \nonumber\\
\qquad{} = \sum M_{mp}\bar{M}_{pm}{\operatorname{d}}w_{pp}\wedge{\operatorname{d}}\bar{w}_{mm} + \sum_{m\not= n}M_{mp}M_{pn}{\operatorname{d}}w_{pp}\wedge{\operatorname{d}}\bar{w}_{mn} \nonumber\\
\qquad\quad{} + \sum_{p\not= q}M_{mp}M_{nq}{\operatorname{d}}w_{pq}\wedge {\operatorname{d}}\bar{w}_{mm}+
\sum_{p\not= q,\, m\not= n}M_{mp}M_{nq} {\operatorname{d}}w_{pq}\wedge{\operatorname{d}}\bar{w}_{mn}\nonumber \\
\qquad{} = (M{\operatorname{d}}W)_{mq}\wedge(\bar{M}{\operatorname{d}}\bar{W})_{qm}.\label{sumL}\end{gathered}$$ We write down as $$\begin{gathered}
\label{omd}
-\ii \frac{2}{k}{\operatorname{d}}\omega_{{{\mathcal{D}}}_n}(W) = {\operatorname{Tr}}(M{\operatorname{d}}W\wedge \bar{M}{\operatorname{d}}\bar{W}).\end{gathered}$$ Now we make the summation $$\begin{gathered}
\sum_{p\le q}h^{\mu}_{i\bar{p}\bar{q}}{\operatorname{d}}z_i\wedge
{\operatorname{d}}\bar{w}_{pq} = \sum h^{\mu}_{i\bar{p}\bar{p}}{\operatorname{d}}z_i\wedge
{\operatorname{d}}\bar{w}_{pp} + \sum_{p< q}h^{\mu}_{i\bar{p}\bar{q}}{\operatorname{d}}z_i\wedge
{\operatorname{d}}\bar{w}_{pq} \\
\hphantom{\sum_{p\le q}h^{\mu}_{i\bar{p}\bar{q}}{\operatorname{d}}z_i\wedge {\operatorname{d}}\bar{w}_{pq}}{}
= \bar{M}_{ip}\eta_p{\operatorname{d}}z_i\wedge {\operatorname{d}}\bar{w}_{pp}+\sum_{p<q}(\bar{M}_{ip}\eta_q+\bar{M}_{iq}\eta_p){\operatorname{d}}z_i \wedge {\operatorname{d}}\bar{w}_{pq} \\
\hphantom{\sum_{p\le q}h^{\mu}_{i\bar{p}\bar{q}}{\operatorname{d}}z_i\wedge {\operatorname{d}}\bar{w}_{pq}}{} = (M{\operatorname{d}}z)_p\wedge({\operatorname{d}}\bar{W}\eta)_p.\end{gathered}$$ We calculate the sum $$\begin{gathered}
H^{\mu}:=\sum_{p\le q,m\le n}h^{\mu}_{pq\bar{m}\bar{n}}{\operatorname{d}}w_{pq}\wedge {\operatorname{d}}\bar{w}_{mn}. \end{gathered}$$ We have $$\begin{gathered}
H^{\mu} = \sum h^{\mu}_{pp\bar{m}\bar{m}}{\operatorname{d}}w_{pp}\wedge \bar{w}_{mm} + \sum_{m<n}h^{\mu}_{pp\bar{m}\bar{n}}{\operatorname{d}}w_{pp}\wedge \bar{w}_{mn} \\
\hphantom{H^{\mu} =}{} + \sum_{p<q}h^{\mu}_{pq\bar{m}\bar{m}}{\operatorname{d}}w_{pq}\wedge \bar{w}_{mm} + \sum_{p< q,\, m< n}h^{\mu}_{pq\bar{m}\bar{n}}{\operatorname{d}}w_{pq}\wedge \bar{w}_{mn} ,\end{gathered}$$ and finally we get $$\begin{gathered}
\label{omdLL}
H^{\mu} = M_{mq}{\operatorname{d}}w_{qp}\wedge {\operatorname{d}}\bar{w}_{mn}\eta_n= (M{\operatorname{d}}W\bar{\eta})_m\wedge ({\operatorname{d}}\bar{W}\eta)_m.\end{gathered}$$ Putting together , and , we have proved .
In order to check out the homogeneity of the Kähler two-form $\omega_{{{\mathcal{D}}}^J_n}$ , we use the formulas $$\begin{gathered}
\label{DW1}
{\operatorname{d}}W_1=(Wq^*+p^*)^{-1}{\operatorname{d}}W (\bar{q}W+\bar{p})^{-1},\\
\label{DM1} M_1=(q\bar{W}+p)M(W q^*+p^*),\\
{\operatorname{d}}z_1= (Wq^*+p^*)^{-1}({\operatorname{d}}W V +{\operatorname{d}}z), \qquad V=-q^*(Wq^*+p^*)^{-1}[\eta+\alpha-W(\bar{\eta}+\bar{\alpha})]-\bar{\alpha}, \nonumber\\
\eta_1=p(\alpha+\eta)+q(\bar{\alpha}+\bar{\eta}),\qquad
{{\mathcal{A}}}_1= (Wq^*+p^*)^{-1}{{\mathcal{A}}}.\nonumber\end{gathered}$$
Now we calculate the “inverse” $h^{-1}$ of the matrix $h$ whose elements are given by .
We have (see, e.g., in [@lutke (1) and (2), p. 30]): if the matrices $h_1$, $h_4-h_3h^{-1}_1h_2$, $h_4$ and respectively $h_1-h_2h^{-1}_4h_3$ are nonsingular, then
\[inv\] $$\begin{gathered}
h^1 = \big(h_1-h_2h^{-1}_4h_3\big)^{-1} = h_1^{-1}+ h_1^{-1}h_2h^4h_3 h_1^{-1},\label{inv11}\\
h^2 = -h^{-1}_1h_2h^4,\label{inv2}\\
h^3 = - h^4 h_3h_1^{-1},\label{inv31}\\
h^4 = (h_4-h_3h^{-1}_1h_2)^{-1}.\label{inv4}\end{gathered}$$
In order to calculate $h^4$, we write as $$\begin{gathered}
\label{NR5}(h^4)^{-1}=h_4-\mu h^I, \qquad\text{where} \quad
h^I:=\frac{1}{\mu}h_3h_1^{-1}h_2.\end{gathered}$$ With formulas –, we get $$\begin{gathered}
\big(h^I\big)_{pq\bar{m}\bar{n}} =h_{pq\bar{i}}(h_1)^{-1}_{ij}h_{j\bar{m}\bar{n}} =
(\bar{\eta}_q\bar{M}_{pi}+\bar{\eta}_p\bar{M}_{qi})\bar{M}^{-1}_{ij}(\eta_n\bar{M}_{jm}+\eta_m\bar{M}_{jn})
f_{pq}f_{mn}\\
\hphantom{\big(h^I\big)_{pq\bar{m}\bar{n}}}{}
=(\bar{\eta}_q\bar{M}_{pi}+\bar{\eta}_p\bar{M}_{qi})(\eta_n\delta_{im}+\eta_m\delta_{im})f_{pq}f_{mn}\\
\hphantom{\big(h^I\big)_{pq\bar{m}\bar{n}}}{}
= [\bar{\eta}_q(\eta_n\bar{M}_{pm}+\eta_m\bar{M}_{pn})+\bar{\eta}_p(\eta_n\bar{M}_{qm}+\eta_m\bar{M}_{qn})]f_{pq}f_{mn},\end{gathered}$$ and with equation we get $$\begin{gathered}
h^I=h^{\mu}.\end{gathered}$$ With , we obtain $$\begin{gathered}
\label{h4h1}
\big(h^4\big)^{-1}=\frac{k}{2}h^k,\end{gathered}$$ and , where $h^k$ has the expression , while $(h^k)^{-1}$ has the expression .
In order to calculate $h^2$, taking into account , we write as $$\begin{gathered}
\label{hjj}
h^2=-(h_1)^{-1}h^J,\qquad\text{where} \quad h^J:=\frac{2}{k}h_2\big(h^k\big)^{-1}. \end{gathered}$$ With formulas , , we have $$\begin{gathered}
\big(h^J\big)_{j\bar{m}\bar{n}} =2\frac{\mu}{k}\sum_{p\leq
q}f_{pq}(\eta_q\bar{M}_{jp}+\eta_p\bar{M}_{jq})k_{pq\bar{m}\bar{n}}\\
\hphantom{\big(h^J\big)_{j\bar{m}\bar{n}}}{}
= 2\frac{\mu}{k}\left[\sum_{p}\eta_p({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{pj}({{\mathbbm{1}}_n}-W\bar{W})_{np}({{\mathbbm{1}}_n}-\bar{W}W)_{pm}\right]\\
\hphantom{\big(h^J\big)_{j\bar{m}\bar{n}}=}{}+
\frac{\mu}{k}\sum_{p<q}\big[\eta_q({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{pj}+\eta_p({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{pj}\big] \\
\hphantom{\big(h^J\big)_{j\bar{m}\bar{n}}=}{}
\times\big[({{\mathbbm{1}}_n}-W\bar{W})_{nq}({{\mathbbm{1}}_n}-\bar{W}W)_{pm}+({{\mathbbm{1}}_n}-W\bar{w})_{np}({{\mathbbm{1}}_n}-\bar{W}W)_{qm}\big]\\
\hphantom{\big(h^J\big)_{j\bar{m}\bar{n}}}{}
= \frac{\mu}{k}\left[\sum_{p\not=q}(\eta_q({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{pj}({{\mathbbm{1}}_n}-W\bar{W})_{nq}({{\mathbbm{1}}_n}-\bar{W}W)_{pm}\right.\\
\hphantom{\big(h^J\big)_{j\bar{m}\bar{n}}=}{}
+\sum_{p\not=q}\eta_p({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{qj}({{\mathbbm{1}}_n}-W\bar{W})_{nq}({{\mathbbm{1}}_n}-\bar{W}{W})_{pm}\\
\left.\hphantom{\big(h^J\big)_{j\bar{m}\bar{n}}=}{}
+2\sum_p\eta_p({{\mathbbm{1}}_n}-W\bar{W})^{-1}_{pj}({{\mathbbm{1}}_n}-W\bar{W})_{np}({{\mathbbm{1}}_n}-\bar{W}W)_{pm}\right]\\
\hphantom{\big(h^J\big)_{j\bar{m}\bar{n}}=}{}
=\frac{\mu}{k}(S_n\delta_{jm}+S_m\delta_{jn}),\qquad \text{where} \quad S_n:=\sum_q\eta_q\bar{M}^{-1}_{qn}.\end{gathered}$$
With , we get and similarly for .
In order to calculate $h^1$, we write as $$\begin{gathered}
h^1=h^{-1}_1-h^Kh^{-1}_1,\qquad\text{where} \quad h^K:= h^2h_3.\end{gathered}$$ With formulas and , we get $$\begin{gathered}
\big(h^K\big)_{ik} =-\frac{\mu}{k}\sum_{m\leq
n}f_{mn}\big(S_n\bar{M}^{-1}_{im}+S_m\bar{M}^{-1}_{in}\big)(\bar{\eta}_n\bar{M}_{mk}+\bar{\eta}_m\bar{M}_{nk})\\
\hphantom{\big(h^K\big)_{ik} }{} =
-\frac{\mu}{k}\left[\frac{1}{2}\sum_m4S_m\bar{\eta}_m\bar{M}^{-1}_{im}\bar{M}_{mk}
+ \sum_{m< n}\big(S_n\bar{M}^{-1}_{im} + S_m\bar{M}^{-1}_{in}\big)\big(\bar{\eta}_n\bar{M}_{mk} + \bar{\eta}_m\bar{M}_{nk}\big)\right]\\
\hphantom{\big(h^K\big)_{ik} }{}= -\frac{2}{k}\mu\alpha\delta_{ik},\end{gathered}$$ where $$\begin{gathered}
\alpha:=\eta^t\bar{M}^{-1}\bar{\eta}=\eta^*M^{-1}\eta=\bar{\eta}_nS_n.\end{gathered}$$ With , we get .
In order to prove , we use the relation $$\begin{gathered}
\det h=\det \left(\begin{matrix} h_1 &h_2\\h_3&
h_4\end{matrix}\right)
=\det h_1 \det \big(h_4-
h_3h^{-1}_1h_2\big).\end{gathered}$$ With and , we get $$\begin{gathered}
\det h =\left(\frac{k}{2}\right)^n\det h_1\det h^k.\end{gathered}$$ is obtained introducing in the above relation the expressions and .
Now we formulate several geometric properties of the Siegel–Jacobi ball relevant for the Berezin quantization of this manifold:
\[geoPL\]
1. The homogeneous Kähler manifold ${{\mathcal{D}}}^J_n$ is contractible.
2. The Kähler potential of the Siegel–Jacobi ball is global. ${{\mathcal{D}}}^J_n$ is a Q.-K. Lu manifold, with normalized Bergman kernel and nowhere vanishing diastasis .
3. The manifold ${{\mathcal{D}}}^J_n$ is a quantizable manifold.
4. ${{\mathcal{D}}}^J_n$ is projectively induced and the Jacobi group $G^J_n$ is a CS-type group.
5. ${{\mathcal{D}}}^J_n$ is a homogeneous reductive space.
$i)$ We have proved in Theorem \[mainTH\] (see also [@JGSP; @nou]) that ${{\mathcal{D}}}^J_n$ is a homogeneous Kähler manifold. We have shown in [@nou Proposition 4.1] that the ${\rm FC}$-transform ${\rm FC}\colon (\eta,W)\rightarrow (z,W)$ expressed by is a homogeneous Kähler diffeomorphism ${{\mathcal{D}}}^J_n\approx{{\mathcal{D}}}_n\times{{\amsbb{C}}}^n$, where the Siegel ball admits the realization . But ${{\mathcal{D}}}_n$ can be achieved as the open ball ${{\mathcal{B}}}_n$ (see, e.g., [@neeb p. 502]) $$\begin{gathered}
{{\mathcal{B}}}_n=\big\{W\in M(n,{{\amsbb{C}}}),\qquad W=W^t\, |\, ||W||<1\big\}.\end{gathered}$$ So, ${{\mathcal{D}}}^J_n$ is diffeomorphic with the product of the contractible spaces ${{\mathcal{D}}}_n$ and ${{\amsbb{C}}}^n$, and consequently, it is itself contractible.
$ii)$ We have for $K(z,W)>0$, $\forall\, (z,W)\in {{\mathcal{D}}}^J_n$. The explicit expressions of the Kähler potential, normalized Bergman kernel and diastasis imply the assertions of $ii)$, but they could be also derived from the Theorem \[LMM\], once we have proved $i)$. Even more, Theorem \[LMM\] asserts that ${{\mathcal{D}}}^J_n$ admits a Berezin quantization.
$iii)$, $iv)$ We have observed in Theorem \[mainTH\] that the Kähler two-form is associated with the balanced metric on ${{\mathcal{D}}}^J_n$, and we apply Remark \[kra\].
$v)$ This assertion was already mentioned in Remark \[geom\], where it was used the explicit form , , of the Jacobi algebra ${{\mathfrak{g}}}^J_n$. We mention it again here because in [@last] (see Remark 1 there for a more precise formulation) we have proved that [*the CS-manifolds are reductive spaces*]{}.
The fact that the Jacobi algebra ${{\mathfrak{g}}}^J_n$ is a CS-algebra is well known, see, e.g., [@lis Theorem 5.2] or [@neeb Example VII.2.3 , p. 294]. Using the explicit form of the base $\Phi(z,w)$ of orthonormal polynomials in which the Bergman kernel $K_{{{\mathcal{D}}}^J_1}(z,w)$ is developed [@jac1], we have proved in [@berr Proposition 2] that the Siegel–Jacobi disk is a CS-manifold. In [@gem] we have determined the base $\Phi(z,W)$ on ${{\mathcal{D}}}^J_n$ in which the reproducing kernel admits a series expansion as in , but in Proposition \[geoPL\] we have used Theorems \[LMN\], \[LMM\] and Remark \[kra\] to prove directly Proposition \[geoPL\].
Ricci form and scalar curvature {#SCR}
-------------------------------
The Ricci form associated to the Kählerian two-form $\omega_M$ is (see [@mor p. 90]) $$\begin{gathered}
\label{RICCI}
\rho_M(z):=\ii \sum_{\alpha,\beta=1}^n\text{Ric}_{\alpha\bar{\beta}}(z){\operatorname{d}}z_{\alpha}\wedge{\operatorname{d}}\bar{z}_{\beta}, \qquad \text{Ric}_{\alpha\bar{\beta}}(z)=
-\frac{{\partial}^2}{{\partial}z_{\alpha}{\partial}\bar{z} _{\beta}} \ln{{\mathcal{G}}}_M(z).\end{gathered}$$ The scalar curvature at a point $p\in M$ of coordinates $z$ is (see [@koba1 p. 294] or [@jost p. 144]) $$\begin{gathered}
\label{scc}
s_M(p):=\sum_{\alpha,\beta=1}^n (h_{\bar{\alpha}{\beta}})^{-1}\text{Ric}_{\alpha\bar{\beta}}(z).\end{gathered}$$ The Bergman metric corresponds to the Kähler two-form (see references in [@berr]) $$\begin{gathered}
\label{bergom1}
\omega^1_M=\ii {\partial}\bar{{\partial}} \ln {{\mathcal{G}}}.\end{gathered}$$
Q.-K. Lu [@LU08] has introduced for a bounded domain the positive definite (1,1)-form $$\begin{gathered}
\label{RIC2}
\tilde{\omega}_M(z):=
\ii\sum_{\alpha,\beta=1} ^n\tilde{h}_{\alpha\bar{\beta}} (z) {\operatorname{d}}z_{\alpha}\wedge
{\operatorname{d}}\bar{z}_{\beta}, \qquad \tilde{h}_{\alpha\bar{\beta}} (z):=
(n+1)h_{\alpha\bar{\beta}} (z)- \text{Ric}_{\alpha,\beta}(z),\end{gathered}$$ which is Kähler, corresponding to the Kähler potential $\tilde{f} = \ln (K _M (z)^{n+1}{{\mathcal{G}}}(z))$.
Now we prove
\[mainPR\]
1. The Siegel–Jacobi ball ${{\mathcal{D}}}^J_n$ is not an Einstein manifold with respect to the homogeneous Kähler metric attached to the Kähler two-form , but it is one with respect to the Bergman metric corresponding to the Bergman Kähler two-form .
2. The Ricci form on the Siegel–Jacobi ball associated to the homogeneous Kählerian two-form has the expression $$\begin{gathered}
\rho_{{{\mathcal{D}}}^J_n}(z,W)=-\omega^1_{{{\mathcal{D}}}^J_n}(z,W) =-\ii (n+2){\operatorname{Tr}}(M{\operatorname{d}}W \wedge \bar{M}{\operatorname{d}}\bar{W}),\end{gathered}$$ while $$\begin{gathered}
\rho_{{{\mathcal{D}}}_n}(W)=-\omega^1_{{{\mathcal{D}}}_n}(W)=-\ii (n+1){\operatorname{Tr}}(M{\operatorname{d}}W \wedge \bar{M}{\operatorname{d}}\bar{W}).\end{gathered}$$
3. The scalar curvature is constant and negative $$\begin{gathered}
s_{{{\mathcal{D}}}^J_n}(z,W)= - \frac{2}{k}n\frac{(n+1)(n+2)}{2}.\end{gathered}$$
4. The Q.-K. Lu Kähler two-form has the expression $$\begin{gathered}
\tilde{\omega}_{{{\mathcal{D}}}^J_n}(z,W)=\frac{(n+1)(n+2)}{2}\omega_{{{\mathcal{D}}}^J_n}(z,W)-\rho_{{{\mathcal{D}}}^J_n}(W).\end{gathered}$$
With formula (or ) and , we find that the only nonzero components of the Ricci tensor are $$\begin{gathered}
\label{RiRi}
(\text{Ric}_{mn\bar{p}\bar{q}})_{{{\mathcal{D}}}^J_n}(z,W)= -(n+2)h^k_{mn\bar{p}\bar{q}}(W).\end{gathered}$$ With calculation , we find $$\begin{gathered}
\rho_{{{\mathcal{D}}}^J_n}(z,W)=-\ii (n+2)\bar{M}_{pi}\bar{M}_{jq}{\operatorname{d}}w_{pj}\wedge
{\operatorname{d}}\bar{w}_{qi}.\end{gathered}$$ Applying for ${{\mathcal{D}}}^J_n$, with formulas , , we get $$\begin{gathered}
s_{{{\mathcal{D}}}^J_n}(z,W) :=\sum_{\alpha,\beta =1}^n
(h)^{-1}_{\bar{\alpha}{\beta}}\text{Ric}_{\alpha\bar{\beta}}(z,W)=\sum_{p\leq
q;\, m\leq n}(h)^{-1}_{mn\bar{p}\bar{q}} (\text{Ric}_{pq\bar{m}\bar{n}}) _{{{\mathcal{D}}}^J_n}\\
\hphantom{s_{{{\mathcal{D}}}^J_n}(z,W)}{}
= -\frac{2}{k}(n+2) \sum_{p\le q;\, m\leq n}(h^k)^{-1}_{mn\bar{p}\bar{q}}
(h^k)_{pq\bar{m}\bar{n}} = -\frac{2}{k}(n+2) \sum_{m\le n}\Delta^{mn}_{mn}\\
\hphantom{s_{{{\mathcal{D}}}^J_n}(z,W)}{}
= -\frac{2}{k}(n+2)\frac{n(n+1)}{2}.\tag*{\qed}\end{gathered}$$
The scalar curvature of the Siegel–Jacobi disk was calculated in [@csg; @berr; @jae].
Laplace–Beltrami operator {#LBO}
-------------------------
We introduce the formulas into the expression of the Laplace–Beltrami operator and we use Remark \[equivv\]. In the formula of the Laplace–Beltrami operator on ${{\mathcal{D}}}^J_n$, we use the expression given in Proposition \[LLB\] for the Laplace–Beltrami operator on ${{\mathcal{D}}}_n$. We introduce the notation (no summation!) $$\begin{gathered}
\label{symbe}
(D_z)_{\mu \nu}=\left(e_{\mu \nu}\frac{{\partial}}{{\partial}z_{\mu\nu}}\right), \qquad e_{\mu\nu}= \frac{1+\delta_{\mu\nu}}{2},~z_{\mu\nu}=z_{\nu\mu}.\end{gathered}$$
We obtain:
\[incaun\]The Laplace–Beltrami operator on the Siegel–Jacobi ball ${{\mathcal{D}}}^J_n$, invariant to the action of the Jacobi group $G^J_n$, has the expression $$\begin{gathered}
\label{LABDL}\Delta_{{{\mathcal{D}}}^J_n}(z,W) = \theta\frac{{\partial}}{{\partial}z^t}N\frac{{\partial}}{{\partial}\bar{z}}-\frac{2}{k}\left\{{\operatorname{Tr}}\left[S D_W
N\frac{{\partial}}{{\partial}\bar{z}}\right]+cc\right\}+\frac{2}{k}\Delta_{{{\mathcal{D}}}_n}(W),\\
N={{\mathbbm{1}}_n}-W\bar{W}.\nonumber\end{gathered}$$
We have also the relations $$\begin{gathered}
\label{lastB}
\Delta_{{{\mathcal{D}}}^J_n}(\ln {{\mathcal{G}}}) = -s_{{{\mathcal{D}}}^J_n} = \frac{2}{k}n\frac{(n+1)(n+2)}{2}.\end{gathered}$$
We apply formula with the matrix elements and we use Remark \[equivv\].
We recall that Theorem 2.5 in Berezin’s paper [@ber74] asserts essentially that $$\begin{gathered}
\Delta_M(z)(\ln({{\mathcal{G}}}(z)))= ct \end{gathered}$$ for the balanced metric plus other 3 conditions. The first equation in $$\begin{gathered}
s=-\Delta\ln \det h \end{gathered}$$ is a general recipe for calculation of scalar curvature, see, e.g., [@jost equation (5.2.24), p. 253].
Also we recall that the Laplace operator on the Siegel–Jacobi ball was calculated in [@Y10] in other coordinates. We have determined the Laplace–Beltrami operator on the Siegel–Jacobi disk in [@berr].
Appendix {#APP1}
========
A remark
--------
\[equivv\] Let $\omega_M(z)$ be the non-degenerate two-form on a complex manifold $M$, invariant to a invertible holomorphic transformation $z'=z'(z)$. Then the differential operator $\Delta_M$ given by is also invariant to the transformation $z\rightarrow z'$. In particular, $M$ may be a homogeneous Kähler manifold and $\Delta_M$ the corresponding Laplace–Beltrami operator.
Let us consider the action $G \times M\rightarrow M$: $g\times z = z'$. Then we have $$\begin{gathered}
-\ii \omega_M(z')=h'_{\lambda\bar{\gamma}}(z'){\operatorname{d}}z'_{\lambda}\wedge {\operatorname{d}}\bar{z}_{\gamma}
=h'_{\lambda\bar{\gamma}}(z'(z))a_{\lambda\alpha}\bar{a}_{\gamma\beta}{\operatorname{d}}z_{\alpha}\wedge {\operatorname{d}}\bar{z}_{\beta},\qquad \text{where} \quad a_{\alpha\lambda}= \frac{{\partial}z'_{\alpha}}{{\partial}z_{\lambda}}.\end{gathered}$$ If $\omega_M(z')=\omega_M(z)$, then $$\begin{gathered}
\label{cnd1}
h_{\alpha\bar{\beta}}(z)=h'_{\lambda\gamma}(z'(z))a_{\lambda\alpha}\bar{a}_{\gamma\beta}.\end{gathered}$$ Now we consider the Laplace–Beltrami operator in the point $z'$. We have $$\begin{gathered}
\frac{{\partial}}{{\partial}z'_{\alpha}}=b_{\gamma\alpha}\frac{{\partial}}{{\partial}z_{\gamma}},\qquad \text{where} \quad b_{\lambda\beta}=\frac{{\partial}z_{\lambda}}{{\partial}z'_{\beta}},\end{gathered}$$ and $$\begin{gathered}
a_{\alpha\lambda}b_{\lambda\beta}=\delta_{\alpha\beta}.\end{gathered}$$ We have $$\begin{gathered}
\Delta_M(z')=(h' (z'))^{-1}_{\alpha\bar{\beta}}\bar{b}_{\gamma\alpha}b_{\rho\beta}\frac{{\partial}^2}{{\partial}\bar{z}_{\gamma}{\partial}z_{\rho}}.\end{gathered}$$ If $\Delta_M(z')=\Delta_M(z)$, then $$\begin{gathered}
\label{cnd2}
h^{-1}_{\gamma\bar{\rho}}(z)=(h' (z'(z)))^{-1}_{\alpha\bar{\beta}}\bar{b}_{\gamma\alpha}b_{\rho\beta}.\end{gathered}$$ We write as $$\begin{gathered}
h=a^th'\bar{a},\end{gathered}$$ which implies $$\begin{gathered}
h^{-1}=(\bar{a})^{-1}(h')^{-1}(a^t)^{-1}=\bar{b}(h')^{-1}b^t,\end{gathered}$$ which is exactly .
A lemma
-------
Formula , modulo the numerical factor, was obtained in [@sbj]. We recall the method to determine the Liouville form used in [@sbj]. We have applied the following technique (see [@hua Chapter IV]):
\[lemmahua\] Let $z'=f (g,z)$ denote the action of the group $G$ on the circular domain $M$. Let us determine the element $g\in G$ such that $z'(z_1)=0$ Then the density of the volume form is $Q= \vert J\vert^2$, where $J$ is the Jacobian $J=\frac{{\partial}z'}{{\partial}z}$.
The transformation with the desired properties is $$\begin{gathered}
z'= U(1-W_1\bar{W}_1)^{1/2}(1-W\bar{W}_1)^{-1}\\
\hphantom{z'= }{} \times \big[z-(1-W\bar{W}_1)(1-W_1\bar{W_1})^{-1}z_1 +
(W-W_1)(1-\bar{W}_1W_1)^{-1}\bar{z}_1\big], \\
W' = U (1-W_1\bar{W}_1)^{-1/2}(W-W_1) (1-\bar{W}_1W)^{-1}(1-\bar{W}_1W_1)^{1/2}U^t, $$ where $U$ is an unitary matrix. We find that $$\begin{gathered}
\frac{{\partial}z'}{{\partial}z}= U(1-W_1\bar{W}_1)^{1/2}(1-W\bar{W}_1)^{-1},\nonumber\\
\label{tt2}
{\operatorname{d}}W'=A {\operatorname{d}}W A^t, \qquad A = U(1-W_1\bar{W}_1)^{1/2}(1-W\bar{W}_1)^{-1}.\end{gathered}$$
In order to calculate the Jacobian of the transformation , we use the following property extracted from Berezin’s paper [@ber75 p. 398]:
[*Let $A$ be a matrix and $L_A$ the transformation of a matrix of the same order $n$, $L_A\xi = A\xi A^t$. If the matrices $A$ and $\xi$ are symmetric, then $\det L_A= (\det A)^{n+1}$*]{}. We find as in [@sbj], in accord with formula before Theorem 4.3.2 in [@hua] $$\begin{gathered}
\label{Qq}
Q_{{{\mathcal{D}}}_n}=\det({{\mathbbm{1}}_n}-W\bar{W})^{-(n+1)},\end{gathered}$$ while $Q_{{{\mathcal{D}}}^J_n}(z,W)$ has the expression .
Appendix\
Laplace–Beltrami operator on the Siegel upper half-plan\
and Siegel ball {#APP2}
========================================================
The Laplace–Beltrami operator on ${{\mathcal{X}}}_n$ has the expression (see [@maa equation (19)]) $$\begin{gathered}
\label{oarecum}\Delta_{{{\mathcal{X}}}_n}=-{\operatorname{Tr}}(Z-W)((Z-W)D_w)^tD_z,
\qquad Z=X+\ii Y \in {{\mathcal{X}}}_n, \qquad W=\bar{Z},\end{gathered}$$ where we use the notation . Formula (43) in [@maa] reads $$\begin{gathered}
\label{DDW}
D_w(Z-W)=-\frac{n+1}{2}{{\mathbbm{1}}_n}+((Z-W)D_w)^t,\end{gathered}$$ where $$\begin{gathered}
D_w(Z-W) =\sum_{\rho=1}^n{{\mbox{\rm e}}}_{\mu\rho}\frac{{\partial}}{{\partial}w_{\mu\rho}}(z_{\rho\nu}-w_{\rho\nu})\nonumber\\
\hphantom{D_w(Z-W) }{} =
\left(\sum_{\rho=1}^n(z_{\mu\rho}-w_{\mu\rho})e_{\rho\nu}\frac{{\partial}}{{\partial}w_{\rho\nu}}\right)^t-\left(\delta_{\mu\nu}\sum_{\rho=1}^ne_{\mu\rho}\right).\label{ma1}\end{gathered}$$
With formula , introducing in , it is obtained $$\begin{gathered}
((Z-W)D_w)^t_{\mu\nu}(f)=\left[\frac{{\partial}f}{{\partial}W}(Z-W)\right]_{\mu\nu}=(z_{\rho \nu}-w_{\rho\nu})e_{\mu\rho}\frac{{\partial}f}{{\partial}w_{\mu\rho}}. \end{gathered}$$ Equation (54) of the Cayley transform [@maa] in the notation reads $$\begin{gathered}
\tilde{Z}=(Z-\ii{{\mathbbm{1}}_n})(Z+\ii {{\mathbbm{1}}_n})^{-1}.\end{gathered}$$ Equation (55) in [@maa] asserts that $$\begin{gathered}
D_z =-\frac{\ii}{2}(\tilde{Z}-{{\mathbbm{1}}_n})((\tilde{Z}-{{\mathbbm{1}}_n})D_{\tilde{z}})^t,\end{gathered}$$ i.e., if now $W$ describes a point in ${{\mathcal{D}}}_n$, we have the formula
\[masH\] $$\begin{gathered}
(D_z)_{\alpha\beta} :=e_{\alpha\beta}\frac{{\partial}}{{\partial}z_{\alpha\beta}} =-\frac{\ii}{2}\big[({{\mathbbm{1}}_n}-W)(({{\mathbbm{1}}_n}-W)D_w)^t\big]_{\alpha\beta} \nonumber\\
\hphantom{(D_z)_{\alpha\beta}}{} =
-\frac{\ii}{2}({{\mathbbm{1}}_n}-W)_{\alpha\gamma}({{\mathbbm{1}}_n}-W)_{\beta\lambda}e_{\lambda\gamma}\frac{{\partial}}{{\partial}w_{\lambda\gamma}}.\label{DZZ1}\end{gathered}$$
We write as $$\begin{gathered}
2\ii {\operatorname{d}}W= A {\operatorname{d}}Z A, \qquad A={{\mathbbm{1}}_n}-W, \qquad W=W^t, \qquad Z=Z^t,\end{gathered}$$ i.e., $$\begin{gathered}
2\ii {\operatorname{d}}w_{pq}=A_{pm}{\operatorname{d}}z_{mn}A_{nq}.\end{gathered}$$ With formula , we get $$\begin{gathered}
2\ii \frac{{\partial}w_{pq}}{{\partial}z_{\alpha\beta}} = A_{pm}A_{nq}\frac{{\partial}z_{mn}}{{\partial}z_{\alpha\beta}}= A_{pm}A_{nq} (\delta_{m\alpha}\delta_{n\beta}+\delta_{m\beta}\delta_{n\alpha}-\delta_{mn}\delta_{\alpha\beta}\delta_{m\alpha})\nonumber\\
\hphantom{2\ii \frac{{\partial}w_{pq}}{{\partial}z_{\alpha\beta}}}{}
= A_{p\alpha}A_{\beta q}+A_{p\beta}A_{\alpha q}-A_{p\alpha}A_{\beta q}\delta_{\alpha\beta}.\label{ref56}\end{gathered}$$ We have the formula $$\begin{gathered}
\frac{{\partial}f}{{\partial}z_{\alpha\beta}}=\sum_{p\le q}\frac{{\partial}f}{{\partial}w_{pq}}\frac{{\partial}w_{pq}}{{\partial}z_{\alpha\beta}}.\end{gathered}$$ For $\alpha\not=\beta$, we get from : $$\begin{gathered}
2\ii \frac{{\partial}f}{{\partial}z_{\alpha\beta}} =
\sum_{p<q}(A_{p\alpha}A_{\beta q}+A_{p\beta}A_{\alpha q})\frac{{\partial}f}{{\partial}w_{pq}}+\sum_{p=q}(A_{p\alpha}A_{\beta p}+A_{p\beta}A_{\alpha
p})\frac{{\partial}f}{{\partial}w_{pq}}\nonumber\\
\hphantom{2\ii \frac{{\partial}f}{{\partial}z_{\alpha\beta}}}{}
= \sum_{p<q}A_{p\alpha}A_{\beta q}\frac{{\partial}f}{{\partial}w_{pq}}+\sum_{q<p}A_{q\beta }A_{\alpha p}\frac{{\partial}f}{{\partial}w_{pq}}+2\sum_{p=q}A_{p\alpha}A_{\beta q}\frac{{\partial}f}{{\partial}w_{pq}}\delta_{pq}\nonumber\\
\hphantom{2\ii \frac{{\partial}f}{{\partial}z_{\alpha\beta}}}{}
= \sum_{pq}A_{p\alpha}A_{\beta q}(1+\delta_{pq})\frac{{\partial}f}{{\partial}w_{pq}}.\label{for1}\end{gathered}$$ Similarly, for $\alpha=\beta$, we have $$\begin{gathered}
2\ii \frac{{\partial}f}{{\partial}z_{\alpha\beta}} = \sum_{p\le q}A_{p\alpha}A_{\beta q}\frac{{\partial}f}{{\partial}w_{pq}}\nonumber\\
\hphantom{2\ii \frac{{\partial}f}{{\partial}z_{\alpha\beta}}}{}
= \frac{1}{2} \sum_{p< q}A_{p\alpha}A_{\beta q}\frac{{\partial}f}{{\partial}w_{pq}}+ \frac{1}{2} \sum_{q< p}A_{q\alpha}A_{\beta p}\frac{{\partial}f}{{\partial}w_{qp}}\frac{1}{2}+ \sum_{p= q}A_{p\alpha}A_{\beta q}\frac{{\partial}f}{{\partial}w_{pq}}\nonumber\\
\hphantom{2\ii \frac{{\partial}f}{{\partial}z_{\alpha\beta}}}{}
= \frac{1}{2} \sum_{p\not= q}A_{p\alpha}A_{\beta q}\frac{{\partial}f}{{\partial}w_{pq}}+ \sum_{p= q}A_{p\alpha}A_{\beta q}\frac{{\partial}f}{{\partial}w_{pq}}= \sum_{p q}A_{p\alpha}A_{\beta q}\frac{1+\delta_{pq}}{2}\frac{{\partial}f}{{\partial}w_{pq}}.\label{for2}\end{gathered}$$ We put together and as in .
With Lemma \[masH\], it is obtained the Laplace–Beltrami operator on the Siegel ball (see [@maa equation (56)])
\[LLB\] The Laplace–Beltrami operator on the Siegel ball ${{\mathcal{D}}}_n$ has the expression $$\begin{gathered}
\label{UndeE}
\Delta_{{{\mathcal{D}}}_n} (W) ={\operatorname{Tr}}\big[N(ND_{\bar{W}})^tD_W\big] =\sum_{p,q,m,n}{{\amsbb{K}}}_{pq\bar{m}\bar{n}}\frac{{\partial}^2}{{\partial}\bar{w}_{mn}{\partial}w_{pq}}, \qquad N:={{\mathbbm{1}}_n}-W\bar{W},
\\
\label{KPQ} {{\amsbb{K}}}_{pq\bar{m}\bar{n}}=e_{mn}e_{pq} K_{pq\bar{m}\bar{n}},\qquad K_{pq\bar{m}\bar{n}} =N_{qn}\bar{N}_{mp},\end{gathered}$$ which is $\operatorname{Sp}(n,{{\amsbb{R}}})_{{{\amsbb{C}}}}$-invariant to the action . We write down the sum as $$\begin{gathered}
\Delta_{{{\mathcal{D}}}_n}(W)=\sum_{m\leq n;\, p\leq q}k_{mn\bar{p}\bar{q}}(W)\frac{{\partial}^2}{{\partial}\bar{w}_{mn}{\partial}w_{pq}},\end{gathered}$$ where, with the expression , we have $$\begin{gathered}
k_{pq\bar{m}\bar{n}} =
\frac{1}{4}(K_{pq\bar{m}\bar{n}}+K_{qp\bar{m}\bar{n}}+K_{pq\bar{n}\bar{m}}+K_{qp\bar{n}\bar{m}}),\qquad m<n, \qquad p<q,\nonumber\\
k_{pq\bar{m}\bar{m}} =\frac{1}{2}(K_{pq\bar{m}\bar{m}}+K_{qp\bar{m}\bar{m}}), \qquad p<q, \qquad n=m,\nonumber\\
k_{pp\bar{m}\bar{n}} =\frac{1}{2}(K_{pp\bar{m}\bar{n}}+K_{pp\bar{n}\bar{m}}), \qquad p=q, \qquad m<n,\nonumber\\
k_{pp\bar{m}\bar{m}} = K_{pp\bar{m}\bar{m}}.\label{sp1}\end{gathered}$$ can be written down as $$\begin{gathered}
k_{mn\bar{p}\bar{q} } = e_{mn}(1-\delta_{pq}) N_{pm}\bar{N}_{nq} + e_{pq}(1-\delta_{mn}) N_{qm}\bar{N}_{np}
+\delta_{pq}\delta_{mn} N_{pm}\bar{N}_{nq}\nonumber\\
\hphantom{k_{mn\bar{p}\bar{q}}}{} =\frac{1}{2}( N_{qn}\bar{N}_{mp}+ N_{qm}\bar{N}_{np}).\label{ofi}\end{gathered}$$
With equation introduced in , it is obtained .
In order to directly (i.e., without Remark \[equivv\]) prove the invariance of $ \Delta_{{{\mathcal{D}}}_n}$ to the group action $g\times W=W_1$, $g\in\operatorname{Sp}(n,{{\amsbb{R}}})_C$, $W, W_1\in{{\mathcal{D}}}_n$, we use the formulas , . We write as $$\begin{gathered}
{\operatorname{d}}w_{pq}= A_{pm}({\operatorname{d}}w_1)_{mn}A_{qn}, \qquad\text{where}\quad A=Wq^*+p^*.\end{gathered}$$ With formula , we have successively $$\begin{gathered}
\frac{{\partial}w_{pq}}{{\partial}(w_1)_{\alpha\beta}} = A_{pm}\frac{{\partial}(w_1)_{mn}}{{\partial}(w_1)_{\alpha\beta}}A_{nq}
= A_{pm}A_{nq} \Delta^{mn}_{\alpha\beta}\nonumber\\
\hphantom{\frac{{\partial}w_{pq}}{{\partial}(w_1)_{\alpha\beta}}}{}
=A_{p\alpha}A_{q\beta}+A_{p\beta}A_{q\alpha}-\delta_{\alpha\beta}A_{p\alpha}A_{q\beta}.\label{PARQ}\end{gathered}$$ With , we calculate $\frac{{\partial}f}{{\partial}(w_1)_{\alpha\beta}}$.
a\) Firstly, we consider the case $\alpha\not=\beta$. We have $$\begin{gathered}
\frac{{\partial}f}{{\partial}( w_1)_{\alpha\beta}}= \left(\sum_{p<q}A_{p\alpha}A_{q\beta} + \sum_{q<p}A_{q\alpha}A_{p\beta}+2\sum_{p=q}A_{p\alpha}A_{q\beta}\right) \frac{{\partial}f}{{\partial}w_{pq}}\nonumber\\
\hphantom{\frac{{\partial}f}{{\partial}( w_1)_{\alpha\beta}}}{} =2 \sum A_{p\alpha}A_{q\beta}e_{pq}\frac{{\partial}f}{{\partial}w_{pq}}.\label{SUR1}\end{gathered}$$
b\) Now we consider the case $\alpha=\beta$. We have $$\begin{gathered}
\label{SUR2}
\frac{{\partial}f}{{\partial}( w_1)_{\alpha\beta} }= 2f_{\alpha\beta}\sum_{p,q}A_{p\alpha}A_{q\beta}e_{pq}.\end{gathered}$$ and are written together as $$\begin{gathered}
\label{LAB}
D_{W_1}=A^tD_WA.\end{gathered}$$ Also we write as $$\begin{gathered}
\label{DM2}
M_1=\bar{A}^tMA.\end{gathered}$$ With and introduced in formula , we prove $$\begin{gathered}
\Delta_{{{\mathcal{D}}}_n}(W_1)=\Delta_{{{\mathcal{D}}}_n}(W), \qquad\text{where}\quad W_1=g\times W.\tag*{{}}\end{gathered}$$
In the notation of [@hua59] $$\begin{gathered}
Z=\left(\begin{matrix} \sqrt{2}
z_{11}& z_{12}& \dots& z_{1n}\\
z_{12}& \sqrt{2}z_{22}&
\dots& z_{2n}\\
\dots&\dots&\dots&\dots\\
z_{1n}&
z_{2n}&\dots&\sqrt{2} z_{nn}\end{matrix}\right),\qquad
{\partial}_Z=\left(\begin{matrix} \sqrt{2}\frac{{\partial}}{{\partial}z_{11}}&\frac{{\partial}}{{\partial}z_{12}}& \dots&\frac{{\partial}}{{\partial}z_{1n}}\vspace{1mm}\\
\frac{{\partial}}{{\partial}z_{12}}& \sqrt{2}\frac{{\partial}}{{\partial}z_{22}}&
\dots&\frac{{\partial}}{{\partial}z_{2n}}\\
\dots&\dots&\dots&\dots\\
\frac{{\partial}}{{\partial}z_{1n}}&\frac{{\partial}}{{\partial}z_{2n}}&\dots&\sqrt{2}\frac{{\partial}}{{\partial}z_{nn}}\end{matrix}\right),\end{gathered}$$ the Laplace–Beltrami operator reads (see [@hua59 equation (2.6.4)]) $$\begin{gathered}
\Delta_{{{\mathcal{D}}}_n}(W)={\operatorname{Tr}}(({{\mathbbm{1}}_n}-\bar{W}W){\partial}_W({{\mathbbm{1}}_n}-W\bar{W})\bar{{\partial}}_W),\end{gathered}$$ i.e., $$\begin{gathered}
\Delta_{{{\mathcal{D}}}_n}(W) =\sum_{\alpha\beta\lambda\mu=1}^n(\delta_{\lambda\mu}-\sum_{\sigma=1}^nP_{\lambda\sigma}P_{\mu\sigma}w_{\lambda\sigma}\bar{w}_{\mu\sigma})\\
\hphantom{\Delta_{{{\mathcal{D}}}_n}(W) =}{} \times(\delta_{\alpha\beta}-\sum_{\gamma=1}^nP_{\alpha\gamma}P_{\beta\gamma}w_{\alpha\gamma}\bar{w}_{\beta\gamma})
P_{\lambda\alpha}P_{\mu\beta}\frac{{\partial}^2}{{\partial}w_{\lambda\alpha}{\partial}\bar{w}_{\mu\beta}},
\end{gathered}$$ where $$\begin{gathered}
P_{\alpha\beta}=1+(\sqrt{2}-1)\delta_{\alpha\beta}.\end{gathered}$$ “$P$” is related with the symbol “$e$” in by the relation $$\begin{gathered}
P_{\alpha\beta} =\sqrt{2}(\sqrt{2}-1)(1+\sqrt{2}e_{\alpha\beta}).\end{gathered}$$
Acknowledgements {#acknowledgements .unnumbered}
----------------
I am grateful to Daniel Beltita for clarifying some aspects of contractibility related with this paper. This research was conducted in the framework of the ANCS project program PN 16 42 01 01/2016 and UEFISCDI-Romania program PN-II-PCE-55/05.10.2011.
[99]{}
Arezzo C., Loi A., Moment maps, scalar curvature and quantization of [K]{}ähler manifolds, [*Comm. Math. Phys.*](http://dx.doi.org/10.1007/s00220-004-1053-3) **246** (2004), 543–559.
Berceanu S., Coherent states and geodesics: cut locus and conjugate locus, [*J. Geom. Phys.*](http://dx.doi.org/10.1016/S0393-0440(96)00011-3) **21** (1997), 149–168, [dg-ga/9502007](http://arxiv.org/abs/dg-ga/9502007).
Berceanu S., On the geometry of complex [G]{}rassmann manifold, its noncompact dual and coherent states, *Bull. Belg. Math. Soc. Simon Stevin* **4** (1997), 205–243, [dg-ga/9509002](http://arxiv.org/abs/dg-ga/9509002).
Berceanu S., Realization of coherent state [L]{}ie algebras by differential operators, in Advances in Operator Algebras and Mathematical Physics, *Theta Ser. Adv. Math.*, Vol. 5, Editors F. Boca, O. Bratteli, R. Longo, H. Siedentop, Theta, Bucharest, 2005, 1–24, [math.DG/0504053](http://arxiv.org/abs/math.DG/0504053).
Berceanu S., A holomorphic representation of the [J]{}acobi algebra, [*Rev. Math. Phys.*](http://dx.doi.org/10.1142/S0129055X06002619) **18** (2006), 163–199, [E]{}rrata, [*Rev. Math. Phys.*](http://dx.doi.org/10.1142/S0129055X12920018) **24** (2012), 1292001, 2 pages, [math.DG/0408219](http://arxiv.org/abs/math.DG/0408219).
Berceanu S., A holomorphic representation of the semidirect sum of symplectic and [H]{}eisenberg [L]{}ie algebras, *J. Geom. Symmetry Phys.* **5** (2006), 5–13.
Berceanu S., A holomorphic representation of the multidimensional [J]{}acobi algebra, in Perspectives in Operator Algebras and Mathematical Physics, *Theta Ser. Adv. Math.*, Vol. 8, Editors F.-P. Boca, R. Purice, S. Stratila, Theta, Bucharest, 2008, 1–25, [math.DG/0604381](http://arxiv.org/abs/math.DG/0604381).
Berceanu S., The [J]{}acobi group and the squeezed states – some comments, [*AIP Conf. Proc.*](http://dx.doi.org/10.1063/1.3275594) **1191** (2009), 21–29, [arXiv:0910.5563](http://arxiv.org/abs/0910.5563).
Berceanu S., A convenient coordinatization of [S]{}iegel-[J]{}acobi domains, [*Rev. Math. Phys.*](http://dx.doi.org/10.1142/S0129055X12500249) **24** (2012), 1250024, 38 pages, [arXiv:1204.5610](http://arxiv.org/abs/1204.5610).
Berceanu S., Coherent states and geometry on the [S]{}iegel–[J]{}acobi disk, [*Int. J. Geom. Methods Mod. Phys.*](http://dx.doi.org/10.1142/S0219887814500352) **11** (2014), 1450035, 25 pages, [arXiv:1307.4219](http://arxiv.org/abs/1307.4219).
Berceanu S., Quantum mechanics and geometry on the [S]{}iegel–[J]{}acobi disk, in Geometric Methods in Physics, Editors P. Kielanowski, P. Bieliavsky, A. Odesski, A. Odzijewicz, M. Schlichenmaier, T. Voronov, [*Trends in Mathematics*](http://dx.doi.org/10.1007/978-3-319-06248-8_8), Birkhäuser, Basel, 2014, 89–98.
Berceanu S., Bergman representative coordinates on the [S]{}iegel–[J]{}acobi disk, *Romanian J. Phys.* **60** (2015), 867–896, [arXiv:1409.0368](http://arxiv.org/abs/1409.0368).
Berceanu S., Gheorghe A., Differential operators on orbits of coherent states, *Romanian J. Phys.* **48** (2003), 545–556, [math.DG/0211054](http://arxiv.org/abs/math.DG/0211054).
Berceanu S., Gheorghe A., Applications of the [J]{}acobi group to quantum mechanics, *Romanian J. Phys.* **53** (2008), 1013–1021, [arXiv:0812.0448](http://arxiv.org/abs/0812.0448).
Berceanu S., Gheorghe A., On the geometry of [S]{}iegel–[J]{}acobi domains, [*Int. J. Geom. Methods Mod. Phys.*](http://dx.doi.org/10.1142/S0219887811005920) **8** (2011), 1783–1798, [arXiv:1011.3317](http://arxiv.org/abs/1011.3317).
Berceanu S., Schlichenmaier M., Coherent state embeddings, polar divisors and [C]{}auchy formulas, [*J. Geom. Phys.*](http://dx.doi.org/10.1016/S0393-0440(99)00075-3) **34** (2000), 336–358, [math.DG/9903105](http://arxiv.org/abs/math.DG/9903105).
Berezin F.A., Quantization in complex bounded domains, *Dokl. Akad. Nauk SSSR* **211** (1973), 1263–1266.
Berezin F.A., Quantization, *Izv. Akad. Nauk SSSR Ser. Mat.* **38** (1974), 1116–1175.
Berezin F.A., General concept of quantization, [*Comm. Math. Phys.*](http://dx.doi.org/10.1007/BF01609397) **40** (1975), 153–174.
Berezin F.A., Quantization in complex symmetric spaces, *Izv. Akad. Nauk SSSR Ser. Mat.* **39** (1975), 363–402.
Berndt R., Sur l’arithmétique du corps des fonctions elliptiques de niveau [$N$]{}, in Seminar on Number Theory, [P]{}aris 1982–83 ([P]{}aris, 1982/1983), *Progr. Math.*, Vol. 51, Birkhäuser Boston, Boston, MA, 1984, 21–32.
Berndt R., B[ö]{}cherer S., Jacobi forms and discrete series representations of the [J]{}acobi group, [*Math. Z.*](http://dx.doi.org/10.1007/BF02570858) **204** (1990), 13–44.
Berndt R., Schmidt R., Elements of the representation theory of the [J]{}acobi group, [*Progr. Math.*](http://dx.doi.org/10.1007/978-3-0348-0283-3), Vol. 163, Birkhäuser Verlag, Basel, 1998.
Borel A., Kählerian coset spaces of semisimple [L]{}ie groups, [*Proc. Nat. Acad. Sci. USA*](http://dx.doi.org/10.1073/pnas.40.12.1147) **40** (1954), 1147–1151.
Cahen B., Global parametrization of scalar holomorphic coadjoint orbits of a quasi-[H]{}ermitian [L]{}ie group, *Acta Univ. Palack. Olomuc. Fac. Rerum Natur. Math.* **52** (2013), 35–48.
Cahen B., Stratonovich–[W]{}eyl correspondence for the [J]{}acobi group, *Commun. Math.* **22** (2014), 31–48.
Cahen M., Gutt S., Rawnsley J., Quantization of [K]{}ähler manifolds. [II]{}, [*Trans. Amer. Math. Soc.*](http://dx.doi.org/10.2307/2154310) **337** (1993), 73–98.
Calabi E., Isometric imbedding of complex manifolds, [*Ann. of Math.*](http://dx.doi.org/10.2307/1969817) **58** (1953), 1–23.
Chern S.S., Complex manifolds without potential theory, 2nd ed., [*Universitext*](http://dx.doi.org/10.1007/978-1-4684-9344-3), Springer-Verlag, New York – Heidelberg, 1979.
Chiribella G., Adesso G., Quantum benchmarks for pure single-mode [G]{}aussian states, [*Phys. Rev. Lett.*](http://dx.doi.org/10.1103/PhysRevLett.112.010501) **112** (2014), 010501, 6 pages, [arXiv:1308.2146](http://arxiv.org/abs/1308.2146).
Donaldson S.K., Scalar curvature and projective embeddings. [I]{}, *J. Differential Geom.* **59** (2001), 479–522.
Dorfmeister J., Nakajima K., The fundamental conjecture for homogeneous [K]{}ähler manifolds, [*Acta Math.*](http://dx.doi.org/10.1007/BF02392294) **161** (1988), 23–70.
Eichler M., Zagier D., The theory of [J]{}acobi forms, [*Progr. Math.*](http://dx.doi.org/10.1007/978-1-4684-9162-3), Vol. 55, Birkhäuser Boston, Inc., Boston, MA, 1985.
Engli[š]{} M., Berezin quantization and reproducing kernels on complex domains, [*Trans. Amer. Math. Soc.*](http://dx.doi.org/10.1090/S0002-9947-96-01551-6) **348** (1996), 411–479.
Gay-Balmaz F., Tronci C., Vlasov moment flows and geodesics on the [J]{}acobi group, [*J. Math. Phys.*](http://dx.doi.org/10.1063/1.4763467) **53** (2012), 123502, 36 pages, [arXiv:1105.1734](http://arxiv.org/abs/1105.1734).
Helgason S., Differential geometry, [L]{}ie groups, and symmetric spaces, *Pure and Applied Mathematics*, Vol. 80, Academic Press, Inc., New York – London, 1978.
Hua L.-K., On the theory of automorphic functions of a matrix level. [I]{}. [G]{}eometrical basis, [*Amer. J. Math.*](http://dx.doi.org/10.2307/2371910) **66** (1944), 470–488.
Hua L.-K., Harmonic analysis of functions of several complex variables in the classical domains, Amer. Math. Soc., Providence, R.I., 1963.
Hua L.-K., Look K.H., Theory of harmonic functions in classical domains, *Scientia Sinica* **8** (1959), 743–806.
Hunziker M., Sepanski M.R., Stanke R.J., Global [L]{}ie symmetries of a system of [S]{}chrödinger equations and the oscillator representation, *Miskolc Math. Notes* **14** (2013), 647–657.
Jost J., Riemannian geometry and geometric analysis, 5th ed., [*Universitext*](http://dx.doi.org/10.1007/978-3-642-21298-7), Springer-Verlag, Berlin, 2008.
K[ä]{}hler E., Raum-[Z]{}eit-[I]{}ndividuum, *Rend. Accad. Naz. Sci. XL Mem. Mat. (5)* **16** (1992), 115–177.
K[ä]{}hler E., Mathematische [W]{}erke/[M]{}athematical works, [Walter de Gruyter & Co.](http://dx.doi.org/10.1515/9783110905434), Berlin, 2003.
Kobayashi S., Geometry of bounded domains, [*Trans. Amer. Math. Soc.*](http://dx.doi.org/10.1090/S0002-9947-1959-0112162-5) **92** (1959), 267–290.
Kobayashi S., Nomizu K., Foundations of differential geometry. [V]{}ol. [I]{}, Interscience Publishers, New York – London, 1963.
Kobayashi S., Nomizu K., Foundations of differential geometry. [V]{}ol. [II]{}, Interscience Publishers, New York – London – Sydney, 1969.
Kostant B., Quantization and unitary representations. [I]{}. [P]{}requantization, in Lectures in Modern Analysis and Applications, [III]{}, [*Lecture Notes in Math.*](http://dx.doi.org/10.1007/BFb0079068), Vol. 170, Springer, Berlin, 1970, 87–208.
Kramer P., Saraceno M., Semicoherent states and the group [${\rm ISp}(2,{\bf
R})$]{}, [*Phys. A*](http://dx.doi.org/10.1016/0378-4371(82)90330-2) **114** (1982), 448–453.
Lee M.H., Theta functions on [H]{}ermitian symmetric domains and [F]{}ock representations, [*J. Aust. Math. Soc.*](http://dx.doi.org/10.1017/S1446788700003256) **74** (2003), 201–234.
Lisiecki W., Coherent state representations. [A]{} survey, [*Rep. Math. Phys.*](http://dx.doi.org/10.1016/0034-4877(96)89292-5) **35** (1995), 327–358.
Loi A., Regular quantizations of [K]{}ähler manifolds and constant scalar curvature metrics, [*J. Geom. Phys.*](http://dx.doi.org/10.1016/j.geomphys.2004.07.006) **53** (2005), 354–364.
Loi A., Mossa R., Berezin quantization of homogeneous bounded domains, [*Geom. Dedicata*](http://dx.doi.org/10.1007/s10711-012-9697-1) **161** (2012), 119–128, [arXiv:1106.2510](http://arxiv.org/abs/1106.2510).
Loi A., Mossa R., Some remarks on homogeneous [K]{}ähler manifolds, [*Geom. Dedicata*](http://dx.doi.org/10.1007/s10711-015-0085-5) **179** (2015), 377–383, [arXiv:1502.00011](http://arxiv.org/abs/1502.00011).
Lu Q.-K., On Kähler manifolds with constant curvature, *Acta. Math. Sin.* **66** (1966), 269–281.
Lu Q.-K., Holomorphic invariant forms of a bounded domain, [*Sci. China Ser. A*](http://dx.doi.org/10.1007/s11425-008-0129-5) **51** (2008), 1945–1964.
L[ü]{}tkepohl H., Handbook of matrices, John Wiley & Sons, Ltd., Chichester, 1996.
Maass H., Die [D]{}ifferentialgleichungen in der [T]{}heorie der [S]{}iegelschen [M]{}odulfunktionen, [*Math. Ann.*](http://dx.doi.org/10.1007/BF01343149) **126** (1953), 44–68.
Mandel L., Wolf E., Optical coherence and quantum optics, Cambridge University Press, Cambridge, 1995.
Marmo G., Michor P.W., Neretin Yu.A., The [L]{}agrangian [R]{}adon transform and the [W]{}eil representation, [*J. Fourier Anal. Appl.*](http://dx.doi.org/10.1007/s00041-013-9315-0) **20** (2014), 321–361, [arXiv:1212.4610](http://arxiv.org/abs/1212.4610).
Molitor M., Gaussian distributions, [J]{}acobi group, and [S]{}iegel–[J]{}acobi space, [*J. Math. Phys.*](http://dx.doi.org/10.1063/1.4903182) **55** (2014), 122102, 40 pages, [arXiv:1409.7917](http://arxiv.org/abs/1409.7917).
Moroianu A., Lectures on [K]{}ähler geometry, [*London Mathematical Society Student Texts*](http://dx.doi.org/10.1017/CBO9780511618666), Vol. 69, Cambridge University Press, Cambridge, 2007.
Neeb K.-H., Holomorphy and convexity in [L]{}ie theory, [*de Gruyter Expositions in Mathematics*](http://dx.doi.org/10.1515/9783110808148), Vol. 28, Walter de Gruyter & Co., Berlin, 2000.
Nomizu K., Invariant affine connections on homogeneous spaces, [*Amer. J. Math.*](http://dx.doi.org/10.2307/2372398) **76** (1954), 33–65.
Perelomov A., Generalized coherent states and their applications, [*Texts and Monographs in Physics*](http://dx.doi.org/10.1007/978-3-642-61629-7), Springer-Verlag, Berlin, 1986.
Quesne C., Vector coherent state theory of the semidirect sum [L]{}ie algebras [${\rm wsp}(2N,{\bf R})$]{}, [*J. Phys. A: Math. Gen.*](http://dx.doi.org/10.1088/0305-4470/23/6/011) **23** (1990), 847–862.
Rawnsley J., Cahen M., Gutt S., Quantization of [K]{}ähler manifolds. [I]{}. [G]{}eometric interpretation of [B]{}erezin’s quantization, [*J. Geom. Phys.*](http://dx.doi.org/10.1016/0393-0440(90)90019-Y) **7** (1990), 45–62.
Rawnsley J.H., Coherent states and [K]{}ähler manifolds, [*Quart. J. Math. Oxford Ser. (2)*](http://dx.doi.org/10.1093/qmath/28.4.403) **28** (1977), 403–415.
Satake I., Algebraic structures of symmetric domains, *Kanô Memorial Lectures*, Vol. 4, Iwanami Shoten, Tokyo, Princeton University Press, Princeton, N.J., 1980.
Sepanski M.R., Stanke R.J., Global [L]{}ie symmetries of the heat and [S]{}chrödinger equation, *J. Lie Theory* **20** (2010), 543–580.
Shuman K.L., Complete signal processing bases and the [J]{}acobi group, [*J. Math. Anal. Appl.*](http://dx.doi.org/10.1016/S0022-247X(02)00653-4) **278** (2003), 203–213.
Siegel C.L., Symplectic geometry, [*Amer. J. Math.*](http://dx.doi.org/10.2307/2371774) **65** (1943), 1–86.
Takase K., A note on automorphic forms, [*J. Reine Angew. Math.*](http://dx.doi.org/10.1515/crll.1990.409.138) **409** (1990), 138–171.
Takase K., On unitary representations of [J]{}acobi groups, [*J. Reine Angew. Math.*](http://dx.doi.org/10.1515/crll.1992.430.139) **430** (1992), 139–149.
Takase K., On [S]{}iegel modular forms of half-integral weights and [J]{}acobi forms, [*Trans. Amer. Math. Soc.*](http://dx.doi.org/10.1090/S0002-9947-99-02168-6) **351** (1999), 735–780.
Vinberg E.B., Gindikin S.G., Kähler manifolds admitting a transitive solvable group of automorphisms, [*Math. USSR Sb.*](http://dx.doi.org/10.1070/SM1967v003n03ABEH002746) **3** (1967), 333–351.
Yang J.-H., The method of orbits for real [L]{}ie groups, *Kyungpook Math. J.* **42** (2002), 199–272, [math.RT/0602056](http://arxiv.org/abs/math.RT/0602056).
Yang J.-H., Invariant metrics and [L]{}aplacians on [S]{}iegel–[J]{}acobi space, [*J. Number Theory*](http://dx.doi.org/10.1016/j.jnt.2006.12.014) **127** (2007), 83–102, [math.NT/0507215](http://arxiv.org/abs/math.NT/0507215).
Yang J.-H., A partial [C]{}ayley transform of [S]{}iegel–[J]{}acobi disk, [*J. Korean Math. Soc.*](http://dx.doi.org/10.4134/JKMS.2008.45.3.781) **45** (2008), 781–794, [math.NT/0507216](http://arxiv.org/abs/math.NT/0507216).
Yang J.-H., Invariant metrics and [L]{}aplacians on [S]{}iegel–[J]{}acobi disk, [*Chin. Ann. Math. Ser. B*](http://dx.doi.org/10.1007/s11401-008-0348-7) **31** (2010), 85–100, [math.NT/0507217](http://arxiv.org/abs/math.NT/0507217).
Yang J.-H., Yong Y.-H., Huh S.-N., Shin J.-H., Min G.-H., Sectional survatures of the [S]{}iegel–[J]{}acobi space, [*Bull. Korean Math. Soc.*](http://dx.doi.org/10.4134/BKMS.2013.50.3.787) **50** (2013), 787–799.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we give the general solution of the functional equation $$\big\{\|f(x)+f(y)\|,\|f(x)-f(y)\|\big\}=\big\{\|x+y\|,\|x-y\|\big\}\qquad(x,y\in X)$$ where $f:X\to Y$ and $X, \,\,Y$ are inner product spaces. Related equations are also considered. Our main tool is a real version of Wigner’s unitary–antiunitary theorem.'
address: 'Institute of Mathematics, University of Debrecen, H-4010 Debrecen, Pf. 12, Hungary'
author:
- Gyula Maksa
- Zsolt Páles
title: 'Wigner’s theorem revisited'
---
Publ. Math. Debrecen **81**(1-2) (2012), 243–249.\
[10.5486/PMD.2012.5359](http://dx.doi.org/10.5486/PMD.2012.5359)
[^1]
Introduction
============
An isometry from a normed space $X$ into another normed space $Y$ is a function $f:X\to Y$ which satisfies the equality This equation implies strong structural properties for the function $f$. A classical result in this direction is a celebrated theorem of Mazur and Ulam [@MazUla32] which states that an isometry $f$ of a real normed space $\emph{onto}$ another normed space is necessarily affine. In other words, for the surjective solutions $f:X\to Y$ of , $x\mapsto f(x)-f(0)$, is a norm preserving linear map. Baker [@Bak71c] showed that the same conclusion remains valid if the surjectivity assumption is replaced by the strict convexity of the target space $Y$. Another important result which is related to linear isometries is Wigner’s theorem [@Wig31] and its generalization obtained by Rätz [@Rat96 Corollary 8(a)]. For further generalizations of this fundamental result, we mention the papers [@AlmSha92], [@Che07], [@Gyo04a], [@Mol96g], [@Mol98a], [@Mol99a], [@Mol00e], [@Mol00d], [@Mol01c], [@Mol02f], and [@SimMukChaSri08].
Assuming that $X$ and $Y$ are *real* inner product spaces, Rätz’s result characterizes functions $f:X\to Y$ that are phase equivalent to a linear isometry (i.e., there exists a function $\varepsilon:X\to\{-1,1\}$ such that $\varepsilon f$ is a norm preserving real linear map) by the property In the complex setting, Wigner’s theorem [@Wig31] (cf. also [@Gyo04a]) says that the solutions of are phase equivalent to a linear or conjugate linear isometry. Without assuming that $X$ and $Y$ are real inner product spaces, we can easily see that all functions $f:X\to Y$ that are phase equivalent to a real linear isometry are also solutions of the functional equation Indeed, if $\varepsilon:X\to\{-1,1\}$ and $g:=\varepsilon f$ is a norm preserving real linear map, then, for all $x,y\in X$, which implies because $\varepsilon(x)\varepsilon(y)$ is either equal to $1$ or to $-1$.
The aim of this short note is to show that the converse also holds provided that $X,Y$ are inner product spaces. That is, in that case, all solutions $f:X\to Y$ of are phase equivalent to a real linear isometry. The main tool in the proof is Rätz’s characterization theorem described above.
The equivalence of some functional equations related to and our main results
============================================================================
Throughout the remaining part of this paper, $X$ and $Y$ denote real or complex inner product spaces. We note that every complex linear space is trivially a real linear space and if $\langle\cdot,\cdot\rangle$ is a complex inner product on $X$ (or on $Y$) then $\langle\!\langle\cdot,\cdot\rangle\!\rangle$ defined as $\langle\!\langle x,y\rangle\!\rangle=\Re\langle x,y\rangle$ is real inner product on $X$ which induces the same norm. (Here $\Re z$ stands for the real part of the complex number $z$.) Therefore, we may assume that $\langle\cdot,\cdot\rangle$ always denotes the real inner product on $X$ and $Y$. A function $f:X\to Y$ is called real linear if $f$ is additive and homogeneous with respect to real numbers. Real linearity does not imply complex linearity in general as it is shown by the following example constructed by Rätz [@Rat96]: Let $X=Y=\C^{2}$ equipped with the usual inner product and $f(x_{1},x_{2})=(x_{1},\overline{x}_{2})$ for $(x_{1},x_{2})\in\C^{2}$. An easy calculation shows that $f$ is norm preserving real linear, but it is not complex-homogeneous, and hence it is not linear.
We begin with a characterization of norm-preserving real linear maps between inner product spaces.
Suppose first that (i) holds. Putting $x=y$, it follows that $f$ is norm-preserving. Now using (i) and the norm preserving property, we get which proves (ii).
Now suppose (ii). Putting $x=y$, the norm preserving property of $f$ follows. Using (ii) three times, for all $x,y,z\in X$, we obtain Applying this identity for $z\in\{x+y,x,y\}$, we get which yields that $f$ is additive.
Finally, assume that $f$ is a norm-preserving real linear map. Then, by the additivity and the norm-preserving property, we get that $\|f(x)+f(y)\|=\|f(x+y)\|=\|x+y\|$ which implies (i).
The equivalence of (i) and (iii) can easily be proved by supposing only that $X$ and $Y$ are normed spaces and $Y$ is strictly convex. Indeed, the substitution $y=x$ in (i) implies that $f$ is norm-preserving. Therefore $\|f(x)+f(y)\|=\|f(x+y)\|$ holds for all $x,y\in X$. Applying a result of Ger [@Ger93b], we obtain that $f$ is additive which implies (iii). On the other hand, (i) follows from (iii) immediately.
In the following theorem, we list four equivalent conditions that are equivalent to .
The statement (i) implies (ii) obviously. With the substitution $y=x$, it follows from (ii) that $f$ is norm preserving, i.e., With $x=0$, this yields $f(0)=0$. Now we square the equation in (ii) to obtain Using , the above equality simplifies to the second equality in (iii). Thus (ii) implies (iii).
Substituting $y=0$ into the second equation in (iii), we get . Squaring the second equation in (iii) and using again, we obtain that This simplifies to which is equivalent to the equation in (iv) proving that (iii) implies (iv).
If (iv) holds, then, by the result of Rätz [@Rat96 Corollary 8(a)] described in the introduction, (v) follows.
Finally, (v) implies (i) as we have seen it in the introduction.
The following corollary describes the continuous solutions of .
Assume that $f$ is a continuous function satisfying any of the conditions (i)–(iv) of . Then there exists a function $\varepsilon:X\to \{-1,1\}$ such that $\varepsilon f$ is norm-preserving and real linear. Thus, by , If $y\neq0$, then there exists an open ball $U$ around $y$ such that This, by the continuity of $f$, shows that $\varepsilon$ is continuous on $U$ and hence it is constant on $U$. The set $X\setminus \{0\}$ is connected (because $X$ is at least two dimensional), therefore $\varepsilon$ is constant on $X\setminus \{0\}$. Thus $f$ must be a norm-preserving real linear map.
In the exceptional nontrivial case when $X$ is one dimensional and real, say $X=\{\lambda a: \, \lambda\in\R\}$ with some $a\in X,\, \|a\|=1$, and $Y$ is at least one dimensional, the above argument shows that $\varepsilon$ is constant on the set of positive reals and constant also on the set of negative reals. Therefore $f$ is either a norm-preserving real linear map or $f(\lambda a)=|\lambda|b$ for all $\lambda\in\R$ and for some $b\in Y$ with $\|b\|=1$.
Finally, we formulate two open problems.
**Problem 1.** What are the solutions $f:X\to Y$ of when $X$ and $Y$ are normed but not necessarily inner product spaces? Under what conditions does it remain valid that, for the solutions of , $\varepsilon f$ is real linear for some function $\varepsilon:X\to\{-1,1\}$?
**Problem 2.** Let $X$ and $Y$ be complex normed spaces. Let $n$ be a fixed positive integer and denote $\beta_{1}, \dots, \beta_{n}$ the $n$th roots of unity. These elements form a multiplicative subgroup of the unit circle in $\C$. Find the solutions $f:X\to Y$ of the following generalization of : Obviously, this is the isometry equation in case $n=1$, and the case $n=2$ was just discussed in this paper. One can also see that if there exists a function $\varepsilon:X\to \{\beta_{1}, \dots, \beta_{n}\}$ such that $\varepsilon f$ is complex linear and norm-preserving, then $f$ satisfies . Under what conditions does it remain valid that, for the solutions of , $\varepsilon f$ is complex linear and norm-preserving for some function $\varepsilon:X\to \{\beta_{1}, \dots, \beta_{n}\}$?
**Acknowledgement.** The authors are indebted to Professor Lajos Molnár, who suggested the investigation of the functional equation in (ii) of Theorem 2 and had also several valuable comments.
[10]{}
D. F. Almeida and C. S. Sharma, *The first mathematical proof of [W]{}igner’s theorem*, J. Natur. Geom. **2** (1992), no. 2, 113–123.
J. A. Baker, *Isometries in normed spaces*, Amer. Math. Monthly **78** (1971), 655–658.
G. Chevalier, *Wigner’s theorem and its generalizations*, Handbook of quantum logic and quantum structures, Elsevier Sci. B. V., Amsterdam, 2007, pp. 429–475.
R. Ger, *On a characterization of strictly convex spaces*, Atti Accad. Sci. Torino Cl. Sci. Fis. Mat. Natur. **127** (1993), no. 3-4, 131–138.
M. Győry, *A new proof of [W]{}igner’s theorem*, Rep. Math. Phys. **54** (2004), no. 2, 159–167.
S. Mazur and S. Ulam, *Sur les transformations isometriques d’espaces vectoriels, normes.*, C. R. Acad. Sci., Paris **194** (1932), 946–948.
L. Molnár, *Wigner’s unitary-antiunitary theorem via [H]{}erstein’s theorem on [J]{}ordan homomorphisms*, J. Natur. Geom. **10** (1996), no. 2, 137–148.
L. Molnár, *An algebraic approach to [W]{}igner’s unitary-antiunitary theorem*, J. Austral. Math. Soc. Ser. A **65** (1998), no. 3, 354–369.
L. Molnár, *A generalization of [W]{}igner’s unitary-antiunitary theorem to [H]{}ilbert modules*, J. Math. Phys. **40** (1999), no. 11, 5544–5554.
L. Molnár, *Generalization of [W]{}igner’s unitary-antiunitary theorem for indefinite inner product spaces*, Commun. Math. Phys. **210** (2000), no. 3, 785–791.
L. Molnár, *Wigner-type theorem on symmetry transformations in type [I]{}[I]{} factors*, Internat. J. Theoret. Phys. **39** (2000), no. 6, 1463–1466.
L. Molnár, *A [W]{}igner-type theorem on symmetry transformations in [B]{}anach spaces*, Publ. Math. Debrecen **58** (2001), no. 1-2, 231–239.
L. Molnár, *Orthogonality preserving transformations on indefinite inner product spaces: generalization of [U]{}hlhorn’s version of [W]{}igner’s theorem*, J. Funct. Anal. **194** (2002), no. 2, 248–262.
J. Rätz, *On [W]{}igner’s theorem: remarks, complements, comments, and corollaries*, Aequationes Math. **52** (1996), no. 1-2, 1–9.
R. Simon, N. Mukunda, S. Chaturvedi, and V. Srinivasan, *Two elementary proofs of the [W]{}igner theorem on symmetry in quantum mechanics*, Phys. Lett. A **372** (2008), no. 46, 6847–6852.
E. Wigner, *Gruppentheorie und ihre Anwendung auf die Quantenmechanik der Atomsprekten*, Vieweg, Braunschweig, 1931.
[^1]: This research has been supported by the Hungarian Scientific Research Fund (OTKA) Grant NK81402 and by the TÁMOP 4.2.1./B-09/1/KONV-2010-0007 project implemented through the New Hungary Development Plan co-financed by the European Social Fund, and the European Regional Development Fund
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The issue of giving an explicit description of the flow of information concerning the time of bankruptcy of a company (or a state) arriving on the market is tackled by defining a bridge process starting from zero and conditioned to be equal to zero when the default occurs. This enables to catch some empirical facts on the behavior of financial markets: when the bridge process is away from zero, investors can be relatively sure that the default will not happen immediately. However, when the information process is close to zero, market agents should be aware of the risk of an imminent default. In this sense the bridge process leaks information concerning the default before it occurs. The objective of this first paper on Brownian bridges on stochastic intervals is to provide the basic properties of these processes.'
author:
- 'M. L. Bedini, R. Buckdahn, H.-J. Engelbert'
date: 'July 13, 2015'
title: Brownian Bridges on Random Intervals
---
\[sec:Introduction\]Introduction
================================
Motivated by the problem of modeling the information concerning the default time of a financial company, we propose a new approach to credit-risk. The problem is to give a mathematical model for the flow of information about the time at which the bankruptcy of a company (or state) occurs. The time at which this crucial event takes place is called *default time* and it is modeled by a strictly positive random variable $\tau$. We tackle the problem by defining the process $\beta=(\beta_{t},\, t\geq0)$, a Brownian bridge between 0 and 0 on the random time interval $[0,\tau]$: $$\beta_{t}:=W_{t}-\frac{t}{\tau\vee t}W_{\tau\vee t},\quad t\geq 0\,, \label{eq:EQ0}$$ where $W=(W_{t},\, t\geq0)$ is a Brownian motion independent of $\tau$. Since we are going to model the information about $\tau$ with such a bridge process, we call $\beta$ the *information process*. The filtration $\mathbb{F}^{\beta}=(\mathcal{F}_{t}^{\beta})_{t\geq0}$ generated by the information process provides partial information on the default time before it occurs. The intuitive idea is that when market agents observe $\beta_{t}$ away from 0, they know that the default will not occur immediately; on the other hand, the period of fear of an imminent default corresponds to the situation in which $\beta_{t}$ is close to 0. In our approach the Bayes formula is of particular importance since it allows to infer the a posteriori probability distribution of the default time $\tau$ given the information carried by $\mathcal{F}_{t}^{\beta}$, for $t\geq0$.
Two classes of credit risk models have greatly inspired our work: the information-based approach and the reduced-form models. The first one has been developed in [@key-20] by Brody, Hughston and Macrina. They model the information concerning a random payoff $D_{T}$, paid at some pre-established date $T>0$, with the natural, completed filtration generated by the process $\xi=(\xi_{t},\,0\leq t\leq T)$, defined by $\xi_{t}:=\beta_{t}^{T}+\alpha tD_{T}$, where $\beta^{T}=(\beta_{t}^{T},\,0\leq t\leq T)$ is a standard Brownian bridge on the deterministic time interval $[0,T]$. The process $\beta^{T}$ is independent of $D_{T}$, and $\alpha>0$ is a positive constant. The idea is that the true information, represented by the component $\alpha tD_{T}$, about the final payoff is disturbed by some noisy information (rumors, mispricing, etc.), represented by the bridge process. In this model for credit risk the information concerning the random cash-flow $D_{T}$ is modeled explicitly but the default time $\tau$ of the company is not. On the other hand, in the models following the reduced-form approach for credit risk the information on the default time is modeled by the natural, completed filtration generated by the single-jump process $H=(H_{t}:=\mathbb{I}_{\{ \tau\leq t\} },\, t\geq0)$ occurring at $\tau$. We refer to the book [@key-21] of Bielecki and Rutkowski and the series of papers [@key-23; @key-24; @key-25] of Jeanblanc and Le Cam, among many other works on the reduced-form approach to credit-risk. Besides the advantages that have made this approach well-known, there is the poor structure of the information concerning $\tau$: people just know if the default has occurred ($H_{t}=1$) or not ($H_{t}=0$). Financial reality is often more complex than this: market agents have indeed more information and there are actually periods in which the default is more likely to happen than in others. We try to reconcile those two approaches considering that in our model the information is carried by the process $\beta$ given by (\[eq:EQ0\]).
This paper is organized as follows. In Section 2 we provide some preliminary facts that are used throughout the paper. In Section 3 we give the definition of a bridge $\beta$ of random length $\tau$. We show that $\tau$ is a stopping time with respect to the natural, completed filtration generated by the process $\beta$. Also, we establish that $\beta$ is a Markov process with respect to the filtration $\mathbb{F}^{0}:=(\sigma(\beta_{s},\,0\leq s\leq t))_{t\geq0}$. In Section 4 we derive Bayesian estimates of the distribution of the default time $\tau$. Thereafter, in Section 5 we extend the results obtained in the previous section to more general situations. Section 6 is devoted to the proof of the Markov property of the information process with respect to the filtration $\mathbb{F}^{\beta}$, the smallest filtration which contains $\mathbb{F}^0$ and satisfies the usual conditions of right-continuity and completeness. In Section 7 we show that the process $\beta$ is a semimartingale and we provide its decomposition. Finally, in Section 8 we consider an application to Mathematical Finance concerned with the problem of pricing a Credit Default Swap in an elementary market model. The paper closes with the Appendix where, for the sake of easy reference, we recall the Bayes Theorem, and we prove a slight extension of the so-called innovation lemma (used in Section \[sec:Semimartingale-decomposition\] for the semimartingale decomposition of the information process $\beta$) and, finally, an auxiliary result.
This paper is based on the thesis [@key-10] and the main objective of it is to introduce and study the information process $\beta$ and to provide its basic properties. Other topics of the thesis [@key-10] are the enlargement of the filtration $\mathbb{F}^{\beta}$ with a reference filtration $\mathbb{F}$, the classification of the default time $\tau$ (with respect to the filtration $\mathbb{F}^{\beta}$ and with respect to the enlarged filtration) and further applications to Mathematical Finance.
Preliminaries
=============
We start by recalling some basic results on Brownian bridges and properties of conditional expectations that will be used in the sequel. The interested reader may find further discussions on Brownian bridges, among others, in the books [@key-1] of Karatzas and Shreve or [@key-3] of Revuz and Yor .
As usual, the set of natural numbers is denoted by $\mathbb{N}$ and the set of real numbers by $\mathbb{R}$. If $A\subseteq\mathbb{R}$, then the notation $A_{+}$ stands for $A_{+}:=A\cap\{ x\in\mathbb{R}:x\geq0\}$. If $E$ is a topological space, then the Borel $\sigma$-algebra over $E$ will be denoted by $\mathcal{B}(E)$. If $A$ is a set, its indicator function will be denoted by $\mathbb{I}_{A}$.
Let $(\Omega,\,\mathcal{F},\,\mathbf{P})$ be a complete probability space. By $\mathcal{N}_{P}$ we denote the collection of $\mathbf{P}$-null sets. If $X$ is a random variable, the symbol $\sigma(X)$ will denote the $\sigma$-algebra generated by $X$.
Let $W:(\Omega,\mathcal{F})\rightarrow(C,\mathcal{C})$ be a map from $\Omega$ into the space of continuous real-valued functions defined on $\mathbb{R}_{+}$ endowed with the $\sigma$-algebra generated by the canonical process, so that with each $\omega\in\Omega$ we associate the continuous function $W(\omega)=(W_{t}(\omega),\, t\geq0)$.
Given a strictly positive real number $r$, the function $\beta^{r}:\Omega\rightarrow C$ defined by $$\beta_{t}^{r}\left(\omega\right):=W_{t}\left(\omega\right)-\frac{t}{r\vee t}W_{r\vee t}\left(\omega\right), \ t\geq0, \ \omega\in\Omega\,,$$ is called *bridge of length $r$ associated with $W$*. If $W$ is a Brownian motion on the probability space $(\Omega,\mathcal{F},\mathbf{P})$, then the process $\beta^{r}$ is called *Brownian bridge of length $r$*.
We have the following fact concerning the measurability of the process $\beta^{r}$.
The map $(r,t,\omega)\mapsto\beta_{t}^{r}(\omega)$ of $((0,+\infty)\times\mathbb{R}_{+}\times\Omega,\,\mathcal{B}((0,+\infty))\otimes\mathcal{B}(\mathbb{R}_{+})\otimes\mathcal{F})$ into $(\mathbb{R},\,\mathcal{B}(\mathbb{R}))$ is measurable. In particular, the $t$-section of $(r,t,\omega)\rightarrow\beta_{t}^{r}(\omega)$: $(r,\omega)\mapsto\beta_{t}^{r}(\omega)$ is measurable with respect to the $\sigma$-algebra $\mathcal{B}((0,+\infty))\otimes\mathcal{F}$, for all $t\geq0$.
From the definition of $\beta^{r}$ we have that (i) the map $\omega\rightarrow\beta_{t}^{r}(\omega)$ is measurable for all $r>0$ and $t\geq0$ and (ii) the map $(r,t)\rightarrow\beta_{t}^{r}(\omega)$ is continuous for all $\omega\in\Omega$. It now suffices to proceed with the discretization of the parameter $(r,t)$ in order to define piecewise constant and measurable functions converging pointwise to $(r,t,\omega)\mapsto\beta_{t}^{r}(\omega)$ and to use standard results on the passage to the limit of sequences of measurable functions. The assertion of the lemma then follows immediately.
\[cor:measyBetaR\] The map $(r,\omega)\rightarrow\beta_{\cdot}^{r}(\omega)$ of $((0,+\infty)\times\Omega,\,\mathcal{B}((0,+\infty))\otimes\mathcal{F})$ into $(C,\,\mathcal{C})$ is measurable.
Note that $\beta^{r}$ is just as the $r$-section of the map $(r,t,\omega)\rightarrow\beta_{t}^{r}(\omega)$.
If the process $W$ is a Brownian motion on $(\Omega,\mathcal{F},\mathbf{P})$, the process $\beta^{r}$ is just a Brownian bridge which is identically equal to 0 on the time interval $[r,\,+\infty)$. If we denote by $p(t,x,y),\, x\in\mathbb{R}$, the Gaussian density with variance $t$ and mean $y$, then the function $\varphi_{t}(r,\cdot)$ given by $$\varphi_{t}\left(r,x\right):=\begin{cases}
p\left(\frac{t\left(r-t\right)}{r},x,0\right), & 0<t<r, \ x\in\mathbb{R},\\
0, & r\leq t, \ x\in\mathbb{R},
\end{cases} \label{eq:bbdensity}$$ is equal to the density of $\beta_{t}^{r}$ for $0<t<r$. Furthermore we have the following properties:
1. (Expectation) $\mathbf{E}\left[\beta_{t}^{r}\right]=0, \ t\geq0.$
2. (Covariance) $\mathbf{E}\left[\beta_{t}^{r}\beta_{s}^{r}\right]=s\wedge t\wedge r-\frac{(s\wedge r)(t\wedge r)}{r}, \ s,t\geq 0$.
3. (Conditional expectation) For $t\in[0,r)$ and $t<u$ we have: If $g(\beta^r _u)$ is integrable, then $$\begin{aligned}
\lefteqn{\mathbf{E}\left[g\left(\beta_{u}^{r}\right)|\beta_{t}^{r}=x\right]}\nonumber\\
&=&\begin{cases}
{\displaystyle \int_{\mathbb{R}}}g\left(y\right)p\left(\frac{r-u}{r-t}\left(u-t\right),\, y,\,\frac{r-u}{r-t}x\right)dy, & t<u<r, \ x\in\mathbb{R},\\
g\left(0\right), & r\leq u, \ x\in\mathbb{R}.
\end{cases}\label{eq:cond exp ext bb}\end{aligned}$$
4. $\beta^{r}$ is a Markov process with respect to its completed natural filtration.
5. $\beta_{t}^{r}=\beta_{t\wedge r}^{r}, \ t\geq0$.
\[BMr\] The process $B^{r}=(B_{t}^{r},\, t\geq0)$ given by $$B_{t}^{r}:=\beta_{t}^{r}+\int_{0}^{t}\frac{\beta_{s}^{r}}{r-s}\, ds\label{eq:BMfrom_extBB}$$ is a Brownian motion stopped at the (deterministic) stopping time $r$, with respect to the completed natural filtration of $\beta^{r}$.
Now we substitute the fixed time $r$ by a random time $\tau$. Let $\tau$ be a strictly positive random variable on $(\Omega,\mathcal{F},\mathbf{P})$. The object which we are interested in is given by the composition of the two mappings $(r,t,\omega)\mapsto\beta_{t}^{r}(\omega)$ and $(t,\omega)\mapsto(\tau(\omega),t,\omega)$. We have the following definition:
The map $\beta=(\beta_{t},\, t\geq0):\,(\Omega,\,\mathcal{F})\rightarrow(C,\,\mathcal{C})$ is defined by $$\beta_{t}\left(\omega\right):=\beta_{t}^{\tau\left(\omega\right)}\left(\omega\right),\quad (t,\omega)\in\mathbb{R}_+\times\Omega\,.\label{eq:beta def 0}$$
The map $\beta:\,(\Omega,\,\mathcal{F})\rightarrow(C,\,\mathcal{C})$ is measurable.
The map $(r,\omega)\rightarrow\beta^{r}(\omega)$ of $((0,+\infty)\times\Omega,\,\mathcal{B}((0,+ \infty))\otimes\mathcal{F})$ into $(C,\,\mathcal{C})$ is measurable because of Corollary \[cor:measyBetaR\]. By definition, the map $\omega\rightarrow(\tau(\omega),\omega)$ of $(\Omega,\mathcal{F})$ into the measurable space $((0,+\infty)\times\Omega,\,\mathcal{B}((0,+\infty))\otimes\mathcal{F})$ is measurable. The statement of the lemma follows from the measurability of the composition of measurable functions.
We now consider the conditional law with respect to $\tau$. As common, by $\mathbf{P}_\tau$ we denote the distribution of $\tau$ with respect to $\mathbf{P}$.
\[lem:If–is\]If $\tau$ is independent of the Brownian motion $W$, then for any measurable function $G$ on $\left(\left(0,+\infty\right)\times C,\mathcal{B}\left(\left(0,+\infty\right)\right)\otimes\mathcal{C}\right)$ such that $G(\tau,\, W)$ is integrable it follows that $$\mathbf{E}\left[G\left(\tau,\, W\right)|\sigma\left(\tau\right)\right]\left(\omega\right)=\left(\mathbf{E}\left[G\left(r,W\right)\right]\right)_{r=\tau\left(\omega\right)},\;\mathbf{P}\textrm{-a.s.}$$
See, e.g., [@key-8 Ch. II.7, p. 221].
\[cor:LEM if–is\] If $h:\,\left((0,+\infty)\times C,\mathcal{B}\left(\left(0,+\infty\right)\right)\otimes\mathcal{C}\right)\mapsto\left(\mathbb{R},\mathcal{B}(\mathbb{R})\right)$ is a measurable function such that $\mathbf{E}\left[|h\left(\tau,\,\beta\right)|\right]<+\infty$, then $$\mathbf{E}\left[h\left(\tau,\,\beta\right)|\tau=r\right]=\mathbf{E}\left[h\left(r,\,\beta^{r}\right)\right],\; \mathbf{P}_{\tau}\textrm{-a.s.},$$ or, equivalently, $$\mathbf{E}\left[h\left(\tau,\,\beta\right)|\sigma\left(\tau\right)\right]\left(\omega\right)=\left(\mathbf{E}\left[h\left(r,\,\beta^{r}\right)\right]\right)_{r=\tau\left(\omega\right)},\; \mathbf{P}\textrm{-a.s.}$$
The last two formulas provide a useful connection between the law of Brownian bridges and the conditional law with respect to $\sigma(\tau)$ of a generic functional involving $\tau$, the Brownian motion $W$ and the map $\beta$ defined in (\[eq:beta def 0\]).
\[sec:Definition-and-First\] Definition and First Properties
============================================================
Let $W=(W_{t},\, t\geq0)$ be a standard Brownian motion (with respect to its completed natural filtration $\mathbb{F}^{W}=(\mathcal{F}_{t}^{W})_{t\geq0}$) starting from 0. Let $\tau:\,\Omega\rightarrow(0,\infty)$ be a strictly positive random time with distribution function denoted by $F$: $F(t):=\mathbf{P}(\tau\leq t),\: t\geq0$. In this paper the following assumption will always be made.
\[ass:assumption1\]The random time $\tau$ and the Brownian motion $W$ are independent.
The random time $\tau$ (which will be interpreted as *default time*) is supposed to be not an $\mathbb{F}^{W}$-stopping time. This means that the case in which $\tau$ is equal to a positive constant $T$ will not be considered.
\[def:obs by MonPon\]The process $\beta=(\beta_{t},\;t\geq0)$ given by $$\beta_{t}:=W_{t}-\frac{t}{\tau\vee t}W_{\tau\vee t},\; t\geq0,\label{eq:betaDEF1}$$ will be called *information process*.
In what follows, we shall use the following filtrations:
1. $\mathbb{F}^{0}=(\mathcal{F}_{t}^{0})_{t\geq0}$, the natural filtration of the process $\beta$:\
$\mathcal{F}_{t}^{0}:=\sigma(\beta_{s},\,0\leq s\leq t),\ t\geq0$.
2. $\mathbb{F}^{P}=(\mathcal{F}_{t}^{P})_{t\geq0}$, the completed natural filtration:\
$\mathcal{F}_{t}^{P}:=\mathcal{F}_{t}^{0}\vee\mathcal{N}_{P},\, t\geq0$.
3. $\mathbb{F}^{\beta}=(\mathcal{F}_{t}^{\beta})_{t\geq0}$, the smallest filtration containing $\mathbb{F}^{0}$ and satisfying the usual hypotheses of right-continuity and completeness: $$\mathcal{F}_{t}^{\beta}:=\sigma\left(\bigcap_{u>t}\mathcal{F}_{u}^{0}\cup\mathcal{N}_{P}\right)=\mathcal{F}_{t+}^{0}\vee\mathcal{N}_{P},\: t\geq0.$$
The aim of this section is to prove that the default time $\tau$ is a stopping time with respect to the completed natural filtration $\mathbb{F}^P$ and that the information process $\beta$ is Markov with respect to $\mathbb{F}^P$.
Although it is possible to prove directly that $\tau$ is an $\mathbb{F}^{\beta}$-stopping time (see [@key-10 Lemma 2.5]) we point out the following result which better involves the role of the probability measure $\mathbf{P}$. For two sets $A, B\in\mathcal{F}$ we shall write $A\subseteq B \ \; \mathbf{P}$-a.s. if $\mathbf{P}(B\setminus A)=0$. If $A\subseteq B \ \; \mathbf{P}$-a.s. and $B\subseteq A \ \; \mathbf{P}$-a.s., then we write $A=B \ \; \mathbf{P}$-a.s.
\[lem:For-all-,\]For all $t>0$, $\{ \beta_{t}=0\} =\{ \tau\leq t\},\ \mathbf{P}$-a.s. In particular, $\tau$ is a stopping time with respect to the filtration $\mathbb{F}^{P}$.
Using the formula of total probability and Corollary \[cor:LEM if–is\], we have that $$\begin{aligned}
\mathbf{P}\left(\beta_{t}=0,\,\tau>t\right) &=\int_{\left(t,+\infty\right)}\mathbf{P}\left(\beta_{t}=0|\tau=r\right)dF(r)\\
&=\int_{\left(t,+\infty\right)}\mathbf{P}\left(\beta_{t}^{r}=0\right)dF\left(r\right)=0\,,\end{aligned}$$ where the latter equality holds because the random variable $\beta_{t}^{r}$ is nondegenerate and normally distributed for $0<t<r$ and, therefore, we obtain $\mathbf{P}(\beta_{t}^{r}=0)=0,\,0<t<r$. Hence, for all $t>0$, $\{ \beta_{t}=0\} \cap\{ \tau>t\} \in\mathcal{N}_{P}$ and, consequently, $\{ \beta_{t}=0\} \subseteq\{ \tau\leq t\} ,\,\mathbf{P}$-a.s. In view of $\{ \tau\leq t\} \subseteq\{ \beta_{t}=0\}$, this yields the first part of the statement: $\{ \tau\leq t\} =\{\beta_{t}=0\} ,\,\mathbf{P}$-a.s. Since $\{ \beta_{t}=0\} \in\mathcal{F}_{t}^{0}$, the event $\{ \tau\leq t\}$ belongs to $\mathcal{F}_{t}^{0}\vee\mathcal{N}_{P}$, for all $t\geq 0$. Hence $\tau$ is a stopping time with respect to $\mathbb{F}^{P}$.
$\tau$ is a stopping time with respect to the filtration $\mathbb{F}^{\beta}$.
The above proposition states that the process $(\mathbb{I}_{\{ \tau\leq t\} },\, t>0)$ is a modification, under the probability measure $\mathbf{P}$, of the process $(\mathbb{I}_{\{ \beta_{t}=0\} },\, t>0)$. On the other hand, these both processes are not indistinguishable, since the process $\beta$ can hit 0 before $\tau$ (in fact, due to the law of iterated logarithm, this happens uncountably many times $\mathbf{P}$-a.s.).
Roughly speaking we can say that in general, if we can observe only $\beta_{t}$, we are sure that $\tau>t$ whenever $\beta_{t}\neq0$. But, if at time $t$ we observe the event $\{ \beta_{t}=0\}$, the information carried by $\mathcal{F}_{t}^{0}$ may be not sufficient to know whether the event $\{ \tau\le t\}$ has occurred or not, we can only say that it occurred $\mathbf{P}$-a.s.
We now show that the information process $\beta$ is a Markov process with respect to its natural filtration $\mathbb{F}^{0}$.
\[TEO:The-information-process\] The information process $\beta$ is an $\mathbb{F}^{0}$-Markov process: $$\mathbf{E}\left[f\left(\beta_{t+h}\right)|\mathcal{F}_{t}^{0}\right]=\mathbf{E}\left[f\left(\beta_{t+h}\right)|\beta_{t}\right],\quad\mathbf{P}\textrm{-a.s., }t\geq0,\label{eq:statementMK1}$$ for all $h\geq0$ and for every measurable function $f$ which is nonnegative or such that $f(\beta_{t+h})$ is integrable.
For $t=0$ the statement is clear. Let us assume $t>0$. On the set $\{ \tau\leq t\}$ we have $\mathbf{P}$-a.s. $$\mathbf{E}\left[f\left(\beta_{t+h}\right)|\mathcal{F}_{t}^{0}\right]\mathbb{I}_{\left\{ \tau\leq t\right\} } =\mathbf{E}\left[f\left(0\right)\mathbb{I}_{\left\{ \tau\leq t\right\} }|\mathcal{F}_{t}^{0}\right]
=f\left(0\right)\mathbb{I}_{\left\{ \tau\leq t\right\}}
=f\left(0\right)\mathbb{I}_{\left\{ \beta_{t}=0\right\}}\,,$$ which is a measurable function with respect to $\sigma(\beta_{t})$ and hence (\[eq:statementMK1\]) is valid on $\{ \tau\leq t\}$ $\mathbf{P}$-a.s. Now we have to prove $$\mathbf{E}\left[f\left(\beta_{t+h}\right)|\mathcal{F}_{t}^{0}\right]\mathbb{I}_{\left\{ \tau>t\right\} }=\mathbf{E}\left[f\left(\beta_{t+h}\right)|\beta_{t}\right]\mathbb{I}_{\left\{ \tau>t\right\} },\;\mathbf{P}\textrm{-a.s.}$$ Both sides being $\mathcal{F}_{t}^{P}$-measurable, it suffices to verify that for all $A\in\mathcal{F}_{t}^{0}$ we have $$\int_{A\cap\left\{ \tau>t\right\} }f\left(\beta_{t+h}\right)d\mathbf{P}=\int_{A\cap\left\{ \tau>t\right\} }\mathbf{E}\left[f\left(\beta_{t+h}\right)|\beta_{t}\right]d\mathbf{P}.\label{eq:equivalentMK1}$$ We observe that $\mathcal{F}_{t}^{0}$ is generated by $$\beta_{t_{n}},\ \xi_n:=\frac{\beta_{t_{n}}}{t_{n}}-\frac{\beta_{t_{n-1}}}{t_{n-1}},\ \xi_{n-1}:=\frac{\beta_{t_{n-1}}}{t_{n-1}}-\frac{\beta_{t_{n-2}}}{t_{n-2}},\ ..., \ \xi_1:=\frac{\beta_{t_{1}}}{t_{1}}-\frac{\beta_{t_{0}}}{t_{0}}\,,$$ $0<t_{0}<t_{1}<...<t_{n}=t$, for $n$ running through $\mathbb{N}$. By the monotone class theorem (see, e.g., [@key-7 I.19, I.21]) it is sufficient to prove (\[eq:equivalentMK1\]) for sets $A$ of the form $A=\{ \beta_{t}\in B,\,\xi_{1}\in B_{1},\ldots,\,\xi_{n}\in B_{n}\}$ with $B, B_{1}, B_{2},\ldots, B_{n}\in \mathcal{B}(\mathbb{R}),\, n\geq1$. Let $g:=\mathbb{I}_{B}$ and $L:=\mathbb{I}_{B_{1}\times B_{2}\times\cdots\times B_{n}}$. Then we have the equality $\mathbb{I}_{A}\!=\!g(\beta_{t})L(\xi_{1},\ldots,\,\xi_{n})$ and, setting $$\eta_{k}:=\frac{W_{t_{k}}}{t_{k}}-\frac{W_{t_{k-1}}}{t_{k-1}},\quad k=1,\ldots, n,$$ we have $\xi_{k}=\eta_{k}$ on $\{ \tau>t\}$, $k=1, \ldots, n$. But, for $r>t$, the random vector $(\eta_{1},\ldots, \eta_{n},\,\beta_{t}^{r},\beta_{t+h}^{r})$ is Gaussian and, denoting by $\textrm{cov}(X,Y)$ the covariance between two random variables $X$ and $Y$, we have that $\textrm{cov}(\eta_{k},\beta_{t}^{r})=\textrm{cov}(\eta_{k},\beta_{t+h}^{r})=0,\; k=1, \ldots, n$. Thus $(\eta_{1},...,\eta_{n})$ is independent of $(\beta_{t}^{r},\beta_{t+h}^{r})$ and, with the notation $H(x,y):=f(x)g(y)$, we also have that $L(\eta_{1},...,\eta_{n})$ is independent of $H(\beta_{t+h}^{r},\beta_{t}^{r})$. Now we can state the following lemma which will allow to complete the proof of Theorem \[TEO:The-information-process\].
\[lem:The-random-variables\] Let $H:\, \mathbb{R}^2\mapsto\mathbb{R}$ be a measurable function. Suppose that $H$ is nonnegative or such that $\mathbf{E}\left[\left| H\left(\beta_{t+h},\beta_{t}\right)\right|\right]<+\infty$. Then the random variables $H(\beta_{t+h},\beta_{t})\mathbb{I}_{\{ \tau>t\} }$ and $L(\eta_{1},...,\eta_{n})$ are uncorrelated: $$\begin{aligned}
\mathbf{E}\left[H\left(\beta_{t+h},\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\}}L\left(\eta_{1},...,\eta_{n}\right)\right] &=\mathbf{E}\left[H\left(\beta_{t+h},\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }\right]\mathbf{E}\left[L\left(\eta_{1},...,\eta_{n}\right)\right]\end{aligned}$$
Using the formula of total probability and Lemma \[lem:If–is\] we get $$\begin{aligned}
\lefteqn{\mathbf{E}\left[H\left(\beta_{t+h},\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }L\left(\eta_{1},...,\eta_{n}\right)\right]}\\
&=\int_{\left(t,+\infty\right)}\mathbf{E}\left[H\left(\beta_{t+h},\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }L\left(\eta_{1},...,\eta_{n}\right)|\tau=r\right]dF\left(r\right)\\
&=\int_{\left(t,+\infty\right)}\mathbf{E}\left[H\left(\beta_{t+h}^{r},\beta_{t}^{r}\right)L\left(\eta_{1},...,\eta_{n}\right)\right]dF\left(r\right)\\
&=\int_{\left(t,+\infty\right)}\mathbf{E}\left[H\left(\beta_{t+h}^{r},\beta_{t}^{r}\right)\right]dF\left(r\right)\mathbf{E}\left[L\left(\eta_{1},...,\eta_{n}\right)\right]\\
&=\mathbf{E}\left[H\left(\beta_{t+h},\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }\right]\mathbf{E}\left[L\left(\eta_{1},...,\eta_{n}\right)\right].\end{aligned}$$ The proof of the lemma is finished.
We now prove (\[eq:equivalentMK1\]) for our special choice of $A$. From Lemma \[lem:The-random-variables\] above we have $$\begin{aligned}
\int_{A\cap\left\{ \tau>t\right\} }f\left(\beta_{t+h}\right)d\mathbf{P} & =\mathbf{E}\left[H\left(\beta_{t+h},\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }L\left(\xi_{1},...,\xi_{n}\right)\right]\\
& =\mathbf{E}\left[H\left(\beta_{t+h},\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }L\left(\eta_{1},...,\eta_{n}\right)\right]\\
& =\mathbf{E}\left[f\left(\beta_{t+h}\right)g\left(\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }\right]\mathbf{E}\left[L\left(\eta_{1},...,\eta_{n}\right)\right]\\
& =\mathbf{E}\left[\mathbf{E}\left[f\left(\beta_{t+h}\right)|\beta_{t}\right]g\left(\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }\right]\mathbf{E}\left[L\left(\eta_{1},...,\eta_{n}\right)\right]\\
& =\mathbf{E}\left[\mathbf{E}\left[f\left(\beta_{t+h}\right)|\beta_{t}\right]g\left(\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }L\left(\eta_{1},...,\eta_{n}\right)\right]\\
& =\mathbf{E}\left[\mathbf{E}\left[f\left(\beta_{t+h}\right)|\beta_{t}\right]g\left(\beta_{t}\right)\mathbb{I}_{\left\{ \tau>t\right\} }L\left(\xi_{1},...,\xi_{n}\right)\right]\\
& =\int_{A\cap\left\{ \tau>t\right\} }\mathbf{E}\left[f\left(\beta_{t+h}\right)|\beta_{t}\right]d\mathbf{P},\end{aligned}$$ which proves that (\[eq:equivalentMK1\]) is true and this ends the proof.
Note that the Markov property is trivially extended to the completed filtration $\mathbb{F}^{P}$.
\[sec:Conditional-Expectations\] Bayes Estimates of the Default Time $\tau$
===========================================================================
The basic aim of the present section is to provide estimates of the a priori unknown default time $\tau$ based on the observation of the information process $\beta$ up to time $t$. For fixed $t\geq 0$, the observation is represented by the $\sigma$-algebra $\mathcal{F}^P_t$ and, because of the Markov property, the observation of $\beta_t$ would be sufficient. To this end it is natural to exploit the Bayesian approach.
The idea is to use the knowledge gained from the observation of the flow $(\beta_t, \ t\geq 0)$ for updating the initial knowledge on $\tau$. At time 0, the market agents have only *a priori* knowledge about $\tau$, represented by its distribution function $F$. As time is increasing, information concerning the default becomes available. Using the Bayes theorem (recalled in the Appendix for easy reference), the *a posteriori* distribution of $\tau$ based on the observation of $\beta$ up to $t$ can be derived and in this way agents can update their initial knowledge obtaining a sharper estimate of the default time $\tau$.
In this section the $\sigma$-algebra generated by the future of $\beta$ at time $t$ is denoted by $\mathcal{F}^P_{t,\infty}:=\sigma(\beta_{s},\, t\leq s\leq+\infty)\vee\mathcal{N}_{P}$. The following is a standard result on Markov processes:
\[lem:basic MKV\]Let $X=(X_{t},\, t\geq0)$ be a stochastic process adapted to a filtration $\mathbb{F}=(\mathcal{F}_{t})_{t\geq0}$. Then the following are equivalent:
1. $X$ is Markov with respect to $\mathbb{F}$.
2. For each $t\geq0$ and bounded (or nonnegative) $\sigma(X_{s},\, s\geq t)$-measurable random variable $Y$ one has $$\mathbf{E}\left[Y|\mathcal{F}_{t}\right]=\mathbf{E}\left[Y|X_{t}\right], \; \mathbf{P}\textrm{-a.s.}$$
See [@key-5 Ch. I, Theorem (1.3)].
The next proposition describes the structure of the a posteriori distribution of $\tau$ based on the observation of $\mathcal{F}^P_t$.
\[prop:HJ1\] For all $t,u\geq0$, it holds $$\label{prop:HJ1second}
\mathbf{P}\left(\tau\leq u|\mathcal{F}_{t}^{P}\right)=\mathbb{I}_{\left\{ \tau\leq t\wedge u\right\} }+\mathbf{P}\left(t<\tau\leq u|\beta_{t}\right)\mathbb{I}_{\left\{ t<\tau\right\} },\;\mathbf{P}\textrm{-a.s.}$$
Obviously, we have $\{ \tau\leq u\} =\{ \tau\leq t\wedge u\} \cup\{ t<\tau\leq u\}$. The first set of the right-hand side of the above decomposition yields the first summand of the statement. Then it suffices to observe that we have the relation $\{ t<\tau\leq u\} =\{ \beta_{t}\neq0,\beta_{u}=0\}$ $\mathbf{P}\textrm{-a.s.}$ where the set on the right-hand side of the above equality belongs to $\mathcal{F}^P_{t,\infty}$. It remains to apply Lemma \[lem:basic MKV\] in order to complete the proof of the statement.
Recalling that $F$ denotes the distribution function of $\tau$ and formula (\[eq:bbdensity\]) for the definition of the function $\varphi_{t}(r,x)$ (which is equal to the density of the Brownian bridge $\beta_{t}^{r}$ at time $t<r$), we have the following result which provides the explicit form of the a posteriori distribution of $\tau$ based on the observation of $\beta$ up to $t$.
\[lem:PRE-theo-4.3\]Let $t>0$. Then, for all $u>0$, $\mathbf{P}$-a.s. $$\begin{aligned}
\mathbf{P}\left(\tau\leq u|\mathcal{F}_{t}^{P}\right) =\mathbb{I}_{\left\{\tau\leq t\wedge u\right\}}+\frac{{\displaystyle \int_{\left(t,u\right]}}\varphi_{t}\left(r,\beta_{t}\right)dF(r)}{{\displaystyle \int_{\left(t,+\infty\right)}}\varphi_{t}\left(v,\beta_{t}\right)dF(v)}\,\mathbb{I}_{\left\{ t<\tau\right\}}\,.\label{eq:condexptau-1}\end{aligned}$$
The result is a consequence of Proposition \[prop:HJ1\] and the Bayes formula (see Corollary \[cor:(Bayes-formula)-a.s.\]).
Theorem \[lem:PRE-theo-4.3\] can be extended to functions $g$ on $\mathbb{R}_+$ as it will be stated in the following corollary.
\[LEM:prop\_cond\_exp\_tau\]Let $t>0,\; g:\mathbb{R}_{+}\rightarrow\mathbb{R}$ be a Borel function such that $\mathbf{E}\left[|g(\tau)|\right]<+\infty$. Then, $\mathbf{P}$-a.s., $$\begin{aligned}
\mathbf{E}\left[g\left(\tau\right)|\mathcal{F}_{t}^{P}\right] & =g\left(\tau\right)\mathbb{I}_{\left\{ \tau\leq t\right\} }+\frac{{\displaystyle \int_{\left(t,+\infty\right)}}g\left(r\right)\varphi_{t}\left(r,\beta_{t}\right)dF(r)}{{\displaystyle \int_{\left(t,+\infty\right)}}\varphi_{t}\left(v,\beta_{t}\right)dF(v)}\,\mathbb{I}_{\left\{ t<\tau\right\}}\,. \label{eq:condexptau2}\end{aligned}$$
If the function $g$ is bounded then the statement immediately follows by an application of the monotone class theorem to simple functions where it is possible to use Theorem \[lem:PRE-theo-4.3\]. In the general case $g$ has to be approximated pointwise by bounded functions and by passing to the limit.
We point out that the function $\phi_{t}$ defined by $$\phi_{t}\left(r,x\right):=\frac{\varphi_{t}\left(r,x\right)}{{\displaystyle \int_{\left(t,+\infty\right)}\varphi_{t}\left(v,x\right)dF\left(v\right)}}, \; \left(r,t\right)\in\left(0,+\infty\right)\times\mathbb{R}_+,\; x\in\mathbb{R}\,, \label{eq:densitybeta2}$$ is, for $t<r$, the a posteriori density function of $\tau$ on $\{\tau>t\}$ based on the observation $\beta_t=x$ (see Corollary \[cor:The-conditional-density\]). Then relation (\[eq:condexptau-1\]), representing the a posteriori distribution of $\tau$ based on the observation of $\mathcal{F}_t^P$, can be rewritten as $$\mathbf{P}\left(\tau\leq u|\mathcal{F}_{t}^{P}\right)=\mathbb{I}_{\left\{ \tau\leq t\right\}} +{\displaystyle \int_{\left(t,u\right]}}\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\mathbb{I}_{\left\{ t<\tau\right\} },\;\mathbf{P}\textrm{-a.s.},$$ while (\[eq:condexptau2\]) is equal to the expression $$\mathbf{E}\left[g\left(\tau\right)|\mathcal{F}_{t}^{P}\right]=g\left(\tau\right)\mathbb{I}_{\left\{ \tau\leq t\right\} }+{\displaystyle \int_{\left(t,+\infty\right)}g\left(r\right)}\,\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\mathbb{I}_{\left\{ t<\tau\right\} },\;\mathbf{P}\textrm{-a.s.}$$ Here it is possible to see how the Bayesian estimate of $\tau$ given above provides a better knowledge on the default time $\tau$ through the observation of the information process $\beta$ at time $t$.
Extensions of the Bayes Estimates
=================================
In this section we shall deal with an extension of the Bayes estimates of $\tau$ provided in Section \[sec:Conditional-Expectations\]. Roughly speaking we shall derive formulas which include the Bayes estimates discussed in Section \[sec:Conditional-Expectations\] as well as the prognose of the information process $\beta$ at some time $u$, the latter being related with the Markov property which has been proven in Section \[sec:Definition-and-First\] (see Theorem \[TEO:The-information-process\]). First we will state a lemma that will be used in the proof of the main results of this section.
\[lem:AUXlemma\]Let $0\leq t<u$ and $g$ be a measurable function on $(0,+\infty)\times\mathbb{R}$ such that $g(\tau,\beta_{u})$ is integrable. Then it holds $\mathbf{P}$-a.s. $$\begin{aligned}
\mathbf{E}\left[g\left(\tau,\beta_{u}\right)\mathbb{I}_{\left\{ t<\tau\right\} }|\mathcal{F}_{t}^{P}\right] & =\mathbf{E}\left[g\left(\tau,\beta_{u}\right)|\beta_{t}\right]\mathbb{I}_{\left\{ t<\tau\right\} }\\
& =\mathbf{E}\left[\mathbf{E}\left[g\left(\tau,\beta_{u}\right)|\sigma\left(\tau\right)\vee\sigma\left(\beta_{t}\right)\right]|\beta_{t}\right]\mathbb{I}_{\left\{ t<\tau\right\} }\\
& =\mathbf{E}\left[\left(\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)|\beta_{t}^{r}\right]\right)_{r=\tau}|\beta_{t}\right]\mathbb{I}_{\left\{ t<\tau\right\} }.\end{aligned}$$
It is clear that the first equality holds true due to the fact that $g(\tau,\beta_{u})\mathbb{I}_{\{ t<\tau\} }$ is $\mathcal{F}_{t,\infty}$-measurable and, hence, Lemma \[lem:basic MKV\] can be applied. The second equality is obvious. Let $h$ be an arbitrary bounded Borel function. Using Lemma \[lem:If–is\] we have that $$\begin{aligned}
\mathbf{E}\left[g\left(\tau,\beta_{u}\right)h\left(\beta_{t}\right)\mathbb{I}_{\left\{ t<\tau\right\} }\right] & =\int_{\left(t,+\infty\right)}\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)h\left(\beta_{t}^{r}\right)\right]dF\left(r\right)\\
& =\int_{\left(t,+\infty\right)}\mathbf{E}\left[\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)h\left(\beta_{t}^{r}\right)|\beta_{t}^{r}\right]\right]dF\left(r\right)\\
& =\int_{\left(t,+\infty\right)}\mathbf{E}\left[\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)|\beta_{t}^{r}\right]h\left(\beta_{t}^{r}\right)\right]dF\left(r\right)\\
& =\mathbf{E}\left[\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)|\beta_{t}^{r}\right]_{r=\tau}h\left(\beta_{t}\right)\mathbb{I}_{\left\{ t<\tau\right\} }\right]\,,\end{aligned}$$ that is, $$\mathbf{E}\left[\left(g\left(\tau,\beta_{u}\right)-\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)|\beta_{t}^{r}\right]_{r=\tau}\right)h\left(\beta_{t}\right)\mathbb{I}_{\left\{ t<\tau\right\} }\right]=0.$$ But $h$ is arbitrary and thus $$\begin{aligned}
\mathbf{E}\left[g\left(\tau,\beta_{u}\right)|\beta_{t}\right] & =\mathbf{E}\left[\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)|\beta_{t}^{r}\right]_{r=\tau}|\beta_{t}\right],\end{aligned}$$ $\mathbf{P}$-a.s. on $\{ t<\tau\} $, for $t<u$.
The next theorem is prepared by the following result.
\[cor:Let–be-BIS\] Let $t\geq0$ and $g$ be a measurable function such that $g(\tau,\beta_{t})$ is integrable. Then, $\mathbf{P}$-a.s., $$\mathbf{E}\left[g\left(\tau,\beta_{t}\right)|\mathcal{F}_{t}^{\beta}\right]=g\left(\tau,0\right)\mathbb{I}_{\left\{ \tau\leq t\right\} }+\int_{\left(t,+\infty\right)}g\left(r,\beta_{t}\right)\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\mathbb{I}_{\left\{ t<\tau\right\} }\,.$$
The statement is clear on the set $\{ \tau\leq t\} $. On the set $\{ t<\tau\}$, we first prove it for bounded measurable functions $g$ and, by a monotone class argument, it suffices to consider functions $g$ of the form $g(\tau,\beta_{t})=g_{1}(\tau)g_{2}(\beta_{t})$ where $g_1$ and $g_2$ are bounded measurable functions. Then we have that $$\begin{aligned}
\mathbf{E}\left[g\left(\tau,\beta_{t}\right)|\mathcal{F}_{t}^{\beta}\right]\mathbb{I}_{\left\{ t<\tau\right\} } & =\frac{\int_{\left(t,+\infty\right)}g_{1}\left(r\right)\varphi_{t}\left(r,\beta_{t}\right)dF(r)}{\int_{\left(t,+\infty\right)}\varphi_{t}\left(v,\beta_{t}\right)dF(v)}\mathbb{I}_{\left\{ t<\tau\right\} }\,g_{2}\left(\beta_{t}\right)\\
& =\int_{\left(t,+\infty\right)}g\left(r,\beta_{t}\right)\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\mathbb{I}_{\left\{ t<\tau\right\} }.\end{aligned}$$ If $g$ is a nonnegative measurable function, we can apply the above result to the functions $g_N:=g\wedge N$ and using the monotone convergence theorem we obtain the asserted equality for $g$. Finally, in the general case, the result is true for the positive and negative parts $g^+$ and $g^-$ of $g$ and for $g=g^+-g^-$ the equality follows from the linearity of both sides.
We can now state the main result of this section.
\[thm:Let–and\] Let $0<t<u$ and $g$ be a measurable function defined on $(0,+\infty)\times\mathbb{R}$ such that $\mathbf{E}\left[|g(\tau,\,\beta_{u})|\right]<+\infty$. Then, $\mathbf{P}$-a.s. $$\begin{aligned}
\lefteqn{\mathbf{E}\left[g\left(\tau,\,\beta_{u}\right)|\mathcal{F}_{t}^{P}\right]=\mathbf{E}\left[g\left(\tau,\,\beta_{u}\right)|\beta_t\right]}\\
=&\;g\left(\tau,0\right)\mathbb{I}_{\left\{ \tau\leq t\right\} }+\int_{\left(t,u\right]}g\left(r,0\right)\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\,\mathbb{I}_{\left\{ t<\tau\right\}}\\
&+\!\!\int\limits_{\left(u,+\infty\right)}\!\!\int\limits_{\mathbb{R}}g\left(r,y\right)p\big(\frac{r-u}{r-t}\left(u-t\right),\, y,\,\frac{r-u}{r-t}\beta_{t}\big)dy\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\mathbb{I}_{\left\{ t<\tau\right\} }\,, \label{eq:prova1}\end{aligned}$$ where $p(t,\cdot,y)$ is the Gaussian density with mean $y$ and variance $t$.
On the set $\{ \tau\leq t\} $ the statement is a consequence of the fact that $\tau$ is an $\mathbb{F}^{P}$-stopping time and that $\beta_{u}=0$ on $\{\tau\leq t<u\}$. On the set $\{ t<\tau\}$, from Lemma \[lem:AUXlemma\] we have $$\begin{aligned}
\mathbf{E}\left[g\left(\tau,\,\beta_{u}\right)|\mathcal{F}_{t}^{P}\right]\mathbb{I}_{\left\{ t<\tau\right\} } & =\mathbf{E}\left[g\left(\tau,\,\beta_{u}\right)|\beta_{t}\right]\mathbb{I}_{\left\{ t<\tau\right\} }\\
& =\mathbf{E}\left[g\left(\tau,0\right)\mathbb{I}_{\left\{ t<\tau\leq u\right\} }|\beta_{t}\right]+\mathbf{E}\left[g(\tau,\beta_{u})\mathbb{I}_{\left\{ u<\tau\right\} }|\beta_{t}\right]\,,\end{aligned}$$ $\mathbf{P}$-a.s. We remark that due to Corollary \[LEM:prop\_cond\_exp\_tau\] $$\mathbf{E}\left[g\left(\tau,0\right)\mathbb{I}_{\left\{ t<\tau\leq u\right\} }|\beta_{t}\right]=\int_{\left(t,u\right]}g\left(r,0\right)\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\,\mathbb{I}_{\{t<\tau\}}\,.$$ On the other hand, from (\[eq:cond exp ext bb\]), for $t<u<r$, $$\begin{aligned}
\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)|\beta_{t}^{r}=x\right]&=\int_{\mathbb{R}}g\left(r,y\right)p\left(\frac{r-u}{r-t}\left(u-t\right),\,y,\,\frac{r-u}{r-t}x\right)dy\nonumber\\
&=:G_{t,u}\left(r,x\right)\,. \label{eq:prova3}\end{aligned}$$ It follows from Lemma \[lem:AUXlemma\] and from (\[eq:prova3\]) that, on $\{ t<\tau\}$ $\mathbf{P}$-a.s., $$\begin{aligned}
\mathbf{E}\left[g(\tau,\beta_{u})\mathbb{I}_{\left\{ u<\tau\right\} }|\beta_{t}\right] & =\mathbf{E}\left[\left(\mathbf{E}\left[g\left(r,\beta_{u}^{r}\right)\mathbb{I}_{\left\{ u<r\right\} }|\beta_{t}^{r}\right]\right)_{r=\tau}|\beta_{t}\right]\\
& =\mathbf{E}\left[\left(G_{t,u}\left(r,\beta_{t}^{r}\right)\right)_{r=\tau}|\beta_{t}\right]\\
& =\mathbf{E}\left[G_{t,u}\left(\tau,\beta_{t}\right)|\beta_{t}\right]\\
& =\int_{\left(u,+\infty\right)}G_{t,u}\left(r,\beta_{t}\right)\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right),\end{aligned}$$ where the latter equality follows from Proposition \[cor:Let–be-BIS\].
As an immediate consequence of Theorem \[thm:Let–and\], for $t<u$, we can calculate the conditional expectation of $\beta_{u}$ given $\beta_{t}$ by $$\mathbf{E}\left[\beta_{u}|\mathcal{F}_{t}^{P}\right]=\beta_{t}\int_{\left(u,+\infty\right)}\frac{r-u}{r-t}\,\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\mathbb{I}_{\left\{ t<\tau\right\} },\; \mathbf{P}\textrm{-a.s.},$$ and the conditional distribution of $\beta_u$ given $\beta_t$: For $\Gamma\in\mathcal{B}(\mathbb{R})$, $\mathbf{P}$-a.s., $$\begin{gathered}
\mathbf{P}\left(\beta_{u}\in\Gamma|\mathcal{F}_{t}^{P}\right)=\mathbb{I}_{\left\{ 0\in\Gamma\right\}}\,\mathbb{I}_{\left\{ \tau\leq t\right\} } +\mathbb{I}_{\left\{ 0\in\Gamma\right\} } \int_{\left(t,u\right]}\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\,\mathbb{I}_{\left\{ \tau\leq t\right\} }\\
+\int_{\left(u,+\infty\right)}\int_{\Gamma}p\big(\frac{r-u}{r-t}\left(u-t\right),\, y,\,\frac{r-u}{r-t}\beta_{t}\big)dy\,\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\mathbb{I}_{\left\{ t<\tau\right\} }.\label{eq:prova2}\end{gathered}$$
\[rem:Strong\_Inhom\_MK\] From the factor $\left(r-u\right)/\left(r-t\right)$ in (\[eq:prova2\]) we see that the process $\beta$ cannot be a homogeneous Markov process because $\mathbf{P}(\beta_{u}\in\Gamma|\mathcal{F}_{t}^{P})$ does not depend only on $u-t$ and $(\beta_{t},\Gamma)$.
Markov Property\[sec:The-Markov-property2\]
===========================================
In this section we are going to strengthen Theorem \[TEO:The-information-process\] on the Markov property of the information process $\beta$. We shall prove that $\beta$ is not only a Markov process with respect to the filtration $\mathbb{F}^0$ (or $\mathbb{F}^P$) but also with respect to $\mathbb{F}^\beta$, the smallest filtration containing $\mathbb{F}^0$ and satisfying the usual conditions. As an important consequence, it turns out that the filtrations $\mathbb{F}^P$ and $\mathbb{F}^\beta$ are equal which amounts to saying that the filtration $\mathbb{F}^P$ is right-continuous. The result is stated in the following theorem.
\[thm:The-process–1\]The process $\beta$ is a Markov process with respect to the filtration $\mathbb{F}^{\beta}$, i.e., $$\mathbf{E}\left[g\left(\beta_{u}\right)|\mathcal{F}_{t}^{\beta}\right]=\mathbf{E}\left[g\left(\beta_{u}\right)|\beta_{t}\right],\;\mathbf{P}\textrm{-a.s.}$$ for all $0\leq t<u$ and all measurable functions $g$ such that $g(\beta_{u})$ is integrable.
The proof is divided into two main parts. In the first one we prove the statement of the above theorem for $t>0$, while in the second part we consider the case $t=0$. Throughout the proof we can assume without loss of generality that the function $g$ is continuous and bounded by some constant $M\in\mathbb{R}_+$. For the first part of the proof, let $t>0$ be a strictly positive real number and let $(t_{n})_{n\in\mathbb{N}}$ be a decreasing sequence converging to $t$ from above: $0<t<...<t_{n+1}<t_{n}<u,\: t_{n}\downarrow t$ as $n\rightarrow\infty$. From the definition of $\mathcal{F}_{t}^{\beta}$ we have that $\mathcal{F}_{t}^{\beta}=\bigcap_{n\in\mathbb{N}}\mathcal{F}_{t_{n}}^{P}$ where we recall that $\mathcal{F}_{v}^{P}=\sigma(\beta_{s},\,0\leq s\leq v)\vee\mathcal{N}_{P}$, $v\geq 0$. Consequently, $$\mathbf{E}\left[g\left(\beta_{u}\right)|\mathcal{F}_{t}^{\beta}\right]=\lim_{n\rightarrow\infty}\mathbf{E}\left[g\left(\beta_{u}\right)|\mathcal{F}_{t_{n}}^{P}\right]\,, \; \mathbf{P}\textrm{-a.s.}$$ From Theorem \[thm:Let–and\] and the definition of $G_{t_n,u}$ by we know that, as $t_{n}<u$, $\mathbf{P}$-a.s. $$\begin{aligned}
\nonumber\lefteqn{\mathbf{E}\left[g\left(\beta_{u}\right)|\mathcal{F}_{t_{n}}^{P}\right]=\mathbf{E}\left[g\left(\beta_{u}\right)|\beta_{t_{n}}\right]}\\
&=&g\left(0\right)\mathbb{I}_{\left\{ \tau\leq t_{n}\right\} }+g\left(0\right)\int_{\left(t_{n},u\right]}\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right)\,\mathbb{I}_{\left\{ t_{n}<\tau\right\} }\label{eq:equiv1}\\
\nonumber&&+\int_{\left(u,+\infty\right)}G_{t_n,u}\left(r,\beta_{t_n}\right)\,\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right)\,\mathbb{I}_{\left\{ t_{n}<\tau\right\} }\,.\end{aligned}$$ We want to prove that $$\lim_{n\rightarrow\infty}\mathbf{E}\left[g\left(\beta_{u}\right)|\mathcal{F}_{t_{n}}^{P}\right]=\mathbf{E}\left[g\left(\beta_{u}\right)|\beta_{t}\right],\;\mathbf{P}\textrm{-a.s.}$$ Using (\[eq:equiv1\]) and Theorem \[thm:Let–and\] we see that this latter relation holds true if the following two identities are satisfied, $\mathbf{P}$-a.s. on $\{t<\tau\}$: $$\begin{aligned}
\lim_{n\rightarrow\infty}\int_{\left(t_{n},u\right]}\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right)&= \int_{\left(t,u\right]}\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\,,\label{eq:star2}\\
\lim_{n\rightarrow\infty}\int\limits_{\left(u,+\infty\right)}\!\!\!G_{t_n,u}\left(r,\beta_{t_{n}}\right)\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right)&=\!\!\! \int\limits_{\left(u,+\infty\right)}\!\!\!G_{t,u}\left(r,\beta_{t}\right)\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\,.\label{eq:star3}\end{aligned}$$ Relation (\[eq:star2\]) can be derived as follows.
*Proof of (\[eq:star2\]).* Recalling that by (\[eq:densitybeta2\]) $$\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)=\frac{\varphi_{t_{n}}\left(r,\beta_{t_{n}}\right)}{{\displaystyle \int_{\left(t_{n},+\infty\right)}}\varphi_{t_{n}}\left(v,\beta_{t_{n}}\right)dF(v)},\quad t_n<r\,,$$ the integral on the left-hand side of (\[eq:star2\]) can be rewritten as $$\int_{\left(t_{n},u\right]}\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right)=\frac{{\displaystyle \int_{\left(t_{n},u\right]}}\varphi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF(r)}{{\displaystyle \int_{\left(t_{n},+\infty\right)}}\varphi_{t_{n}}\left(v,\beta_{t_{n}}\right)dF(v)}\,.$$ The plan is to apply Lebesgue’s bounded convergence theorem to the numerator and the denominator of the above expression. To this end we have to prove $\mathbf{P}$-a.s. pointwise convergence and uniform boundedness of the integrand $\varphi_{t_n}(\cdot, \beta_{t_n})$ $\mathbf{P}$-a.s. We begin by focusing our attention on the function $(t,r,x)\mapsto\varphi_{t}(r,x)$ defined by , which is continuous on $(0,+\infty)\times[0,+\infty)\times\mathbb{R}\backslash\{ 0\}$. Setting $\varphi_{t}(+\infty,x):=p(t,x,0)$ for every $t>0$ and $x\in\mathbb{R},$ we see that the resulting function $(t,r,x)\mapsto\varphi_{t}(r,x)$, now defined on $(0,+\infty)\times[0,+\infty]\times\mathbb{R}\backslash\{ 0\}$, is continuous, too. Hence $\lim_{n\rightarrow \infty}\varphi_{t_{n}}(r,\beta_{t_{n}})=\varphi_{t}(r,\beta_{t})$, $\mathbf{P}$-a.s. on $\tau>t$, providing pointwise convergence. For this we note that $\beta_{t}\neq0$ $\mathbf{P}$-a.s. on the set $\{t<\tau\}$. Now we fix $\omega\in\Omega$ such that $t<\tau(\omega)$ and $\beta_{t}(\omega)\neq0$. Then the set $\{ t_{n}:\, n\in\mathbb{N}\} \times(t,+\infty]\times\{ \beta_{t_{n}}(\omega):\, n\in\mathbb{N}\} $ is obviously contained in a compact subset of $(0,+\infty)\times[0,+\infty]\times\mathbb{R}\backslash\{0\}$ (depending on $\omega$). This implies that $\varphi_{t_{n}}(r,\beta_{t_{n}}(\omega))$ is bounded (by a constant depending on $\omega$). Using Lebesgue’s bounded convergence theorem we can conclude $$\lim_{n\rightarrow\infty}\int_{\left(t_{n},u\right]}\varphi_{t_{n}}\left(r,\beta_{t_{n}}\left(\omega\right)\right)dF\left(r\right)=\int_{\left(t,u\right]}\varphi_{t}\left(r,\beta_{t}\left(\omega\right)\right)dF\left(r\right).$$ Consequently, we have proven that $$\lim_{n\rightarrow\infty}\int_{\left(t_{n},u\right]}\varphi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right)=\int_{\left(t,u\right]}\varphi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\,\textrm{on }\left\{ t<\tau\right\} \;\mathbf{P}\textrm{-a.s.}$$ This is also valid for intervals $(t_{n},+\infty)$and $(t,+\infty)$ (instead of $(t_{n},u]$ and $(t,u]$). Relation (\[eq:star2\]) follows immediately.
Let us conclude the first part of the proof by showing that equality (\[eq:star3\]) is indeed true.
*Proof of (\[eq:star3\]). * Recall that $t<t_{n}<u$, $n\in\mathbb{N}$. We start by noting that the function $$y\mapsto p\left(\frac{r-u}{r-t_{n}}\left(u-t_{n}\right),\, y,\,\frac{r-u}{r-t_{n}}\beta_{t_{n}}\right)$$ is a density on $\mathbb{R}$ for all $n$. Denoting by $N(\mu,\sigma^{2})$ the normal distribution with expectation $\mu$ and variance $\sigma^{2}$, it follows that the probability measures $N(\frac{r-u}{r-t_{n}}\beta_{t_{n}}(\omega),\frac{r-u}{r-t_{n}}(u-t_{n}))$ converge weakly to $N(\frac{r-u}{r-t}\beta_{t}(\omega),\frac{r-u}{r-t}(u-t))$. Since the function $g$ is bounded by $M$, we have that $$\begin{aligned}
|G_{t_n,u}(r,\beta_{t_n}(\omega))|&=\left|\int_{\mathbb{R}}g\left(y\right)p\left(\frac{r-u}{r-t_{n}}\left(u-t_{n}\right),\, y,\,\frac{r-u}{r-t_{n}}\beta_{t_{n}\left(\omega\right)}\right)dy\right|\nonumber\\
&\leq M<+\infty\,.\label{eq:HJ1}\end{aligned}$$ Furthermore, $$\begin{aligned}
\lefteqn{\lim_{n\rightarrow\infty} G_{t_n,u}\left(r,\beta_{t_n}(\omega)\right)\nonumber}\\
&=\lim_{n\rightarrow\infty}\int_{\mathbb{R}}g\left(y\right)p\left(\frac{r-u}{r-t_{n}}\left(u-t_{n}\right),\, y,\,\frac{r-u}{r-t_{n}}\beta_{t_{n}}\left(\omega\right)\right)dy\nonumber\\
&=\int_{\mathbb{R}}g\left(y\right)p\left(\frac{r-u}{r-t}\left(u-t\right),\,y,\,\frac{r-u}{r-t}\beta_{t}\left(\omega\right)\right)dy
=G_{t,u}\left(r,\beta_{t}(\omega)\right)\label{eq:HJ2}\end{aligned}$$ immediately follows from the assumption that $g$ is bounded and continuous combined with the weak convergence of the Gaussian measures stated above. Now (\[eq:star3\]) can be derived using Lebesgue’s bounded convergence theorem and the properties of $G_{t_n,u}(r,\beta_{t_n}(\omega))$ and $\varphi_{t_{n}}(r,\beta_{t_{n}}(\omega))$ (and hence $\phi_{t_{n}}(r,\beta_{t_{n}}(\omega))$) of $\mathbf{P}$-a.s. boundedness and pointwise convergence verified above.
The first part of the proof of the theorem is now finished. In the second part of the proof we consider the case $t=0$ which is divided into two steps. In the first step we assume that there exists $\varepsilon>0$ such that $\mathbf{P}(\tau>\varepsilon)=1$ and in the second step we will drop this condition.
Let us assume that there exists $\varepsilon>0$ such that $\mathbf{P}(\tau>\varepsilon)=1$. Let $(t_{n})_{n\in\mathbb{N}}$ be a decreasing sequence of strictly positive real numbers converging to $0$: $0<...<t_{n+1}<t_{n},\: t_{n}\downarrow0$ as $n\rightarrow\infty$. Without loss of generality we assume $t_{n}<\varepsilon$ for all $n\in\mathbb{N}$. Then (\[eq:densitybeta2\]) can be rewritten as follows: $$\label{Phi-rewritten}
\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)=\frac{\frac{1}{\sqrt{2\pi t_{n}}}\sqrt{\frac{r}{r-t_{n}}}\exp\left[-\frac{\beta_{t_{n}}^{2}r}{2t_{n}\left(r-t_{n}\right)}\right]}{{\displaystyle \int_{\left(\varepsilon,+\infty\right)}}\frac{1}{\sqrt{2\pi t_{n}}}\sqrt{\frac{s}{s-t_{n}}}\exp\left[-\frac{\beta_{t_{n}}^{2}s}{2t_{n}\left(s-t_{n}\right)}\right]dF\left(s\right)}\,\mathbb{I}_{\left(t_n,+\infty\right)}\left(r\right)\,.$$ We have the following auxiliary result:
\[lem:On-the-set-2\] Suppose that $\mathbf{P}(\tau>\varepsilon)=1$. Then the function $r\mapsto\phi_{t_{n}}(r,\beta_{t_{n}})$ is $\mathbf{P}_\tau$-a.s. uniformly bounded by some constant $K=K(\varepsilon,\omega)<+\infty$ and, for all $r>0$, $$\lim_{n\rightarrow\infty}\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)=1\,,\;\mathbf{P}\textrm{-a.s.}$$
See Appendix \[sec:Proofs-of-Section\].
Now we turn to the proof of the Markov property of $\beta$ with respect to $\mathbb{F}^\beta$ at $t=0$ under the additional assumption that $\mathbf{P}(\tau>\varepsilon)=1$. Since $\mathbf{E}\left[g(\beta_u)|\mathcal{F}^\beta_0\right]= \lim_{n\rightarrow\infty} \mathbf{E}\left[g(\beta_u)|\mathcal{F}^P_{t_n}\right]$, as in the first part, it is sufficient to verify that $$\lim_{n\rightarrow\infty}\mathbf{E}\left[g\left(\beta_{u}\right)|\mathcal{F}^P_{t_{n}}\right]=\mathbf{E}\left[g\left(\beta_{u}\right)|\beta_{0}\right],\;\mathbf{P}\textrm{-a.s.}\label{eq:equiv 2}$$ Using the formula of total probability, Corollary \[cor:LEM if–is\] and , we can calculate $$\begin{aligned}
\mathbf{E}\left[g\left(\beta_{u}\right)|\beta_{0}\right]&=\mathbf{E}\left[g\left(\beta_{u}\right)\right]\\
&=g\left(0\right)\,F\left(u\right)+\int_{\left(u,+\infty\right)}\int_{\mathbb{R}}g\left(y\right)p\left(\frac{r-u}{r}u,y,0\right)dy\,dF\left(r\right).\end{aligned}$$ Consequently, recalling (\[eq:equiv1\]) for computing the left-hand side of (\[eq:equiv 2\]) and the definition of $G_{t_n,u}$ and $G_{t,u}$ by and noting the obvious relation $\lim_{n\rightarrow\infty}\mathbb{I}_{\{ \tau\leq t_{n}\} }=0$, it is sufficient to prove the following two equalities, $\mathbf{P}$-a.s.: $$\begin{aligned}
\lim_{n\rightarrow\infty}\int_{\left(t_{n},u\right]}\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right) &=F\left(u\right)\,,\label{eq:star 2 bis}\\
\lim_{n\rightarrow\infty}\int\limits_{\left(u,+\infty\right)}G_{t_n,u}\left(r,\beta_{t_n}\right)\,\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)dF\left(r\right)&=\int\limits_{\left(u,+\infty\right)}G_{t,u}\left(r,\beta_t\right)\,dF\left(r\right)\,.\label{eq:star 3 bis}\end{aligned}$$ The first equality relies on Lemma \[lem:On-the-set-2\] and Lebesgue’s bounded convergence theorem. For verifying the second equality we use Lebesgue’s bounded convergence theorem combined with Lemma \[lem:On-the-set-2\] and the boundedness and pointwise convergence of the sequence $G_{t_n,u}(\cdot,\beta_{t_n})$ (see (\[eq:HJ1\]) and (\[eq:HJ2\])).
We now turn to the general case where $\tau>0$. Let $\varepsilon>0$ be arbitrary, but fixed. In what follows the process $^{\varepsilon}\!\beta=(^{\varepsilon}\!\beta_{t},\, t\geq0)$ will denote the process defined by $^{\varepsilon}\!\beta_{t}:=
(\beta_{t}^{r})_{r=\tau\vee\varepsilon}$. The natural filtration of the process $^{\varepsilon}\!\beta$ will be denoted by $\mathbb{F}^\varepsilon=(\mathcal{F}^\varepsilon_{t})_{t\geq0}$, where $\mathcal{F}_{t}^\varepsilon :=\sigma(^{\varepsilon}\!\beta_{s},\,0\leq s\leq t)$. Obviously, for proving the Markov property of $\beta$ with respect to $\mathbb{F}^\beta$ at $t=0$ it is sufficient to show that $\mathcal{F}^\beta_0:=\mathcal{F}^0_{0+}\vee\mathcal{N}_P$ is $\mathbf{P}$-trivial. In order to show that $\mathcal{F}_{0+}^{0}\vee\mathcal{N}_{P}$ is indeed the trivial $\sigma$-algebra, we consider a set $A\in\mathcal{F}_{0+}^{0}$ and we show that if $\mathbf{P}(A)>0$, then $\mathbf{P}(A)=1$.
If $\mathbf{P}(A)>0$, then there exists an $\varepsilon>0$ such that $\mathbf{P}(A\cap\{ \tau>\varepsilon\})>0$. Since $A\in\mathcal{F}_{0+}^{0}$, it follows that $A\in\mathcal{F}_{u}^{0}$ for all $0<u\leq\varepsilon$ and, consequently, $A\cap\{ \tau>\varepsilon\} \in\mathcal{F}_{u}^{0}|_{\{ \tau>\varepsilon\} }\vee\mathcal{N}_{P}$ where $\mathcal{F}_{u}^{0}|_{\{ \tau>\varepsilon\}}$ denotes the trace $\sigma$-field of $\mathcal{F}_{u}^{0}$ on the set $\tau>\varepsilon$. Moreover, on the set $\{\tau>\varepsilon\} $, $\beta_{t}={}^{\varepsilon}\!\beta_{t}$ for all $t\geq0$, i.e., $\beta$ and $^{\varepsilon}\beta$ generate the same trace filtration on $\{\tau>\varepsilon\}$ and, consequently, $A\cap\{ \tau>\varepsilon\} \in\mathcal{F}_{u}^\varepsilon|_{\{ \tau>\varepsilon\} }\vee\mathcal{N}_{P}$. Hence, there exists a set $A_{u}\in\mathcal{F}_{u}^\varepsilon$ such that $$A\cap\left\{ \tau>\varepsilon\right\} =A_{u}\cap\left\{ \tau>\varepsilon\right\} \label{eq:star_epsilon}$$ $\mathbf{P}$-a.s., for all $0<u\leq\varepsilon$. Replacing $u$ by $1/n$, for $n\in\mathbb{N}$ sufficiently large, and defining $A_{0}:=\limsup_{n\rightarrow\infty}A_{1/n}$, we obtain that $A_{0}\in\mathcal{F}_{0+}^\varepsilon$. We know from the first step of the second part of the proof that the $\sigma$-field $\mathcal{F}_{0+}^\varepsilon$ is $\mathbf{P}$-trivial and, consequently, $\mathbf{P}(A_{0})\in\{ 0,1\}$. However, from equality (\[eq:star\_epsilon\]) we have that, $\mathbf{P}$-a.s., $A\cap\{ \tau>\varepsilon\} =A_{0}\cap\{\tau>\varepsilon\}$. By hypothesis we have $\mathbf{P}(A\cap\{ \tau>\varepsilon\})>0$ and thus $\mathbf{P}(A_{0} \cap\{\tau>\varepsilon\})=\mathbf{P}(A\cap\{ \tau>\varepsilon\})>0$, which implies that $\mathbf{P}(A_{0})=1$ and, consequently, we get that $\mathbf{P}(A\cap\{ \tau>\varepsilon\})=\mathbf{P}(\{ \tau>\varepsilon\})$. Since $\varepsilon$ is arbitrary we can take the limit for $\varepsilon\downarrow0$, and we obtain that $\mathbf{P}(A)=\mathbf{P}(\Omega)=1$, which ends the proof.
\[rem:equalitySTANDARDwithCOMPLETED\] The filtration $\mathbb{F}^{P}$ satisfies the usual conditions of right-continuity and completeness.
See, e.g., [@key-5 (Ch. I, (8.12))].
As a consequence of Corollary \[rem:equalitySTANDARDwithCOMPLETED\], the filtrations $\mathbb{F}^{\beta}$ and $\mathbb{F}^{P}$ coincide and, in particular, the $\sigma$-algebra $\mathcal{F}^{\beta}_0$ is $\mathbf{P}$-trivial.
It is worth to mention that in fact the statement of Corollary \[rem:equalitySTANDARDwithCOMPLETED\] combined with Theorem \[TEO:The-information-process\] is equivalent to the statement of Theorem \[thm:The-process–1\].
\[sec:Semimartingale-decomposition\] Semimartingale Decomposition of the Information Process
============================================================================================
This section deals with the semimartingale decomposition of $\beta$ with respect to $\mathbb{F}^{\beta}$. To begin with, we recall the notion of the optional projection of a general measurable process $X$.
Let $X$ be a nonnegative measurable process and $\mathbb{F}$ a filtration satisfying the usual conditions. There exists a unique (up to indistinguishability) $\mathbb{F}$-optional process $^{o}\!X$ such that $$\mathbf{E}\left[X_{T}\,\mathbb{I}_{\left\{ T<+\infty\right\} }|\mathcal{F}_{T}\right]=\,^{o}\!X_{T}\,\mathbb{I}_{\left\{ T<+\infty\right\} },\;\mathbf{P}\textrm{-a.s.}$$ for every $\mathbb{F}$-stopping time $T$.
See, for example, [@key-3 Ch. IV, (5.6)].
\(i) The process $^{o}\!X$ is called the *optional projection* of the nonnegative measurable process $X$ with respect to $\mathbb{F}$.
\(ii) Let $X$ be an arbitrary measurable process. Then we define the *optional projection $^{o}\!X$ of $X$ with respect to $\mathbb{F}$* as $$^{o}\!X_{t}\left(\omega\right):=\begin{cases}
^{o}\!X_{t}^{+}\left(\omega\right)-\,^{o}\!X_{t}^{-}\left(\omega\right), & \textrm{if }^o\!X_{t}^{+}\left(\omega\right)\wedge ^o\!\!X_{t}^{-}\left(\omega\right)<+\infty,\\
+\infty, & \textrm{otherwise},
\end{cases}$$ where $^{o}\!X^{+}$ (resp. $^{o}\!X^{-}$) is the optional projection of the positive part $X^+$ (resp. the negative part $X^{-}$) of $X$ with respect to $\mathbb{F}$.
\[rem:If–for rem 63\] Let $\xi$ be an arbitrary random variable and $\mathcal{G}$ a sub-$\sigma$-field of $\mathcal{F}$. Then the conditional expectations $\mathbf{E}\left[\xi^+|\mathcal{G}\right]$ and $\mathbf{E}\left[\xi^-|\mathcal{G}\right]$ always exist and in analogy to the above definition we agree to define the conditional expectation $\mathbf{E}\left[\xi|\mathcal{G}\right]$ by $$\begin{aligned}
\mathbf{E}\left[\xi|\mathcal{G}\right]
&:=\begin{cases}
\mathbf{E}\left[\xi^+|\mathcal{G}\right]-\mathbf{E}\left[\xi^-|\mathcal{G}\right], & \textrm{on } \{\mathbf{E}\left[\xi^+|\mathcal{G}\right]\wedge \mathbf{E}\left[\xi^-|\mathcal{G}\right]<+\infty\},\\
+\infty, & \textrm{otherwise}\,.
\end{cases}\end{aligned}$$ Now let $X$ be an arbitrary measurable process and $\mathbb{F}$ a filtration satisfying the usual conditions. We emphasize that then for every $\mathbb{F}$-stopping time $T$ we have $$\mathbf{E}\left[X_{T}\,\mathbb{I}_{\left\{ T<+\infty\right\} }|\mathcal{F}_{T}\right]=\,^{o}\!X_{T}\,\mathbb{I}_{\left\{ T<+\infty\right\} },\;\mathbf{P}\textrm{-a.s.}$$ In particular, $\mathbf{E}\left[X_{t}|\mathcal{F}_{t}\right]=\,^{o}\!X_{t}$, $\mathbf{P}\textrm{-a.s.}$, for all $t\geq 0$.
Next we are going to state a slight extension of a well-known result from filtering theory which will be used in the sequel. The reader may find useful references, e.g., in [@key-4 Proposition 5.10.3.1], or in [@key-1000 Ch. VI, (8.4)].
First we will introduce the following definition.
Let $B$ be a continuous process, $\mathbb{F}$ a filtration and $T$ an $\mathbb{F}$-stopping time. Then $B$ is called an $\mathbb{F}$-Brownian motion stopped at $T$ if $B$ is an $\mathbb{F}$-martingale with square variation process $\langle B,B\rangle$: $\langle B,B\rangle_t=t\wedge T$, $t\geq 0$.
\[pro:innovationLEMMA\] Let $\mathbb{F}=(\mathcal{F}_{t})_{t\geq0}$ be a filtration satisfying the usual conditions, $T$ an $\mathbb{F}$-stopping time and $B$ an $\mathbb{F}$-Brownian motion stopped at $T$. Let $Z=(Z_{t},\, t\geq0)$ be an $\mathbb{F}$-optional process such that $$\label{ass:innovation-lemma}
\mathbf{E}\left[\int_{0}^{t}\left|Z_{s}\right|\, ds\right]<+\infty,\quad t\geq0.$$ Let the process $X=(X_{t},\, t\geq0)$ be given by $$X_{t}:=\int_{0}^{t}Z_{s}\, ds+B_{t},\quad t\geq 0\,.$$ Denote by $^{o}\!Z$ the optional projection of $Z$ with respect to $\mathbb{F}^{X}=(\mathcal{F}_{t}^{X})_{t\geq0}$. Then the process $b$, $$b_{t}:=X_{t}-\int_{0}^{t}{}^{o}\!Z_{s}\, ds,\quad t\geq0,$$ is an $\mathbb{F}^{X}$-Brownian motion stopped at $T$.
For the sake of easy reference a proof of this result is provided in Appendix \[sec:Proof of Innovation lemma\].
In the remainder of this section we shall make use of the filtration $\mathbb{G}=(\mathcal{G}_{t})_{t\geq0}$ defined as $$\mathcal{G}_{t}:=\bigcap_{u>t}\mathcal{F}_{u}^{\beta}\vee\sigma\left(\tau\right),\quad t\geq 0\,,\label{eq:FF D}$$ which is equal to the initial enlargement of the filtration $\mathbb{F}^{\beta}$ by the $\sigma$-algebra $\sigma(\tau)$. In the sequel, the process $Z=(Z_{t},\, t\geq0)$ is defined by $$Z_{t}:=\frac{\beta_{t}}{\tau-t}\,\mathbb{I}_{\{t<\tau\}}\,,\quad t\geq 0\,.
\label{eq:processo Z}$$ The following auxiliary results will be used to prove the semimartingale property of the process $\beta$.
\[lem:sqrt-tau-1\] We have $$\mathbf{E}\left[\int_{0}^{t}\left|Z_{s}\right|\, ds\right]<+\infty \mbox{ for all } t\geq0\,.$$
Using the formula of total probability, Corollary \[cor:LEM if–is\] and , we can calculate $$\begin{aligned}
\mathbf{E}\left[\int_{0}^{t}\left|Z_{s}\right|\, ds\right] & =
\int_0^{+\infty}\int_0^{t\wedge r}E\left|\beta^r_s\right|/\left(r-s\right)\,ds\,dF\left(r\right)\\
&=(2/\pi)^{1/2}\int_{0}^{+\infty}r^{-1/2}\int_{0}^{t\wedge r}s^{1/2}\left(r-s\right)^{-1/2}\,ds\,dF(r).\end{aligned}$$ The outer integral on the right-hand side of the above expression can be split into two integrals, the first one over $(0,t]$ and the second one over $(t,+\infty)$. For the first integral we see that $$\begin{aligned}
\lefteqn{\int\limits_{(0,t]}r^{-1/2}\int_{0}^{t\wedge r}s^{1/2}\left(r-s\right)^{-1/2}\,ds\,dF(r)}\\
& \leq\int\limits_{(0,t]}\int_{0}^{t\wedge r}\left(r-s\right)^{-1/2}\,ds\,dF(r)=\int\limits_{(0,t]}2\,r^{1/2}\, dF\left(r\right)\leq2\,t^{1/2}<+\infty\,.\end{aligned}$$ The second integral can be estimated as follows: $$\begin{aligned}
\lefteqn{\int_{\left(t,+\infty\right)}r^{-1/2}\int_{0}^{t\wedge r}s^{1/2}\left(r-s\right)^{-1/2}\,ds\,dF(r)}\\
&\leq& t^{1/2}\int_{\left(t,+\infty\right)}r^{-1/2}\int_{0}^{t}\left(r-s\right)^{-1/2}\, ds\,dF(r)\\
&\leq& t^{1/2}\int_{\left(t,+\infty\right)}r^{-1/2}\,2\,r^{1/2}\,dF\left(r\right)\leq2\,t^{1/2}<+\infty\,.\end{aligned}$$ The statement of the lemma is now proved.
\[cor:The-process-Z\] $Z$ defined by (\[eq:processo Z\]) is integrable with respect to the Lebesgue measure $\mathbf{P}$-a.s.: $$\mathbf{P}\left(\int_{0}^{t}\left|Z_{s}\right|\, ds<+\infty\right)=1, \textrm{ for all } t\geq0\,.$$
\[Z-modification\] In view of Corollary \[cor:The-process-Z\], there can be found a process $\overline{Z}$ which is indistinguishable from $Z$ such that $\int_0^t\left|\overline{Z}_s\right|\, ds<+\infty$ for all $t\geq 0$ *everywhere* (and not only $\mathbf{P}$-a.s.). Without loss of generality we can assume that $Z$ has this property. If this would not be the case we could modify the paths of $Z$ on a negligible set. By this modification, the optional projection $^o\!Z$ with respect to any filtration $\mathbb{F}$ will stay in the same class of indistinguishable processes. We can also modify $\beta$ putting $\beta=0$ on the negligible set where the above integrals are not finite for all $t$. In this way the desired property for $Z$ would be fulfilled automatically.
\[rem:We-make-the QWERTY\] (i) The process $Z$ is optional with respect to the filtration $\mathbb{G}$ because $Z$ is right-continuous and $\mathbb{G}$-adapted.
\(ii) Let $l_+$ be the Lebesgue measure on $\mathbb{R}_+$. Because of Lemma \[lem:sqrt-tau-1\], using Fubini’s theorem, there is a measurable subset $\Lambda$ of $\mathbb{R}_+$ such that $l_+(\mathbb{R}_+\setminus\Lambda)=0$ and $E[|Z_t|]<+\infty$ for all $t\in\Lambda$. For later use, we fix a set $\Lambda$ with these properties.
\(iii) Formulas obtained in Section \[sec:Conditional-Expectations\] allow to compute the optional projection $^o\!Z$ of $Z$ with respect to the filtration $\mathbb{F}^\beta$ on the set $\Lambda$: For all $t\in\Lambda$ we have $\mathbf{P}$-a.s. $$\begin{aligned}
^o\!Z_t&=\mathbf{E}\left[Z_{t}|\mathcal{F}_{t}^{\beta}\right] =\mathbf{E}\left[\frac{\beta_{t}}{\tau-t}\,\mathbb{I}_{\left\{ t<\tau\right\} }|\beta_{t}\right]\\
&=\beta_{t}\int_{\left(t,+\infty\right)}\frac{1}{r-t}\phi_{t}\left(r,\beta_{t}\right)dF\left(r\right)\,\mathbb{I}_{\left\{ t<\tau\right\} }\,, \end{aligned}$$ where the first equality follows from Remark \[rem:If–for rem 63\], the second from the Markov property of the process $\beta$ and Definition (\[eq:processo Z\]) of the process $Z$, while the third equality follows directly from Proposition \[cor:Let–be-BIS\] and the measurability of $\beta_{t}$ with respect to $\sigma(\beta_{t})$. Note that in the above equality all terms are well-defined for *every* $t\geq 0$, the condition that $t\in\Lambda$ is only needed for the second and third equality.
\[B\] The process $B=(B_{t},\, t\geq0)$ defined by $$B_{t}:=\beta_{t}+\int_{0}^{t}Z_{s}\, ds\,,\quad t\geq 0\,,\label{eq:BIGBM}$$ is a $\mathbb{G}$-Brownian motion stopped at $\tau$.
Note that by Corollary \[cor:The-process-Z\] and Remark \[Z-modification\], the process $(Z_{t},\, t\geq0)$ is integrable with respect to the Lebesgue measure, hence $B$ is well-defined.
It is clear that the process $B$ is continuous and $\mathbb{G}$-adapted. In order to prove that it is indeed a $\mathbb{G}$-Brownian motion stopped at $\tau$, it suffices to prove that the process $B$ and the process $X$ defined by $X_t:=B_{t}^{2}-(t\wedge\tau)$, $t\geq0$, are both $\mathbb{G}$-martingales. To this end we shall use Corollary \[cor:LEM if–is\].
First we show that $B$ is a $\mathbb{G}$-martingale. Let $n\in\mathbb{N}$ be an arbitrary but fixed natural number, $0<t_{1}<...<t_{n-1}<t_{n}:=t,\: h\geq0$ and $g$ an arbitrary bounded Borel function. Recalling the definition of $B^r$ in , we have that $$\begin{aligned}
\lefteqn{\mathbf{E}\left[\left(B_{t+h}-B_{t}\right)g\left(\beta_{t_{1}},\dots,\beta_{t_{n}},\tau\right)\right]}\\
&=\int_{\left(0,+\infty\right)}\mathbf{E}\left[\left(B_{t+h}-B_{t}\right)g\left(\beta_{t_{1}},...,\beta_{t_{n}},\tau\right)|\tau=r\right]dF\left(r\right)\\
& =\int_{\left(0,+\infty\right)}\mathbf{E}\left[\left(B_{t+h}^{r}-B_{t}^{r}\right)g\left(\beta_{t_{1}}^{r},\ldots,\beta_{t_{n}}^{r},r\right)\right]dF(r)=0,\end{aligned}$$ because $B^{r}$ is a Brownian motion and hence a martingale with respect to the filtration generated by $\beta^{r}$ (cf. Lemma \[BMr\]). Using a monotone class argument, from this it can easily be derived that $B$ is a martingale with respect to $\mathbb{G}$.
It remains to prove that the process $X$ is a $\mathbb{G}$-martingale. Putting $X_{t}^{r}:=(B_{t}^{r})^{2}-(t\wedge r)$, $t\geq0$, and repeating the same arguments used above, we see that $$\begin{aligned}
\lefteqn{\mathbf{E}\left[\left(X_{t+h}-X_{t}\right)g\left(\beta_{t_{1}},\dots,\beta_{t_{n}},\tau\right)\right]}\\
&=\int_{\left(0,+\infty\right)}\mathbf{E}\left[\left(X_{t+h}-X_{t}\right)g\left(\beta_{t_{1}},\ldots,\beta_{t_{n}},\tau\right)|\tau=r\right]dF\left(r\right)\\
& =\int_{\left(0,+\infty\right)}\mathbf{E}\left[\left(X_{t+h}^{r}-X_{t}^{r}\right)g\left(\beta_{t_{1}}^{r},\ldots,\beta_{t_{n}}^{r},r\right)\right]dF(r)=0,\end{aligned}$$ since $X^{r}$ is a martingale with respect to the filtration generated by $\beta^{r}$. As above from this follows that $X$ is a martingale with respect to $\mathbb{G}$. This completes the proof of Lemma \[B\].
We are now ready to state the main result of this section.
\[TEO:semimartFb\] The process $b=(b_{t},\, t\geq0)$ given by $$\begin{aligned}
\nonumber b_{t}&:=\beta_{t}+\int_{0}^{t}\mathbf{E}\left[\frac{\beta_{s}}{\tau-s}\,\mathbb{I}_{\{s<\tau\}}\big|\beta_{s}\right]\, ds\\
&=\beta_t+\int_0^t \beta_s \int_{(s,+\infty)}\frac{1}{r-s}\,\phi_s(r,\beta_s)\, dF(r)\,\mathbb{I}_{\{s<\tau\}}\, ds\,,
\label{eq:semimart-dec-1}\end{aligned}$$ where the second equality holds $\mathbf{P}$-a.s., is an $\mathbb{F}^{\beta}$-Brownian motion stopped at $\tau$. Thus the information process $\beta$ is an $\mathbb{F}^\beta$-semimartingale whose decomposition is determined by .
First we notice that by Lemma \[lem:sqrt-tau-1\] and Remarks \[rem:If–for rem 63\] and \[rem:We-make-the QWERTY\] (iii), the integrands of the second and third term of are well-defined for all $s\geq 0$ and they are equal $l_+$-a.e. and integrable on $[0,t]\times\Omega$ with respect to $l_+\times\mathbf{P}$. This yields the second equality $\mathbf{P}$-a.s. Now we can apply Proposition \[pro:innovationLEMMA\] to the processes $(-Z)$ and $X$ where $X$ is defined by $X_t=\int_{0}^{t}(-Z_{s})\,ds+B_{t}$, $t\geq0$. By Proposition \[B\] we know that the process $B$ with $B_{t}:=\beta_{t}+\int_{0}^{t}Z_{s}\, ds$, $t\geq 0$, is a $\mathbb{G}$-Brownian motion stopped at $\tau$ (see Proposition \[B\]). Note that $X=\beta$. According to Proposition \[pro:innovationLEMMA\] the process $b$ with $$b_t=X_t-\int_0^{t}{^o(-Z)}_s\,ds=\beta_t+ \int_0^{t}{^o\!Z}_s\,ds=\beta_t+\int_{0}^{t}\mathbf{E}\left[\frac{\beta_{s}}{\tau-s}\,\mathbb{I}_{\{s<\tau\}}\big|\beta_{s}\right]\, ds\,,$$ is a Brownian motion stopped at $\tau$, with respect to $\mathbb{F}^X=\mathbb{F}^\beta$, where we have used Remark \[rem:We-make-the QWERTY\] (iii) for the third equality. This completes the proof of Theorem \[TEO:semimartFb\].
\[rem:QUA COV 1\] Note that the quadratic variation $\langle \beta,\beta\rangle$ of the information process $\beta$ is given by $$\langle \beta,\beta\rangle _{t}=\langle b,b\rangle _{t}=t\wedge\tau,\quad t\geq0\,.$$ This follows immediately from the semimartingale decomposition of $\beta$, see, for example, [@key-3 Ch. IV, (1.19)] (the quadratic variation does not depend on the filtration, provided that the process is a continuous semimartingale).
\[sec:Local-time-of\] Example: pricing a Credit Default Swap
============================================================
A rather common financial contract that is traded in the credit market is the Credit Default Swap (CDS). A CDS with maturity $T>0$ is a contract between a buyer who wants to protect against the possibility that the default of a financial asset will take place before $T$, and a seller who provides such insurance. If the default has not occurred before $T$, the buyer will pay to the seller a fee until the maturity. But if the default time $\tau$ occurs before the maturity, the fee will be paid until the default and then the seller will give immediately to the buyer a pre-established amount of money $\delta$, called *recovery*. The recovery may depend on the time at which the default occurs and, hence, it is modeled by a positive function $\delta:[0,T]\rightarrow\mathbb{R}_{+}$. We follow the approach developed in [@key-29], and the reader can find some details also in [@key-4 Ch. 7.8]. We assume that the default-free spot interest rate $r$ is constant and that the fee that the buyer must pay to the seller is paid continuously in time according to some rate $\kappa>0$, that is to say, the buyer has to pay an amount $\kappa dt$ during the time $dt$ until $\tau\wedge T$. In case the default time $\tau$ occurs before the maturity $T$, the seller will pay to the buyer a recovery $\delta(\tau)$ at time $\tau$. If the pricing measure is $\mathbf{P}$ and the market filtration is $\mathbb{G}=(\mathcal{G}_{t})_{\geq0}$, the price $S_{t}(\kappa,\delta,T,r)$ at time $t$ of the CDS is given by $$S_{t}\left(\kappa,\delta,r,T\right):=e^{rt}\,\mathbf{E}\left[e^{-r\tau}\delta\left(\tau\right)\mathbb{I}_{\left\{ t<\tau\leq T\right\} }-\int_{t\wedge\tau}^{T\wedge\tau}e^{-rv}\kappa \, dv|\mathcal{G}_{t}\right].\label{eq:CDS}$$ We would like to make a comparison between the result obtained in our model with the one presented in [@key-29]. However, differently from [@key-29], we reduce ourselves to the simple situation where the market filtration can be either the minimal filtration $\mathbb{H}$ that makes $\tau$ a stopping time or the filtration $\mathbb{F}^{\beta}$ generated by the information process $\beta$. It is worth to note that the minimal filtration that makes $\tau$ a stopping time is of particular importance in the theory of enlargement of filtrations and its applications to mathematical models of credit risk. We refer to the series of papers [@key-23; @key-24; @key-25] and to the book [@key-4] for the topics of mathematical finance, where the filtration $\mathbb{H}$ is used, and to the books [@key-18] and [@key-7] for a discussion of the subject from a purely mathematical point of view.
In the following, let us make the further assumption that $r=0$. We first recall the pricing formula for the market filtration $\mathbb{G}=\mathbb{H}$.
If the market filtration $\mathbb{G}$ is the minimal filtration $\mathbb{H}=(\mathcal{H}_{t})_{t\geq0}$ that makes $\tau$ a stopping time, then the price $S_{t}(\kappa,\delta,0,T)$ at time $t$ of a CDS is given by $$\begin{aligned}
S_{t}\left(\kappa,\delta,0,T\right) & =\mathbf{E}\left[\delta\left(\tau\right)\mathbb{I}_{\left\{ t<\tau\leq T\right\} }-\kappa\left(\left(\tau\wedge T\right)-t\right)\mathbb{I}_{\left\{t<\tau\right\} }|\mathcal{H}_{t}\right]\nonumber \\
& =\mathbb{I}_{\left\{t<\tau\right\} }\frac{1}{G\left(t\right)}\left(-\int_{t}^{T}\delta\left(v\right)dG\left(v\right)-\kappa\int_{t}^{T}G\left(v\right)dv\right)\label{eq:CDSeasy}\end{aligned}$$ where $G(u):=\mathbf{P}(u<\tau)=1-F(u)$ is supposed to be strictly greater than 0 for every $u\in[0,T]$.
See [@key-29 Lemma 2.1].
In our model, i.e., if the market filtration is the filtration $\mathbb{F}^{\beta}=(\mathcal{F}_{t}^{\beta})_{t\geq0}$ generated by the information process $\beta$, the pricing formula is given in the following proposition.
\[lem:If-,-i.e.,\]If $\mathbb{G}=\mathbb{F}^{\beta}$, then the price $S_{t}(\kappa,\delta,0,T)$ at time $t$ of a CDS is given by $$\begin{aligned}
S_{t}\left(\kappa,\delta,0,T\right) & =\mathbf{E}\left[\delta\left(\tau\right)\mathbb{I}_{\left\{ t<\tau\leq T\right\} }-\kappa\left(\left(\tau\wedge T\right)-t\right)\mathbb{I}_{\left\{t<\tau\right\} }|\mathcal{F}_{t}^{\beta}\right]\label{price-formula} \\
& =\mathbb{I}_{\left\{t<\tau\right\} }\left(-\int_{t}^{T}\delta\left(v\right)d_{v}\Psi_{t}\left(v\right)-\kappa\int_{t}^{T}\Psi_{t}\left(v\right)dv\right)\label{eq:CDSmy}\end{aligned}$$ where $\Psi_{t}(u):=\mathbf{P}(u<\tau|\mathcal{F}_{t}^{\beta}) =\int_{u}^{+\infty}\phi_{t}(v,\beta_{t})\,dF(v)$ and the writing $d_{v}\Psi_{t}(v)$ in the above formula means that the integral is computed using $v$ as integrating variable.
Concerning the first term in , in view of Corollary \[LEM:prop\_cond\_exp\_tau\] and , we have, $\mathbf{P}$-a.s.: $$\begin{aligned}
\mathbf{E}\left[\delta\left(\tau\right)\mathbb{I}_{\left\{ t<\tau\leq T\right\} }|\mathcal{F}_{t}^{\beta}\right]&=\mathbb{I}_{\left\{t<\tau\right\} }\int_{t}^{T}\delta\left(v\right)\phi_{t}\left(v,\beta_{t}\right)dF\left(v\right)\\
&=\mathbb{I}_{\left\{t<\tau\right\} }\int_{t}^{T}\delta\left(v\right)d_{v}\Phi_{t}\left(v\right)\end{aligned}$$ where $\Phi_{t}(v):=\mathbf{P}(\tau\leq v|\mathcal{F}_{t}^{\beta})=1-\Psi_{t}(v)$ and hence $$\begin{aligned}
\label{Ex1}
\mathbf{E}\left[\delta\left(\tau\right)\mathbb{I}_{\left\{ t<\tau\leq T\right\} }|\mathcal{F}_{t}^{\beta}\right]&=
\mathbb{I}_{\left\{t<\tau\right\} }\int_{t}^{T}\delta\left(v\right)d_{v}\Phi_{t}\left(v\right)\\
&=-\mathbb{I}_{\left\{t<\tau\right\} }\int_{t}^{T}\delta\left(v\right)d_{v}\Psi_{t}\left(v\right)\,.\end{aligned}$$ Concerning the second term in , again in view of Corollary \[LEM:prop\_cond\_exp\_tau\] and , we obtain $\mathbf{P}$-a.s., $$\begin{aligned}
\nonumber\mathbf{E}\left[T\wedge\tau|\mathcal{F}_{t}^{\beta}\right]\mathbb{I}_{\left\{t<\tau\right\} } &=&\left(\int_{t}^{T}vd_{v}\Phi_{t}\left(v\right)+T\left(1-\Phi_{t}\left(T\right)\right)\right)\mathbb{I}_{\left\{t<\tau\right\} }\\
\nonumber&=&\left(-\int_{t}^{T}vd_{v}\Psi_{t}\left(v\right)+T\Psi_{t}\left(T\right)\right)\mathbb{I}_{\left\{t<\tau\right\}}\\
&=&\left(\int_{t}^{T}\Psi_{t}\left(v\right)dv+t\Psi_{t}\left(t\right)\right)\mathbb{I}_{\left\{t<\tau\right\}}\,. \label{Ex2}\end{aligned}$$ Inserting and into the price formula and noting that $\Psi_t(t)=1$ on $\{t<\tau\}$ $\mathbf{P}$-a.s., we obtain the asserted result . The proof of the proposition is completed.
Although there is a formal analogy between and , the pricing formulas are quite different. The second formula is much more informative because it uses the observation of $\beta_t$ and the Bayes estimate of $\tau$, namely, the a posteriori distribution of $\tau$ after observing $\beta_t$. Given that $t<\tau$, the price in is a deterministic value, while in the price depends on the observation $\beta_t$.
The so-called *fair spread* of a CDS at time $t$ is the value $\kappa^{*}$ such that $S_{t}(\kappa^{*},\delta,0,T)=0$. In our model we have that $$\kappa^{*}=-\frac{\int_{t}^{T}\delta\left(r\right)d_{r}\Psi_{t}\left(r\right)}{\int_{t}^{T}\Psi_{t}\left(r\right)dr}$$ while, in the simpler situation where the market filtration is $\mathbb{H}$, the fair spread is given by $$\kappa^{*}=-\frac{\int_{t}^{T}\delta\left(r\right)dG\left(r\right)}{\int_{t}^{T}G\left(r\right)dr}.$$
When dealing with problems related to credit risk, it is of interest to consider the case where the market filtration is a filtration $\mathbb{G}$ obtained by progressively enlarging a reference filtration $\mathbb{F}$ with another filtration, $\mathbb{D}=(\mathcal{D}_{t})_{t\geq0}$, which is responsible for modeling the information associated with the default time $\tau$. Traditionally, the filtration $\mathbb{D}$ has been settled to be equal to the filtration $\mathbb{H}$ generated by the single-jump process occurring at $\tau$. We intend to consider a different setting where the reference filtration $\mathbb{F}$ will be enlarged with the filtration $\mathbb{F}^{\beta}$ and we will provide the relative pricing formulas of credit instruments like the CDS. However, this is beyond the scope of the present paper.
The Bayes Formula
=================
Here the basic results on the Bayes formula are recalled without proofs. For further details we refer to [@key-8 Ch. II.7.8] or any book on Bayesian Statistics.
Let $\tau$ and $X$ be random variables on a probability space $(\Omega,\,\mathcal{F},\,\mathbf{P})$ with values in measurable spaces $(E_{1},\,\mathcal{E}_{1})$ and $(E_{2},\,\mathcal{E}_{2})$, respectively. Let $\mathbf{P}_{r}$ be a regular conditional distribution of $X$ with respect to $\tau=r$, i.e., for $B\in\mathcal{E}_{2}$, $\mathbf{P}_{r}(B)=\mathbf{P}(X\in B|\tau=r)$, $\mathbf{P}_{\tau}$-a.s. By $\mathbf{P}_{\tau}$ we denote the distribution of $\tau$ on $(E_{1},\,\mathcal{E}_{1})$ (called the *a priori* distribution). Moreover, for $C\in\mathcal{E}_{1}$, let $\mathbf{G}_{C}$ be defined as follows: $$\begin{aligned}
\mathbf{G}_{C}\left(B\right) & :=\int_{C}\mathbf{P}_{r}\left(B\right)\mathbf{P}_{\tau}\left(dr\right),\, B\in\mathcal{E}_{2}\,.\label{eq:starUAN}\end{aligned}$$ We are interested in the *a posteriori* probability $\mathbf{Q}(x,C):=\mathbf{P}(\tau\in C|X=x)$, for $x\in E_{2}$ and $C\in\mathcal{E}_{1}$. By $\mathbf{P}_{X}$ we denote the law of $X$.
We have $\mathbf{G}_{C}\ll\mathbf{P}_{X}$ and $$\mathbf{Q}\left(x,C\right)=\frac{d\mathbf{G}_{C}}{d\mathbf{P}_{X}}\left(x\right),\quad x\in E_{2},\; C\in\mathcal{E}_{1},\;\mathbf{P}_{X}\textrm{-a.s.}$$
Now we assume that there exists a $\sigma$-finite measure $\mu$ on $(E_{2},\,\mathcal{E}_{2})$ such that $\mathbf{P}_{r}\ll\mu$, for all $r\in E_1$. Furthermore, we assume that there is a measurable function $p$ on $(E_{1}\times E_{2},\,\mathcal{E}_{1}\otimes\mathcal{E}_{2})$ such that $$p\left(r,x\right)=\frac{d\mathbf{P}_{r}}{d\mu}\left(x\right),\quad \mu\textrm{-a.e.}, \; r\in E_{1}\,.$$
We have that
1. $\mathbf{G}_{C}\ll\mu$ and $\displaystyle{\frac{d\mathbf{G}_{C}}{d\mu}(x)=\int_{C}p(r,x)\mathbf{P}_{\tau}(dr),\;\mu\textrm{-a.e.}}$,
2. $\mathbf{P}_{X}\ll\mu$ and $\displaystyle{\frac{d\mathbf{P}_{X}}{d\mu}(x)=\int_{E_{1}}p(r,x)\mathbf{P}_{\tau}(dr),\;\mu\textrm{-a.e.}}$
\[cor:(Bayes-formula)-a.s.\] $$\mathbf{P}\left(\tau\in C|X=x\right)=
\mathbf{Q}\left(x,C\right)=\frac{{\displaystyle \int_{C}p\left(r,x\right)\mathbf{P}_{\tau}\left(dr\right)}}{{\displaystyle \int_{E_{1}}p\left(v,x\right)\mathbf{P}_{\tau}\left(dv\right)}},\quad \mathbf{P}_{X}\textrm{-a.s.},\; C\in\mathcal{E}_{1}\,.$$
\[cor:The-conditional-density\] The a posteriori density $q(r,x)$ of $\tau$ under $X=x$ with respect to $\mathbf{P}_{\tau}$ is given by $$q\left(r,x\right):=p\left(r,x\right) \left(\int_{E_{1}}p\left(v,x\right)\mathbf{\mathbf{P}_{\tau}}\left(dv\right)\right)^{-1},\quad r\in E_1,\; x\in E_2\,.$$ Thus $$\mathbf{P}\left(\tau\in C|X=x\right)=\int_{C}q\left(r,x\right)\mathbf{\mathbf{P}_{\tau}}\left(dr\right),\quad \mathbf{P}_{X}\textrm{-a.s.},\; C\in\mathcal{E}_{1}\,.$$
For the proof of Theorem \[lem:PRE-theo-4.3\], we have to choose $(E_1,\mathcal{E}_1)=(\mathbb{R}_+,\mathcal{B}(\mathbb{R}_+))$, $(E_2,\mathcal{E}_2)=(\mathbb{R},
\mathcal{B}(\mathbb{R}))$, $X=\beta_t$, $\tau$ as the default time and the $\sigma$-finite measure $\mu$ as the measure $\delta_0+l$ where $\delta_0$ is the Dirac measure at $0$ and $l$ is the Lebesgue measure on $\mathbb{R}$. Then Corollary \[cor:(Bayes-formula)-a.s.\] can be applied to the second term on the right-hand side of which yields the second term on the right hand side of .
\[sec:Proof of Innovation lemma\] Proof of Proposition \[pro:innovationLEMMA\] (Innovation Lemma)
=================================================================================================
We start by observing that the asssumption of Proposition \[pro:innovationLEMMA\] implies $$\int_{0}^{t}\left|Z_{s}\right|\, ds<+\infty,\quad t\geq0, \;\mathbf{P}\textrm{-a.s.}$$ Putting $Z$ equal to zero on the negligible set where $\int_0^t |Z_s|\, ds=+\infty$ for some $t\geq0$, we can assume without loss of generality that this property holds everywhere. Hence $\int_{0}^{t}Z_{s}\, ds$ is well defined and finite everywhere.
By Remark \[rem:If–for rem 63\] we know that for the optional projection $^o\!Z$ of $Z$ with respect to $\mathbb{F}^X$ we have that $^{o}\!Z_{s}=\mathbf{E}\left[Z_{s}|\mathcal{F}_{s}^{X}\right]$, $\mathbf{P}$-a.s., for all $s\geq0$. This yields $\mathbf{E}\left[\int_{0}^{t}\left|^{o}\!Z_{s}\right|\, ds\right]\leq \mathbf{E}\left[\int_{0}^{t}\left|Z_{s}\right|\, ds\right]<+\infty$, $t\geq0$, and hence $\int_{0}^{t}|{^o\!Z}_{s}|\, ds<+\infty$, for all $t\geq0$, $\mathbf{P}$-a.s. As above, without loss of generality we can modify $^{o}\!Z$ on a $\mathbf{P}$-negligible set such that $\int_{0}^{t}\left|^{o}\!Z_{s}\right|\, ds<+\infty$, $t\geq0$, everywhere, while being still the optional projection of $Z$. Hence $\int_{0}^{t}\,^{o}\!Z_{s}\, ds$ is well-defined and finite everywhere, for all $t\geq0$. Since $^o\!Z$ is $\mathbb{F}^X$-optional, it is clear that the process $\int_{0}^{\cdot}\,^{o}\!Z_{s}\, ds$ is $\mathbb{F}^{X}$-adapted. Therefore, the process $b$ with $b_t:=X_t-\int_0^t{^o\!Z}_s\, ds$, $t\geq0$, is $\mathbb{F}^{X}$-adapted. Furthermore, $Z$ being $\mathbb{F}$-optional, the process $\int_{0}^{\cdot}Z_{s}\, ds$ is $\mathbb{F}$-adapted and hence the process $X$ is also $\mathbb{F}$-adapted. This yields the inclusion $\mathbb{F}^X\subseteq\mathbb{F}$.
Next we show that the process $b$ is an $\mathbb{F}^{X}$-martingale. Obviously, $b_{t}$ is integrable for all $t\geq0$ and, as shown above, $b$ is $\mathbb{F}^{X}$-adapted. Let $0\leq s<t$. For showing the martingale property, using Fubini’s theorem, we first notice that there exists a Borel set $\Lambda\subseteq\mathbb{R}_+$ such that $l_+(\mathbb{R}_+\setminus \Lambda)=0$ and $Z_s$, and hence also $^o\!Z_s$, is integrable for all $s\in\Lambda$. Now $$\begin{aligned}
\lefteqn{\mathbf{E}\left[b_{t}-b_{s}|\mathcal{F}_{s}^{X}\right]}\\
&=\mathbf{E}\left[B_{t}-B_{s}|\mathcal{F}_{s}^{X}\right]+\mathbf{E}\left[\int_{s}^{t}\left(Z_{u}-{}^{o}\!Z_{u}\right)du|\mathcal{F}_{s}^{X}\right]\\
&=\mathbf{E}\left[\mathbf{E}\left[B_{t}-B_{s}|\mathcal{F}_{s}\right]|\mathcal{F}_{s}^{X}\right]+\mathbf{E}\left[\int_{s}^{t}\mathbb{I}_{\Lambda}\left(u\right)\,\left(Z_{u}-{}^{o}\!Z_{u}\right)du|\mathcal{F}_{s}^{X}\right]\\
& =\int_{0}^{t}\mathbb{I}_{\Lambda}\left(u\right)\,\mathbf{E}\left[Z_{u}-{}^{o}\!Z_{u}|\mathcal{F}_{s}^{X}\right]du\\
&=\int_{0}^{t}\mathbb{I}_{\Lambda}\left(u\right)\,\mathbf{E}\left[\mathbf{E}\left[Z_{u}-{}^{o}\!Z_{u}|\mathcal{F}^X_u\right]|\mathcal{F}_{s}^{X}\right]du=0\end{aligned}$$ where we have used that $\mathbb{F}^X\subseteq\mathbb{F}$ and that $B$ is an $\mathbb{F}$-martingale, Fubini’s theorem and properties of the optional projection (see Remark \[rem:If–for rem 63\]). This proves that $b$ is a continuous $\mathbb{F}^X$-martingale.
Finally, since $B$ is an $\mathbb{F}$-Brownian motion stopped at $T$, from the definition of $b$ it is clear that $b$ is a continuous $\mathbb{F}$-semimartingale with square variation process $\langle b,b\rangle$: $\langle b,b\rangle_t=t\wedge T$, $t\geq 0$. In view of the well-known fact that the square variation of continuous semimartingales does not depend on the filtration, this is also true with respect to $\mathbb{F}^X$. This proves that $b$ is an $\mathbb{F}^X$-Brownian motion stopped at $T$.
\[sec:Proofs-of-Section\] Proof of Lemma \[lem:On-the-set-2\]
=============================================================
First we recall that $(t_{n})_{n\in\mathbb{N}}$ is a strictly decreasing sequence converging to $0$ such that $0<t_{n+1}<t_{n}<\ldots t_{1}<\varepsilon,\, t_{n}\downarrow0$. For proving that $\lim_{n\rightarrow\infty}\phi_{t_{n}}(r,\beta_{t_{n}})=1$, using the assumption that $\mathbf{P}(\tau>\varepsilon)=1$, and , we can write $$\begin{aligned}
\nonumber\lefteqn{\phi_{t_{n}}\left(r,\beta_{t_{n}}\right)}\\
=&\frac{\left(2\pi t_{n}\right)^{-1/2}\,{r}^{1/2}\left(r-t_{n}\right)^{-{1/2}}\,\exp\big[-\beta_{t_{n}}^{2}r/2t_{n}\left(r-t_{n}\right)\big]\mathbb{I}_{\left(t_n,+\infty\right)}\left(r\right)}{\!\!\!\!\int\limits_{\left(\varepsilon,+\infty\right)}\left(2\pi t_{n}\right)^{-1/2}\,{s}^{1/2}\left(s-t_{n}\right)^{-1/2}\, \exp\big[-\beta_{t_{n}}^{2}s/2t_{n}\left(s-t_{n}\right)\big]\, dF\left(s\right)}\\
=&\frac{r^{1/2}\left(r-t_{n}\right)^{-1/2}\; \mathbb{I}_{\left(t_n,+\infty\right)}\left(r\right)}{\int\limits_{\left(\varepsilon,+\infty\right)}{s}^{1/2}\left(s-t_{n}\right)^{-1/2}\exp\big[\beta_{t_{n}}^{2}\left(s-r\right)/2\left(r-t_{n}\right)\left(s-t_{n}\right)\big]\,dF\left(s\right)}\,.\label{eq:boh}\end{aligned}$$ First we note that for $s\in\left(\varepsilon,+\infty\right)$ $$\label{Estimate:square-root}
1\leq s^{1/2}\left(s-t_{n}\right)^{-1/2}\leq\varepsilon^{1/2} \left(\varepsilon-t_{n}\right)^{-1/2}\leq\varepsilon^{1/2}\left(\varepsilon-t_{1}\right)^{-1/2}\,.$$ Secondly, if $s\in(\varepsilon,r)$, then $\exp\left[\beta_{t_{n}}^{2}(s-r)/2(r-t_{n})(s-t_{n})\right]\leq1,$ while for $s\in[r,+\infty)$ we have that $(s-r)/(s-t_{n})\leq1$, and thus $$\exp\left[\beta_{t_{n}}^{2}(s-r)/2(r-t_{n})(s-t_{n})\right]\leq\exp\left[\beta_{t_{n}}^{2}/2(r-t_{n})\right].$$ We note that the right-hand side of this inequality is bounded with respect to $n$, since $t_n\downarrow0$ and $\beta_{t_n}\rightarrow0$. Furthermore, $$\lim_{n\rightarrow\infty}s^{1/2}\left(s-t_{n}\right)^{-1/2}\exp\left[\beta_{t_{n}}^{2}\left(s-r\right)/2\left(r-t_{n}\right)\left(s-t_{n}\right)\right]=1\,.$$ Thus we can apply Lebesgue’s bounded convergence theorem and it follows that $$\lim_{n\rightarrow\infty}
\int\limits_{\left(\varepsilon,+\infty\right)}\!\!\!\!s^{1/2}\left(s-t_{n}\right)^{-1/2}\exp\left[\beta_{t_{n}}^{2}\left(s-r\right)/2\left(r-t_{n}\right)\left(s-t_{n}\right)\right]\,dF\left(s\right)=1\,.$$ Finally, as for the numerator in , $$\lim_{n\rightarrow\infty}r^{1/2}(r-t_{n})^{-1/2}\, \mathbb{I}_{(t_n,+\infty)}(r)=1,\; r>0\,,$$ we have $\lim_{n\rightarrow\infty}\phi_{t_{n}}(r,\beta_{t_{n}})=1$.
In order to prove the $\mathbf{P}_\tau$-a.s. uniform boundedness of $\phi_{t_{n}}(\cdot,\beta_{t_{n}})$, first we notice that in view of the numerator in is uniformly bounded in $r\in(\varepsilon,+\infty)$ and using the assumption that $\mathbf{P}(\tau>\varepsilon)=1$ it follows that it is $\mathbf{P}_\tau$-a.s. uniformly bounded in $r\in(0,+\infty)$, too. It remains to verify that the denominator in is bounded from below by a strictly positive constant only depending on $\varepsilon$ and $\omega$. Using , for $s\in(\varepsilon,+\infty)$ the integrand in the denominator of can be estimated from below by $$\begin{aligned}
\lefteqn{{s}^{1/2}\left(s-t_{n}\right)^{-1/2}\exp\big[\beta_{t_{n}}^{2}\left(s-r\right)/2\left(r-t_{n}\right)\left(s-t_{n}\right)\big]}\\
&\geq&\exp\big[\beta_{t_{n}}^{2}\left(s-r\right)/2\left(r-t_{n}\right)\left(s-t_{n}\right)\big]\\
&\geq&\exp\left[-\beta_{t_{n}}^{2}/\left(\varepsilon-t_{1}\right)\right]\, \mathbb{I}_{\left(\varepsilon,r\right)}+\mathbb{I}_{\left[r,+\infty\right)}\geq e^{-\gamma}\end{aligned}$$ where $\gamma=\gamma(\varepsilon,\omega)=\sup_{n\geq1}\beta_{t_{n}}^{2}(\omega)/(\varepsilon-t_{1})<+\infty$. Hence for the denominator in it follows $$\begin{aligned}
\lefteqn{e^{-\gamma}=\int_{\left(\varepsilon,+\infty\right)}e^{-\gamma}\, dF\left(s\right)}\\
&\leq&\int_{\left(\varepsilon,+\infty\right)}{s}^{1/2}\left(s-t_{n}\right)^{-1/2}\exp\big[\beta_{t_{n}}^{2}\left(s-r\right)/2\left(r-t_{n}\right)\left(s-t_{n}\right)\big]\,dF\left(s\right)\,.\end{aligned}$$ The proof of Lemma \[lem:On-the-set-2\] is completed.
Acknowledgment {#acknowledgment .unnumbered}
--------------
This work has been financially supported by the European Community’s FP 7 Program under contract PITN-GA-2008-213841, Marie Curie ITN Controlled Systems.
[99]{} Bedini M. L. *Information on a Default Time: Brownian Bridges on Stochastic Intervals and Enlargment of Filtrations*. PhD Thesis, Friedrich Schiller University of Jena (Germany), 2012.
Bielecki T.R., Jeanblanc M., M. Rutkowski. *Hedging of basket of credit derivatives in a credit default swap market*. Journal of Credit Risk, 3: 91–132, 2007.
Blumenthal R.M., Getoor R.K. *Markov Processes and Potential Theory*. Academic Press, 1968.
Brody D., Hughston L., Macrina A. *Beyond Hazard Rates: A New Framework for Credit-Risk Modeling*. Advances in Mathematical Finance: Festschrift Volume in Honour of Dilip Madan, pp. 231–257, Basel: Birkhäuser, 2007.
Bielecki T.R., Rutkowski M. *Credit risk: Modelling Valuation and Hedging*. Springer Verlag, Berlin 2001.
Dellacherie C. *Capacités et processus stochastiques*. Volume 67 of Ergebnisse. Springer 1972.
Dellacherie C., Meyer P.-A. *Probabilities and Potential*. North-Holland, 1978.
Jeanblanc M., Le Cam Y. *Immersion property and Credit Risk Modelling*. Optimality and Risk - Modern Trends in Mathematical Finance, pp. 99–132. Springer Berlin Heidelberg, 2010.
Jeanblanc M., Le Cam Y. *Progressive enlargement of filtrations with initial times*. Stoch. Process. Appl., 2009, v. 119, No. 8, 2523–2543.
Jeanblanc M., Le Cam Y. *Reduced form modelling for credit risk*. Preprint 2007, availabe at: http://ssrn.com/abstract=1021545.
Jeanblanc M., Yor M., Chesney M. *Mathematical Methods for Financial Markets*. Springer, First edition, 2009.
Kallenberg O. *Foundations of Modern Probability*. Springer-Verlag, New-York, Second edition, 2002.
Karatzas I., Shreve S. *Brownian Motion and Stochastic Calculus*. Springer- Verlag, Berlin, Second edition, 1991.
Revuz D., Yor M. *Continuous Martingales and Brownian Motion*. Springer-Verlag, Berlin, Third edition, 1999.
Rogers L.C.G., Williams D. *Diffusions, Markov Processes and Martingales. Vol. 2: Itô Calculus*. Cambridge University Press, Second edition, 2000.
Shiryaev A.N. *Probability*. Springer Verlag, Second Edition, 1991.
Matteo Ludovico Bedini, Intesa Sanpaolo, Milano, Italy;\
e-mail: [email protected]
Rainer Buckdahn, Laboratoire de Mathématiques CNRS-UMR 6204,\
Université de Bretagne Occidentale, Brest, France; School of Mathematics, Shandong University, Jinan, Shandong Province, P.R.China;\
e-mail: [email protected]
Hans-Jürgen Engelbert, Friedrich-Schiller-Universität Jena, Jena, Germany; e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Absolute separable states is a kind of separable state that remain separable under the action of any global unitary transformation. These states may or may not have quantum correlation and these correlations can be measured by quantum discord. We find that the absolute separable states are useful in quantum computation even if it contains infinitesimal quantum correlation in it. Thus to search for the class of two-qubit absolute separable states with zero discord, we have derived an upper bound for $Tr(\varrho^{2})$, where $\varrho$ denoting all zero discord states. In general, the upper bound depends on the state under consideration but if the state belong to some particular class of zero discord states then we found that the upper bound is state independent. Later, it is shown that among these particular classes of zero discord states, there exist sub-classes which are absolutely separable. Then we construct a ball for $2\otimes d$ quantum system described by $Tr(\rho^{2})\leq Tr(X^{2})+2\lambda_{min}(X)\lambda_{min}(Z)+Tr(Z^{2})$, where the $2\otimes d$ quantum system is described by the density operator $\rho$ which can be expressed by $d\otimes d$ block matrices $X,Y$ and $Z$ with $X,Z\geq 0$ and $\lambda_{min}(X)$, $\lambda_{min}(Z)$ denoting the minimum eigenvalues of the block matrices $X$ and $Z$ respectively. In particular, we show that the newly constructed ball contain class of absolute separable states described by two-qubit density operator $\rho$ that are lying outside the ball described by $Tr(\rho^{2})\leq \frac{1}{3}$. Also we have discussed an example of a class of absolute separable states in $2\otimes 3$ system where we find that most of the absolute separable states are residing inside the new ball and few of them are lying outside the ball.'
author:
- Satyabrata Adhikari
title: 'Constructing a ball of separable and absolutely separable states for $2\otimes d$ quantum system'
---
\[theorem\][Lemma]{} \[theorem\][Corollary]{} \[theorem\][Proposition]{} \[theorem\][Conjecture]{}
\[theorem\][Definition]{}
Introduction
============
Quantum correlation can be considered as a necessary ingredient for the development of quantum information theory and quantum computation. A remarkable application of quantum correlation can be found in different areas of quantum communication such as quantum teleportation [@bennett], quantum dense coding [@wiesner], quantum remote state preparation [@pati] , quantum cryptography [@gisin] etc. Till few years ago, it has been thought that this non-local feature in terms of quantum correlation only exist in the entangled state and responsible for the computational speed-up in the known quantum algorithms [@ekert]. Later, Lloyd [@lloyd] showed that there are quantum search methods which does not require entanglement to provide a computational speed-up over classical methods. In this modern line of research, Ahn et al. [@ahn] have shown that instead of entanglement, quantum phase is an essential ingredient for the computational speed-up in the Grover’s quantum search algorithm [@grover]. Meyer [@meyer] was able to reduce the number of queries in a quantum search compared to classical search of a database using only interference, not entanglement. Gottesman-Knill theorem also declare the fact that entanglement is not only a factor to for a quantum computers to outperform classical computers [@nielsen]. In 2004, E. Biham et.al. [@biham] then conclude that entangled state is not a compulsory ingredient for quantum computing and discovered that there exist quantum state lying arbitrarily close to the maximally mixed states, which are enough to increase the computational speed-up in quantum algorithms. To characterize the nature of mixed density matrices lying in the sufficiently small neighbourhood of maximally mixed state, it has been shown that all such states are separable states [@Zyczkowski; @braunstein; @vidal].\
Separable states can be defined as the mixture of locally indistinguishable states. Mathematically, a bipartite separable state described by the density operator $\rho$ in a composite Hilbert space $H_{1}\otimes H_{2}$ can be expressed as $$\begin{aligned}
\rho=\sum_{k} p_{k} \rho_{1}^{k}\otimes \rho_{2}^{k},~~0\leq p_{k}\leq 1
\label{sepstate}\end{aligned}$$ where $\rho_{1}$ and $\rho_{2}$ represents two density operators in two Hilbert spaces $H_{1}$ and $H_{2}$ respectively. These states can be prepared using local quantum operation and classical communication (LOCC). Thus prescription of its preparation is different from entangled states, which cannot be prepared with the help of LOCC. Since separable states are prepared by performing quantum operation within the structure of LOCC on quantum bit so they can exhibit quantum correlation [@modi]. Therefore, it can be inferred that not only entaglement but also this non-classical feature exhibited by some separable states.\
The entanglement measures [@wootters; @vidal1] cannot quantify the quantum correlation present in the separable state due to the reason that any entanglement measure gives value zero for all separable states. Thus, Ollivier and Zurek [@ollivier] proposed a measure for quantum correlation which can be defined as the difference between the quantum mutual information and the measurement-induced quantum mutual information. This measure is commonly known as quantum discord. If a separable state has no quantum correlation then they are termed as zero discord state. The quantum discord of two-qubit maximally mixed marginals and two qubit X-state has been calculated in [@luo; @ali]. It has been found that quantum correlation plays a vital role in mixed-state quantum computation speed-up and it is due to the correlation present in the separable states [@dutta2005; @dutta2007].\
We should note an important fact that there exist separable states (with or without quantum correlation) that can be converted to entangled state under the action of global unitary operation. The class of separable states that remain separable state after performing global unitary operation are known as absolutely separable states [@kus]. The necessary and sufficient condition for the absolute separability of a state in $2\otimes 2$ system described by the density operator $\sigma$ is given by [@verstraete] $$\begin{aligned}
\lambda_{1}\leq \lambda_{3}+2\sqrt{\lambda_{2}\lambda_{4}}
\label{abssepcond}\end{aligned}$$ where $\lambda_{i},i=1,2,3,4$ denoting the eigenvalues of $\sigma$ arranged in descending order as $\lambda_{1}\geq \lambda_{2}\geq\lambda_{3}\geq\lambda_{4}$. Further, Johnston [@johnston] generalize the absolute separability condition for $H_{2}\otimes H_{d}$ system and showed that a state $\sigma \in H_{2}\otimes H_{d}$ is absolute separable if and only if $$\begin{aligned}
\lambda_{1}\leq \lambda_{2d-1}+2\sqrt{\lambda_{2d-2}\lambda_{2d}}
\label{abssepcondgen}\end{aligned}$$ Since the absolute separability conditions (\[abssepcond\]) and (\[abssepcondgen\]) depends on the eigenvalues of the state under investigation so sometimes it is also known as separability from spectrum. Recently, the absolutely separable states are detected and characterized in [@nirman; @halder].\
This work is motivated by two earlier results and they are described as follows: Firstly, we can observe that the state used in solving the Deutsch-Jozsa (DJ) problem [@deutsch] is a pseudo-pure state (PPS) [@gershenfeld] which can be expressed as $$\begin{aligned}
\rho_{PPS}^{(2)}=\epsilon |\psi\rangle\langle \psi|+\frac{1-\epsilon}{4}I_{4}, 0\leq \epsilon \leq 1
\label{pps}\end{aligned}$$ where $|\psi\rangle$ is any two-qubit pure state. If $|\psi\rangle$ represent any two-qubit pure maximally entangled state then it reduces to the two-qubit Werner state. The sufficient condition that the state $\rho_{PPS}^{(2)}$ is separable whenever [@braunstein] $$\begin{aligned}
\epsilon < \frac{1}{9}
\label{sepcondpps}\end{aligned}$$ The quantum algorithm of Deustch and Jozsa solves DJ problem with a single query while classical algorithm uses 3 queries if initially two-qubit PPS with parameter $\epsilon$ $(0\leq \epsilon \leq 1)$ is used in the algorithm. As the number of qubit increases in the initial PPS, the number of queries increases exponentially for classical algorithm while the quantum algorithm of Deustch and Jozsa still require single query and this provide the quantum advantage over classical algorithm. It has been shown that if $\epsilon \leq \frac{1}{33}$ then the initial separable PPS with which the computation has started remain separable throughout the entire computation [@biham]. This may imply that the initial PPS is absolutely separable for $0\leq \epsilon \leq \frac{1}{33}$. To verify this statement, let us consider the PPS described by the density operator $\rho_{PPS}$ given by $$\begin{aligned}
\rho_{PPS}=\epsilon |\psi^{-}\rangle\langle \psi^{-}|+\frac{1-\epsilon}{4}I_{4}, 0\leq \epsilon \leq \frac{1}{33}
\label{newpps}\end{aligned}$$ where $|\psi^{-}\rangle=\frac{1}{\sqrt{2}}(|01\rangle-|10\rangle)$.\
The eigenvalues of $\rho_{PPS}$ can be arranged in descending order as $\mu_{1}\geq \mu_{2}\geq \mu_{3}\geq \mu_{4}$, where $$\begin{aligned}
\mu_{1}=\frac{1+3\epsilon}{4} ,\mu_{2}=\frac{1-\epsilon }{4}, \mu_{3}=\frac{1-\epsilon }{4} , \mu_{4}=\frac{1-\epsilon }{4}
\label{eigvalnew}\end{aligned}$$ We find that the state $\rho_{PPS}$ is absolutely separable if and only if $$\begin{aligned}
0\leq \epsilon\leq \frac{1}{33}
\label{abssepex}\end{aligned}$$ Again, the quantum discord of the state $\rho_{PPS}$ is given by [@luo] $$\begin{aligned}
D(\rho_{PPS})&=&\frac{1-\epsilon}{4}log_{2}(1-\epsilon)-\frac{1+\epsilon}{2}log_{2}(1+\epsilon)\nonumber\\&&+
\frac{1+3\epsilon}{4}log_{2}(1+3\epsilon)
\label{discordluo}\end{aligned}$$ In particular, if we choose the value of the parameter $\epsilon$ very close to zero, say $\epsilon=0.001$ then $D(\rho_{PPS})\simeq 1.441255\times 10^{-6}$. Therefore, the discord $D(\rho_{PPS})$ is very negligible and can be approximated to zero. Although the quantum correlation of the absolute separable state $\rho_{PPS}$ measured by quantum discord is very near to zero but still it is useful in quantum computation. This fact motivate us to search for the class of absolute separable state with zero discord.\
Secondly, the largest ball constructed for $d\otimes d$ quantum system is given by $Tr(\rho^{2})\leq \frac{1}{d^{2}-1}$, where the density operator $\rho$ representing either separable or absolutely separable states [@Zyczkowski; @gurvit]. In particular, it was shown that the largest ball for $2\otimes 2$ quantum system centered at maximally mixed state neither contain all separable states nor absolutely separable states [@kus]. To illustrate this, let us consider a state described by the density operator $\sigma_{1}$ given by $$\begin{aligned}
\sigma_{1}=(\frac{1}{5}|0\rangle\langle 0|)-\frac{4}{5}|1\rangle\langle 1|)\otimes \frac{1}{2}I_{2}
\label{abs1}\end{aligned}$$ The eigenvalues of $\sigma_{1}$ are given by $\frac{1}{10}, \frac{1}{10}, \frac{2}{5}, \frac{2}{5}$. It can be easily verified that the eigenvalues of $\sigma_{1}$ satisfy the condition (\[abssepcond\]). Thus the state $\sigma_{1}$ is absolutely separable state. Next, our task is to verify whether the state $\sigma_{1}$ satisfies the inequality $Tr(\sigma_{1}^{2})\leq \frac{1}{3}$. We find that $Tr(\sigma_{1}^{2})=\frac{17}{50}$ which is greater than $\frac{1}{3}$. This implies that the state $\sigma_{1}$ lies outside the ball described by $Tr(\rho^{2})\leq \frac{1}{3}$. To deal with this problem, we construct a ball for $2\otimes d$ quantum system and in particular, we have shown that the constructed ball contain almost all absolute separable states in $2\otimes 2$ dimensional Hilbert space.\
This work is organised as follows: In section-II, we derive the upper bound of $Tr(\varrho^{2})$, where the two-qubit zero discord states are described by the density operator $\varrho$. It is shown that the upper bound is state independent for certain classes of two-qubit zero discord state. These specific classes of two-qubit zero discord states satisfy the condition for the separability from spectrum. In particular, we unearth the class of two-qubit product states which are absolutely separable. Also we find that there exist states from the class of absolute separable zero discord states are not lying within and on a ball described by $Tr(\rho^{2})\leq \frac{1}{3}$. In section-III, we have constructed a ball of separable as well as absolute separable state in $2\otimes d$ dimensional system. In section-IV, we give few examples to support that the newly constructed ball is larger in size. It is evident from the fact that in $2\otimes 2$ dimensional system, it contain two-qubit absolute separable states which are lying not only inside but also outside the ball described by $Tr(\rho^{2})\leq \frac{1}{3}$. Further, we provided the example of absolute separable states in $2\otimes 3$ quantum system. In section-V, we end with concluding remarks.
Identification of a class of absolutely separable states that does not contain quantum correlation
==================================================================================================
In this section, we first derive the upper bound of $Tr(\rho_{ZD}^{2})$, where $\rho_{ZD}$ represent the zero discord state and thereby constructing a ball in which the zero discord state is lying. Then we show that there exist a class of zero discord state residing in the region within the ball, which is separable from spectrum.
Construction of a ball that contain zero discord state
------------------------------------------------------
We construct a ball of zero discord state and to accomplish this task we derive an inequality in terms of $Tr(\rho_{ZD}^{2})$, where the density matrix $\rho_{ZD}$ denoting the zero discord state lying in $2\otimes 2$ dimensional Hilbert space. In general, the derived upper bound of the inequality is state dependent but we found some particular class of zero discord state for which the upper bound is independent of the state.\
To start with, let us consider a $2\otimes 2$ dimensional zero discord state $\rho_{ZD}$ that can be expressed as [@dakic] $$\begin{aligned}
\rho_{ZD}= p|\psi\rangle\langle\psi|\otimes \rho_{1}+(1-p)|\psi_{\perp}\rangle\langle\psi_{\perp}|\otimes \rho_{2},~~0\leq p\leq 1
\label{zdstate}\end{aligned}$$ where the pure states $|\psi\rangle$ and $|\psi_{\perp}\rangle$ are orthogonal to each other i.e. $\langle\psi|\psi_{\perp}\rangle=0$. The single qubit density operator $\rho_{i} (i=1,2)$ are given by $$\begin{aligned}
\rho_{i}= \frac{1}{2}I_{2}+\vec{r}_{i}.\vec{\sigma},~~i=1,2
\label{singlequbitdensity}\end{aligned}$$ $I_{2}$ represent a $2\times 2$ identity matrix, $\vec{r}_{i}=(r_{i1},r_{i2},r_{i3})$ denote the Bloch vector and the component of $\vec{\sigma}=(\sigma_{1},\sigma_{2},\sigma_{3})$ are usual Pauli matrices.\
**Theorem-1:** A two-qubit zero discord state $\rho_{ZD}$ satisfies the inequality $$\begin{aligned}
Tr(\rho_{ZD}^{2})\leq min\{\frac{1}{2}+2|\vec{r}_{1}|^{2},\frac{1}{2}+2|\vec{r}_{2}|^{2}\}
\label{th1}\end{aligned}$$ **Proof:** Let us start with the expression of $Tr(\rho_{ZD}^{2})$, which is given by $$\begin{aligned}
Tr(\rho_{ZD}^{2})= p^{2}Tr(\rho_{1}^{2})+(1-p)^{2}Tr(\rho_{2}^{2})
\label{zdexpression}\end{aligned}$$ It can be seen that the value of $Tr(\rho_{ZD}^{2})$ is changing by varying the values of the parameter $p$ in the range $0\leq p \leq 1$, and the block vectors $\vec{r_{i}},i=1,2$ satisfying $|\vec{r_{i}}|^{2}\leq 1$. Thus one can ask for the upper bound of $Tr(\rho_{ZD}^{2})$. To probe this question, we assume that the zero discord state $\rho_{ZD}$ satisfies the inequality given by $$\begin{aligned}
Tr(\rho_{ZD}^{2})\leq \alpha(\vec{r}_{i}),~~i=1,2
\label{assumption}\end{aligned}$$ regardless of the parameter $p$, where $\alpha(\vec{r}_{i})$ denote the parameter depend on the state parameter $\vec{r}_{i},i=1,2$.\
Our task is to find $\alpha(\vec{r}_{i})$. To search for $\alpha(\vec{r}_{i})$, we need to combine (\[zdexpression\]) and (\[assumption\]). Thus, we obtain $$\begin{aligned}
&&p^{2}Tr(\rho_{1}^{2})+(1-p)^{2}Tr(\rho_{2}^{2})\leq \alpha(\vec{r}_{i}) \nonumber\\&&
\Rightarrow p^{2}(Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2}))-2pTr(\rho_{2}^{2})+\nonumber\\&&(Tr(\rho_{2}^{2})-\alpha(\vec{r}_{i}))\leq 0
\label{step1}\end{aligned}$$ Solving the inequality (\[step1\]) for the parameter $p$, we get $$\begin{aligned}
a\leq p\leq b
\label{step2a}\end{aligned}$$ where $a$ and $b$ are given by $$\begin{aligned}
a=\frac{Tr(\rho_{2}^{2})-\sqrt{\alpha(\vec{r}_{i})(Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2}))-Tr(\rho_{1}^{2})Tr(\rho_{2}^{2})}}{Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2})}
\label{step2b}\end{aligned}$$ and $$\begin{aligned}
b=\frac{Tr(\rho_{2}^{2})+\sqrt{\alpha(\vec{r}_{i})(Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2}))-Tr(\rho_{1}^{2})Tr(\rho_{2}^{2})}}{Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2})}
\label{step2c}\end{aligned}$$ We impose the condition on $a$ and $b$ in such a way so that $0\leq p\leq 1$ is satisfied. The required conditions are given below $$\begin{aligned}
a\geq 0
\label{acond1}\end{aligned}$$ $$\begin{aligned}
b\leq 1
\label{bcond2}\end{aligned}$$ The first condition (\[acond1\]) gives $$\begin{aligned}
&&\frac{Tr(\rho_{2}^{2})-\sqrt{\alpha(\vec{r}_{i})(Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2}))-Tr(\rho_{1}^{2})Tr(\rho_{2}^{2})}}{Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2})}\geq 0\nonumber\\&&
\Rightarrow \alpha(\vec{r}_{i})\leq Tr(\rho_{2}^{2})
\label{cond1}\end{aligned}$$ The second condition (\[bcond2\]) gives $$\begin{aligned}
&&\frac{Tr(\rho_{2}^{2})+\sqrt{\alpha(\vec{r}_{i})(Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2}))-Tr(\rho_{1}^{2})Tr(\rho_{2}^{2})}}{Tr(\rho_{1}^{2})+Tr(\rho_{2}^{2})}\leq 1\nonumber\\&&
\Rightarrow \alpha(\vec{r}_{i})\leq Tr(\rho_{1}^{2})
\label{cond2}\end{aligned}$$ Therefore, (\[cond1\]) and (\[cond2\]) can be expressed jointly as $$\begin{aligned}
\alpha(\vec{r}_{i}) \leq min\{Tr(\rho_{1}^{2}),Tr(\rho_{2}^{2})\}
\label{condcomp}\end{aligned}$$ Now we calculate $Tr(\rho_{i}^{2})$ by recalling (\[singlequbitdensity\]), and it is given by $$\begin{aligned}
Tr(\rho_{i}^{2})=\frac{1}{2}+2|\vec{r}_{i}|^{2}, ~~~i=1,2
\label{trrho}\end{aligned}$$ Using (\[condcomp\]) and (\[trrho\]), we get $$\begin{aligned}
\alpha(\vec{r}_{i})\leq min\{\frac{1}{2}+2|\vec{r}_{1}|^{2},\frac{1}{2}+2|\vec{r}_{2}|^{2}\}
\label{redineq}\end{aligned}$$ Combining the inequalities (\[assumption\]) and (\[redineq\]), we arrive at the required result given by $$\begin{aligned}
Tr(\rho_{ZD}^{2}) \leq min\{\frac{1}{2}+2|\vec{r}_{1}|^{2},\frac{1}{2}+2|\vec{r}_{2}|^{2}\}
\label{th1a}\end{aligned}$$ Geometrically, the inequality given by (\[th1a\]) represent a region within and on a ball containing zero discord state. From (\[th1a\]), it can be easily seen that the upper bound of $Tr(\rho_{ZD}^{2})$ depends on the local bloch vector $\vec{r}_{i}$ and hence the upper bound is state dependent. The state independent bound of $Tr(\rho_{ZD}^{2})$ can be obtained for particular classes of zero discord state and it is given in the corollary below.\
**Corollary-1:** The density operators $\rho_{ZD}^{(1)}$ and $\rho_{ZD}^{(2)}$ satisfy the inequality $$\begin{aligned}
Tr([\rho_{ZD}^{(i)}]^{2})\leq \frac{1}{2},~~i=1,2
\label{cor1}\end{aligned}$$ where $\rho_{ZD}^{(1)}$ and $\rho_{ZD}^{(2)}$ denote the particular class of zero discord state given by $$\begin{aligned}
\rho_{ZD}^{(1)}= p|\psi\rangle\langle\psi|\otimes \frac{1}{2}I_{2}+(1-p)|\psi_{\perp}\rangle\langle\psi_{\perp}|\otimes \rho_{2},\nonumber\\ \rho_{ZD}^{(2)}= p|\psi\rangle\langle\psi|\otimes \rho_{1}+(1-p)|\psi_{\perp}\rangle\langle\psi_{\perp}|\otimes \frac{1}{2}I_{2},
\label{zdstate1}\end{aligned}$$ $0\leq p\leq 1$.\
**Proof:** To prove it, consider the following two cases: (i) $|\vec{r}_{1}|^{2}\leq |\vec{r}_{2}|^{2}$, (ii) $|\vec{r}_{2}|^{2}\leq |\vec{r}_{1}|^{2}$.\
**Case-I:** If $|\vec{r}_{1}|^{2}\leq |\vec{r}_{2}|^{2}$ then theorem-1 gives $$\begin{aligned}
Tr(\rho_{ZD}^{2}) \leq \frac{1}{2}+2|\vec{r}_{1}|^{2}
\label{ineq1}\end{aligned}$$ In particular, the inequality (\[ineq1\]) holds even if we take the minimum value of the expression $\frac{1}{2}+2|\vec{r}_{1}|^{2}$ over all $\vec{r}_{1}$. Therefore, we have $$\begin{aligned}
Tr(\rho_{ZD}^{2}) \leq min_{\vec{r}_{1}}[\frac{1}{2}+2|\vec{r}_{1}|^{2}]
\label{ineq2}\end{aligned}$$ We obtain $min_{\vec{r}_{1}}[\frac{1}{2}+2|\vec{r}_{1}|^{2}]=\frac{1}{2}$ and the minimum value is attained when $\vec{r}_{1}=\vec{0}$. Thus, the minimum value is obtained when the state $\rho_{ZD}$ reduces to $\rho_{ZD}^{(1)}$. Hence the inequality (\[ineq2\]) reduces to $$\begin{aligned}
Tr([\rho_{ZD}^{(1)}]^{2})\leq \frac{1}{2}
\label{cor1a}\end{aligned}$$ **Case-II:** If $|\vec{r}_{2}|^{2}\leq |\vec{r}_{1}|^{2}$ then we can proceed in a similar way as in case-I and obtain $Tr([\rho_{ZD}^{(2)}]^{2})\leq \frac{1}{2}$.\
Therefore, we have obtained the particular classes of zero discord states described by the density operators $\rho_{ZD}^{(i)} (i=1,2)$ given in (\[zdstate1\]) satisfy the inequality $Tr([\rho_{ZD}^{(i)}]^{2})\leq \frac{1}{2} (i=1,2)$. Thus, the upper bound does not depend on the state $\rho_{ZD}^{(i)} (i=1,2)$.
Class of zero discord state which is separable from spectrum
------------------------------------------------------------
Let us consider a class of zero discord state either described by the density operator $\rho_{ZD}^{(1)}$ or $\rho_{ZD}^{(2)}$ given in (\[zdstate1\]). Recalling $\rho_{ZD}^{(1)}= p|\psi\rangle\langle\psi|\otimes \frac{1}{2}I_{2}+(1-p)|\psi_{\perp}\rangle\langle\psi_{\perp}|\otimes \rho_{2}$ with the single qubit density operator $\rho_{2}$ given by (\[singlequbitdensity\]) and a pair of orthogonal pure states $|\psi\rangle$ and $|\psi_{\perp}\rangle$, where $|\psi\rangle=\alpha|0\rangle+\beta|1\rangle$ and $|\psi_{\perp}\rangle=\beta|0\rangle-\alpha|1\rangle$. We assume that the parameters $\alpha$ and $\beta$ are real number satisfying $\alpha^{2}+\beta^{2}=1$. Therefore, the density matrix for $\rho_{ZD}^{(1)}$ is given by $$\begin{aligned}
\rho_{ZD}^{(1)}=
\begin{pmatrix}
a_{11} & a_{12} & a_{13} & a_{14} \\
a_{12}^{*} & a_{22} & a_{23} & a_{24} \\
a_{13}^{*} & a_{23}^{*} & a_{33} & a_{34} \\
a_{14}^{*} & a_{24}^{*} & a_{34}^{*} & a_{44}
\end{pmatrix}, \sum_{i=1}^{4}a_{ii}=1\end{aligned}$$ where $$\begin{aligned}
&&a_{11}=p\frac{\alpha^{2}}{2}+(1-p)\beta^{2}(\frac{1}{2}+r_{23}),\nonumber\\&& a_{12}=(1-p)\beta^{2}(r_{21}-ir_{22}),\nonumber\\&& a_{13}=p\frac{\alpha\beta}{2}-(1-p)\alpha\beta(\frac{1}{2}+r_{23}),\nonumber\\&&
a_{14}=-(1-p)\alpha\beta(r_{21}-ir_{22}), \nonumber\\&& a_{22}=p\frac{\alpha^{2}}{2}+(1-p)\beta^{2}(\frac{1}{2}-r_{23}),\nonumber\\&& a_{23}=-(1-p)\alpha\beta(r_{21}+ir_{22}),\nonumber\\&&
a_{24}=p\frac{\alpha\beta}{2}-(1-p)\alpha\beta(\frac{1}{2}-r_{23}),\nonumber\\&&
a_{33}=p\frac{\beta^{2}}{2}+(1-p)\alpha^{2}(\frac{1}{2}+r_{23}),\nonumber\\&&
a_{34}=(1-p)\alpha^{2}(r_{21}-ir_{22}),\nonumber\\&&
a_{44}=p\frac{\beta^{2}}{2}+(1-p)\alpha^{2}(\frac{1}{2}-r_{23})
\label{matrix}\end{aligned}$$ The eigenvalues of $\rho_{ZD}^{(1)}$ are given by $$\begin{aligned}
&&\lambda_{1}=\frac{1-p}{2}(1+2|\vec{r}_{2}|),\lambda_{2}=\frac{1-p}{2}(1-2|\vec{r}_{2}|)\nonumber\\&&
\lambda_{3}=\lambda_{4}=\frac{p}{2}
\label{eigval}\end{aligned}$$ The state $\rho_{ZD}^{(1)}$ satisfy the positive semi-definiteness property if $$\begin{aligned}
|\vec{r}_{2}|\leq \frac{1}{2}
\label{possem}\end{aligned}$$ Now our task reduces to the following; (i) verify whether the class of states $\rho_{ZD}^{(1)}$ satisfy the condition of separability from spectrum and (ii) if the class of states verified as absolute separable states then find out whether they lying within the ball described by $Tr([\rho_{ZD}^{(1)}]^{2})\leq\frac{1}{3}$. In this context, a table is constructed by taking different ranges of the parameter $p$ and some values of $|\vec{r}_{2}|$ for which we find that the zero discord state described by the density operator $\rho_{ZD}^{(1)}$ satisfy the inequality (\[abssepcond\]). This means that there exist classes of two-qubit zero discord states that are absolutely separable also. We call these classes of two-qubit states as Absolutely Separable Zero Discord Class $(ASZDC)$. Further, we have constructed another table which reveals the fact that whether the class of states given by $ASZDC$ satisfies the inequality $Tr([\rho_{ASZDC}]^{2})\leq\frac{1}{3}$. Without any loss of generality, we have verified the above two tasks by considering the values of the parameter $p$ in $[0,\frac{1}{2}]$ and taking few values of $|\vec{r_{2}}|$. Similar analysis can be done for other range the parameter $p\in [\frac{1}{2},1]$ and other values of $|\vec{r_{2}}|\leq \frac{1}{2}$.\
------------------- ---------------- ----------------------------------- --------------------
$Parameter$ $Parameter$ $\lambda_{1}-\lambda_{3}$ $Nature~of~state$
$(|\vec{r}_{2}|)$ (p) $-2\sqrt{\lambda_{2}\lambda_{4}}$
0 \[0, 0.15) positive Separable
0 \[0.15, 0.5\] Negative Absolute separable
0.1 \[0, 0.213) positive Separable
0.1 \[0.213, 0.5\] Negative Absolute separable
0.2 \[0, 0.291) positive Separable
0.2 \[0.291, 0.5\] Negative Absolute separable
0.3 \[0, 0.38) positive Separable
0.3 \[0.38, 0.5\] Negative Absolute separable
0.4 \[0, 0.483) positive Separable
0.4 \[0.483, 0.5\] Negative Absolute separable
0.5 \[0, 0.5\] positive Separable
------------------- ---------------- ----------------------------------- --------------------
: Table verifying whether the state $\rho_{ZD}^{(1)}$ satisfy (\[abssepcond\])
------------------- ------------------- -------------------------- ------------
$Parameter$ $Parameter$ $Tr([\rho_{ASZDC}]^{2})$ $State$
$(|\vec{r}_{2}|)$ (p) $\leq\frac{1}{3}$ $residing$
0 \[0.15, 0.211) Violated Outside
0 \[0.211, 0.5\] Satisfied Inside
0.1 \[0.213, 0.2325) Violated Outside
0.1 \[0.2325, 0.5\] Satisfied Inside
0.2 \[0.291, 0.29205) Violated Outside
0.2 \[0.29205, 0.5\] Satisfied Inside
0.3 \[0.38, 0.38056) Violated Outside
0.3 \[0.38056, 0.5\] Satisfied Inside
0.4 \[0.483, 0.49) Violated Outside
0.4 \[0.49, 0.5\] Satisfied Inside
------------------- ------------------- -------------------------- ------------
: Table shows that whether the absolute separable state given in Table-I is residing inside or outside the ball described by $Tr([\rho_{ASZDC}]^{2})\leq\frac{1}{3}$
Since the maximal ball described by $Tr([\rho_{ASZDC}]^{2})\leq\frac{1}{3}$ does not contain all states from the class ASZDC and such states lying outside the ball so we investigate in the next section that whether it is possible to increase the size of the maximal ball.
Constructing the bigger ball of separable as well as absolutely separable states around maximally mixed state
=============================================================================================================
In this section, we will show that it is possible to construct a ball which is larger than the earlier constructed ball described by $Tr(\rho^{2})\leq\frac{1}{3}$ where the state $\rho$ represent either separable or absolutely separable states around maximally mixed state. This means that there is a possibility for the new ball, constructed in this work, to contain those separable as well as absolute separable states which are lying outside the ball described by $Tr(\rho^{2})\leq\frac{1}{3}$.
A Few Definitions and Results
-----------------------------
Firstly, we recapitulate a few definitions and earlier obtained results which are required to construct a new ball.\
**Definition-1:** *p*-norm of a matrix $A$ is defined as $$\begin{aligned}
(\|A\|_{p})^{p}=Tr(A^{\dagger}A)^{\frac{p}{2}}
\label{pnorm}\end{aligned}$$ In particular for $p=2$ and $A=\rho$, where $\rho$ denoting a quantum state, we have $$\begin{aligned}
(\|\rho\|_{2})^{2}=Tr(\rho^{2})
\label{pnorm1}\end{aligned}$$ **Definition-2: [@hilderbrand]** A quantum state $\rho \in H_{2}\otimes H_{d}$ is absolutely separable if $U\rho U^{\dagger}$ remain a separable state for all global unitary operator $U \in U(2d)$.\
If we denote $\rho'=U\rho U^{\dagger}$ then it can be easily shown that $Tr[(\rho')^{2}]=Tr[(\rho)^{2}]$, i.e. $Tr[(\rho)^{2}]$ is invariant under unitary transformation.\
**Result-1 [@king]:** Let $M$ be a $2d\times 2d$ positive semi-definite matrix expressed in the block form as $$\begin{aligned}
M=
\begin{pmatrix}
A & C \\
C^{\dag} & B
\end{pmatrix}\end{aligned}$$ where $A,B,C$ are $d\times d$ matrices.\
If we define the $2\times 2$ matrix as $$\begin{aligned}
m=
\begin{pmatrix}
\|A\|_{p} & \|C\|_{p}\\
\|C\|_{p} & \|B\|_{p}
\end{pmatrix}\end{aligned}$$ then the following inequalities hold:\
(a) for $1\leq p \leq 2$,\
$$\begin{aligned}
\|M\|_{p}\geq \|m\|_{p}
\label{inequalitya}\end{aligned}$$ (b) for $2\leq p < \infty$,\
$$\begin{aligned}
\|M\|_{p}\leq \|m\|_{p}
\label{inequalityb}\end{aligned}$$ Thus for $p=2$, we have $$\begin{aligned}
\|M\|_{2}= \|m\|_{2}
\label{inequality}\end{aligned}$$ **Result-2 [@johnston]:** Let us choose $d\times d$ matrices $A,B,C$ such that $A$ and $B$ are positive semi-definite matrices. Then the block matrix $$\begin{aligned}
X=
\begin{pmatrix}
A & C \\
C^{\dag} & B
\end{pmatrix}\end{aligned}$$ is separable if $\|C\|_{2}^{2}\leq \lambda_{min}(A)\lambda_{min}(B)$, where $\lambda_{min}(A)$ and $\lambda_{min}(B)$ denoting the minimum eigenvalue of the matrices $A$ and $B$ respectively.\
Construction of a new ball that contain separable as well as absolutely separable states
----------------------------------------------------------------------------------------
Let us consider a quantum state described by the density matrix $\rho \in H_{2}\otimes H_{d}$. The density matrix can be written in the block form as $$\begin{aligned}
\rho=
\begin{pmatrix}
X & Y \\
Y^{\dagger} & Z
\end{pmatrix}\end{aligned}$$ where $X,Y,Z$ denoting $d\times d$ matrices with $X,Z\geq 0$.\
Using Result-1, we have $$\begin{aligned}
\|\rho\|_{2}=
\|\begin{pmatrix}
X & Y\\
Y^{\dagger} & Z
\end{pmatrix}\|_{2}=\|\begin{pmatrix}
\|X\|_{2} & \|Y\|_{2}\\
\|Y\|_{2} & \|Z\|_{2}
\end{pmatrix}\|_{2}\end{aligned}$$ Let us now calculate the value of $Tr(\rho^{2})$. It is given by $$\begin{aligned}
Tr(\rho^{2})&=&\|\rho\|_{2}^{2}= \|X\|_{2}^{2}+2\|Y\|_{2}^{2}+\|Z\|_{2}^{2}\nonumber\\&=&
Tr(X^{2})+2\|Y\|_{2}^{2}+Tr(Z^{2})\nonumber\\&\leq&
Tr(X^{2})+2\lambda_{min}(X)\lambda_{min}(Z)\nonumber\\&+&Tr(Z^{2})
\label{trace}\end{aligned}$$ where $\lambda_{min}(X)$ and $\lambda_{min}(Z)$ denoting the minimum eigenvalues of the block matrices $X$ and $Z$ respectively. The last inequality follows from Result-2. Therefore, the state $\rho$ is separable if $$\begin{aligned}
Tr(\rho^{2})\leq Tr(X^{2})+2\lambda_{min}(X)\lambda_{min}(Z)+Tr(Z^{2})
\label{trace1}\end{aligned}$$ The state described by the density operator $\rho$ is absolutely separable if for any global unitary transformation $U \in U(2d)$, the inequality $$\begin{aligned}
Tr[(U\rho U^{\dagger})^{2}]=Tr(\rho^{2})&\leq& Tr(X^{2})+2\lambda_{min}(X)\lambda_{min}(Z)\nonumber\\&+&Tr(Z^{2})
\label{abscond}\end{aligned}$$ holds.\
It can be observed that the upper bound of the inequality (\[abscond\]) depends on the parameter of the state under consideration. Thus the upper bound is state dependent and it can be maximized over the given range of the parameter of the state. We grasp this idea to show that there is a possibility to increase the size of the ball that contains more separable as well as absolutely separable state compared to $Tr(\rho^{2})\leq \frac{1}{3}$.
Illustrations
=============
In this section, we will show with examples that the new ball constructed in this work described by (\[abscond\]) contains more two-qubit absolutely separable states than the ball descibed by $Tr(\rho^{2})\leq \frac{1}{3}$. Also, we discuss about the absolute separable states in $2\otimes 3$ quantum system.
### Two-qubit class of sates from ASZDC
Let us consider a subclass of the two-qubit quantum state belong to ASZDC described by the density operator $\rho^{(1)}$ as $$\begin{aligned}
\rho^{(1)}&=& (p|0\rangle\langle 0|-(1-p)|1\rangle\langle 1|)\otimes \frac{1}{2}I_{2},\nonumber\\&&
0\leq p\leq 1
\label{class1a}\end{aligned}$$ where $I_{2}$ represent the identity matrix of order 2. The state $\rho^{(1)}$ is a product state and thus separable for $0\leq p\leq 1$.\
The matrix representation of $\rho^{(1)}$ is given by $$\begin{aligned}
\rho^{(1)}=
\begin{pmatrix}
X & Y \\
Y^{\dagger} & Z
\end{pmatrix}\end{aligned}$$ where $Y$ is a null matrix and the matrices $X$ and $Z$ are given by $$\begin{aligned}
&&X=
\begin{pmatrix}
\frac{p}{2}& 0 \\
0 & \frac{p}{2}
\end{pmatrix},\nonumber\\&& Z=\begin{pmatrix}
\frac{1-p}{2}& 0 \\
0 & \frac{1-p}{2}
\end{pmatrix}\end{aligned}$$ The eigenvalues of $\rho^{(1)}$ are given by $\frac{p}{2}, \frac{p}{2},\frac{1-p}{2},\frac{1-p}{2}$.\
Case-I: When the parameter $p$ is lying in the interval $[0,\frac{1}{2}]$ then the eigenvalues are arranged in descending order as $\lambda_{1}\geq \lambda_{2}\geq\lambda_{3}\geq\lambda_{4}$, where $$\begin{aligned}
\lambda_{1}=\frac{1-p}{2}, \lambda_{2}=\frac{1-p}{2},\lambda_{3}=\frac{p}{2},\lambda_{4}=\frac{p}{2}
\label{eigenvalue1}\end{aligned}$$ The state $\rho^{(1)}$ is separable from spectrum if $$\begin{aligned}
p+\sqrt{p(1-p)}\geq \frac{1}{2}
\label{eigenvalue11}\end{aligned}$$ The inequality (\[eigenvalue11\]) holds if $\frac{3}{20}\leq p\leq 1/2$. Therefore, the state $\rho^{(1)}$ is absolutely separable for $p \in [\frac{3}{20},\frac{1}{2}]$.\
Now, $Tr[(\rho^{(1)})^{2}]$ can be calculated as $$\begin{aligned}
Tr[(\rho^{(1)})^{2}]=\frac{p^{2}}{2}+\frac{(1-p)^{2}}{2}
\label{trace11}\end{aligned}$$ From Fig.1, it can be seen that there exist absolutely separable states for $p \in [\frac{3}{20}, \frac{21}{100}]$ that are lying outside the ball described by $Tr[(\rho^{(1)})^{2}]\leq\frac{1}{3}$. Thus, it is interesting to see whether the newly constructed ball contain all the absolutely separable states for $p \in [\frac{3}{20}, \frac{21}{100}]$. To probe this, we calculate the upper bound of $Tr[(\rho^{(1)})^{2}]$ using the inequality (\[abscond\]). The upper bound is given by $$\begin{aligned}
&&Tr(X^{2})+2\lambda_{min}(X)\lambda_{min}(Z)+Tr(Z^{2})\nonumber\\&&=\frac{1}{2}[1-p(1-p)]
\label{ub}\end{aligned}$$ Again, Fig.1 shows that the newly constructed ball described by $Tr[(\rho^{(1)})^{2}]\leq \frac{1}{2}[1-p(1-p)]$ contains all absolutely separable belong to the class described by the density operator $\rho^{(1)}$.
![Plot of $Tr[(\rho^{(1)})^{2}]$ versus the state parameter $p$ ](figure-1.pdf)
Case-II: In a similar fashion, the case where $p \in [\frac{1}{2},1]$ can be analyzed.
### $2\times 2$ isotropic state
Let us consider a $2\otimes 2$ isotropic state given by $$\begin{aligned}
\rho_{2\otimes 2}^{(iso)}(f)=
\begin{pmatrix}
\frac{1+2f}{6} & 0 & 0 & \frac{4f-1}{6} \\
0 & \frac{1-f}{3} & 0 & 0 \\
0 & 0 & \frac{1-f}{3} & 0 \\
\frac{4f-1}{6} & 0 & 0 & \frac{1+2f}{6}
\end{pmatrix}, 0\leq f\leq 1\end{aligned}$$ It is known that the state described by the density operator $\rho_{2\otimes 2}^{(iso)}$ is separable for $0\leq f\leq \frac{1}{2}$. Further, it can be easily verified that all separable states in the class represented by $\rho_{2\times 2}^{(iso)}$ are also absolute separable states.\
The matrix of $2\otimes 2$ isotropic state can be re-expressed in terms of block matrices of order $2\times 2$ as $$\begin{aligned}
\rho_{2\otimes 2}^{(iso)}(f)=
\begin{pmatrix}
X & Y \\
Y^{\dagger} & Z
\end{pmatrix}\end{aligned}$$ where $2\times 2$ block matrices $X,Y$ and $Z$ are given by $$\begin{aligned}
X=
\begin{pmatrix}
\frac{1+2f}{6} & 0\\
0 & \frac{1-f}{3}
\end{pmatrix}, Y=\begin{pmatrix}
0 & \frac{4f-1}{6}\\
0 & 0
\end{pmatrix}, Z=\begin{pmatrix}
\frac{1-f}{3} & 0\\
0 & \frac{1+2f}{6}
\end{pmatrix}\end{aligned}$$ The minimum eigenvalue of the block matrices $X$ and $Z$ are given by $$\begin{aligned}
\lambda_{min}(X)=\lambda_{min}(Z)&=&\frac{1+2f}{6},~~0\leq f\leq \frac{1}{4}\nonumber\\&=&
\frac{1-f}{3},~~\frac{1}{4}\leq f\leq \frac{1}{2}
\label{mineigen}\end{aligned}$$ We now discuss two cases based on different ranges of the parameter $f$.\
**Case-I:** When $0\leq f \leq \frac{1}{4}$ $$\begin{aligned}
Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq \frac{2f^{2}+1}{3}
\label{trace1}\end{aligned}$$ Since $f \in [0,\frac{1}{4}]$ so (\[trace1\]) can be re-expressed as $$\begin{aligned}
Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq Max_{0\leq f\leq\frac{1}{4}}\frac{2f^{2}+1}{3}
\label{trace11}\end{aligned}$$ Since $\frac{2f^{2}+1}{3}$ is an increasing function of the parameter $f$ so its maximum value is attained at $f=\frac{1}{4}$. Therefore, $$\begin{aligned}
Max_{0\leq f\leq\frac{1}{4}}\frac{2f^{2}+1}{3}=\frac{3}{8}
\label{max1}\end{aligned}$$ Thus, the state $\rho_{2\otimes 2}^{(iso)}(f)$ satisfies the inequality given by $$\begin{aligned}
Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq \frac{3}{8}
\label{case1}\end{aligned}$$ **Case-II:** When $\frac{1}{4}\leq f \leq \frac{1}{2}$ $$\begin{aligned}
Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq \frac{4f^{2}-4f+3}{6}
\label{trace2}\end{aligned}$$ Since $f \in [\frac{1}{4},\frac{1}{2}]$ so (\[trace2\]) can be reexpressed as $$\begin{aligned}
Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq Max_{\frac{1}{4}\leq f\leq\frac{1}{2}}\frac{4f^{2}-4f+3}{6}
\label{trace22}\end{aligned}$$ Since $\frac{4f^{2}-4f+3}{6}$ is a decreasing function of the parameter $f$ so its maximum value is attained at $f=\frac{1}{4}$. Therefore, $$\begin{aligned}
Max_{\frac{1}{4}\leq f\leq\frac{1}{2}}\frac{4f^{2}-4f+3}{6}= \frac{3}{8}
\label{max2}\end{aligned}$$ Thus, in this case also the state $\rho_{2\otimes 2}^{(iso)}(f)$ obey the inequality given by $$\begin{aligned}
Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq \frac{3}{8}
\label{case2}\end{aligned}$$ Combining the above two cases, it can be concluded that the state $\rho_{2\otimes 2}^{(iso)}(f)$ satisfy the inequality $$\begin{aligned}
Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq\frac{3}{8}, 0\leq f\leq\frac{1}{2}
\label{abssep1}\end{aligned}$$ Since $\frac{3}{8}>\frac{1}{3}$ so the new ball described by (\[abssep1\]) is bigger in size compared to the ball described by $Tr[(\rho_{2\otimes 2}^{(iso)}(f))^{2}]\leq \frac{1}{3}$ and hence the new ball contains more absolutely separable state.
### Class of states in $2\otimes 3$ quantum system
Let us consider a class of states in $2\otimes 3$ quantum system parameterized with two parameters $\alpha$ and $\gamma$, which is given by [@chi] $$\begin{aligned}
&&\rho_{\alpha,\gamma}^{2\otimes 3}=\alpha(|02\rangle\langle 02|+|12\rangle\langle 12|)+\frac{4\gamma+2\alpha-1}{3} |\psi^{-}\rangle\langle \psi^{-}|+\nonumber\\&&
\frac{1-\gamma-2\alpha}{3}(|00\rangle\langle 00|+|01\rangle\langle 01|+|10\rangle\langle 10|+|11\rangle\langle 11|),~~\nonumber\\&&
0\leq \alpha \leq \frac{1}{2},~~ 0\leq \gamma \leq 1
\label{state23}\end{aligned}$$ where $|\psi^{-}\rangle=\frac{1}{\sqrt{2}}(|01\rangle-|10\rangle)$. The state is separable if and only if $\alpha+\gamma\leq \frac{1}{2}$.\
To simplify the calculation, let us choose $\gamma=\frac{1}{3}$. For this particular case, the state $\rho_{\alpha,\frac{1}{3}}^{2\otimes 3}$ is separable if and only if $0\leq \alpha\leq \frac{1}{6}$. Therefore, with this chosen value of $\gamma$, we can re-express the state $\rho_{\alpha,\frac{1}{3}}^{2\otimes 3}$ in terms of block matrices as $$\begin{aligned}
\rho_{\alpha,\frac{1}{3}}^{2\otimes 3}=
\begin{pmatrix}
X_{1} & Y_{1} \\
Y_{1}^{\dagger} & Z_{1}
\end{pmatrix}\end{aligned}$$ where $3\times 3$ block matrices $X_{1},Y_{1}$ and $Z_{1}$ are given by $$\begin{aligned}
&&X_{1}=
\begin{pmatrix}
\frac{2-6\alpha}{9} & 0 & 0\\
0 & \frac{5-6\alpha}{18} & 0\\
0 & 0 & \alpha
\end{pmatrix}, Y_{1}=\begin{pmatrix}
0 & -\frac{1+6\alpha}{18} & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{pmatrix},\nonumber\\&& Z_{1}=\begin{pmatrix}
\frac{5-6\alpha}{18} & 0 & 0\\
0 & \frac{2-6\alpha}{9} & 0\\
0 & 0 & \alpha
\end{pmatrix}, 0\leq \alpha \leq \frac{1}{6}\end{aligned}$$ The eigenvalues of the state $\rho_{\alpha,\frac{1}{3}}^{2\otimes 3}$ arranged in descending order $(\varepsilon_{1}\geq \varepsilon_{2} \geq \varepsilon_{3}\geq \varepsilon_{4}\geq \varepsilon_{5}\geq \varepsilon_{6})$ for different ranges of $\alpha$ as\
(i) When $0\leq \alpha \leq 0.134$ $$\begin{aligned}
\varepsilon_{1}=\frac{1}{3}, \varepsilon_{2}=\varepsilon_{3}=\varepsilon_{4}=\frac{2-6\alpha}{9}, \varepsilon_{5}=\varepsilon_{6}=\alpha
\label{eigenvsl23a}\end{aligned}$$ (ii) When $0.134 \leq \alpha \leq \frac{1}{6}$ $$\begin{aligned}
\varepsilon_{1}=\frac{1}{3}, \varepsilon_{2}=\varepsilon_{3}=\alpha, \varepsilon_{4}=\varepsilon_{5}=\varepsilon_{6}=\frac{2-6\alpha}{9}
\label{eigenvsl23b}\end{aligned}$$ It can be easily verified using (\[abssepcondgen\]) that the state $\rho_{\alpha,\frac{1}{3}}^{2\otimes 3}$ represent absolute separable state for $0.019\leq \alpha \leq \frac{1}{6}$.\
Further, we find that the class of absolute separable states described by the density operator $\rho_{\alpha,\frac{1}{3}}^{2\otimes 3}$ lying within the ball described by (\[abscond\]) for $0.084\leq \alpha \leq \frac{1}{6}$. Thus, the new ball contain most of the absolute separable states but does not contain all absolute separable states.
Conclusion
==========
To summarize, we have characterize the absolute separable states in terms of quantum correlation which can be measured by quantum discord. We found an instance of absolute separable states with such negligible amount of quantum correlation that can be approximated to zero but still it is useful in quantum algorithm to solve Deutsch-Jozsa problem. Since these absolute separable states have approximately zero quantum correlation so we expect that it can be prepared in the experiment easily and not only that these states give quantum advantage over classical with respect to the running time of the algorithms. This prompted us to investigate about the structure of the class of absolute separable states with zero discord. We found the class of absolute separable zero discord state which are residing within the ball described by $Tr(\rho^{2})\leq \frac{1}{3}$. Further, we find that there exist classes of absolute separable zero discord state that falls outside the ball. To fill this gap, we have constructed a new ball that holds most of the absolute separable states lying in $2\otimes d$ dimensional Hilbert space. In particular, we have shown that the absolute separable states that lying outside the ball described by $Tr(\rho^{2})\leq \frac{1}{3}$, now residing inside the newly constructed ball. Thus, we conclude that the new ball is bigger in size and this fact is illustrated by giving few examples.
[90]{}
C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres, and W. K. Wooters, Phys. Rev. Lett. **70** 1895 (1993). C. H. Bennett, and S. J. Wiesner, Phys. Rev. Lett. **69** 2881 (1992). A. K. Pati, Phys.Rev. A **63**, 014320 (2000). N. Gisin, G. Ribordy, W. Tittel, and H. Zbinden, Rev. Mod. Phys. **74**, 145 (2002). A. Ekert, and R. Jozsa, Phil. Trans. R. Soc. Lond. A **356**, 1769 (1998). S. Lloyd, Phys. Rev. A **61**, 010301(R) (1999). J. Ahn, T. C. Weinacht, and P. H. Bucksbaum, Sci. Rep. **287**, 463 (2000). L. K. Grover, in Proceedings of the 28th Annual ACM Symposium on the Theory of Computing, Philadelphia, 1996 (ACM, New York, 1996), p. 212. D. A. Meyer, Phys. Rev. Lett. **85**, 2014 (2000). M. Nielsen and I. Chuang, Quantum Computation and Quantum Information Cambridge University Press, Cambridge, England, 2000. E. Biham, G. Brassardb, D. Kenigsberga, and T. Mor, Theo. Comp. Sci. **320**, 15 (2004). K. Zyczkowski, P. Horodecki, A. Sanpera, M. Lewenstein, Phys. Rev. A **58**, 883 (1998). S. L. Braunstein, C. M. Caves, R. Jozsa, N. Linden, S. Popescu, and R. Schack, Phys. Rev. Lett. **83**, 1054 (1999). G. Vidal, R. Terrach, Phys. Rev. A **59**, 141 (1999). K. Modi, T. Paterek, W. Son, V. Vedral, and M. Williamson, Phys. Rev. Lett. **104**, 080501 (2010). W. K. Wootters, Phys. Rev. Lett. **80**, 2245 (1998). G. Vidal, and R.F. Werner, Phys. Rev. A **65**, 032314 (2002). H. Ollivier, and W. H. Zurek, Phys. Rev. Lett. **88**, 017901 (2001). S. Luo, Phys. Rev. A **77**, 042303 (2008). M. Ali, A. R. P. Rau, and G. Alber, Phys. Rev. A **81**, 042105 (2010). A. Datta, S. T. Flammia, and C. M. Caves, Phys. Rev. A **72**, 042316 (2005). A. Datta, G. Vidal, Phys. Rev. A **75**, 042310 (2007). M. Kus and K. Zyczkowski, Phys. Rev. A **63**, 032307 (2001). F. Verstraete, K. Audenaert, and B. D. Moor, Phys. Rev. A **64**, 012316 (2001). N. Johnston, Phys. Rev. A **88**, 062330 (2013). N. Ganguly, J. Chatterjee, and A. S. Majumdar, Phys. Rev. A **89**, 052304 (2014). S. Halder, S. Mal, A. Sen(De), arxiv:1911.13145 \[quant-ph\]. D. Deutsch, R. Jozsa, Proc. Roy. Soc. London A **439**, 553 (1992). N.A. Gershenfeld, I.L. Chuang, Science **275**, 350 (1997). L. Gurvits and H. Barnum, Phys. Rev. A **66**, 062311 (2002). B. Dakic, V. vedral, and C. Brukner, Phys. Rev. Lett. **105**, 190502 (2010). R. Hilderbrand, Phys. Rev. A **76**, 052325 (2007). C. King, Commun. Math. Phys. **242**, 531 (2003). D. P. Chi and S. Lee, J. Phys. A: Math. Theor. **36**, 11503 (2003).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[We find the symmetry algebras of cosets which are generalizations of the minimal-model cosets, of the specific form $\frac{SU(N)_{k} \times SU(N)_{\ell}}{SU(N)_{k+\ell}}$. We study this coset in its free field limit, with $k,\ell \rightarrow \infty$, where it reduces to a theory of free bosons. We show that, in this limit and at large $N$, the algebra $\W^e_\infty[1]$ emerges as a sub-algebra of the coset algebra. The full coset algebra is a larger algebra than conventional $\W$-algebras, with the number of generators rising exponentially with the spin, characteristic of a stringy growth of states. We compare the coset algebra to the symmetry algebra of the large $N$ symmetric product orbifold CFT, which is known to have a stringy symmetry algebra labelled the ‘higher spin square’. We propose that the higher spin square is a sub-algebra of the symmetry algebra of our stringy coset.]{}'
author:
- Dushyant Kumar
- Menika Sharma
title: Symmetry Algebras of Stringy Cosets
---
Introduction
============
The papers [@Gaberdiel:2012] established a duality between the CFT of the coset model \[specialcoset\] , and three-dimensional higher-spin Vasiliev theory [@vas1], in the large $N,\ell$ limit. This duality is characterized by a large symmetry algebra $W_\infty[\mu]$ on the CFT side which is interpreted as the asymptotic symmetry algebra on the bulk side [@Campoleoni:2010]. The parameter $\mu=\frac{N}{\ell+N}$ is the ’t Hooft coupling on the CFT side, while on the bulk side it determines the mass of the scalar field. The $\W_\infty[\mu]$ algebra consists of generators with spin $2$ to $\infty$ with each generator having multiplicity one and the commutation relations of these generators depend on the parameter $\mu$.
Our aim is to find the symmetry algebra for the coset \[gencoset\] which is a generalization of the coset in [Eq. (\[specialcoset\])]{}. The central charge for this coset is . The coset in [Eq. (\[gencoset\])]{} has three independent parameters and two interesting limits. The first limit arises on taking $N$ and $\ell$ to infinity while holding $k$ and $\mu=\frac{N}{N+\ell}$ fixed. The central charge reduces to \~k N (1- \^2). Since, the central charge scales as $\sim N$, in this limit the coset in [Eq. (\[gencoset\])]{} is usually referred to as a vector coset model. This vector coset model was studied in Refs. [@Creutzig:2013], where a related, but different, coset model $SU(k+\ell)_N/SU(\ell)_N$ was proposed as the CFT dual to a bulk Vasiliev theory with a matrix extension.
There exist other limiting procedures which result in a central charge of the coset in [Eq. (\[gencoset\])]{} which scales as $N^2$. One way is to take $N$ and $k, \ell$ to infinity while holding $k-\ell$ and $\mu=\frac{N}{k+\ell+N}$ fixed. In this case, the central charge scales as $\sim N^2(1- \frac{1}{\mu})$. A variation of this, which is the limiting procedure we use in this paper, is to take $N$ and $k, \ell$ to infinity with $N/\ell$ set to zero and \[lambdac\] = fixed. The coset central charge is then . Since the central charge scales as a matrix model, in this limit we expect the coset in [Eq. (\[gencoset\])]{} to have a string dual and we refer to it in the text as the stringy $SU(N)$ coset. In this paper, we will study the symmetry algebra of this stringy $SU(N)$ coset in the limit where $k$ and $\ell$ go to infinity, but keep $N$ finite. Thus, we are not determining the algebra explicitly at large $N$. However, from the general behavior of the algebra at finite $N$, we can infer many of its properties at infinite $N$ which we will elaborate on in the text. In particular, this method tells us the properties of the large $N$ coset algebra at $\lambda=0$ with $\lambda$ defined as in [Eq. (\[lambdac\])]{}.
Historically, the algebra for the coset model in [Eq. (\[specialcoset\])]{} was first studied at finite $N$ before the infinite $N$ case was dealt with. The finite $N$ algebra is called $\W_N$ [@Zamolodchikov:1985] and has generators ranging from $2$ to $N$ with multiplicity one. Indeed, it has taken many years to completely understand the large $N$ limit of the $\W_N$ algebra [@Bakas:1990; @Gaberdiel:2012t; @Linshaw:2017]. The algebra of the coset in [Eq. (\[gencoset\])]{} for small $N$ and large $k,\ell$ has been studied before in Refs. [@Bouwknegt:1992wg; @Bais:1987a; @Bais:1987b], although it has only attracted a fraction of the attention that the $\W_N$ algebra has and perhaps rightly so. $\W_N$ algebras, which are extensions of the Virasoro algebra, have complicated commutation relations but a simple spectrum of fields. In contrast, the symmetry algebras of the coset theories in [Eq. (\[gencoset\])]{} have a spectrum of generators with the multiplicity climbing at a exponential rate with the spin (the algebra still has, of course, a finite number of generators at finite $N$). Unlike their $\W_N$ counterparts, these algebras belong to a class of algebras which are finitely non-freely generated [@deBoer:1993] and thus are less tractable. On the other hand, the fact that this coset model and supersymmetric generalizations have generators whose multiplicity increases with spin makes them prime candidates to be dual to string theories in AdS. It is with this motivation that we study them in this paper.
We exclusively work with the bosonic coset in [Eq. (\[gencoset\])]{}, so that we can study the symmetry algebra in its simplest form. A $\N=2$ supersymmetric generalization of the coset in [Eq. (\[gencoset\])]{} was studied in [@Gopakumar:2012]. Related work for a coset with $\N=1$ supersymmetry appears in Refs. [@Ahn:1990; @Ahn:2012]. However, a crucial distinction between our analysis and the supersymmetric cases studied is that we are working in the limit of zero coupling, with a free theory.
The coset theories that we study are similar to $SU(N)$ gauge theories in four dimensions which are known to have string duals on the $AdS_5$ background. However, string theories on $AdS_3$ are expected to be dual to a different family of CFTs: symmetric product orbifolds. In this paper, we explore the relation between the symmetry algebra of the bosonic symmetric product orbifold theory and the coset theory. To be able to do this, we explicitly write down the currents of the coset theory. For the coset in [Eq. (\[specialcoset\])]{} with level $k=1$, the currents of the $\W$-algebra correspond to Casimir operators of $SU(N)$. For the more general coset theory, currents of the $\W$-algebra can be generated from the Casimir operators by sprinkling additional derivatives on the constituent currents. We construct these currents in Section \[secCurrents\]. However, as we will see, the coset theory also has additional currents that cannot be constructed from the Casimir operators.
Information about the generators of any coset theory resides in the vacuum character of the partition function of the theory. In Section \[secGrowth\], we write down the vacuum character of the coset in [Eq. (\[gencoset\])]{}. Unlike the case $k=1$, it is not possible to formulate this character in closed form for general $k$ and $N$ and it can only be expressed in terms of string functions. We therefore resort to numerical techniques to find the generators of the algebra from its vacuum character for low values of $N$, following Ref. [@Bouwknegt:1992wg]. Later in Section \[secCurrents\], when we explicitly construct the currents for finite $N$, the calculation in Section \[secGrowth\] serves as a touchstone for our results.
This paper is organized as follows. In [Sec. \[secGrowth\]]{} we find the low lying spectrum of the symmetry algebra of the coset in [Eq. (\[gencoset\])]{} for small values of $N$ in the large $k,\ell$ limit. In [Sec. \[secCurrents\]]{} we construct the currents for this same coset for the special values $N=2$ and $N=3$. The $N=3$ case is especially important for understanding the structure of the currents at general $N$ and we present this case in some detail. In [Sec. \[secCurrents\]]{}, we also work out the relation of the coset algebra to the algebra $\W_\infty[\mu]$ and also its relation to the higher spin square. The algebra of the symmetric product orbifold at general values of $N$ is worked out in Appendix C.
Perturbative growth of states {#secGrowth}
=============================
In this section, we compute the vacuum character of the coset model in [Eq. (\[gencoset\])]{} at finite $N$, with $k,\ell \rightarrow \infty$. This computation will tell us at what rate the perturbative states of the current algebra grow with the spin. The density of states of a CFT partition function in the regime of large spin $s$ but $s< c$ determines the dual holographic theory. Since, in this paper, we are only interested in the symmetry algebra of the coset in [Eq. (\[gencoset\])]{}, we will focus on the vacuum character. We will not be able to determine the vacuum character exactly but will compute the low-lying spectrum of the symmetry algebra. We carry this out in Sec. \[growthN\] and our results appear in Table \[t:1\]. It is also of interest to compute the vacuum character at fixed $k$ at finite $N$, since this helps us to understand the nature of the symmetry algebra of our stringy coset. We do this in Sec. \[growthk\] and the results appear in Tables \[t:2\] and \[t:3\]. In Sec. \[asymptoticN\], we determine the asymptotic growth of states of the vacuum character. In the following, we describe the method we use to compute the vacuum character.
It is well known that the $\W$-algebra of the coset $\mathfrak{g}_k/\mathfrak{g}$ is the same as the $\W$-algebra of the coset $(\mathfrak{g}_k \oplus \mathfrak{g}_\ell)/\mathfrak{g}_{k+\ell}$ in the $\ell \rightarrow \infty$ limit. Therefore, to find the symmetry algebra for the coset as $\ell \rightarrow \infty$, we find the algebra of the coset model \[coset2\] . We will use the so-called “character technique”. To find the coset symmetry algebra we need to look at the vacuum character, so we work out the branching function $b_\l^\L(q) $ for the weights: $\L=(k,0,\cdots)$ and $\l=(0,0,\cdots)$. This will give us a series in the variable $q$. We can rearrange this series as \[fvac\] (1- j q\^n ), where F\_[s]{} \_[k=s]{}\^ (1-q\^k). [Eq. (\[fvac\])]{} is the general form of the vacuum character for an algebra with fields of spins $s_i$, where $i$ ranges from $1$ to $l$ and with $j$ null states starting at order $n$. Here, $l$ is the total number of generators of the algebra. Note that the character-technique is not fool-proof. The actual algebra may have additional currents, since we can always add currents to the denominator of [Eq. (\[fvac\])]{}, while at the same time increasing the number of null states to keep the vacuum character unchanged. Nevertheless, studying the vacuum character gives us a good indication of the nature of the algebra. We also restrict our attention to the vacuum character and ignore any extensions of the coset algebra at specific values of the level $k$.
The branching function for the coset in [Eq. (\[coset2\])]{} is given by \[branching1\] b\_ł\^Ł(q) \_[L\_[Ł,ł]{}]{} q\^[L\_0 -c/24]{} = \_[wW]{} (w) c\^Ł\_[w(ł+)-+ kŁ\_0]{}(q) q\^[ |w(ł+)-|\^2]{}, where the $c^\L_\l(q)$ are the Kac-Peterson string functions defined in [Eq. (\[stringfunction\])]{}.
Algebra for small $N$ {#growthN}
-----------------------
We now take $k$ to $\infty$. Then the branching function in [Eq. (\[branching1\])]{} reduces to \[branching2\] b\_ł\^Ł(q) = \_[wW]{} (w) c\^Ł\_[w(ł+)-+ kŁ\_0]{}(q). The string functions are given by \[stringfunction\] c\^Ł\_ł(q) = [q\^[-c/24]{}(1-q\^n)\^[[dim]{}]{} ]{} \_[w]{} (w) q\^[h\_[w\*Ł,ł]{} ]{} \_[ { n\_| \_[\_+]{} n\_= w\*Ł-ł} ]{} (\_[\_+]{} \_[n\_]{}(q) ), where we have introduced \_n = \_[m0]{} (-1)\^m q\^[m(m+1) +nm]{},\_[-n]{}(q) = q\^n \_n(q), h\_[Ł,ł]{} = [ (Ł,Ł+2) 2(k+h\^V)]{} - [(ł,ł)2k]{}, and $w * \L = w(\L+ \rho) - \rho$. Here $\Delta_+$ denotes the set of positive roots of $\mathfrak{g}$. In the large $k$ limit, the sum in [Eq. (\[stringfunction\])]{} over the affine Weyl group elements will reduce to a sum over the finite Weyl group elements. Thus the expression in [Eq. (\[stringfunction\])]{} simplifies to \[stringfunction1\] c\^Ł\_ł(q) = [q\^[-c/24]{}(1-q\^n)\^[[dim]{}]{} ]{} \_[wW]{} (w) \_[ { n\_| \_[\_+]{} n\_= w\*Ł-ł} ]{} (\_[\_+]{} \_[n\_]{}(q) ).
### Algebra for stringy $SU(2)$ coset
In this case, there are two Weyl group elements: $1$ and $w_{\alpha_1}$. Thus, the branching function in [Eq. (\[branching2\])]{} becomes b\_[(0)]{}\^[(k,0)]{}(q) =c\^[(k,0)]{}\_[(k,0)]{}(q) - c\^[(k,0)]{}\_[(k-2,2)]{}(q) . In the limit $k\rightarrow \infty$, the string functions appearing in the RHS of the above expression are given by c\^[(k,0)]{}\_[(k,0)]{}(q) = [q\^[-c/24]{} (1-q\^n)\^[3]{} ]{} {\_0(q) - \_[-1]{}(q) } and c\^[(k,0)]{}\_[(k-2,2)]{}(q) = [q\^[-c/24]{}(1-q\^n)\^[3]{} ]{} {\_[-1]{}(q) - \_[-2]{}(q) }. These expressions can be read off from [Eq. (\[stringfunction1\])]{}. Working out (and rearranging) the branching function we get b\_[(0)]{}\^[(k,0)]{}(q) = (1-q\^[13]{} -3 q\^[14]{} -7 q\^[15]{}- ) Thus the coset $\frac{SU(2)_k}{SU(2)}$ has a symmetry algebra with generators of spin \[SU2currents\] 2,4,6\^2,8\^2,9,10\^2, 12 in the large $k$ limit.
### Algebra for stringy $SU(3)$ coset
The branching function in [Eq. (\[branching2\])]{} for the case of the coset $\frac{SU(3)_k}{SU(3)}$ becomes b\_[(0,0)]{}\^[(k,0,0)]{}(q) =c\^[(k,0,0)]{}\_[(k,0,0)]{}(q) - 2 c\^[(k,0,0)]{}\_[(k-2,1,1)]{}(q) + c\^[(k,0,0)]{}\_[(k-3,3,0)]{}(q)+ c\^[(k,0,0)]{}\_[(k-3,0,3)]{}(q) - c\^[(k,0,0)]{}\_[(k-4,2,2)]{}(q) . Defining (q)= q\^[-c/24]{} , and calculating the string functions in the limit $k\rightarrow \infty$ we find $$\begin{aligned}
c^{(k,0,0)}_{(k,0,0)}(q) &=\zeta(q) \sum_{n\in\mathbb{Z}}\phi_{-n}(q) \big\{ 2\phi_{n-1}(q)\phi_{n-2}(q)-2\phi_{n-1}(q)\phi_{n}(q)-\phi_{n-2}(q)^2+\phi_{n}(q)^2 \big\}\,,\nonumber \\
c^{(k,0,0)}_{(k-2,1,1)}(q) &=\zeta(q) \sum_{n\in\mathbb{Z}}\phi_{-n}(q) \big\{ 2\phi_{n-2}(q)\phi_{n-3}(q)-2\phi_{n-1}(q)\phi_{n-2}(q)-\phi_{n-3}(q)^2+\phi_{n-1}(q)^2 \big\}\,,\nonumber\\
c^{(k,0,0)}_{(k-3,3,0)}(q) & = \zeta(q)\sum_{n\in\mathbb{Z}}\phi_{-n}(q) \big\{\phi_{n-4}(q)\phi_{n-2}(q)-\phi_{n-4}(q)\phi_{n-3}(q)-\phi_{n-3}(q)\phi_{n-1}(q) \nonumber\\[-3\jot] &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~+\phi_{n-3}(q)^2+ \phi_{n-2}(q)\phi_{n-1}(q)-\phi_{n-2}(q)^2 \big\}\,,\nonumber\\
\nonumber\\
c^{(k,0,0)}_{(k-4,2,2)}(q) &= \zeta(q) \sum_{n\in\mathbb{Z}}\phi_{-n}(q) \big\{2\phi_{n-3}(q)\phi_{n-4}(q)-2\phi_{n-2}(q)\phi_{n-3}(q)-\phi_{n-4}(q)^2+\phi_{n-2}(q)^2 \big\}\,.\nonumber\\
$$ In addition, c\^[(k,0,0)]{}\_[(k-3,0,3)]{}(q) = c\^[(k,0,0)]{}\_[(k-3,3,0)]{}(q).
$N$ Vacuum Character Algebra
----- ----------------------------------------------------------------------- ---------------------------------------------
$2$ $1 + q^2 + q^3 + 3 q^4 + 3 q^5 + 8 q^6 + 9 q^7 + 19 q^8 + \cdots$ $2, 4, 6^2,8^2,\cdots\,.$
$3$ $1 + q^2 + 2 q^3 + 4 q^4 + 6 q^5 + 15 q^6 + 22 q^7 + 46 q^8 + \cdots$ $2, 3, 4, 5, 6^4, 7^2, 8^7,\cdots$
$4$ $1 + q^2 + 2 q^3 + 5 q^4 + 7 q^5 + 18 q^6 + 29 q^7 + 64 q^8 + \cdots$ $2, 3, 4^2, 5, 6^5, 7^4 , 8^{12},\cdots $
$5$ $1 + q^2 + 2 q^3 + 5 q^4 + 8 q^5 + 19 q^6 + 32 q^7 + 71 q^8 + \cdots$ $2, 3, 4^2, 5^2, 6^5, 7^5, 8^{14},\cdots $
$6$ $1 + q^2 + 2 q^3 + 5 q^4 + 8 q^5 + 20 q^6 + 33 q^7 + 74 q^8+ \cdots$ $2, 3, 4^2, 5^2, 6^6, 7^5, 8^{15},\cdots $
$7$ $1 + q^2 + 2 q^3 + 5 q^4 + 8 q^5 + 20 q^6 + 34 q^7 + 75 q^8+ \cdots$ $2, 3, 4^2, 5^2, 6^6, 7^6, 8^{15},\cdots $
: \[t:1\]The vacuum character for the stringy coset model for small values of $N$. The corresponding algebra appears in the third column. The central charge of the coset is related to $N$ by $c=N^2-1$. Note that the vacuum character (and hence the algebra) stabilizes till order $q^N$: which means that the generators up to spin $N$ do not change on further increasing $N$.
Working out the branching function we get \[n3char\] b\_[(0)]{}\^[(k,0)]{}(q) = (1 - 24 q\^[17]{} -137 q\^[18]{}- 404 q\^[19]{}-). The spin of the generators of the algebra and their multiplicity can now be read off from the denominator. Our results for the $N=2$ and $N=3$ cases agree with those in Ref. [@Bouwknegt:1992wg].
We can work out the vacuum character for the coset $\frac{SU(4)_k}{SU(4)}$, in a similar fashion, and we find that the algebra has generators of spin: \[VCSU4\] 2, 3, 4\^2, 5, 6\^5, 7\^4, 8\^[12]{}, 9\^[15]{}, 10\^[28]{}, 11\^[41]{}, 12\^[75]{}, 13\^[103]{}, 14\^[166]{}, 15\^[235]{}, 16\^[313]{}, 17\^[362]{}, 18\^[310]{}. Using Mathematica, we have worked out the symmetry algebra for the coset $\frac{SU(N)_k}{SU(N)}$ in the large $k$ limit till $N=7$ and up to generators of spin $8$. The results appear in Table \[t:1\]. As we can see from the table, the currents up to spin $N$ stop changing as $N$ is further increased. We, therefore, expect the algebra at $N=\infty$ to have the following low-lying spectrum of generators: \[VCSUN\] 2,3,4\^2,5\^2,6\^6,7\^6,.
Algebra for cosets with finite $k$ {#growthk}
-----------------------------------
To better understand the coset algebra in the infinite $k, \ell$ limit, it is instructive to find the algebra of the coset when $\ell$ is large but $k$ is fixed to a given value. The vacuum character for such a coset is given by [Eq. (\[branching1\])]{} and the string functions continue to be given by [Eq. (\[stringfunction\])]{}. The string functions for a fixed level can be calculated in Mathematica using the package affine.m [@Nazarov:2011mv].
In Table \[t:2\] we list the algebra for various $N$ for $k=3$. As can be seen, even for this low value of the level, the number of generators of the algebra grow quickly with the spin. In fact, for a fixed value of $N$, the coset algebra stabilizes for a small value of the level $k$ — that is the coset generators do not change after a certain level. This fact was earlier reported in [@Blumenhagen:1991]. As we show in Table \[t:3\], for $N=3$ the algebra has stabilized at level $8$. The field content at this level is identical to the field content at level $k=\infty$ calculated in [Eq. (\[n3char\])]{}. Note that the number of null states continue to change and specifically decrease as we increase the level to infinity. The null states, however, never disappear from the spectrum and are present even in the infinite level limit.
Note that the growth rate of currents at finite $k$ is sharper than what might expect from the $T$-dual coset $SU(k+\ell)_N/(SU(k)_N \times SU(\ell)_N )$. The symmetry algebra of this dual coset is expected to be a (truncation of) the matrix extension of $\W_N$ in the large $\ell, N$ limit for fixed $k$ — so the multiplicity for a given spin would be at most $k^2$. However, it is possible for dual cosets to have different symmetry algebras [@Bowcock:1988].
Asymptotic growth of vacuum character {#asymptoticN}
--------------------------------------
The asymptotic growth of the vacuum character can be determined from the general formula for the asymptotic behaviour of branching functions [@Kac:1988]. Let us write the general branching function as b\_ł\^Ł(q) = \_s a\_s q\^s. Then, asymptotically as $s\rightarrow \infty$, a\_s \~ (c/6)\^[1/4]{} b(Ł,ł) s\^[-]{} ${\pi \sqrt{2/3\, c\, s}}$ where, $ b(\L,\l)$ is a positive real number and $c$ is the central charge. Thus, in our case: a\_s \~s\^[-]{} $ \pi\sqrt{{\tfrac{2}{3}N^2 s}}$ where we have dropped the constants.
In spite of the fact that there is an exponential increase in the number of currents, for small spin $s$, a large number of null states occur as $s$ becomes greater than $N^2$ for the vacuum branching function, as we saw in the previous section. Hence, the vacuum character has Cardy growth at large $s$. Note that this asymptotic behaviour holds only for finite $N$ and may change for infinite $N$.
Generalized Casimir Current Algebra {#secCurrents}
===================================
In this section, we write down the explicit form of the currents for the stringy coset in [Eq. (\[coset2\])]{}. This coset and the associated current algebra have been studied in Refs. [@Bouwknegt:1992wg; @Bais:1987a]. We review the known facts about the current algebra for this coset at arbitrary level $k$ in [Sec. \[sec:generalk\]]{}. We are interested in the current algebra in the limit of large level $k$. We show in [Sec. \[sec:largek\]]{} that in this limit, the coset theory reduces to a theory of free bosons. We also demonstrate that the number of currents grows with spin as expected from the vacuum character calculation in the previous section. We do this for the $N=2$ and $N=3$ cases in Secs. \[sec:su2\] and \[sec:su3\] respectively. Extrapolating from these results, we write down the general form of the current algebra generators in the large $N$ limit in [Sec. \[sec:largeN\]]{}. We will identify the simplest of these generators with the free field realization of the $\W^e_\infty[1]$ algebra. The algebra $\W^e_\infty[\mu]$ is an infinite-dimensional sub-algebra of $\W_\infty[\mu]$, which consists of fields of even spin only [@Candu]. In [Sec. \[sec:Relation\]]{}, we will show that a subset of the generators of our stringy coset can be arranged in representations of $\W^e_\infty[\mu]$. This is evocative of the higher spin square which is the symmetry algebra of the large $N$ symmetric product orbifold theory [@Gaberdiel:2015]. We will remark on this correspondence between the generators of the higher spin square and the coset theory in [Sec. \[sec:Relation\]]{}.
Here, we first establish background facts that we need to determine the currents for the coset $SU(N)_k/SU(N)$. For any coset algebra $G/H$, the generators are the currents of $G$ that commute with that of $H$. The generators of our coset algebra are composed from the $SU(N)_k$ generators: $J^a$, where $a$ varies from $1$ to $N^2-1$. These affine algebra generators, that transform in the adjoint representation of $SU(N)$, satisfy the following operator product expansion: \[Jope\] J\^a(z)J\^b(w) = + + . Repeated indices will always imply summation, regardless of the placement of the indices. The generators of the coset algebra are those operators of $SU(N)_k$ that commute with the $SU(N)$ currents, which are given by the zero modes of the affine currents. Thus, if $Q(z)$ is a generator of the coset algebra, then $$J^0_{a}, Q(z)$$=0. As is shown in Sec. 7.2.1 of Ref. [@Bouwknegt:1992wg], this implies that $Q(z)$ must be a differential polynomial invariant in the $SU(N)$ currents. The first such invariant is the stress-energy tensor T(z)\~ [::]{} which is the quadratic Casimir of $SU(N)$ defined up to an overall normalization and $:\cdots:$ symbol denotes normal ordering. It is a quasi-primary field of conformal dimension two. We will find that the coset algebra currents, in general, take a simple form in the quasi-primary basis. A quasi-primary field is defined as having the following commutator with the Virasoro modes of the stress-energy tensor \[qpdef\] $$L_m, Q_n(z)$$ ={ n- (d-1)m}Q\_[n+m]{} where $m\in\{-1,0,1\}$. Here, $d$ is the conformal dimension of the field. A primary field on the other hand obeys [Eq. (\[qpdef\])]{} for all mode numbers $m$.
As is well-known, the Casimir invariants are independent symmetric polynomial invariants of $SU(N)$. In general, the number of polynomial [*differential*]{} invariants for a group, which is the set of possible currents for the coset CFT, is much larger.
General $k$ and $k=1$ {#sec:generalk}
---------------------
At general level $k$ the stress energy tensor is given by \[Tdef\] T=[::]{} The coset currents for spin $3$ and $4$ for general level $k$ have been written down in Ref. [[@Bais:1987a]]{}, which we now review. At any level $k$ and for any $N$ there is always a single spin $3$ current of the form \[Q3def\] Q\_3 = d\_[abc]{} [::]{}. Here, $\alpha$ is a normalization factor that is given up to a constant by \^2 = , and $d_{abc}$ is the third-order invariant symmetric tensor for $SU(N)$. The above operator is, therefore, proportional to the third order Casimir of $SU(N)$. The normal ordering for three fields is defined as [::]{} = [::]{} and in a similar manner for operators consisting of more fields. Two primary spin $4$ currents were found for general $k$. The first field is \[Q41\] {[::]{} - 3 [::]{}} + [::]{} + [::]{}, where $\beta$ and $\gamma$ are numerical factors dependent on $N$ and $k$. The second field is given by \[Q4\] Q\_4 = (k+N) d\_[abcd]{} [::]{} , where $d_{abcd}$ is the fourth-order invariant symmetric traceless tensor of $SU(N)$. In general, for $SU(N)$, there are $N-1$ primitive $d$-tensors of order $2,\cdots,N$. Each of these corresponds to a current. As we will show below, however, there are other $SU(N)$ tensor invariants relevant to constructing coset currents.
The spin $4$ field in [Eq. (\[Q4\])]{} occurs in the OPE of $Q_3(z)$ and $Q_3(w)$: $$\begin{aligned}
\label{Q3ope}
Q_3(z)Q_3(w) =& \frac{c/3}{(z-w)^6} + \frac{2T(w)}{(z-w)^4} + \frac{\partial T(w)}{(z-w)^3} \nn \\
& + \frac{1}{(z-w)^2} \bigg\{\frac{32}{22+5c} {:\mathrel{\mkern2mu T(w)T(w) \mkern2mu}:} + \frac{3(c-2)}{10c+44}\partial^2 T(w) + Q_4(w) \bigg\} \nn\\
&+ \frac{1}{z-w} \bigg\{\frac{16}{22+5c} {:\mathrel{\mkern2mu \partial(T(w)T(w)) \mkern2mu}:} + \frac{c-6}{15c+66}\partial^3 T(w) + \frac{1}{2} \partial Q_4(w) \bigg\} + \cdots\end{aligned}$$
A well-known result is that for the coset $SU(N)_1/SU(N)$, primary fields with spin higher than $N$ are either null or vanish and that there is only a single field at a given spin, so that the algebra becomes identical to the $W_N$ algebra. Let us show this for $N=3$, for the spin $4$ operators. The field in [Eq. (\[Q4\])]{} vanishes for $SU(3)$ as the tensor $d_{abcd}$ collapses to zero for $N<4$. The second spin $4$ field in [Eq. (\[Q41\])]{} can be written as \[Q34\] Q\^[N=3]{}\_[4]{} = [::]{}- [::]{}- \^2 T . While in Ref. [[@Bais:1987a]]{} it was shown that the field in [Eq. (\[Q34\])]{} vanishes upon using an explicit realization of the Kac-Moody algebra in terms of free boson vertex operators (so that the Sugwara construction for $\W_N$ maps to the free field Miura realization), we can also directly compute the two-point function for $Q^{N=3}_{4}(z)$. This is given by , omitting overall numerical coefficients. The only non-zero integer value for which this vanishes and the primary field $Q^{N=3}_{4}(z)$ becomes null is $k=1$.
The limit $k\rightarrow \infty$ {#sec:largek}
-------------------------------
In this section, we will construct the coset currents in the large $k$ limit, which is the main objective of this paper. We will work at finite $N$ and then extrapolate our results to large $N$. The coset currents for finite $k$ explicitly depend on $k$. To remove this dependence, we will redefine the $SU(N)_k$ generators as follows J\^a J\^a . As a result of this redefinition the stress-energy tensor in [Eq. (\[Tdef\])]{} becomes \[Tinfdef\] T=- J\^a J\^a in terms of the new generators. Similarly the spin $3$ current in [Eq. (\[Q3def\])]{} (and other higher spin currents) become $k$-independent under this redefinition. The OPE in [Eq. (\[Jope\])]{} between the Kac-Moody currents becomes J\^a(z)J\^b(w) = , since the single-pole term is now suppressed by $\sqrt{k}$. The currents therefore, become essentially free in the large $k$ limit and the theory behaves like a theory of $N^2-1$ free bosons.
To maintain continuity with finite $k$, we will continue to take the coset currents to be $SU(N)$ invariants. The behavior of the theory in the large $k$ limit can also be motivated as follows. The $SU(N)_k$ current algebra can be written in terms of $N-1$ bosons and so-called “generating” parafermions [@Gepner:1987]. The bosonic fields are denoted by $\phi_i$, where $1\leq i \leq N-1$ and expressed as a vector $\boldsymbol{\phi}$. The parafermions $\psi_{\a}$ are fractional spin fields associated with the root lattice of $SU(N)$. Thus, here $\a$ labels the roots of $SU(N)$. The conformal dimension of these fields is given by (\_) = 1- = 1- . where we have used the normalization $\a^2=2$. In terms of these parafermions $\psi_{\a}$ and the bosonic field $\boldsymbol{\phi}$ the $SU(N)_k$ generators take the form \[ve1\] J\_i(z) \~ \_i \_z when $J_i(z) $ belongs to the Cartan sub-algebra and the form \[ve2\] J\_(z) \~ \_ for the rest of the generators corresponding to the $N^2-N$ roots of $SU(N)$. For level $k=1$, the parafermions $\psi_{\a}$ have vanishing dimension and the generators reduce to the usual vertex operator representation of the current algebra in terms of $N-1$ bosonic fields. As $k\rightarrow \infty$, the parafermions are promoted to bosons (of spin one). The $\exp$ term in [Eq. (\[ve2\])]{} reduces to one and the form of the generators $J_{\a}(z)$ become similar to $J_i(z)$. This is often referred to in the literature [@Bakas:1990] as flattening of the $SU(N)_k$ algebra in the large level limit to a $U(1)^{N^2-1}$ algebra.
In the following, we will write down the generators of the coset algebra in the infinite level limit for the cases $N=2$ and $N=3$. We will denote a quasi-primary field of spin $s$ by $Q_s$ and a primary field by $P_s$. The associated $N$ value should be clear from context. We will always define the fields up to an overall normalization. To find the quasi-primary and primary currents, we have used the Mathematica package OPEdefs [@Thielemans:1991].
### $N=2$ {#sec:su2}
The $N=2$ case was studied in Ref. [@deBoer:1993]. A set of classical currents for the stringy $SU(2)$ coset can be obtained by acting on the Casimir invariant $\Tr(J J)$ by derivatives. These currents are of the form \[bilinear\] (\^J \^J). In addition to the Casimir invariant, $SU(2)$ has cubic invariants given by \[trilinear\] (\[\^J ,\^J\]\^J) .
We can count the number of these invariants. The number of bilinear terms is the number of ways one can divide an integer into exactly two parts. The number of trilinear terms is given by the generating series for the number of ways to divide an integer into three [*distinct*]{} parts. The number of ways to divide an integer into $p$ distinct parts is given by the generating function: \[kdistinct\] . To get the independent terms we remove the total derivatives. Then the generating function for the classical currents is given by + = q\^2+ q\^4+ 2 q\^6 + 2 q\^8 + q\^9 + 2 q\^[10]{} + q\^[11]{} + 2 q\^[12]{} + . To find the quantum algebra of the coset, we have to find the primary completion of the classical currents. Relations between the invariant tensors of $SU(2)$ will make some of these currents vanish. This together with the presence of null states will truncate the set of infinite currents to the finite set listed in [Eq. (\[SU2currents\])]{}, derived from the vacuum character. In fact, it can be shown that the first vanishing state arises at the order at which a syzygy (a relation between the invariants) is present. For more details regarding this, the reader is referred to [@deBoer:1993].
We now write down the explicit form of the currents up to an overall normalization. The stress energy tensor of the coset is given by [Eq. (\[Tinfdef\])]{}. There is no spin $3$ current. We have a single primary field of dimension $4$ which is of the form \[n2p4\] P\_4 = (:\^2 J\^a J\^a: - :J\^a J\^a: )- :T T: + \^2 T . Note that the first term in brackets is a quasi-primary field of dimension $4$ while the rest are correction terms to make the field primary. Next we have two primary fields of spin $6$. The first is of the form in [Eq. (\[bilinear\])]{} and is given by $$\begin{aligned}
\label{n2p61}
P_{6,1} = {:\mathrel{\mkern2mu \partial^3 J^a \partial J^a \mkern2mu}:} &-{:\mathrel{\mkern2mu \partial^2 J^a \partial^2 J^a \mkern2mu}:} - \tfrac{1}{10}{:\mathrel{\mkern2mu \partial^4 J^a J^a \mkern2mu}:}\ \nn \\
&+ ~\alpha {:\mathrel{\mkern2mu TTT \mkern2mu}:} + ~\beta{:\mathrel{\mkern2mu \partial^2T T \mkern2mu}:} +~ \gamma {:\mathrel{\mkern2mu \partial T \partial T \mkern2mu}:} +~ \delta {:\mathrel{\mkern2mu Q_4 T \mkern2mu}:} +~\epsilon \,\partial^2 Q_4+ \zeta\, \partial^4 T \,.\end{aligned}$$ The term in the first line of the RHS is the associated quasi-primary field. The coefficients $\a,\b,\cdots$ are given in Appendix B. The second field is of the form [Eq. (\[trilinear\])]{} and is given by \[prim3\] P\_[6,2]{} = \_[abc]{} [::]{} where $\epsilon_{abc}$ is the Levi-Civita tensor. The expression as written above is already a primary field of dimension $6$ and does not need any correction terms. Continuing in this manner, one can write down all operators of the algebra. As stated above, the primary fields start becoming null from spin $11$ and the algebra thus consists of a finite number of fields.
### $N=3$ {#sec:su3}
For the stringy $SU(3)$ coset, the classical currents related to the Casimir invariants are of the form $$\begin{aligned}
\label{su3casimirs}
&\Tr(\partial^\mu J\partial^\nu J) \equiv \partial^\mu J^a \partial^\nu J^a \,,\nn\\
&\Tr(\{\partial^\mu J, \partial^\nu J\} \partial^\gamma J) \equiv d_{abc} \partial^\mu J^a \partial^\nu J^b \partial^\gamma J^c \,.\end{aligned}$$ Below we list all the currents up to spin $6$. The lowest spin operators having the structure in [Eq. (\[su3casimirs\])]{} are the stress-energy tensor and the spin $3$ Casimir current: P\_3 = d\_[abc]{} [::]{}. The lowest spin currents with additional derivatives are the spin $4$ current of the form \[n3q4\] Q\_4= [::]{} - [::]{} and the spin $5$ current of the form Q\_5 = d\_[abc]{} ( [::]{} - [::]{} ) . At the next level there are two additional spin $6$ currents of the form \[b6\] Q\_[6,1]{} =[::]{} -[::]{} - [::]{} and \[s6\] Q\_[6,2]{} = d\_[abc]{} ( [::]{} -6 [::]{} + 6[::]{} ) . The currents as written are all quasi-primary, save for the spin $3$ current $P_3$ which is primary. The primary completion of these fields appear in Appendix B.
Apart from the Casimir invariants, there are other invariants as well for $SU(3)$. The tensor $f_{abc}$ is a skewsymmetric invariant and it leads to the currents of the form f\_[abc]{} \^J\^a \^J\^b \^J\^c . Indeed the first such current is \[as6\] P\_[6,3]{} = f\_[abc]{}[::]{}, which is a primary field of dimension $6$. This accounts for three of the four spin $6$ currents predicted by the vacuum character.
To write the final spin $6$ current it is useful to look at the primary fields in the theory. Primary fields of the coset CFT can be divided into two categories: those that are $SU(3)$ singlets and those that are not. The vacuum character contains information about primary fields that are also $SU(3)$ singlets. Taking the vacuum character in [Eq. (\[n3char\])]{} and expanding in Virasoro characters gives \[VirExp\] (1-q)V\_0(q) + V\_2(q) + V\_3(q) + V\_4(q) + V\_5(q) + 5 V\_6(q) + 3 V\_7(q) + 11 V\_8(q) + . where $V_h(q)$ is the character of the Virasoro algebra Verma module V\_h(q) = q\^h \_[j=1]{}\^ . and we have assumed that the Virasoro characters are irreducible. From [Eq. (\[VirExp\])]{}, we can see that there should be a total of five primary fields of spin $6$, that are also $SU(3)$ singlets. These include the fields that are composite operators. Composite operators can also be divided into two categories: those that are composed of $SU(3)$ singlets and those that are composed of fields transforming non-trivially under $SU(3)$. An example of the first kind of operator is the following composite spin $6$ operator: P\_[6,4]{}=[::]{}+ [::]{} + [::]{}+ [::]{} + [::]{} + \^2 Q\_4 + \^4 T. This is a primary field if the coefficients $\a,\b,\gamma,\delta,\epsilon,\zeta$ take the values = , = - , =, =, =, =-.
To find the other $SU(3)$ invariants we look for primary fields that are not $SU(3)$ singlets. A spin $2$ field of this nature was introduced in Ref. [@Bais:1987a], and takes the form P\^a\_2 = d\^[a]{}\_[bc]{}[::]{}. This is not the only possible primary field transforming in the $SU(3)$ adjoint rep. Fields of the schematic form $d_{abc} {:\mathrel{\mkern2mu \partial^\mu J^a \partial^\nu J^b \mkern2mu}:}$ can be potential primaries. The operator P\^a\_4 = d\^[a]{}\_[bc]{}$ {:\mathrel{\mkern2mu \partial^2 J^b J^c \mkern2mu}:} - \tfrac{3}{2} {:\mathrel{\mkern2mu \partial J^b \partial J^c \mkern2mu}:}$ + [::]{} - [::]{} . is a (non-null) spin $4$ primary. In fact, we can generate primaries from the skew-symmetric tensor invariant $f_{abc}$ in the same manner. The field P\^a\_3 = f\^[a]{}\_[bc]{}[::]{} is a primary operator of dimension three.
We can construct new $SU(3)$ singlets from such primary fields. It will not, however, always be the case that the operators generated in this way are distinct to the singlets already constructed or are not null. The composite spin $4$ primary [::]{} - \^2 T - [::]{} is the same as the primary completion of the spin $4$ field in [Eq. (\[n3q4\])]{}. However, the composite field \[p65\] P\_[6,5]{} =[::]{} +[::]{} - [::]{} + [::]{} + [::]{} + \^4 T - \^2 Q\_4 is a new spin $6$ primary field. It can be easily verified, using Mathematica, that the operators $P_{6,1}, P_{6,2}, P_{6,3}, P_{6,4}$ and $P_{6,5}$ are linearly independent. Note that we can construct another spin $6$ singlet of the form ${:\mathrel{\mkern2mu P_3^a P_3^a \mkern2mu}:}$, but this operator is a linear combination of $P_{6,1}, P_{6,2}$ and $P_{6,5}$.
The primary field in [Eq. (\[p65\])]{} is a composite field as far as the CFT is concerned, but is an independent $SU(N)$ invariant. It will, therefore, be counted by the vacuum character of the coset CFT. The spin $6$ fields that contribute to the vacuum character are thus: $P_{6,1}, P_{6,2}, P_{6,3}$ and $P_{6,5}$. Note that the field $P_{6,4}$ is a “double-trace” $SU(N)$ invariant and its contribution to the vacuum character has already been accounted for by $P_3$. The number of independent currents at spin $6$ is thus four as predicted by the vacuum character. We see that unlike for the $N=2$ case, the $N=3$ stringy coset also needs composite currents to generate the full set of currents. Of course, the currents $T,P_4$ etc are also composites of the primary field $J^a$, but we use the word composite here to mean operators that are composites of primary fields that are themselves composite in $J^a$. It is not difficult to estimate the number of such composite currents, although it is hard to discern which of them are non-redundant [@Dittner:1972]. For $SU(3)$, invariant tensors that can be formed out of the composite operators, denoted here by $A^a,B^a,C^a$, take the form d\_[ab]{} A\^a B\^b, d\_[abc]{}A\^a B\^b C\^c. The growth rate for such composite operators (with increasing spin) exceeds the growth rate for generators predicted by the vacuum character in [Eq. (\[n3char\])]{}. However, it has to be checked on a case-by-case basis which of these fields are independent and contribute to the vacuum character.
All the generators that we have constructed for infinite $k$ are also present at finite $k$ (for $k\geq3$). It is worthwhile to write down the exact form of some of these generators for finite $k$. At finite $k$, the spin $4$ primary field takes the form P\_4(k)= Q\_4 - T + [::]{}. The bilinear spin $6$ primary current takes the form $$\begin{aligned}
P_{6,1}(k)= Q_{6,1} &- \tfrac{21 \a ^2(-129+106k +55k^2)}{5 \b\g\delta} {:\mathrel{\mkern2mu \partial^2T T \mkern2mu}:} - \tfrac{42 \a^2(1+k)(6+k)}{ \b\g\delta} {:\mathrel{\mkern2mu \partial T \partial T \mkern2mu}:} + \tfrac{21 \a }{5\b} {:\mathrel{\mkern2mu Q_4 T \mkern2mu}:} + \nn \\ &+ \tfrac{2 \a^2 (-219+ 1126 k +790 k^2) }{5 \b\g\delta}\partial^4 T - \tfrac{49 \a }{10\b} \partial^2 Q_4 - \tfrac{42 \a^3 (3-7k)}{\b\g\delta}{:\mathrel{\mkern2mu T T T \mkern2mu}:}\,\end{aligned}$$ where $\a= 3+k, \b= 9+4k, \g= 1+5k, \delta= 51+31 k$. In the above equations, the stress-energy tensor is defined as in [Eq. (\[Tdef\])]{} and the the currents $J^a$ are rescaled as $\sqrt{k}J^a$. The form of the quasi-primary operators is thus independent of $k$ for the bilinear currents. This is not true for the generator $P_{6,3}$. This current is not primary at finite $k$ and has to be modified in the following way to stay primary $$\begin{aligned}
P_{6,3}(k)=& f_{abc} {:\mathrel{\mkern2mu \partial^2 J^a \partial J^b J^c \mkern2mu}:} -\tfrac{21 \a^2(1353+6518k +2225k^2)}{5 \b\g\delta} {:\mathrel{\mkern2mu \partial^2T T \mkern2mu}:} - \tfrac{42 \a^2 (159+160k+61k^2)}{5 \b\g\delta} {:\mathrel{\mkern2mu \partial T \partial T \mkern2mu}:} + \nn \\& + \tfrac{23 \a }{5\b} {:\mathrel{\mkern2mu Q_4 T \mkern2mu}:} + \tfrac{2 \a (-2529 +8379k +9668 k^2 +2150 k^3) }{5 \b\g\delta}\partial^4 T + \tfrac{474 +157k}{30\b} \partial^2 Q_4 - \tfrac{2 \a^3 (3+193k)}{\b\g\delta}{:\mathrel{\mkern2mu T T T \mkern2mu}:}\,.\end{aligned}$$ Note that as $k \rightarrow \infty$ the fields $P_4(k)$ and $P_{6,1}(k)$ reduce to their corresponding counterparts in [Eq. (\[An3p4\])]{} and [Eq. (\[An3p61\])]{}, while $P_{6,3}(k)$ becomes identical to [Eq. (\[as6\])]{}.
### Large $N$ {#sec:largeN}
![\[rectfig\] The operators in the top-most row correspond to the $SU(N)$ Casimir invariants. A subset of generators for the stringy coset algebra is generated by acting by derivatives on the constituent terms of each Casimir operator. The cross-hatched area denotes the operators that are null and not part of the algebra for a given level $k$ and $N$. For $k=1$, only the top row of operators is not null and corresponds to the $W_N$ algebra. Increasing the level $k$ reduces the cross-hatched area but does not eliminate it completely even as $k$ tends to $\infty$. Increasing $N$ corresponds to adding more columns and also reducing the cross-hatched area. As we take $N$ to $\infty$, with $k$ already taken to $\infty$, null states disappear from the first column and the algebra becomes $\W_\infty^e[1]$. []{data-label="figrect"}](rectangle.pdf)
At large $N$, the currents that follow from the symmetric $d$-tensor invariants are given by \[genericC\] d\_[abc]{}\^J\^a \^J\^b \^J\^c . For $SU(N)$ there is a single Casimir invariant at orders $2,\cdots,N$. Hence for infinite $N$, the generating function for the generalized Casimir currents is \[SCF\] \_[p=2]{}\^ = \_[n=2]{}\^ -q. Since, the tensors $d_{abc\cdots}$ are totally symmetric, the generating function for an order $p$ current of the form in [Eq. (\[genericC\])]{} is given by the number of ways to divide an integer into $p$ parts. Classically, thus, the number of independent currents grows at least as fast as $\exp({\sqrt{n})}$. If one assumes that at large $N$, there are no null fields in this set of currents, then the growth of this set of currents matches that of the higher spin square algebra.
Clearly as we saw from the above examples of $SU(2)$ and $SU(3)$, these are not the only possible currents. For the group $SU(N)$, there are skew-symmetric invariant tensors $f_{abc\cdots}$ of order $3,5,\cdots,2N-1$ which lead to antisymmetric currents. In the literature, these skew-symmetric tensors are also known as $\Omega$ tensors [@deAzcarraga:1997; @deAzcarraga:2000]. The total number of such states as $N \rightarrow \infty$, assuming no null states is given by \[ASCF\] \_\^ = {\_[n=1]{}\^ (1+q\^n) - \_[n=1]{}\^ (1-q\^n) } -q.
As we saw for finite $N$, composite currents are also present for the coset CFT. In general, it is hard to count these currents. Invariants composed of primary fields that transform non-trivially under $SU(N)$ do not always lead to new or non-null currents. This happens because of the identities that exist between $SU(N)$ tensors (see, for example, Appendix B of [@Bais:1987a]).
Relation with $\W_\infty^e[1]$ and the higher spin square {#sec:Relation}
---------------------------------------------------------
We now focus on the currents bilinear in the $J^i$’s. As we saw for finite $N$, the independent bilinear currents at any value of the level $k$ have even spin only with multiplicity one at each spin. In the quasi-primary basis, these currents do not change with increasing $k$ or $N$, since only the overall normalization changes. (In the primary basis, this is no longer true). In the large level limit and at finite $N$, the coset theory reduces to that of $N^2-1$ free bosons, hence the bilinear currents can be identified with a realization of the $\W_\infty^e[1]$ algebra at central charge $N^2-1$. As is well known, such a realization is a finite truncation of the $\W_\infty^e[1]$ algebra [@Blumenhagen:1994]. As we take $N\rightarrow \infty$, we recover the full $\W_\infty^e[1]$ algebra.
It is natural to ask whether the higher-order generators of the stringy coset algebra as arranged in Fig. (\[figrect\]) can be identified with representations of the $\W_\infty^e[1]$ algebra. The OPE of any $\W_\infty^e[1]$ operator, which are the bilinear operators, with a generator of order $p$ gives rise to operators of the same order. For example, the OPE of any bilinear term of the form ${:\mathrel{\mkern2mu \partial^\mu J^a \partial^\nu J^a \mkern2mu}:}$ with a generic trilinear term is $$\begin{aligned}
&{:\mathrel{\mkern2mu \partial^\mu J^a \partial^\nu J^a (z) \mkern2mu}:}{:\mathrel{\mkern2mu d_{bcd} \partial^\a J^b \partial^\b J^c \partial^\gamma J^d(w) \mkern2mu}:} \nn\\
\sim\, &\frac{\delta^{ab} \delta^{ac} d_{bcd} \partial^\gamma J^d(w)}{(z-w)^{\mu+\nu+\a+\b+4}}~+~\frac{\delta^a_{b} d_{bcd} {:\mathrel{\mkern2mu \partial^\nu J^a(z) \partial^\b J^c(w) \partial^\gamma J^d(w) \mkern2mu}:}}{(z-w)^{\mu+\a+2}} \nn\\
\sim\, &\frac{ d_{bcd} {:\mathrel{\mkern2mu \partial^\nu J^b(w) \partial^\b J^c(w) \partial^\gamma J^d(w) \mkern2mu}:}}{(z-w)^{\mu+\a+2}} + \frac{ d_{bcd} {:\mathrel{\mkern2mu \partial^{\nu+1} J^b(w) \partial^\b J^c(w) \partial^\gamma J^d(w) \mkern2mu}:}}{(z-w)^{\mu+\a+1}} +\cdots\end{aligned}$$ We have written the OPE schematically omitting numerical coefficients in all terms and writing only a [*single*]{} representative term for different possible ways of contracting. The first term in the second line vanishes because the tensor $d_{bcd}$ is traceless. Thus we are only left with trilinear operators in the OPE. The same logic applies to the OPE of any $p$-th order operator with a bilinear operator. The operators in each column of Fig. (\[figrect\]) thus fall into a representation of $\W_\infty^e[1]$ .
Let us identify the $\W_\infty^e[1]$ representation that corresponds to each column of operators in Fig. (\[figrect\]). We use the standard coset notation for representations of the $\W_\infty^e[1]$ alegbra. Conventional $\W$-algebras that we deal with in this paper are the symmetry algebras of cosets of the form $\frac{\mathfrak{g}_{k} \otimes \mathfrak{g}_{1}}{\mathfrak{g}_{k+1}}$. The notation $(\Lambda_+; \Lambda_-)$ is used to denote a representation of a $\W$-algebra where $\Lambda_+$ is a representation of $\mathfrak{g}_{k}$ and $\Lambda_-$ is a representation of $\mathfrak{g}_{k+1}$. Then, a order $p$ column corresponds to the representation ${([0^{p-1},1,0,\ldots,0];0)}$ of $\W_\infty^e[1]$. The wedge character of a representation $R$ of $\W_\infty^e[\mu]$ is given by q\^[ B(R\^T) ]{}\^[U()]{}\_[R\^T]{}, where $R^T$ is the transpose of the representation $R$, $B(R^T)$ is the number of boxes in the Young tableaux of $R^T$ and $\chi^{U(\infty)}$ is the associated Schur function (See [@Gaberdiel:2011; @Gaberdiel:2015] for details). Thus the wedge character of the representation ${([0^{p-1},1,0,\ldots,0];0)}$ is given by \[bwedge1\] b\^[([wedge]{}) \[=1\]]{}\_[(\[0\^[p-1]{},1,0,…,0\];0)]{} = which is the same as the generating function of a column of operators with order $p$. Note that these are the same representations that constitute the higher spin square, the algebra of the large $N$ symmetric product orbifold.
The operators of the coset algebra look very different to corresponding operators of the higher spin square algebra. Nevertheless, the number of operators (and hence the character of the representation) is the same in a column of the coset algebra whose highest weight state is an operator of the form $d_{ab..}J^a J^b ..$ of order $p$ and in a column of the higher spin square (See Fig. \[squafig\]) whose highest weight state is of the form $J^a J^a..$ of the same order $p$. Identical characters imply identical representations, since a representation of an algebra has a unique character. Thus, the subset of generators of the coset algebra that are present in Fig. (\[figrect\]) must be isomorphic to the higher spin square.
Generators of the higher spin square can also be organized in terms of a $\W_{\infty}[0]$ algebra which is called the horizontal sub-algebra in Ref. [@Gaberdiel:2015]. In the coset case, this means that there should exist a change of basis for the $SU(N)$ currents, such that in the new basis the generators in the top row of Fig. (\[figrect\]) close at infinite $k$. It is possible that such a basis exists, since the coset theory also has a free fermion formulation at $c=N^2-1$. As is well known, the $\W_{\infty}[0]$ algebra can be expressed in terms of free fermions [@Pope:1991]. Next we look at operators of the generic form ${:\mathrel{\mkern2mu f_{bcd\cdots} \partial^\a J^b \partial^\b J^c \partial^\gamma J^d(w)\cdots \mkern2mu}:}$. The generating function for this set of operators is given by [Eq. (\[kdistinct\])]{}. Interpreting this as a wedge character of $\W_\infty^e[1]$, we find = = b\^[([wedge]{}) \[=1\]]{}\_[(\[p,0,…,0\];0)]{} . that it corresponds to the representation ${([p,0,\ldots,0];0)}$. All the “elementary operators” of the stringy coset algebra can thus be organized into representations of $\W_\infty^e[1]$.
Discussion
==========
In this paper, we have examined the coset in [Eq. (\[gencoset\])]{} in the free field limit, which is equivalent to its zero coupling limit. On the basis that the central charge of this coset scales as $N^2$, it has generally been expected that this coset is dual to a string theory in the bulk. We computed the vacuum character for the coset at finite $N$ and found that the currents of the symmetry algebra indeed exhibit an exponential growth with the spin. The $\W$-algebra of the coset in [Eq. (\[gencoset\])]{} in the free field limit is the same as the $\W$-algebra of the coset in [Eq. (\[coset2\])]{} in the large $k$ limit. We have written down the explicit form of the low-dimension currents for the second coset model. As we saw in the text, the currents of this coset are simply $SU(N)$ invariants composed out of differential operators of the form $\partial^\mu J^a$, where the $J^a$ are $SU(N)_k$ currents. As such, this coset theory is an exact analog in two dimensions of $SU(N)$ Yang-Mills theories in higher dimensions, with the addition of Virasoro symmetry.
The zero coupling limit of supersymmetric Yang-Mills theory in four dimensions is expected to be dual to tensionless string theory on the $AdS_5$ background [@Sundborg:2000]. It is, therefore, of interest to ask what the bulk dual of our coset model in the free field limit is. We know, from considerations of the D1-D5 system in Type IIB string theory, that it is a symmetric product orbifold theory that is expected to be dual to string theory on the $AdS_3$ background. Indeed, there has been much recent work in this direction clarifying aspects of this duality between tensionless strings and the free symmetric product orbifold theory [@Eberhardt:2018]. It is not clear where the coset theories, we have considered in this paper, fit into this picture. A more detailed understanding of the moduli space of $AdS_3$ string theories would be useful [@OhlssonSax:2018]. Further hints may come from integrability [@Sax:2014].
How does the more general $\W$-algebra of the stringy coset models relate to the $\W_\infty[\mu]$ symmetry of the vector coset models? We found that in the free field limit, the algebra $\W^e_\infty[1]$ is a sub-algebra of the full coset algebra. Further, there is a distinguished set of generators of the coset algebra that can be arranged into representations of $\W^e_\infty[1]$. Operators that are directly derived from the symmetric tensor invariants of $SU(N)$ can be arranged in the ${([0^{p-1},1,0,\ldots,0];0)}$ representations of $\W^e_\infty[1]$. Operators that are related to antisymmetric invariants of $SU(N)$ can be arranged in the ${([p,0,\ldots,0];0)}$ representations of $\W^e_\infty[1]$. Since the first set of operators can be organized in the same set of representations of $\W^e_\infty[1]$ as the operators of the higher spin square, we propose that this set of operators is identical to the higher spin square. The higher spin square also has a $\W_{\infty}[0]$ horizontal algebra, in addition to the vertical $\W^e_\infty[1]$ algebra. We have not explicitly identified this horizontal algebra in the coset case. It is important to do so, in order to cement the identification of the coset sub-algebra with the higher spin square.
In addition to the “elementary” generators, the coset theory also has a large number of composite operators at general values of $N$. In this paper, we have not attempted to classify them in representations of $\W^e_\infty[1]$. It is obviously of interest to understand the nature of these generators to comprehend the full symmetry algebra of the stringy coset theory. The coset in [Eq. (\[gencoset\])]{} is $T$-dual to a coset which is holographically dual to Vasiliev theory with a matrix extension. It would be interesting to explore the exact relation between this stringy coset and the matrix cosets [@Creutzig:2018]. In this paper, we have worked in the limit of zero coupling. However, in general the coset algebra depends on the parameter $\lambda$. It would be interesting to find how the algebra changes once the coupling is switched on. At certain values of non-zero $\lambda$, for example at $\lambda=1$, the coset theory has a formulation in terms of free bosons/fermions. This is also the point, where the symmetry algebra of the coset enhances to an $\N=1$ supersymmetric algebra. It would be nice to do an analysis similar to this paper for the $\lambda=1$ theory. More generally, Wolf space coset generalizations of [Eq. (\[gencoset\])]{} can be studied in a similar manner.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Rajesh Gopakumar, Yang-Hui He, Bogdan Stefanski and Alessandro Torrielli for discussions. We thank Biswajit Ransingh for running a computer program for us at HRI, Allahabad.
The $SU(N)$ tensor invariants
=============================
The tensors $d_{abcd..}$ are totally symmetric tensors which we have also chosen to be traceless. Suitable traceless symmetric tensors are defined in [@deAzcarraga:1997], where they are referred to as $t$-tensors.
We use the notation $\d$ for the standard symmetric invariant tensors of $SU(N)$. They are defined recursively [@Sudbery:1990] starting from the standard third-order symmetric tensor $\d_{ijk}$ . One can construct the tensor \[recursion\] \^[(r+1)]{}\_[[i\_1]{} … [i\_[r+1]{}]{}]{}= \^[(r)]{}\_[[i\_1]{} … [i\_[r-1]{}]{}j]{} \^[(3)]{}\_[j[i\_r]{} [i\_[r+1]{}]{}]{} , r=3,4, …. For $r \geq 3$, the above construction does not define totally symmetric tensors. The $\d$-family of symmetric tensors is obtained by symmetrising over all free indices in (\[recursion\]).
The $SU(n)$ $d$ tensors are related to members of the $\d$-family in the following way $$\begin{aligned}
\label{tdefs}
d{}_{ij} & \sim \delta_{ij} \,, \nn \\
d{}_{ijk} & \sim \d_{ijk} \,, \nn \\
d{}_{ijkl} & \sim n(n^2+1) \d^{(4)}{}_{(ijkl)} -2(n^2-4) \delta_{(ij}\delta_{kl)} \,, \nn \\
d{}_{ijklm} & \sim n(n^2+5) \d^{(5)}{}_{(ijklm)}-2(3n^2-20) \d_{(ijk}\delta_{lm)} \,, \cdots\end{aligned}$$ up to numerical coefficients dependent on $n$. The $d$-tensors vanish when their order is larger than $n$. The $d$-tensors are totally symmetric and are orthogonal to all other $d$-tensors of different order. For instance, for the fourth-order tensor this means d\_[ijkl]{} \_[ij]{}=0 , d\_[ijkl]{} d\_[ijk]{}=0 . Thus, the maximal contraction of the indices of two $d$-tensors of [*different*]{} order is zero. Using trace formulas for $\d$-tensors \_[(ijkl)]{} \_[ijm]{}= \_[klm]{} , we can find the contraction of two indices for the third-order and fourth-order $d$-tensor d\_[ijkl]{} d\_[ijm]{} \~d\_[klm]{} . Combinations of $d$-tensors provide a basis for the vector space of symmetric invariant polynomials of $SU(n)$.
For $N=3$, the tensor $d_{abc}$ takes the following values [1.5]{}[1.6]{}
[lllll]{} d\_[118]{}= & d\_[228]{}= & d\_[338]{}= & d\_[888]{}= &\
d\_[448]{}= & d\_[558]{}= & d\_[668]{}= & d\_[778]{}= &\
d\_[146]{}= & d\_[157]{}= & d\_[247]{}=- & d\_[256]{}= &\
d\_[344]{}= & d\_[355]{}= & d\_[366]{}=- & d\_[377]{}=- &
The anti-symmetric tensor $f_{abc}$ takes the following values [1.5]{}[1.6]{}
[lll]{} f\_[123]{} = 1 & f\_[147]{} = & f\_[156]{} = -\
f\_[246]{} = & f\_[257]{} = & f\_[345]{} =\
f\_[367]{} = - & f\_[458]{} = & f\_[678]{} =
Primary fields
==============
Here, we write down the full primary operators corresponding to the quasi-primary operators in [Sec. \[sec:su2\]]{} and [Sec. \[sec:su3\]]{}. Note that all fields are defined only up to an overall normalization factor.
$N=2$ primaries {#n2-primaries .unnumbered}
---------------
The primary field of [Eq. (\[n2p61\])]{} is $$\begin{aligned}
P_{6,1}= \big({:\mathrel{\mkern2mu \partial^3 J^a \partial J^a \mkern2mu}:} &-{:\mathrel{\mkern2mu \partial^2 J^a \partial^2 J^a \mkern2mu}:} - \tfrac{1}{10}{:\mathrel{\mkern2mu \partial^4 J^a J^a \mkern2mu}:}\big)+ \tfrac{15288}{16465} {:\mathrel{\mkern2mu \partial^2 T T \mkern2mu}:} +
\tfrac{5838}{16465} {:\mathrel{\mkern2mu \partial T \partial T \mkern2mu}:}+ \nn \\
& +
\tfrac{ 653}{16465} \partial^4 T- \tfrac{56}{135} \partial^2 Q_4 +
\tfrac{112}{45} {:\mathrel{\mkern2mu Q_4 T \mkern2mu}:} - \tfrac{22176}{16465} {:\mathrel{\mkern2mu T T T \mkern2mu}:}\,.
\end{aligned}$$
$N=3$ primaries {#n3-primaries .unnumbered}
---------------
The primary completion of the quasi-primary field $Q_4$ is \[An3p4\] P\_4 = Q\_4 - \^2 T+ [::]{}. The spin $5$ primary field is \[An3p5\] P\_5 = Q\_5 + \^2 Q\_3 - [::]{} . The bilinear spin $6$ primary is \[An3p61\] P\_[6,1]{} = Q\_[6,1]{} - [::]{} - [::]{} + [::]{} + \^4 T - \^2 Q\_4 + [::]{} . The trilinear spin $6$ primary is \[An3p62\] P\_[6,2]{} = Q\_[6,2]{} - \^3 P\_3 + [::]{} - [::]{} .
Algebra of symmetric product orbifold CFT
=========================================
The most straightforward way to find the spin and multiplicity of the generators of the symmetry algebra for the symmetric product orbifold theory is by looking at its vacuum character. Let us denote the chiral vacuum character of a seed theory by \[seed\] \_1(q) = \_[m=0]{}\^a\_m q\^m. Then the vacuum character of the $N$’th symmetric product orbifold can be read off from the following plethystic exponential [@elliptic] (,q) = \_[m=0]{}\^ = $\sum_{k=0}^{\infty} \frac{1}{k} \chi_1(q^k)\nu^k $. Expanding the exponential in the RHS gives a series in powers of $\nu$, we get: $$\begin{aligned}
\label{symmProdN}
\chi(\nu,q) &= \sum_{N=0}^{\infty} \chi_N(q) \nu^N \nn \\
&= 1 + \chi_1(q) \, \nu + \frac{{\chi_1(q)}^2 + \chi_1(q^2)} {2} \, \nu^2 +
\frac{\chi_1(q)^3 + 3 \chi_1(q) \chi_1(q^2) + 2 \chi_1(q^3)}{6} \, \nu^3 + \cdots\,.\end{aligned}$$ We can find the vacuum character for the $N$’th symmetric orbifold CFT by reading off the coefficients of $\nu^N$. In our case, the seed theory is the single boson theory whose chiral character is given by \[seedboson\] \_1(q) = \_[m=0]{}\^a\_m q\^m = \_[n=1]{}\^ . Using this expression for $\chi_1$, we can compute the the characters and the corresponding symmetry algebra of the symmetric product orbifold CFT using [Eq. (\[symmProdN\])]{}. These chiral characters agree with standard results for $S_N$ orbifolds [@Bantay:1999]. From these characters, we can compute the spectrum of the algebra for small values of $N$: $$\begin{aligned}
N=2: \qquad& 1,2,4\,.\nn\\
N=3: \qquad&1, 2, 3, 4, 5, 6^2 \,.\nn\\
N=4: \qquad&1,2, 3, 4^2, 5, 6^3, 7^2, 8^3, 9 \,. \nn\\
N=5: \qquad& 1, 2, 3, 4^2, 5^2, 6^3, 7^3, 8^5,9^4,10^5,11\,.\nn \\
N=6: \qquad &1, 2, 3, 4^2, 5^2, 6^4, 7^3, 8^6, 9^6, 10^8, 11^7, 12^8, 13\,.\nn\end{aligned}$$ As for the coset case, there can be more generators present. Despite the initial exponential growth in the number of operators with spin, at finite $N$, null states start appearing is the spectrum at some finite value of the spin and thus the algebra truncates. This is reflected in the asymptotic growth of the vacuum character: n\^[-]{}${\pi \sqrt{\tfrac{2}{3} n N}}$. which exihibits Cardy growth as $n\rightarrow \infty$.
As $N\rightarrow \infty$, the vacuum character is given by \_[1]{}(1-) (,q) =\_[m=0]{}\^ , which is again the plethystic exponential of [Eq. (\[seedboson\])]{}. This can be rewritten as \_[n = 1]{}\^\_\^ \_[n = m]{}\^. The generators of the infinite $N$ algebra are thus enumerated by the generating function \[bosonpart\] \_[n=2]{}\^ . along with a spin one field.
![\[squafig\] The higher spin square is generated by acting by derivatives on the operators in the top row. The second column corresponds to the $\W^e_\infty[1]$ algebra while subsequent columns correspond to its representations.](square.pdf)
We now write down the exact form of the currents. In the large $N$ limit, the “single-particle” generators for the symmetric product orbifold are symmetrized products of the form \[spgen\] \_[i=1]{}\^[N]{} (\^[m\_1]{} \_i) (\^[m\_p]{} \_i ) , m\_1,…, m\_p1 . Because of the symmetrization over $N$, this set of generators is in one-to-one correspondence with the chiral sector of a single boson. Removing the terms that are total derivatives, and in the $N\rightarrow \infty$ limit, they also constitute a set of linearly independent operators. Out of these generators, the subset of generators of [Eq. (\[spgen\])]{} of the form \[bgen\] \_[i=1]{}\^[N]{} (\^[m\_1]{} \_i) (\^[m\_2]{} \_i) , m\_1,m\_21 , define quasiprimary generators of spin $s=m_1+m_2$, in specific linear combinations and when $s$ is even. In fact, only one independent current can be constructed at each even spin, meaning that it is not a linear combination of derivatives of lower-spin currents, and there are no independent odd-spin currents. This set of independent currents generate the even spin $\W$-algebra ${\cal W}^{e}_\infty[1]$. The generators in [Eq. (\[bgen\])]{} are of order two, i.e., they are bilinear in the $\phi$s. The currents in [Eq. (\[spgen\])]{} are of arbitrary order $p\geq 1$. However, it turns out that the currents of a fixed order $p$, suitably corrected by lower-order terms, form a representation of the wedge algebra of ${\cal W}^{e}_\infty[1]$. This is captured in Fig. (\[squafig\]) where currents of a given order correspond to columns. The operators of the symmetric product orbifold algebra are, therefore, organized into representations of ${\cal W}^{e}_\infty[1]$. The statement that operators of the higher spin square can be organized in representations of ${\cal W}^{e}_\infty[1]$ is captured by the following identity: \[VIfull\] \_[n=1]{}\^ = 1+\_[p=1]{}\^ b\^[([wedge]{}) \[=1\]]{}\_[(\[0\^[p-1]{},1,0,…,0\];0)]{} , where $b^{({\rm wedge}) [\lambda=1]}_{([0^{p-1},1,0,\ldots,0];0)} $ denotes the wedge character of the $([0^{p-1},1,0,\ldots,0];0)$ representation of ${\cal W}^{e}_\infty[1]$. The $b^{({\rm wedge}) [\lambda=1]}_{([0^{p-1},1,0,\ldots,0];0)} $ character is defined in [Eq. (\[bwedge1\])]{}.
The LHS of [Eq. (\[VIfull\])]{} is the normalized partition function for a single boson. Combinatorially, the LHS is just the generating function for the number of ways one can partition an integer varying from $1$ to $\infty$ into an arbitrary number of parts. Each term in the sum on the RHS is number of ways one can partition an integer into exactly $p$ parts. In terms of the operators in [Eq. (\[spgen\])]{}, this is the spin $s$ of the operator, varying from $p$ to $\infty$, being partitioned into $m_1,m_2,\ldots,m_p$ at fixed $p$. An alternate way to organize the operators of the higher spin square is in terms of representations of ${\cal W}_{1+\infty}[0]$.
[99]{} M. R. Gaberdiel and R. Gopakumar, J. Phys. A [**46**]{}, 214002 (2013) doi:10.1088/1751-8113/46/21/214002 \[arXiv:1207.6697 \[hep-th\]\].\
M. R. Gaberdiel and R. Gopakumar, Phys. Rev. D [**83**]{}, 066007 (2011) doi:10.1103/PhysRevD.83.066007 \[arXiv:1011.2986 \[hep-th\]\].\
M. R. Gaberdiel, R. Gopakumar and A. Saha, JHEP [**1102**]{}, 004 (2011) doi:10.1007/JHEP02(2011)004 \[arXiv:1009.6087 \[hep-th\]\].
E. S. Fradkin and M. A. Vasiliev, Annals Phys. [**177**]{}, 63 (1987). doi:10.1016/S0003-4916(87)80025-8\
M. P. Blencowe, Class. Quant. Grav. [**6**]{}, 443 (1989). doi:10.1088/0264-9381/6/4/005\
S. F. Prokushkin and M. A. Vasiliev, Nucl. Phys. B [**545**]{}, 385 (1999) doi:10.1016/S0550-3213(98)00839-6 \[hep-th/9806236\]. M. Henneaux and S. J. Rey, JHEP [**1012**]{}, 007 (2010) doi:10.1007/JHEP12(2010)007 \[arXiv:1008.4579 \[hep-th\]\]. A. Campoleoni, S. Fredenhagen, S. Pfenninger and S. Theisen, JHEP [**1011**]{}, 007 (2010) doi:10.1007/JHEP11(2010)007 \[arXiv:1008.4744 \[hep-th\]\].\
A. Campoleoni, S. Fredenhagen and S. Pfenninger, JHEP [**1109**]{}, 113 (2011) doi:10.1007/JHEP09(2011)113 \[arXiv:1107.0290 \[hep-th\]\]. T. Creutzig, Y. Hikida and P. B. Ronne, JHEP [**1311**]{}, 038 (2013) doi:10.1007/JHEP11(2013)038 \[arXiv:1306.0466 \[hep-th\]\].\
C. Candu and C. Vollenweider, JHEP [**1404**]{}, 145 (2014) doi:10.1007/JHEP04(2014)145 \[arXiv:1312.5240 \[hep-th\]\].
A. B. Zamolodchikov, Theor. Math. Phys. [**65**]{}, 1205 (1985) \[Teor. Mat. Fiz. [**65**]{}, 347 (1985)\]. doi:10.1007/BF01036128 I. Bakas and E. Kiritsis, Nucl. Phys. B [**343**]{}, 185 (1990) Erratum: \[Nucl. Phys. B [**350**]{}, 512 (1991)\]. doi:10.1016/0550-3213(90)90600-I, 10.1016/0550-3213(91)90269-4 M. R. Gaberdiel and R. Gopakumar, JHEP [**1207**]{}, 127 (2012) doi:10.1007/JHEP07(2012)127 \[arXiv:1205.2472 \[hep-th\]\]. A. R. Linshaw, arXiv:1710.02275 \[math.RT\]. P. Bouwknegt and K. Schoutens, Phys. Rept. [**223**]{}, 183 (1993) doi:10.1016/0370-1573(93)90111-P \[hep-th/9210010\]. F. A. Bais, P. Bouwknegt, M. Surridge and K. Schoutens, Nucl. Phys. B [**304**]{}, 348 (1988). doi:10.1016/0550-3213(88)90631-1 F. A. Bais, P. Bouwknegt, M. Surridge and K. Schoutens, Nucl. Phys. B [**304**]{}, 371 (1988). doi:10.1016/0550-3213(88)90632-3 J. de Boer, L. Feher and A. Honecker, Nucl. Phys. B [**420**]{}, 409 (1994) doi:10.1016/0550-3213(94)90388-3 \[hep-th/9312049\]. R. Gopakumar, A. Hashimoto, I. R. Klebanov, S. Sachdev and K. Schoutens, Phys. Rev. D [**86**]{}, 066003 (2012) doi:10.1103/PhysRevD.86.066003 \[arXiv:1206.4719 \[hep-th\]\]. C. h. Ahn, K. Schoutens and A. Sevrin, Int. J. Mod. Phys. A [**6**]{}, 3467 (1991). doi:10.1142/S0217751X91001684 C. Ahn, JHEP [**1304**]{}, 033 (2013) doi:10.1007/JHEP04(2013)033 \[arXiv:1211.2589 \[hep-th\]\].\
C. Ahn, Phys. Rev. D [**94**]{}, no. 12, 126014 (2016) doi:10.1103/PhysRevD.94.126014 \[arXiv:1604.00756 \[hep-th\]\].\
C. Ahn, JHEP [**1307**]{}, 141 (2013) doi:10.1007/JHEP07(2013)141 \[arXiv:1305.5892 \[hep-th\]\]. M. R. Gaberdiel and R. Gopakumar, J. Phys. A [**48**]{}, no. 18, 185402 (2015) doi:10.1088/1751-8113/48/18/185402 \[arXiv:1501.07236 \[hep-th\]\]. A. Nazarov, Comput. Phys. Commun. [**183**]{}, 2480 (2012) doi:10.1016/j.cpc.2012.06.014 \[arXiv:1107.4681 \[math.RT\]\]. R. Blumenhagen, “W algebras in conformal quantum theory,” BONN-IR-91-06. P. Bowcock and P. Goddard, Nucl. Phys. B [**305**]{}, 685 (1988). doi:10.1016/0550-3213(88)90122-8
V. G. Kac and M. Wakimoto, Adv. Math. [**70**]{}, 156 (1988). doi:10.1016/0001-8708(88)90055-2 M. R. Gaberdiel and C. Vollenweider, JHEP [**1108**]{}, 104 (2011) doi:10.1007/JHEP08(2011)104 \[arXiv:1106.2634 \[hep-th\]\].\
C. Candu, M. R. Gaberdiel, M. Kelm and C. Vollenweider, JHEP [**1301**]{}, 185 (2013) doi:10.1007/JHEP01(2013)185 \[arXiv:1211.3113 \[hep-th\]\].\
C. Candu and C. Vollenweider, JHEP [**1311**]{}, 032 (2013) doi:10.1007/JHEP11(2013)032 \[arXiv:1305.0013 \[hep-th\]\].\
D. Kumar and M. Sharma, Phys. Rev. D [**95**]{}, no. 6, 066015 (2017) doi:10.1103/PhysRevD.95.066015 \[arXiv:1606.00791 \[hep-th\]\]. K. Thielemans, Int. J. Mod. Phys. C [**2**]{}, 787 (1991). doi:10.1142/S0129183191001001
J. A. de Azcarraga, A. J. Macfarlane, A. J. Mountain and J. C. Perez Bueno, Nucl. Phys. B [**510**]{}, 657 (1998) doi:10.1016/S0550-3213(97)00609-3 \[physics/9706006\]. J. A. de Azcarraga and A. J. Macfarlane, Int. J. Mod. Phys. A [**16**]{}, 1377 (2001) doi:10.1142/S0217751X01003111, 10.1142/S0217751X0100311X \[math-ph/0006026\]. D. Gepner, Nucl. Phys. B [**290**]{}, 10 (1987). doi:10.1016/0550-3213(87)90176-3
T. Procházka, JHEP [**1509**]{}, 116 (2015) doi:10.1007/JHEP09(2015)116 \[arXiv:1411.7697 \[hep-th\]\]. P. Dittner, Commun. Math. Phys. [**27**]{}, 44 (1972). doi:10.1007/BF01649658 R. Blumenhagen, W. Eholzer, A. Honecker, K. Hornfeck and R. Hubel, Int. J. Mod. Phys. A [**10**]{}, 2367 (1995) doi:10.1142/S0217751X95001157 \[hep-th/9406203\].\
R. Blumenhagen, W. Eholzer, A. Honecker, K. Hornfeck and R. Hubel, Phys. Lett. B [**332**]{}, 51 (1994) doi:10.1016/0370-2693(94)90857-5 \[hep-th/9404113\].
M. R. Gaberdiel, R. Gopakumar, T. Hartman and S. Raju, JHEP [**1108**]{}, 077 (2011) doi:10.1007/JHEP08(2011)077 \[arXiv:1106.1897 \[hep-th\]\]. C. N. Pope, hep-th/9112076. B. Sundborg, Nucl. Phys. Proc. Suppl. [**102**]{}, 113 (2001) doi:10.1016/S0920-5632(01)01545-6 \[hep-th/0103247\].\
P. Haggi-Mani and B. Sundborg, JHEP [**0004**]{}, 031 (2000) doi:10.1088/1126-6708/2000/04/031 \[hep-th/0002189\]. L. Eberhardt, M. R. Gaberdiel and R. Gopakumar, arXiv:1812.01007 \[hep-th\].\
M. R. Gaberdiel and R. Gopakumar, JHEP [**1805**]{}, 085 (2018) doi:10.1007/JHEP05(2018)085 \[arXiv:1803.04423 \[hep-th\]\].\
G. Giribet, C. Hull, M. Kleban, M. Porrati and E. Rabinovici, JHEP [**1808**]{}, 204 (2018) doi:10.1007/JHEP08(2018)204 \[arXiv:1803.04420 \[hep-th\]\].\
M. R. Gaberdiel, R. Gopakumar and C. Hull, JHEP [**1707**]{}, 090 (2017) doi:10.1007/JHEP07(2017)090 \[arXiv:1704.08665 \[hep-th\]\].
O. Ohlsson Sax and B. Stefanski, JHEP [**1805**]{}, 101 (2018) doi:10.1007/JHEP05(2018)101 \[arXiv:1804.02023 \[hep-th\]\].
O. Ohlsson Sax, A. Sfondrini and B. Stefanski, JHEP [**1506**]{}, 103 (2015) doi:10.1007/JHEP06(2015)103 \[arXiv:1411.3676 \[hep-th\]\].\
M. Baggio, O. Ohlsson Sax, A. Sfondrini, B. Stefanski and A. Torrielli, JHEP [**1704**]{}, 091 (2017) doi:10.1007/JHEP04(2017)091 \[arXiv:1701.03501 \[hep-th\]\].
T. Creutzig and Y. Hikida, arXiv:1812.07149 \[hep-th\]. A. Sudbery, “Computer-friendly d-tensor identities for SU(n),” PRINT-90-0325 (YORK), Journal of Physics A: Mathematical and General [**23**]{} [15]{} [L705]{} (1990)
R. Dijkgraaf, G. W. Moore, E. P. Verlinde and H. L. Verlinde, Commun. Math. Phys. [**185**]{}, 197 (1997) doi:10.1007/s002200050087 \[hep-th/9608096\].\
S. Benvenuti, B. Feng, A. Hanany and Y. H. He, JHEP [**0711**]{}, 050 (2007) doi:10.1088/1126-6708/2007/11/050 \[hep-th/0608050\].\
B. Feng, A. Hanany and Y. H. He, JHEP [**0703**]{}, 090 (2007) doi:10.1088/1126-6708/2007/03/090 \[hep-th/0701063\].\
P. Bantay, Nucl. Phys. B [**633**]{}, 365 (2002) doi:10.1016/S0550-3213(02)00198-0 \[hep-th/9910079\].\
A. Jevicki and J. Yoon, J. Phys. A [**49**]{}, no. 20, 205401 (2016) doi:10.1088/1751-8113/49/20/205401 \[arXiv:1511.07878 \[hep-th\]\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This paper determines the sharp asymptotic order of the following reverse Hölder inequality for spherical harmonics $Y_n$ of degree $n$ on the unit sphere $\sph$ of $\RR^d$ as $n\to \infty$: $$\|Y_n\|_{L^q(\sph)}\leq C n^{\a(p,q)}\|Y_n\|_{L^p(\sph)},\ \
0<p<q\leq \infty.$$ In many cases, these sharp estimates turn out to be significantly better than the corresponding estimates in the Nilkolskii inequality for spherical polynomials. Furthermore, they allow us to improve two recent results on the restriction conjecture and the sharp Pitt inequalities for the Fourier transform on $\RR^d$.
address:
- |
Department of Mathematical and Statistical Sciences\
University of Alberta\
Edmonton, Alberta T6G 2G1, Canada.
- |
Department of Mathematical and Statistical Sciences\
University of Alberta\
Edmonton, Alberta T6G 2G1, Canada.
- |
ICREA, Centre de Recerca Matemàtica\
Campus de Bellaterra, Edifici C 08193 Bellaterra (Barcelona), Spain, and Universitat Autònoma de Barcelona.
author:
- Feng Dai
- Han Feng
- Sergey Tikhonov
title: ' Reverse Hölder’s inequality for spherical harmonics '
---
[^1]
Introduction
============
Let $\sph=\{ x\in \mathbb{R}^{d}: \|x\|=1\}$ denote the unit sphere of $\RR^d$ endowed with the usual Haar measure $d\sa(x)$, where $\|\cdot\|$ denotes the Euclidean norm of $\RR^d$. Given $0<p\leq \infty$, we denote by $L^p(\sph)$ the usual Lebesgue $L^p$-space defined with respect to the measure $d\s(x)$ on $\sph$, and by $\|\cdot\|_p$ the norm of $L^p(\sph)$. Throughout the paper, unless otherwise stated, all functions on $\sph$ will be assumed to be real-valued and measurable, and the notation $ A\sim B$ means that there exists an inessential constant $c>0$, called the constant of equivalence, such that $ c^{-1} A \leq B \leq c A.$
Let $ \Pi_n^d$ denote the space of all spherical polynomials of degree at most $n$ on $\sph$ (i.e., restrictions on $\sph$ of polynomials in $d$ variables of total degree at most $n$), and $\HH_n^d$ the space of all spherical harmonics of degree $n$ on $\sph$. As is well known (see, for instance, [@BOOK chapter 1]), $\HH_n^d$ and $\Pi_n^d$ are all finite dimensional spaces with $\text{dim} \HH_n^d\sim n^{d-2}$ and $\text{dim} \Pi_n^d \sim n^{d-1}$ as $n\to \infty$. Furthermore, the spaces $\HH_k^d$, $k=0,1,\cdots$ are mutually orthogonal with respect to the inner product of $L^2(\sph)$, and each space $\Pi_n^d$ can be written as a direct sum $\Pi_n^d =\sum_{j=0}^n \HH_j^d$. Since the space of spherical polynomials is dense in $L^2(\sph)$, each $f\in L^2(\sph)$ has a spherical harmonic expansion, $
f=\sum_{k=0}^\infty \proj_k f,$ where $\proj_k$ is the orthogonal projection of $L^2(\sph)$ onto the space $\HH_k^d$ of spherical harmonics. The orthogonal projection $\proj_k$ has an integral representation: $$\label{1-1-0}
\proj_kf(x) =C_{k,d}\int_{\sph} f(y) P_k^{(\f {d-3}2, \f
{d-3}2)}(x\cdot y)\, d\sa(y),\ \ \ x\in\sph,$$ where $$C_{k,d}:= \f {\Ga(\f d2)\Ga(\f {d-1}2) }{2\pi^{d/2}\Ga(d-1) } \f { (2k+d-2)\Ga(k+d-2)}{\Ga(k+\f {d-1}2)},$$ and $P_k^{(\al,\be)}$ denotes the usual Jacobi polynomial of degree $k$ and indices $\al, \be$, as defined in [@Sz Chapter IV].
Our goal in this paper is to find a sharp asymptotic order of the quantity $\sup_{Y_n\in\HH_n^d}\f{ \|Y_n\|_q}{\|Y_n\|_p}$ for $0<p<q\leq \infty$ as $n\to \infty$. The background of this problem is as follows. In 1986, Sogge [@So1] proved that for $d\ge 3$, and $\l:=\f {d-2}2$, $$\label{1-1}
\sup_{Y_n\in\mathcal{H}_n^d}\f{\|Y_n\|_{L^q(\sph)}}{\|Y_n\|_{L^2(\sph)}} \sim \begin{cases}
n^{\l(\f 12-\f 1q)}, & \qquad \quad 2\leq q\le 2(1+\f 1\l),\\
n ^{2\lambda (\f 12-\f 1q)-\f1q}, & \qquad\quad 2(1+\f 1\l)\le q\le \infty,\end{cases}$$ which confirms a conjecture of Stanton–Weinstein [@Stan] in the case of $d=3$ and $q=4$. Here and throughout the paper, it is agreed that $0/0=0$. Recently, De Carli and Grafakos [@grafakos] proved that if $1\leq p\leq q\leq 2$ and $Y_n\in \HH_n^d$ can be written in the form $$\label{1-3}Y_n(x) = e^{ i m_{d-2} x_{d-1}} \prod_{k=0}^{d-2}
(\sin x_{k+1})^{m_{k+1}} P_{m_k-m_{k+1}}^{(m_{k+1}+\f {d-2-k}2, m_{k+1}+\f {d-2-k}2)} (\cos x_{k+1}),$$ with $n=m_0 \ge m_1\ge \cdots m_{d-2}\ge 0$ being integers, then $$\label{1-4}
\frac{ \|Y_n\|_{L^q(\mathbb{S}^{d-1})} }{ \|Y_n\|_{L^p(\mathbb{S}^{d-1})} }\leqs C n^{\frac{d-2}{2} (\f {1}{p}-\f 1q)},\ \ \ \ 1\leq p<q\leq 2,$$ which was further applied in [@grafakos] to prove the restriction conjecture for the class of functions consisting of products of radial functions and spherical harmonics that are in the form . Note that the set of functions $Y_n$ in with $n=m_0 \ge m_1\ge \cdots m_{d-2}\ge 0$ forms a linear basis of the space $\HH_n^d$. It is therefore natural to ask whether or not holds for all spherical harmonics $Y_n$ of degree $n$. A related work in this direction was done recently by De Carli, Gorbachev and Tikhonov in [@DeC], where the following weaker estimate was obtained for all spherical harmonics and applied to study a sharp Pitt inequality for the Fourier transform on $\RR^d$: $$\label{1-2}
\sup_{Y_n\in\HH_n^d} \f { \|Y_n\|_{p'}} { \|Y_n\|_p} \leq C n^{ (d-1) (\f 1p-\f 12)},\ \ \ \ \f1p +\f 1{p'}=1,\ \ \ 1\leq p\leq 2,$$ Finally, let us recall the following well-known result of Kamzolov [@Kam] on the Nikolskii inequality for spherical polynomials: $$\label{Nikol}
\|P_n\|_q \leqs C n^{(d-1) (\f 1p-\f1q)}\|P_n\|_p,\qquad \forall P_n\in\Pi_n^d,\ \ 0<p<q\leq \infty.$$ Since $\HH_n^d\subset \Pi_n^d$, the Nikolskii inequality is applicable to every spherical harmonics $Y_n\in \HH_n^d$. It turns out, however, that the resulting estimates are not sharp for spherical harmonics in many cases (see, for instance, , and ).
In this paper, we will prove the following result, which, in particular, shows that holds for all spherical harmonics $Y_n\in\HH_n^d$, and the upper bound on the right hand side of can be improved to be $ C n^{ (d-2) (\f 1p-\f 12)}$.
\[cor2\] Assume that $d\ge 3$ and $\f 1p+\f 1{p'}=1$ if $p\ge 1$. Set $\l:=\f {d-2}2$.
1. If either $0<p\leq 1$ and $p<q\leq \infty$, or $1\leq p\leq 2$ and $ p<q \leq
(1+\f 1\l)p'$, then $$\label{1-6}
\sup_{\sub{Y_n\in\mathcal{H}_n^d}}\f{\|Y_n\|_q}{\|Y_n\|_p} \sim n^{\l(\f 1p-\f 1q)}.$$
2. If either $1\leq p\leq 2$ and $q\ge (1+\f 1\l)p'$, or $2\leq p< 2+\f 1\l$ and $q> 2+\f 2\l$, then $$\sup_{\sub{Y_n\in\mathcal{H}_n^d}}\f{\|Y_n\|_q}{\|Y_n\|_p}\sim n ^{2\lambda (\f 12-\f 1q)-\f1q}.$$
3. If $ 2+\f 1\l<p <q \leq \infty$, then $$\sup_{\sub{Y_n\in\mathcal{H}_n^d}}\f{\|Y_n\|_q}{\|Y_n\|_p}\sim n^{(2\l+1) (\f 1p-\f1q)}.$$
4. If $d=3$ and $2\leq p< 4=2+\f1\l$, then for $q\ge 3 p'=(1+\f 1\l) p'$, $$\sup_{\sub{Y_n\in\mathcal{H}_n^d}}\f{\|Y_n\|_q}{\|Y_n\|_p}\sim
n^{\f 12 -\f 2q},$$ whereas for $p<q\leq 3p'$, $$\sup_{\sub{Y_n\in\mathcal{H}_n^d}}\f{\|Y_n\|_q}{\|Y_n\|_p}
\sim n^{\f 12 (\f 1p-\f1q)}.$$
Of particular interest is the case when $1\leq p\leq 2$ and $q=p'$, where our result can be stated as follows:
\[cor-1-2\]If $Y_n\in \mathcal{H}_n^d$ and $1\leq p\leq 2$, then $$\label{1-7} \|Y_n\|_{p'}\leq C n^{(d-2)(\f 1p-\f12)} \|Y_n\|_p,\ \ \ 1\leq p\leq 2.$$ Furthermore, this estimate is sharp.
Several remarks are in order.
Estimate for $p=p_\l:=1+\f{\l}{\l+2}$ follows directly from the well-known result of Sogge [@So1] on the orthogonal projection $\proj_n : L^2(\sph)\to \HH_n^d$. However, for $1\leq p<2$ and $p\neq p_\l$, the sharp estimate in Corollary \[cor-1-2\] is nontrivial and cannot be deduced from the result of Sogge [@So1]. Indeed, it was shown in [@So1] that for $1\leq p\leq p_\l:=1+\f{\l}{\l+2}$, $$\label{lemma-sogge}
\|\proj_n f\|_{2}\leq C n^{\l(\f{1}{p}-\f{1}{2} )+ \f{1}{2p(\l+2)}(p_\l-p)}
\|f\|_{p},\ \ \ \ \forall f\in L^{p}(\sph),$$ and this estimate is sharp. Since $\proj_n f=f$ for $f\in \HH_n^d$, this leads to the inequality $$\|Y_n\|_{2} \leq C n^{\l(\f{1}{p}-\f{1}{2} )+ \f{1}{2p(\l+2)}(p_\l-p)}\|Y_n\|_{p},\ \ \ \ \forall\, Y_n\in\HH_n^d,\ \ 1\leq p\leq p_\l,$$ which, according to Corollary \[cor-1-2\], is not sharp unless $p=p_\l$.
Interesting reverse Hölder inequalities for spherical harmonics, $$\sup_{Y_n\in\HH_n^d} \f{\|Y_n\|_q}{\|Y_n\|_p} \leq C(n,q)$$ with the constant $C(n,q)$ being independent of the dimension $d$ but dependent on the degree $n$ of spherical harmonics, were obtained in [@Duo] for some pairs of $(p,q)$, $0<p<q<\infty$. The general constants $C$ in our paper are dependent on the dimension $d$, but independent of the degree $n$.
For $d\ge 4$, it remains open to find the asymptotic estimate of the supremum on the left hand side of for $2<p<1+\f 1\l$ and $p<q< 2+\f 2\l$.
This paper is organized as follows. In Section 2, we construct a sequence of convolution operators $\{T_n\}_{n=0}^\infty$ on $L^1(\sph)$ with the properties that $T_n f=f$ for $f\in\HH_n^d$, $|T_n f|\leq C \sup_{0\leq j \leq d} |\proj_{n+2j} f|$ and $\|T_n f\|_\infty \leq C n^\l \|f\|_1$ for all $f\in L^1(\sph)$. These operators play an indispensable role in the proof of Theorem \[cor2\], which is given in the third section. Finally, in Section 4, we give two applications of our main result, improving a recent result of [@grafakos] on restriction conjecture and a result of [@DeC] on sharp Pitt’s inequality.
A sequence of convolution operators
=====================================
We start with the following well-known result of Sogge [@So1] on the operator norms of the orthogonal projections $\proj_n: L^2(\sph)\to \HH_n^d$.
[@So1]\[lem-4-3\] Let $n\in\NN$, $d\ge 3$ and $\l=\f{d-2}2$. Then the following statements hold:
1. If $1\leq p\leq p_\l:=1+\f \l {\l+2}$, then $$\|\proj_n f\|_2\leq C n^{(2\l+1)(\f 1p-\f12)-\f12}\|f\|_p.$$
2. If $p_\l\leq p\leq 2$, then $$\|\proj_n f\|_2\leq C n^{\l(\f 1p-\f12)}\|f\|_p.$$
3. If $2+\f 2\l\leq q \leq \infty$, then $$\|\proj_n f\|_q\leq C n^{(2\l+1)(\f12-\f1q)-\f12}\|f\|_2.$$
4. If $2\leq q \leq 2+\f 2\l$, then $$\|\proj_n f\|_q \leq C n^{\l(\f 12-\f 1q)} \|f\|_2.$$ Here, the letter $C$ denotes a general positive constant independent of $n$ and $f$.
As was pointed out in the introduction, Lemma \[lem-4-3\] will not be enough for the proof of our main result. The crucial step in the proof of Theorem \[cor2\] is to construct a sequence of linear operators $\{T_n\}_{n=0}^\infty$ with the properties that $T_n f=f$ for $f\in\HH_n^d$, $|T_n f|\leq C \sup_{0\leq j \leq d} |\proj_{n+2j} f|$ and $\|T_n f\|_\infty \leq C n^\l \|f\|_1$ for all $f\in L^1(\sph)$ .
To define the operators $T_n$, we need to recall several notations. First, given $h\in\NN$, and a sequence $\{a_n\}_{n=0}^\infty$ of real numbers, define $$\tr_h a_n =a_n-a_{n+h},\ \
\tr_h^{\ell+1}=\tr_h \tr_h^\ell,\ \ \ell=1,2,\ldots.$$ Next, let $$R_n^\l (\cos\t):=\f {P_n^{(\l-\f12, \l-\f12)}(\cos\t) }{P_n^{(\l-\f12, \l-\f12)}(1)},\ \ \ \t\in [0,\pi]$$ denote the normalized Jacobi polynomial, and for a step $h\in \NN$, define $$\tr_h^\ell R_n^\l(\cos\t)=\tr_h^\ell a_n,\ \ \ell=1,2,\ldots,\ \ n=0,1,\cdots,$$ with $a_n:= R_n^{\l}(\cos\t)$. Here and throughout, the difference operator in $\tr_h^\ell R_n^\l(\cos\t)$ is always acting on the integer $n$. In the case when the step $h=1$, we have the following estimate ([@BOOK Lemma B.5.1], [@DD]): $$\label{0-1}
\Bl|\tr_1^{\ell} R_n^\l (\cos\t)
\Br|\leq C \t^\ell (1+n\t)^{-\ld},\ \ \ \t\in [0, \pi/2],\ \ \ell\in \NN.$$ On the other hand, however, the $\ell$-th order difference $\tr_1^{\ell} R_n^\l (\cos\t)$ with step $h=1$ does not provide a desirable upper estimate when $\t$ is close to $\pi$, and as will be seen in our later proof, estimate itself will not be enough for our purpose.
To overcome this difficulty, instead of the difference with step $1$, we consider the $\ell$-th order difference $\tr_2^{\ell} R_n^\l (\cos\t)$ with step $h=2$. Since $\tr_2^\ell a_n =\sum_{j=0}^\ell\binom{\ell} j \tr_1^\ell a_{n+j},$ on one hand, implies that $$\Bl|\tr_2^{\ell} R_n^\l (\cos\t)
\Br|\leq C \t^\ell (1+n\t)^{-\ld},\ \ \t\in [0,\pi/2].$$ On the other hand, however, since $$\tr_2^{\ell} R_n^\l (\cos\t) =\sum_{j=0}^{\ell} (-1)^j \binom{\ell} j R_{n+2j}^\l (\cos \t),$$ and since $R_{n+2j}^\l (-z) = (-1)^n R_{n+2j}^\l (z)$, we have $\tr_2^{\ell} R_n^\l (\cos(\pi -\t))=(-1)^n \tr_2^{\ell} R_n^\l (\cos\t).$ It follows that $$\label{0-2}
\Bl|\tr_2^{\ell} R_n^\l (\cos\t)
\Br|\leq C \begin{cases}
\t^\ell (1+n\t)^{-\ld},\ \ \t\in [0, \pi/2],\\
(\pi-\t)^{\ell} (1+n (\pi-\t))^{-\ld},\ \ \t\in [\pi/2, \pi].\end{cases}$$
By , we obtain that for every $P\in \HH_n^d$, $$P(x)=c_n \int_{\sph} P(y) R_n^\l (x\cdot y) \, d\s(y),\ \ x\in\sph,$$ where $$c_n:=\f{\Ga(\f d2)}{2\pi^{d/2}}\f{ d+2n-2}{d+n-2}\f{\Ga( d+n-1)}{\Ga(n+1)\Ga(d-1)}\sim n^{d-2},$$ and $x\cdot y$ denotes the dot product of $x, y\in\RR^d$. Since $R_j^\l (x\cdot)\in\HH_j^d$ for any fixed $x\in\sph$, it follows by the orthogonality of spherical harmonics that for any $P\in \mathcal{H}_n^d$, and any $\ell\in\NN$, $$\begin{aligned}
P(x)&
= c_n \sum_{j=0}^\ell (-1)^j \binom {\ell}j \int_{\sph} P(y) R_{n+2j}^\l (x\cdot y)\, d\s(y)\notag \\
&=c_n \int_{\sph} P(y) \tr_2^\ell
R_n^\l (x\cdot y) \, d\s(y).\label{2-1}
\end{aligned}$$ For the rest of the paper, we will choose $\ell$ to be an integer bigger than $\l$ (for instance, we may set $\ell=d-2$), so that by , we have $$\label{2-4-1}
\Bl|\tr_2^{\ell} R_n^\l (\cos\t)
\Br|\leq C n^{-\l}.$$
Now we are in a position to define the operators $T_n$.
For $f\in L(\sph)$, we define $$\label{2-2}
T_n f(x):=\int_{\sph} f(y) \Phi_n(x\cdot y)\, d\s(y),\
\ x\in\sph,$$ where $$\Phi_n(\cos \t):=
c_n \sum_{j=0}^{d-2} (-1)^j \binom {d-2}jR_{n+2j}^\l(\cos\t).$$
By , we have $$\label{2-6-1}
|\Phi_n(\cos\t)|\leq C n^\l,\ \ \
\t\in [0,\pi],$$ whereas by $$\label{2-7-1}
T_n P(x) =P(x),\ \ \ \forall P\in \HH_n^d, \ \ \forall x\in \sph.$$
The main result of this section can now be stated as follows.
\[thm-2-3\]
1. If $1\leq p\leq 2$ and $p'\leq q\leq (1+\f 1\l) p'$, then $$\label{2-5}
\|T_nf\|_{q}\leq C n^{\l(\f 1p-\f1q)} \|f\|_p,\ \ \forall f\in L^p(\sph).$$
2. If $1\leq p\leq 2$ and $q\ge (1+\f 1\l) p'$, then $$\begin{aligned}
\|T_n f\|_q \leq C n^{\l-\f{2\l+1}q}\|f\|_p,\ \ \forall f\in L^p(\sph).
\end{aligned}$$
First, we prove the assertion (i). Note that by definition, for each $f\in L^2(\sph)$, $$\label{3-4-0}T_n f =
\sum_{j=0}^{d-2} (-1)^j \binom {d-2}j \f{c_n}{c_{n+2j}}
\proj_{n+2j} f,$$ which implies that $$\label{2-4}
\|T_n f\|_2\leq C \|f\|_2,\ \ \forall f\in L^2(\sph).$$ On the other hand, however, using , we have $$\label{2-3}
\|T_n f\|_\infty\leq C n^\l\|f\|_1,\
\ \forall f\in L^1(\sph).$$ Thus, applying the Riesz-Thorin interpolation theorem, and using and , we deduce that for $1\leq p\leq 2$, $$\label{2-5-0}
\|T_nf\|_{p'}\leq C n^{(d-2)(\f 1p-\f12)} \|f\|_p,\ \ \forall f\in L^p(\sph).$$
Next, by (iv) of Lemma \[lem-4-3\], and using , we obtain that for $2\leq r\leq 2(1+\f 1\l)$, $$\label{2-13}
\|T_n f\|_{r} \leq C n^{\l( \f 12-\f 1{r})}\|f\|_2,\ \ \forall f\in L^2(\sph).$$ Assume that $1\leq p\leq 2$ and $p'\leq q\leq (1+\f 1\l) p'$. Let $\t=\f 2{p'}\in [0,1]$, and let $r=\t q=\f 2{p'} q$. Then $2\leq r\leq 2(1+\f 1\l)$, and $$\f 1p = 1-\t +\f \t 2,\ \ \f 1q =\f{1-\t}{\infty}+\f \t r.$$ Thus, by , and applying the Riesz-Thorin interpolation theorem, we obtain that $$\|T_n f\|_q \leq
C n^{\l (1-\t)} n^{\l (\f 12-\f 1r)\t}\|f\|_p= C n^{\l (\f 1p-\f 1q)}\|f\|_p.$$ This completes the proof of the assertion (i).
Assertion (ii) can be proved similarly. Indeed, using and (iii) of Lemma \[lem-4-3\], we have that for $r\ge 2(1+\f 1\l)$, $$\label{2-14}
\|T_n f\|_{r} \leq C n^{2\l (\f 12 -\f 1r)-\f 1r}\|f\|_2,\ \ \forall f\in L^2(\sph).$$ Assume that $1\leq p\leq 2$ and $ q\ge (1+\f 1\l) p'$. Let $\t=\f 2{p'}$ and $r=\t q =\f {2}{p'} q$. Then $r\ge 2(1+\f 1\l)$. Using , and applying the Riesz-Thorin interpolation theorem, we deduce that $$\begin{aligned}
\|T_n f\|_q &\leq C n^{\l(1-\t)} n ^{ (d-2)\t (\f 12 -\f 1r)-\f \t r} \|f\|_p=C n^{\l-\f{2\l+1}q}\|f\|_p\\
&=C n^{(d-2) (\f 12-\f1q)-\f1q}\|f\|_p.
\end{aligned}$$ This completes the proof of (ii).
Proof of Theorem \[cor2\]
=========================
The stated lower estimates of Theorem \[cor2\] follow directly from the following two known lemmas.
[@So1]\[lem-4-1\] Let $$f_n(x)=(x_1+ix_2)^n,\ \ x\in\sph.$$ Then $f\in\mathcal{H}_n^d$ and $$\|f_n\|_p\sim n^{-\ld /p},\ \ 0< p<\infty.$$
[@Sz p.391]\[lem-4-2\] Let $$g_n(x) =P_n^{(\f {d-3}2, \f {d-3}2)}(x\cdot e)$$ for a fixed point $e\in\sph$. Then $g_n\in\mathcal{H}_n^d$, and $$\|g_n\|_p \sim \begin{cases}
n^{\f {d-3}2} n^{-\f {d-1}p},\ & p> \f { 2(d-1)}{d-2},\\
n^{-\f12}(\log n)^{\f1p}, \ &p=\f { 2(d-1)}{d-2},\\
n^{-\f12}, & p<\f { 2(d-1)}{d-2}.\end{cases}$$
For the proof of the upper estimates, we let $P\in\HH_n^d$. The crucial tool in our proof is Theorem \[thm-2-3\], where we recall that $T_n P=P$ for all $P\in\HH_n^d$. We consider the following cases:\
[*Case 1.*]{} $1\leq p\leq q\leq p'$.\
In this case, $1\leq p\leq 2\leq p'$, and the stated upper estimate for $q=p'$ follows directly from Theorem \[thm-2-3\]. In general, for $p\leq q \leq p'$, let $\t\in [0,1]$ be such that $\f 1q=\f \t p+\f {1-\t}{p'}.$ Then by the log-convexity of the $L^p$-norm, we have $$\|P\|_q\leq \|P\|_p^\t \|P\|_{p'}^{1-\t} \leq C n^{\l(\f 1p-\f1{p'})(1-\t)}\|P\|_p\leq C n^{\l( \f 1p-\f1q)}\|P\|_p,$$ which is as desired in this case.\
[*Case 2.*]{} $0<p\leq 1$ and $p<q$.\
In this case, note that $$\|P\|_1\leq \|P\|_p^p \|P\|_{\infty} ^{1-p}\leq
C n^{\l(1-p)} \|P\|_p^p \|P\|_1^{1-p}.$$ It follows that $$\|P\|_1\leq C n^{\l(\f 1p-1)}\|P\|_p,\ \ 0<p\leq 1,$$ which, in turn, implies that for $p< q$ and $\f 1q= \f {1-\t}p$, $$\|P\|_q \leq \|P\|_\infty^\t \|P\|_p^{1-\t}\leq
C n^{\l \t } \|P\|_1^\t \|P\|_p^{1-\t}\leq C n^{\l (\f 1p-\f1q)}\|P\|_p.$$
[*Case 3.*]{} $1\leq p\leq 2$ and $q\ge p'$.\
The desired estimate in this case follows directly from the first and the second parts of Theorem \[thm-2-3\] since $T_n P=P$ for all $P\in\HH_n^d$.\
[*Case 4.*]{} $2\leq p\leq 2+\f 1\l$ and $q\ge 2+\f 2\l$.\
For $P\in\HH_n^d$, by the already proven cases it follows that $$\begin{aligned}
\|P\|_q&\leq C n^{(d-2) (\f 12-\f1q)-\f1q}\|P\|_2\leq
C n^{(d-2) (\f 12-\f1q)-\f1q}\|P\|_p.
\end{aligned}$$
[*Case 5.*]{} $2+\f 1\l<p<q\leq \infty$.\
The reverse Hölder inequality in this case follows directly from the corresponding Nikolskii inequality for spherical polynomials given by (\[Nikol\]).\
[*Case 6.*]{} $d=3$ and $2\leq p< 4=2+\f1\l$.\
The proof in this case relies on the following result of Sogge [@So1]:
\[4-1-lem\]If $d=3$, $\f 43<p<4$ and $q=3p'=p'(1+\f 1\l)$, then $$\|\proj_n f\|_q \leq C n^{\f 12-\f 2q}\|f\|_p.$$
Now we return to the proof in Case 6. Again, in view of Lemmas \[lem-4-1\] and \[lem-4-2\], it is enough to prove the upper estimates. Assume first that $q\ge 3p'$. Let $2\leq p< p_1 < 4$ and let $\t\in [0,1]$ be such that $$\f 1p = \f {1-\t}{p_1}+\f \t 2.$$ Set $q_1=3p_1'$. Then by Lemma \[4-1-lem\], $$\label{4-1}
\|Tf\|_{q_1} \leq C n ^{ \f 12 - \f 2 {q_1}}\|f\|_{p_1}.$$ For $q\ge 3p'>3p_1'=q_1$, let $q_2\ge q$ be such that $$\f 1q = \f {1-\t}{q_1} +\f \t{q_2}.$$ Then $$\begin{aligned}
\f13\ge \f 1{3p} +\f 1q =\t(\f 16 +\f 1{q_2}) + \f 13(1-\t)=\t (\f 1q_2-\f16)+\f13.\end{aligned}$$ This implies that $q_2\ge 6=2+\f 2\l$, hence by (ii) of Theorem \[thm-2-3\], $$\begin{aligned}
\label{4-22}
\|T_n f\|_{q_2}\leq C n^{\f 12 -\f 2{q_2}} \|f\|_2.\end{aligned}$$ Thus, using , , and the Riesz-Thorin theorem, we obtain $$\|T_n f\|_q \leq C n^{\f 12 -\f 2q}\|f\|_p,$$ which implies the desired estimate for the case of $q\ge 3p'$.
The case of $p<q< 3p'$ can be treated similarly. In fact, let $p_1, q_1$ and $\t$ be as above. Observing that $\f12 -\f 2 {q_1} =\f 12 (\f 1{p_1}-\f 1{q_1})$, we may rewrite as $$\|Tf\|_{q_1} \leq C n ^{ \f 12 (\f 1{p_1}-\f 1{q_1})}\|f\|_{p_1}.$$ Furthermore, we may choose $p_1>p$ to be very close to $p$ so that $q < q_1=3p_1'<3p'$. Let $q_3\leq q $ be such that $$\f 1q = \f {1-\t}{q_1} +\f \t{q_3}.$$ Then $$\begin{aligned}
\f13< \f 1{3p} +\f 1q =\t(\f 16 +\f 1{q_3}) + \f 13(1-\t)=\t (\f 1q_3-\f16)+\f13.\end{aligned}$$ Hence $2<q_3 <6$, and using (i) of Theorem \[thm-2-3\], we deduce $$\begin{aligned}
\|T_n f\|_{q_3}\leq C n^{\f 12 ( \f 12 -\f 1{q_3})} \|f\|_2.\end{aligned}$$ The stated estimate for $p<q<3p'$ then follows by the Riesz-Thorin interpolation theorem.
Applications: Fourier inequalities
==================================
The restriction conjecture.
-----------------------------
One of the most challenging problems in classical Fourier analysis is the restriction conjecture, which states that if $1 \le p < \frac{2d}{d+1}$ and $q \leq \frac{ d-1}{
d+1} p'$, then there exists a constant $C$ depending only on $p, q, d$ such that $$\label{rc0}
\frac{\| \widehat{F} \|_{L^q(\mathbb{S}^{d-1})}}
{\| {F} \|_{L^p(\mathbb{R}^{d})}}
\le C,\ \ \ \forall F \in C^\infty_0(\mathbb{R}^d),$$ where $\hat{F}(\xi): =\int_{\RR^d} F(x) e^{-2\pi i x\cdot \xi}\, dx,$ $ \xi\in\RR^d.$ This conjecture has been completely proved only in the case of $d=2$. We refer to the book [@stein Chapter IX] for more background information of this problem.
De Carli and Grafakos [@grafakos] recently proved that the restriction conjecture is valid for all functions $F$ that can be expressed in the form $$F(x)=f(\|x\|) \|x\|^n g_n\bl(\f{x}{\|x\|}\br),\ \ \ n=0,1,\cdots$$ with $f(\|\cdot\|)\in C_0^\infty(\RR^d) $ and $g_n\in\HH_n^d$ being given in . Using [Theorem \[cor2\]]{} (i), and following the argument of [@grafakos], we may conclude here that the restriction conjecture holds for a wider class of functions $$F\in \bigcup_{n=0}^\infty \Bl\{ f(\|x\|) \|x\|^n Y_n\bl(\f{x}{\|x\|}\br):\ \
f(\|\cdot\|)\in C_0^\infty(\RR^d),\ \ Y_n\in\HH_n^d\Br\}.$$ Indeed, it was shown in [@grafakos] that for $F(x)=f(\|x\|)\|x\|^n
Y_n(x/\|x\|)$ with $f\in C_0^\infty (\RR^d)$ and $Y_n \in\HH_n^d$, $$\begin{aligned}
\frac{\| \widehat{F} \|_{L^q(\mathbb{S}^{d-1})}}
{\| {F} \|_{L^p(\mathbb{R}^{d})}}
&=
\frac
{
\Big|
\int_0^{\infty} f (r) J_{\frac d2-1+n}(r)r^{\frac d2+n} dr\Big|
}
{
\Big(
\int_0^{\infty} |f(r)|^p r^{d-1+np} dr\Big)^{1/p}
}
\frac{ \|Y_n\|_{L^q(\mathbb{S}^{d-1})} }{ \|Y_n\|_{L^p(\mathbb{S}^{d-1})} }
\nonumber
\\
&\leq
C n^{(d-1)(\f 1 2-\f 1 p)+\f 1 {p'}}\frac{ \|Y_n\|_{L^q(\mathbb{S}^{d-1})} }{ \|Y_n\|_{L^p(\mathbb{S}^{d-1})} }, \label{5-2}\end{aligned}$$ where $J_n(r)$ is the Bessel function of the first kind. However, according to (i) of [Theorem \[cor2\]]{} , we obtain that for $1 \le p < \frac{2d}{d+1}$ and $q \leq \frac{ d-1}{
d+1} p'$, $$\begin{aligned}
\text{RHS of \eqref{5-2}} \leq C
\sup\limits_{m\ge 1}
m^{(d-1)(\frac12-\frac1p)+\frac{1}{p'}+\frac{d-2}{2} (\f {1}{p}-\f 1q)} \leq C.\end{aligned}$$
The sharp Pitt inequality
-------------------------
The following sharp Pitt inequality has been recently proved in [@DeC]:
\[T-Pitt2\] If $1\leq p\leq 2$ and $s=
(d-1)\left(\frac 12-\frac 1p\right)$, then for every $Y_k\in \CH_k^d$ and every radial $f\in {\mathcal S}(\R^d)$, the Pitt inequality $$\label{e-newpitt}
\|\,|y|^{-s} {\widehat}{f Y_k} \|_{L^{p'} (\R^d)} \leq C \|\,|x|^s fY_k \|_{L^p(\R^d)}$$ holds with the best constant $$\label{e-B}
C=(2\pi)^{\frac d2} 2^{\frac{1}{2}-\frac{1}{p'}} \frac{p^{\frac{(2 k+d-1)p+2}{4p}} \Gamma \left(\frac{(2 k+d-1)
p'+2}{4}\right)^{\frac{1}{p'}}}{(p')^{\frac{(2 k+d-1)p'+2}{4p'}} \Gamma
\left(\frac{(2 k+d-1) p+2}{4}\right)^{\frac 1p}}\sup_{Y_k\in \CH_k^d} \frac{\|Y_k\|_{L^{p'}(\SS^{n-1})}}{\|Y_k\|_{L^ p(\SS^{n-1})}}.$$
According to [Theorem \[cor2\]]{}, we have $$\sup_{Y_k\in \CH_k^d} \frac{\|Y_k\|_{L^{p'}(\SS^{n-1})}}{\|Y_k\|_{L^ p(\SS^{n-1})}}\sim
k^{(d-2)(\frac{1}{p}-\frac{1}{2})},$$ whereas only the weaker estimate was obtained in [@DeC].
[999]{}
F. Dai and Y. Xu, [*Approximation Theory and Harmonic Analysis on Spheres and Balls*]{}, Springer Monographs in Mathematics, 2013.
F. Dai and Z. Ditzian, Combinations of multivariate averages, [*J. Approx. Theory*]{} [**131**]{} (2004), no. 2, 268–283.
L. De Carli, D. Gorbachev, S. Tikhonov, [Pitt and Boas inequalities for Fourier and Hankel transforms]{}, [*J. Math. Anal. Appl.*]{} **408** (2013), no. 2, 762–774.
L. De Carli and G. Grafakos, On the restriction conjecture, *Michigan Math. J.* **52** (2004), no. 1, 163–180.
J. Duoandikoetxea, Reverse Hölder inequalities for spherical harmonics, *Proc. Am. Math. Soc.* **101** (1987), 487–491.
A. I. Kamzolov, Approximation of functions on the sphere $\mathbb{S}^n$, *Serdica* **84** (1984), no. 1, 3–10.
C. D. Sogge, Oscillatory integrals and spherical harmonics, [*Duke Math. J.*]{} [**53**]{} (1986), 43–65.
R. J. Stanton, A. Weinstein, On the $L^4$ norm of spherical harmonics, [*Math. Proc. Camb. Phil. Soc.*]{} [**89**]{} (1981), 343–358.
E. M. Stein, [*Harmonic Analysis,*]{} Princeton Univ. Press, Princeton, 1993.
G. Szegö, [*Orthogonal Polynomials*]{}, Amer. Math. Soc., New York, 1967.
[^1]: The first and the second authors were partially supported by the NSERC Canada under grant RGPIN 311678-2010. The third author was partially supported by MTM 2011-27637, 2014 SGR 289, RFFI 13-01-00043 and the Alexander von Humboldt Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the past few years, there has been a huge growth in Twitter sentiment analysis having already provided a fair amount of research on sentiment detection of public opinion among Twitter users. Given the fact that Twitter messages are generated constantly with dizzying rates, a huge volume of streaming data is created, thus there is an imperative need for accurate methods for knowledge discovery and mining of this information. Although there exists a plethora of twitter sentiment analysis methods in the recent literature, the researchers have shifted to real-time sentiment identification on twitter streaming data, as expected. A major challenge is to deal with the Big Data challenges arising in Twitter streaming applications concerning both Volume and Velocity. Under this perspective, in this paper, a methodological approach based on open source tools is provided for real-time detection of changes in sentiment that is ultra efficient with respect to both memory consumption and computational cost. This is achieved by iteratively collecting tweets in real time and discarding them immediately after their process. For this purpose, we employ the Lexicon approach for sentiment characterizations, while change detection is achieved through appropriate control charts that do not require historical information. We believe that the proposed methodology provides the trigger for a potential large-scale monitoring of threads in an attempt to discover fake news spread or propaganda efforts in their early stages. Our experimental real-time analysis based on a recent hashtag provides evidence that the proposed approach can detect meaningful sentiment changes across a hashtag’s lifetime.'
author:
- '\'
bibliography:
- 'mybib.bib'
title: |
Real Time Sentiment Change Detection of\
Twitter Data Streams\
---
Twitter, Change Detection, Data Stream Mining
Introduction
============
In the last decade, there has been a huge growth in the use of microblogging platforms such as Twitter [@Kouloumpis] which is overwhelmed by amazing statistics. People send more than 500 million tweets per day (last update: 1/24/2017), 300 million are the total number of monthly active Twitter users (last update: 1/1/2018) and 100 million are the number of Twitter daily active users (last update: 1/24/2017). This wealth of information has attracted the interest of the research community, focusing on day-to-day emotion analysis that can be proven to be of great value in analyzing opinions about events, products, persons or political stances. It is well documented that people can feel and express emotions through Computer-Mediated Communication (CMC) even if it is asynchronous and text-based [@Walther]. Sentiment analysis is a growing area of Natural Language Processing with research ranging from document level classification [@Pang2008] to learning the polarity of words and phrases [@Esuli06]. The sudden spurt of Twitter has enabled the emotion identification of several people at the same time for a specific subject, hence a plethora of studies have focused on Twitter Sentiment Analysis (TSA) research field. Nowadays, companies, media organizations and politicians strategy are affected by their Twitter popularity, since they can hear the common opinion on a daily basis [@Kouloumpis; @Wang2012]. Twitter data streams are generated continuously at each trice offering the opportunity for a more realistic society’s reflection on various issues. First steps towards this direction were made by Kalucki [@Kalucki] providing a publicly available twitter streaming Application Programming Interface (API). Thenceforth, the ’Twitter Streaming API’ has been created by Twitter, allowing anyone to retrieve at most one percent sample of all the data by providing some parameters [^1]. Twitter data streams pose several challenges for the data mining field, such as managing the limited resources (time and memory) and dealing with data shifts across time [@bifet11].
Sentiment analysis software reads a text and uses an algorithm to produce an estimate of its sentiment content. This estimate can be in several different forms: binary - either positive/negative or objective/subjective; trinary - positive/neutral negative; scale - e.g. -5 (strongly negative) to 5 (strongly positive); dual scale - e.g. 1 (no positivity) - 5 (strong positivity) and -1 (no negativity) - -5 (strong negativity); and multiple - e.g. happiness (0-100), sadness (0-100), fear (0-100). Sentiment analysis algorithms tend to use either a machine learning or a lexical approach [@Thelwall2017].
Related Work
============
Recently, studies for Twitter sentiment analysis have focused on design and implementation of scalable systems. These systems can be categorized in real-time systems and systems for batch processing [@karanasou2016scalable]. Towards the direction of real-time systems is the work of Wang et al. [@wang2012system] where they proposed a system for real-time sentiment analysis on Twitter streaming data towards presidential candidates (US 2012). However, their system is based on a crowd-sourcing approach to do sentiment annotation. In [@karanasou2016scalable] a real-time architecture for scalable twitter sentiment analysis is presented, dealing with the dynamic content using a feedback mechanism in the sentiment analysis process, also incorporating supervised learning methods for sentiment analysis and additionally an off-line phase for feature extraction. Similarly, in [@calais2011bias] the authors introduced a transfer learning approach to performing real-time sentiment analysis. It is considered as the first study which measures the bias of social media users toward a topic providing evidence that user bias tends to be more consistent over time although the possible changes in the dynamic context (newcomer terms or old terms meaning change)
A crucial step in sentiment analysis is also to identify the users’ sentiment changes across the time. MOA-TweetReader, performs a stream mining from Twitter stream tweets, highlighting the sentiment changes [@bifet2011detecting]. It utilizes a feature generation filter to vectors of attributes or machine learning instances based on an incremental term frequency–inverse document frequency (tf-idf) weighting scheme. The system also applies the SPACE SAVING Algorithm [@metwally2005efficient] for mining and storing the frequency of the most frequent terms. More specifically, the system initialization is done by the first k distinct elements and their counts as they are stored in memory per k pairs elements (item and count). Then, it follows the rule which checks every time if the upcoming data has already been monitored. If the answer is yes or no, data its count is incremented by one or it replaces the least in count item and initializing its count by one. Afterward, MOA-TweetReader uses ADWIN [@bifet2007learning] as a change detector, an Adaptive sliding window algorithm, thus memory requirements may grow significantly depending on the window size. In addition, although the authors manage to collect author-provided sentiment indicators to build classifiers for sentiment analysis there is still an off-line training stage.
Methodology {#sec:meth}
===========
In this work, in an attempt to provide a lightweight solution that requires no training (no off-line phase) we employ the lexicon approach to characterize tweets. Then, we use change detection algorithms to detect opinion changes in a time series of sentiment scores taking advantage of the fact that such time series is bounded by nature (there is a limit in the number of words appearing in a post) and thus specifying appropriate parameters for control charts is relatively straightforward. In brief, the proposed methodology is constituted by two core parts:
- Collecting and characterizing tweets according to their sentiment.
- Detect significant changes in the time series of sentiment scores.
Constructing the Sentiment Time Series {#sec:stream}
--------------------------------------
Tweets are collected by the Twitter stream Application Program Interfaces (API) continuously constituting a data stream. The API only requires a valid Twitter account for authentication, while our analysis is based on the open source tools provided by the R-project [@R]. Using the “rtweet” package [@rtweet-package] we connect to the API to stream tweets filtered by keywords, for example, Twitter hashtags. Data are retrieved in JSON format and are easily parsed using the “rtweet” package. Once we retrieve text for each tweet we need to “clean” it removing hashtags, spaces, numbers, punctuations, URLs etc, employing functions from the “stringr” and “glue” packages respectively [@stringr; @glue]. Next, we process the resulting text by extracting tokens (tokenization) and retrieving only sentiment words using the “tidyverse” package [@tidyverse]. Tokenization is a process of creating a bag-of-words from the text where the incoming string gets broken into comprising words using white space in separating individual words. Usually, tokenization of social-media data is considerably difficult but the aforementioned cleaning process has proven to be very successful. Each word from the bag-of-words is compared against the lexicon and if the word is found, we update the sentiment score of the post accordingly. Examples of the existing lexicons include: Opinion Lexicon [@lexicon] which categorizes words in a binary fashion into positive and negative categories, AFINN Lexicon[@Nielsen] which assigns words with a score that runs between $-5$ and $5$, with negative scores indicating negative sentiment and positive scores indicating positive sentiment and SentiWordNet which assigns to each synset of WordNet three sentiment scores: positivity, negativity, objectivity [@Esuli06].
Change Detection using CUSUM
----------------------------
The cumulative sum (CUSUM) algorithm was first proposed by Page in [@page1954] for on-line and off-line change detection and it has been shown to be more efficient than Shewhart charts in detecting small shifts in the mean of a process. The CUSUM control chart have received a great deal of attention in modern industries while being an active research topic with various recent applications [@Phuong; @TASOULIS201387; @Georgakopoulos] and proposed variations or extensions [@Abujiya; @Perry; @Wangcusum].
In this work, we will utilize the online version of the CUSUM algorithm focusing on a technique connected to a simple integration of signals with adaptive threshold [@Basseville93detectionof]. To describe the change detection algorithm, we consider a sequence of independent random variables $y_k$, where $y_k$ is a sensor signal at the current time instant $k$ (discrete time), with a probability density $p_\theta(y)$ depending only upon one scalar parameter $\theta$. Before the unknown change time $t_0$, the parameter $\theta$ is equal to $\theta_0$, and after the change it is equal to $\theta_1 \neq \theta_0$. Then, the problem is to detect and estimate this parameter change. In this work, our goal is to detect the change assuming that the parameters $\theta_0$ and $\theta_1$ are known, which is a quite unrealistic assumption for practical applications. Usually parameters $\theta_0$ and $\theta_1$ can be experimentally estimated using test data, which we also not consider as available for the application at hand. However, for strongly bounded problems parameter values can be assumed relatively easy. For example, we may consider $\theta_0$ the state of a neutral conversation with sentiment scores gathered around $0$ and $\theta_1 = 1 $ when positive comments dominate the stream. As such we may use our prior knowledge about the signal to correctly set *the change magnitude*.
Next, we introduce the basic idea used in quality control. Samples are iteratively taken and at the end of each arrival a decision rule is computed to test the two following hypotheses concerning parameter $\theta$: $$\begin{split}
H_0:\theta = \theta _0,\\
H_1:\theta = \theta _1.
\end{split}$$ As long as the decision is in favor of $H_0$, the sampling and test continue. Sampling is stopped after the first sample of observations for which the decision is in favor of $H_1$. This sample also determines the stopping time.
In our case, the samples (tweets) are arriving at each time instant and the decision rule is computed. We will use the following notation. Let $$S_{k}=\sum_{i=1}^{k}s_i, \quad {\rm where} \quad
s_i=\ln \frac{p_{\theta _1}(y_i)}{p_{\theta _0}(y_i)},$$ is the log-likelihood ratio for the observations from $y_i$ to $y_k$ and $k$ be the current time instant. We refer to $s_i$ as sufficient statistic. Let us now consider the particular case where the distribution is Gaussian with mean value $\mu$ and constant variance $\sigma$. In this case, the changing parameter $\theta$ is $\mu$. The probability density is $$p_\theta (y)= \frac{1}{\sigma \sqrt{2\pi}} e^{-\frac{(y-\theta
)^2}{2\sigma ^2}} ,$$ and the sufficient statistic $s_i$ is $$s_i=\frac{\theta_1 - \theta_0}{\sigma ^2} \left(y_i - \frac{\theta_0 + \theta
_1}{2}\right).$$ The corresponding decision rule is then, at each time instant, to compare this difference to a threshold as follows: $$g_k=S_k - m_k \geqslant h, \quad\rm{where}\quad m_k= \min_{1\geqslant j \geqslant k} S_j.$$ The stopping time is $$t_a=\min\{k:g_k \geqslant h \},$$ which can be rewritten as $$t_a=\min\{k:S_k \geqslant m_k + h \}.$$ This decision rule is a comparison between the cumulative sum $S_k$ and an adaptive threshold $m_k + h$. Because of $m_k$, this threshold not only is modified on-line but also keeps the complete memory of the entire information contained in the past observations. Moreover, in the case of a change in the mean of a Gaussian sequence, $S_k$ is a standard integration of the observations.
The detection threshold $h$ is a user-defined tuning parameter in which the appropriate form for its determination is based on the average run length function, which is defined as the expected number of samples before an action is taken [@page1954]. More precisely one has to set the mean time between false alarms $ARL_0$ and the mean detection delay $ARL_1$. These two specific values of the ARL function depend on the detection threshold $h$, and can thus be used to set the performance of the CUSUM algorithm to the desired level for a particular application [@Granjon14thecusum].
### Two-sided Algorithm and Resets
The described algorithm, which is called one-sided CUSUM, focuses on change detections in one direction only. However, in our application, it is necessary to detect changes in each direction discovering both positive and negative sentiment change of twitter posts. For this purpose, two one-sided algorithms were used, one to detect an increase and the other to detect a decrease in the parameter $\theta$. This leads to two different instantaneous log-likelihood ratios. As such, two cumulative sums and two decision functions are computed, while a change is detected by testing both decision functions simultaneously.
When a change is triggered for any of the two-sided functions, the CUSUM algorithm reset to zero and a re-initialization takes place. The algorithm restarts with a new value for $\theta _0$ equal to the average of the few last observations. As such we define a new control state representing the current sentiment at time $t$, while $\theta _1$ is recalculated based on a fixed *change magnitude*. The rest of the parameters remain the same. Under this perspective, we deal with the major challenge in twitter streaming data, which is the sentiment trend detection under a dynamic estimation. This is crucial since a hashtag trend could change several times across its lifetime and thus changes should be estimated based on its current sentiment state rather than its initial state. Towards this direction, our methodological framework can imprint on-line, the change detections of a hashtag by updating its initial state. A flowchart diagram of the methodology is presented in Figure \[fig:flow\]
Experimental Results
====================
The described methodological framework was applied on a Twitter dataset streamed from 15-03-2018 to 24-03-2018 using the hashtag “theresamay”. Obviously, this term refers to Theresa Mary May, Prime Minister of the United Kingdom and Leader of the Conservative Party since 2016. Our choice lies in the fact that politic-related Twitter hashtags offer a satisfying opportunity for testing sentiment changes since these are not one-sided (supporters are on both sides while critics are involved as well). In addition, this hashtag was selected considering that during this period the “Brexit” news topic attracted attention due to further discussions amongst high-level politicians regarding the relationships between the United Kingdom and European Union. $15491$ posts are included in the streamed dataset after discarding non-English language posts. After sequentially applying the procedure described in Section \[sec:stream\] we retrieved the time series of sentiment score presented in Figure \[fig:sent\]. To calculate the sentiment score for each tweet we used the “bing” lexicon [@lexicon] applied to the extracted words.
In an attempt to improve a visual inspection of the possible changes in sentiment, the corresponding moving average of sentiment scores was employed. Figure \[fig:ma\] illustrates the calculated moving average for a window of total size $200$. As shown, we may visually discriminate the areas that could be characterized by changing distributions.
We observe that the whole thread is characterized by a slightly negative sentiment, which can be confirmed by the histogram of sentiments (see Figure \[fig:hist\]).
In what follows, the results of the change detection algorithm are reported. For the initialization of parameters values we set $\theta_0 = -0.5$, while the *change magnitude* is set to $0.5$; thus, we consider $\theta_1^{pos} = 0$ and $\theta_1^{neg} = -1$ for the two-sided CUSUM, respectively. In Figure \[fig:cusum\] we show the CUSUM functions and the corresponding changes retrieved by the algorithm with vertical lines. We may recall that each time a change is detected the algorithm restarts with an updated initialization. Selecting the $h$ parameter value is usually subject to question and depends on the user requirements. For this particular topic of study in our analysis we assume that a small number of reported changes per day are sufficient and thus we define $h=20$.
To visually investigate the reported change points we employed the calculated moving average presented in Figure \[fig:ma\]. Vertical lines are depicted in the plot (see Figure \[fig:cusumma\]) with the corresponding colors from Figure \[fig:cusum\], where the blue lines correspond to negative changes, while the yellow lines correspond to positive changes. At this point, we can conclude that reported change points agree with a simple visual investigation of possible changes. In order to further verify this outcome, a very successful off-line methodology for change detection was employed. It is capable to discover multiple change points [@Killick; @changepoint] and estimate their number in the time series automatically if required. The penalty parameter of this algorithm used for this test was the default value $2*\log(n)$, where $n$ is the total number of samples. The $Q$ parameter specifying the maximum number of estimated change points is set equal to the total number of change points retrieved by the two sided CUSUM algorithm. The retrieved change points are also reported in Figure \[fig:cusumma\] with red vertical lines. We observe that this result highly agrees with the change points retrieved by CUSUM with only minor report of delays between the estimations.
The last step of the proposed methodology contains the content of two consecutive separate parts of the time series split by change points declaring the transition to a more negative sentiment and subsequently to a more positive sentiment. The first area ranges from the beginning of the time series until the first reported negative change point (sample count $2142$ and corresponding date “2018-03-15 23:59:29 UTC”) while the second part covers the area between the aforementioned point and the first positive change point (sample count $4427$ and corresponding date “2018-03-16 17:42:08 UTC”). For each part, we report word-clouds plotting the frequency of words appearing in the corresponding collection of Twitter posts. The text is being preprocessed using the Term Frequency-Inverse Document Frequency (TF-IDF) methodology [@tfidf], where the frequency of words is rescaled by how often they appear in all tweets of the category, penalizing most frequent words (see Figure \[fig:wc1\]). It appears that there is discrimination between the word-clouds, while, we interestingly observe that in a hashtag about Teresa May the word-cloud that corresponds to the more positive part (before the negative change) is characterized by the appearance of words like “corbyn” and “labour”.
Concluding Remarks
==================
Twitter keeps attracting the interest of research community as in contrast to other social networks like Facebook or LinkedIn, its extensive data availability provides great potential for Machine Learning research. Sentiment analysis is a growing area of Natural Language Processing, which taking advantage of Twitter user activity has enabled the real-time emotion identification of large groups of people for a specific subject. In this work, based on open source tools, we propose a complete methodology for continues real-time detection of sentiment changes in Twitter conversations. Focusing on simplicity, our methodology does not require training or an off-line phase, while memory and computational requirements are minimal. The experimental analysis provides enough justification of the usefulness of the methodology and exposes a great potential for further use. In our future research, we indent to examine approaches that enhance the robustness of the change detection algorithm, while further testing methodologies for sentiment characterization.
Acknowledgment {#acknowledgment .unnumbered}
==============
This research has been financially supported by the National Strategic Reference Framework (NSRF) Program with title: “Researcher Support with Emphasis on New Researches”, co-financed by the European Union (European Social Fund – ESF) and Greek national funds.
[^1]: [
https://dev.twitter.com/docs/streaming-apis](
https://dev.twitter.com/docs/streaming-apis)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove an adiabatic theorem for the ground state of the Dicke model in a slowly rotating magnetic field and show that for weak electron-photon coupling, the adiabatic time scale is close to the time scale of the corresponding two level system–without the quantized radiation field. There is a correction to this time scale which is the Lamb shift of the model. The photon field affects the rate of approach to the adiabatic limit through a logarithmic correction originating from an infrared singularity characteristic of QED.'
address: 'Department of Physics, Technion, 32000 Haifa, Israel'
author:
- 'J. E. Avron and A. Elgart'
title: 'An Adiabatic Theorem without a Gap Condition: Two level system coupled to quantized radiation field'
---
Introduction
============
In this work we investigate the relation between adiabatic theorems for models that, like QED, allow for the creation and annihilation of photons, and the corresponding quantum mechanical models where the electron is decoupled from the photon field. We study this problem in the context of a specific and essentially soluble model: The Dicke model [@d]. The corresponding quantum mechanical model is a two level system, such as a spin in an adiabatically rotating magnetic field, which is a basic paradigm of adiabatic theory [@berry].
In the usual quantum adiabatic theorem [@bf; @kato] the gap between eigenvalues plays an important role: It fixes the adiabatic time scale and determines the rate at which the adiabatic limit is approached. There is no such gap in the corresponding QED models so the nature of the adiabatic theorem in the two cases has qualitatively different features. For example, there is no gap in the spin-boson and Dicke models (for weak coupling) both of which describe a two level system in a radiation field.
The first problem we address is whether there is an adiabatic theorem for the ground state in a radiation field. Assuming a positive answer, the second question is, what property of the QED model, plays the role of the gap in the adiabatic theorem. Another way of phrasing this question is how does the adiabatic time scale of the two level system compare to that of the QED model? Are the two close in the limit of small fine structure constant, $\alpha$, and if so, how close? The third question compares the rate of approach to the adiabatic limit in the two models.
Consider a two level system, such as a spin or a twofold Zeeman split atomic level, in an external magnetic field pointing in the z direction. When radiation effects are neglected, the corresponding Hamiltonian is $$H= m\sigma_z,\quad m=\mu B .$$ The corresponding Dicke model is $$H_D=H\otimes {\bf 1} + \alpha^{-1}\, {\bf 1}\otimes E +
\sqrt{\alpha}\,\sigma_+\otimes a^{\dagger}(f) +
\sqrt{\alpha}\,\sigma_-\otimes a(f),$$ where $$E=\int{|k|\, a^{\dagger}(k)a(k)d^dk},$$ and[^1] $$f(k)=\,\sqrt{\frac{2\pi}{|k|}}
\,\langle\psi_1\vert\,\{e^{-i{k}\cdot {x}}, p\}\,\vert\psi_2\rangle,
\label{fk}$$ with $\psi_j(x)$ the atomic wave functions of the two level system. Polarization indices are omitted since the helicity of the photon does not play an interesting role in the questions we study. We use atomic units where $e=\hbar=1$ so $\alpha=1/c=1/137$ is small. $\mu$, the magnetic moment, is also of order $\alpha$ in these units.
Following M. Berry’s [@berry] let us consider the case where the magnetic field changes its direction adiabatically and has fixed magnitude. The time dependent Hamiltonian for the two level system is $$H(s)=\mu B(s)\cdot\sigma = U(s)HU^*(s),$$ with $U(s)\in SU(2)$ the appropriate rotation. The corresponding adiabatic Dicke model has the time dependent Hamiltonian: $$H_D(s)=\left(U(s)\otimes {\bf 1}\right)\, H_D\,
\left(U^*(s)\otimes {\bf 1}\right).$$ Our aim is to compare the adiabatic evolution of the ground state of $H(s)$ with that of $H_D(s)$.
Adiabatic theorems for quantum systems coupled to a field have been studied in [@nt; @ds]. In [@nt] Narnhofer and Thirring give characterization of extremal KMS states by adiabatic invariance. When applicable, this result says, in particular, that the ground state is adiabatic invariant. The characterization depends on a condition of asymptotic abelianess which does not hold for the models we consider. In [@ds] Davis and Spohn give a derivation of linear response theory for a system coupled to a bath in the adiabatic limit. The notion of adiabaticity in this work is such that the coupling between the field and the quantum system vanishes in the adiabatic limit. This is not a standard notion of adiabaticity.
Let us now describe our results. First, we show that there is an adiabatic theorem for the ground state of the Dicke model, even though the model has no spectral gap to protect the ground state. Second, we show that the the distance to a nearby resonance in the Dicke model plays the role of a gap. Third, we show that the adiabatic time scale for the Dicke model and the two level system agree in the limit of small $\alpha$. The (inverse) of two time scales differ by the Lamb shift of the Dicke model. And finally, we show that the approach to the adiabatic limit in the two models is different: While in the two level system the approach to the adiabatic limit is with an error $O\left(\frac{1}{\tau}\right)$, the approach to the limit in the Dicke model is with an error of $O\left(\frac{\sqrt{\log \tau}}{\tau}\right)$. The logarithm comes from an infrared divergence characteristic of QED.
Although the results we derive here are for a rather special model we suggest that something similar happens also for more realistic models. The success of the quantum in numerous applications that depend on a correct prediction of the adiabatic time scale is evidence that at least the time scale aspect of our results may well carry over to more realistic models. It would be interesting to know if this is indeed the case for the Spin-Boson model [@hs; @dg; @ms]. The spin-boson model is a more realistic QED version of a two level system, which, unlike the Dicke model, is not explicitly soluble. However, as much progress in the spectral analysis of the spin-boson problem has been recently made, the problem we pose here may be a reasonable challenge.
The Adiabatic Theorem and A Commutator Equation
===============================================
In this section we explain what we mean by “adiabatic theorem”, and give a condition for an adiabatic theorem to hold. This condition is that the commutator equation, Eq. (\[commutators\]) below, has solutions $X$, $Y$ which are bounded operators[^2]. We also introduce notation, terminology, and collect known facts that we need. To simplify the presentation, we shall stay away from making optimal assertions.
We consider Hamiltonians that are bounded from below, and choose the origin of the energy axis so that the spectrum begins at zero. Let $H(s)\ge 0$ be a family of such self-adjoint Hamiltonians. The unitary evolution generated by the Hamiltonian, $U_\tau(s)$, is the solution of the initial value problem: $$i\,\dot U_\tau (s) = \tau H(s) U_\tau(s),\quad
U_\tau(0)=1,
\quad s\in [0,1].\label{schrodinger}$$ $\tau$ is the adiabatic time scale, and we are concerned with the limit of large $\tau$. The physical time is $t=\tau s\in[0,\tau]$. Since $\tau$ is large $H(s)=H(t/\tau)$ varies adiabatically. We assume that all operators are defined on some fixed dense domain in the Hilbert space.
The (instantaneous) ground state is in the range of the kernel of $H(s)$ and we assume that the kernel is smooth and one-dimensional. Let $P(s)\neq 0$ be the projection on the kernel of $H(s)$, i.e. $H(s)\,
P(s)=0$, $dim\,P= Tr\,P=1$. By smoothness we mean that $\dot P(s)$ a bounded operator.
The adiabatic theorems we consider are concerned with the large time behavior of the evolution of the ground state where $t=O(\tau)$ or, equivalently,$s=O(1)$. The smoothness of the kernel implies that there is a natural candidate for an adiabatic theorem for the ground state, which is independent of whether $H(s)$ does or does not have a gap in it spectrum. Namely, that if $\psi(0)\in Range \, P(0)$ at time $s=0$, then it evolves in time so that, $\psi_\tau(s)=U_\tau (s)\ \psi(0)$ lies in $Range\, P(s)$ at time $s$ in the adiabatic limit, $\tau\to\infty$.
To formulate the adiabatic theorem with error estimates we need to get hold of [*adiabatic phases*]{} [@berry]. To do that we introduce the adiabatic evolution of Kato [@kato]: Let $U_A(s)$ be the solution of the evolution equation $$\dot U_A(s) = [\dot P(s),P(s)]\, U_A(s),\quad U_A(0)=1,
\quad s\in [0,1].\label{kato}$$ It is known that $$U_A(s)\, P(0) =P(s)\, U_A(s).$$ That is $U_A(s)$ maps $Range\ P(0)$ onto $Range\ P(s)$. We can now formulate the basic adiabatic theorem :\
\
[**[Theorem II.1]{}**]{} Let $H(s)P(s)=0$ for all $0\le s\le 1$, with $P$ differentiable projection on the ground state, with $\Vert
\dot P(s)\Vert\le D$. Suppose that the commutator equation $$[\dot P(s), P(s)] = [H(s),X(s)] +Y(s),\label{commutators}$$ has operator valued solutions, $X(s)$ and $Y(s)$ so that for $\varepsilon\searrow 0$ $$\Vert X(s)\Vert\ + \Vert \dot X(s)\Vert\ \le\ C\left\{ \begin{array}{c}
\,\varepsilon^{-\nu}\\ \vert\log \varepsilon\vert
\end{array} \right. ,\quad \Vert Y(s)\Vert \le \hat C\,\varepsilon^{\mu},$$ with $\mu,\nu\ge 0$. Then $$\Vert (U_\tau (s)-U_A(s))P(0)\Vert \le \tilde C \left\{
\begin{array}{c}\tau^{-\frac{\mu}{\nu+\mu}}
\\ \frac{\log \tau}{\tau}\end{array}\right. ,\quad s\in[0,1].$$
Remarks: 1. In the case that there is a gap in the spectrum, one can always find $X(s)$ bounded so $\nu=0$, and $Y=0$, see [@asy]. $X$, and therefor also $\tilde C$, is of the order of (gap)$^{-1}$. This gives error of $1/\tau$, and generalizes the adiabatic theorem of Born and Fock and Kato for discrete spectra, to more complicated spectra provided there is a gap.
2\. The theorem says that the physical evolution clings to the instantaneous spectral subspace. In particular, if $P$ is one dimensional, it says that the physical evolution of the ground state remains close to the instantaneous ground state.
3\. Here, and throughout, we are concerned only with the adiabatic theorem to lowest order. If $s$ is chosen outside the support of $\dot P$ then much stronger results can be obtained. See e.g. [@ks].
4\. The adiabatic time scale $\tau_0$ set by this theorem is $\tau_0=
O((2+D)C)$.
Proof: Let $W(s) = U_A^\dagger(s) U_\tau(s))$, with $W(0)=1$. From the equation of motion, and the commutator equation, Eq. (\[commutators\]), $$\begin{aligned}
P(0)\,\dot W(s)\,&=& -P(0) U^\dagger_A(s)\Big(i\,\tau\,H(s)+[\dot
P(s),P(s)]\Big) U_\tau(s)\nonumber
\\ &=&-U^\dagger_A(s)\,P(s)\Big(i\,\tau\,H(s)+[\dot
P(s),P(s)]\Big)U_\tau(s)
\nonumber
\\ &=&\,-U^\dagger_A(s)\,P(s)\,[\dot P(s),P(s)]\, U_\tau(s)
\nonumber
\\ &=&\,-U^\dagger_A(s)\,P(s)\,\Big([H(s),X(s)]+Y(s)\Big)\,U_\tau(s)
\nonumber
\\ &=&\,-U^\dagger_A(s)\,P(s)\,\Big(-X(s)\,H(s)+Y(s)\Big)\,U_\tau(s)
\nonumber
\\ &=&\frac{i}{\tau}\,P(0)\, U^\dagger_A(s)\,X(s)\,
\dot U_\tau(s)-P(0)\,U^\dagger_A(s)\,Y(s)\, U_\tau(s).\end{aligned}$$ To get rid of derivatives of $U_\tau$, which are large by the equation of motion, we rewrite the first term on the rhs (up to the $P(0)$ on the right) as : $$\begin{aligned}
U^\dagger_A(s)\,X(s)\,\dot U_\tau(s) &=&\dot{\left( U^\dagger_A(s)\,X(s)
U_\tau(s)\right)}- U^\dagger_A(s)\,\dot X(s)\, U_\tau(s)-
\dot U^\dagger_A(s)\,X(s)\,U_\tau(s) \\&=&
\dot{\left( U^\dagger_A(s)\,X(s)\, U_\tau(s)\right)} - U^\dagger_A(s)\,\dot
X(s)\, U_\tau(s)+ U^\dagger_A(s)\,[\dot P(s),P(s)]\,X(s)\,
U_\tau(s).\nonumber\end{aligned}$$ From this it follows, by integrating, that for $s\in[0,1]$ $$\begin{aligned}
\Vert (U_\tau(s)-U_A(s))\,P(0)\Vert &=& \Vert
P(0)(U^\dagger_\tau(s)-U^\dagger_A(s))\,\Vert\nonumber \\
\Vert P(0)(1-W(s))\,\Vert&\le& \hat C\varepsilon^\mu +\frac{(2+D)\,
C}{\tau}\left\{
\begin{array}{c}{\varepsilon^{-\nu}}
\\ \vert\log\varepsilon\vert.\end{array}\right.\end{aligned}$$ Choosing $\varepsilon= \tau^{-\frac{1}{\mu+\nu}}$ gives $$\Vert (U_\tau(s)-U_A(s))\,P(0)\Vert \le \tilde C \left\{
\begin{array}{c}\tau^{-\frac{\mu}{\nu+\mu}}
\\ \frac{\log \tau}{\tau}\end{array}\right. .$$ This concludes the proof of the theorem.$\hfill\vbox{\hrule height
0.6pt
\hbox{\vrule width 0.6pt height 1.8ex \kern 1.8ex\vrule width 0.6pt}
\hrule height 0.6pt}$
It is convenient to rewrite this solvability condition in a way that one needs to solve for a fixed $X$ and $Y$ rather than functions $X(s)$ and $Y(s)$. This is accomplished by\
\
[**[Corollary II.1.1]{}**]{} Let $P(s)$ be the family $$P(s)=V(s)\, P\,V^\dagger(s),\quad V(s)=\exp
(i\,s\,\sigma).$$ It is enough to solve for the commutator equation $$i\, K=[H, X] +Y, \quad K= \{\sigma,P\} -2 P\sigma P,\label{commute}$$ for fixed $X$ and $Y$ so that for $\varepsilon\searrow 0$ $$\Vert X\Vert\ \le\ C\left\{ \begin{array}{c}
\,\varepsilon^{-\nu}\\ \vert\log \varepsilon\vert
\end{array} \right. ,\quad \Vert Y\Vert
\le \hat C\,\varepsilon^{\mu},$$ with $\mu,\nu\ge 0$, and $\Vert \dot P(s)\Vert\le D$. $X(s)$ and $Y(s)$ are then determined by the obvious unitary conjugation. Proof: Since $P(s)=V(s)\, P\,V^\dagger(s)$, we have $$\dot P(s)=i\,V(s)\, [\sigma,P]\,V^\dagger(s),$$ and $$\begin{aligned}
[\dot P(s),P(s)]&=&i\,V(s)\,
\Big[[\sigma,P],P\Big]\,V^\dagger(s)\nonumber \\ &=&i\,V(s)\,
\Big(\{\sigma,P\}-2P\,\sigma\,P\Big)\,V^\dagger(s).\end{aligned}$$ $\hfill\vbox{\hrule height 0.6pt\hbox{\vrule width 0.6pt height 1.8ex
\kern 1.8ex \vrule width 0.6pt}\hrule height 0.6pt}$
An Adiabatic Theorem for a Threshold State: The Friedrichs Model
=================================================================
As a warmup, and as a preparation for the analysis of the Dicke model, let us prove an adiabatic theorem for the Friedrichs model which has a bound state at the threshold of the continuum.
There is an inherent difficulty in the situation of a bound state at threshold in general, and in the Friedrichs model [@friedrichs; @friedrichs1; @howland] in particular, namely, that a bound state at threshold is not a stable situation. Under a small deformation of the Hamiltonian, the ground state will, generically, split away from the absolutely continuous spectrum and a gap develops. Since our aim is to study families related by a unitary, this problem does not appear. That is, we consider the family $H_F(s)=V(s)\,H_F\,V^\dagger(s)$ where $H_F$ has a bound state at threshold and $V(s)$ is a smooth family of unitaries.
The Friedrichs Model
--------------------
We shall consider a family of Hamiltonians, closely related to the standard Friedrichs model [@friedrichs], parameterized by the scaled time $s$, a real number $d>0$ that plays the role of dimension, and a function $f$ that describes the deformation of the family. Since we are only interested in the low energy behavior of the family we shall introduce an “ultraviolet cutoff” to avoid inessential difficulties.
The Hilbert space of the Friedrichs model (with an ultraviolet cutoff) is ${\cal H}=\kern.1em{\raise.47ex\hbox{ $\scriptscriptstyle
|$}}\kern-.40em{\rm C}\oplus L^2([0,1],k^{d-1}\,dk)$. A vector $\psi\in{\cal H}$ is normalized by $$\psi=\left(\begin{array}{c}\beta\\ f(k)
\end{array}\right)\, \quad \Vert \psi\Vert^2= |\beta\vert^2+\int_0^1 \vert
f(k)\vert^2 k^{d-1}\,dk,\quad
\beta\in\kern.1em{\raise.47ex\hbox{ $\scriptscriptstyle
|$}}\kern-.40em{\rm C}.$$ We choose a special, and trivial, case of a diagonal Hamiltonian whose action on a vector $\psi$ is as follows: $$H_F\,\psi=\left(\begin{array}{ll} 0&0\\ 0&k
\end{array}\right)\,\left(\begin{array}{c}\beta\\ f(k)
\end{array}\right)=\left(\begin{array}{c} 0\\ k\,f(k)
\end{array}\right).$$ $H$ has a ground state at zero energy with projection $$P=\left(\begin{array}{ll} 1&0\\ 0&0
\end{array}\right).$$ The rest of the spectrum is the unit interval $[0,1]$, and is absolutely continuous. The density of states in this model is proportional to $E^{d-1}$.
We construct the family $H(s)$ by conjugating $H$ with a family of unitaries: $$V_f(s)=\exp is\sigma(f), \quad
\sigma(f)=\left(\begin{array}{ll} 0&\langle f\vert \\ \vert f \rangle&0
\end{array}\right),$$ where $f$ is a vector in $L^2([0,1],k^{d-1}\, dk)$.\
\
[**[Theorem III.1]{}**]{} Let $H_F(s;d,f)$ be the family of Friedrichs models with a ground state at threshold for all s $$H_F(s;d,f)= V_f(s)\, H_F\,V^\dagger_f(s).$$ Suppose that $$g(k)=i\, k^{-1}\, f(k)\in L^2([0,1],k^{d-1} dk), \quad
V_f(s)=\exp\, i\,s\,\sigma(f)\, ,$$ then the quantum evolution of the ground state of $H_F(s;d,f)$ is adiabatic and its deviation from the instantaneous ground state is, at most, $O(1/\tau)$. Remarks: 1. Note that if the conditions in the theorem hold in dimension $d_0$, then they hold in all dimensions $d\ge d_0$. The physical interpretation of that is that the density of states at low energies decreases with $d$. So, even though there is spectrum near zero, there is only very little of it.
2\. If $g$ is not in $L^2$ there may still be an adiabatic theorem with slower falloff in $\tau$ by accommodating $Y\neq 0$. An example will be discussed in the next section.
3\. The Friedrichs model is vanilla: $H_F$ has no interesting energy scale to fix the adiabatic time scale. The scale is set by the perturbation alone: $\tau_0= O((1+\Vert f\Vert^2)\,\Vert g\Vert )$. This is quite unlike the case in the usual adiabatic theorem and unlike what we shall show for the Dicke model.
Proof: In this case $K$ of Corollary 2.1 is $K=\sigma(f)$. With $g\in
L^2$, $\sigma(g)$ is a bounded (in fact, finite rank) operator and an easy calculation gives $$[H_F, \sigma(g)]=\left(\begin{array}{lr}0&\langle -kg|\\
|kg\rangle&0\end{array}\right)=i\sigma(f).$$ Hence $$X= \sigma(g) ,\quad Y=0\, ,$$ solve the commutator equation, Eq. (\[commute\]), with a bounded $X(s)$ and $Y(s)=0$.$\hfill\vbox{\hrule height 0.6pt\hbox{\vrule
width 0.6pt height 1.8ex \kern 1.8ex \vrule width 0.6pt} \hrule height
0.6pt}$
Adiabatic Theorem for the Dicke Model
=====================================
In this section we describe an adiabatic theorem for the Dicke model [@d] that says that the an adiabatic rotation of a two level system evolves the ground state so that it adheres to the instantaneous ground state and the time scale, at least in three dimensions, is essentially the time scale fixed by Quantum Mechanics without photons. The rate of approach to the adiabatic limit is different from that of a two level system and has a logarithmic correction in three dimensions. This section also collects known facts about the Dicke model that we need.
The Dicke Model
---------------
The Spin-Boson Hamiltonian in the canonical QED version of a two level system [@hs; @dg; @ms]. The Dicke model is a simplified version of the Spin-Boson Hamiltonian in the rotating wave approximation. The rotating wave approximation, can indeed be motivated in the single-mode Dicke model. In the multi-mode case we consider the rotating wave approximation is a name that describes which terms in the Spin Boson Hamiltonian are kept and which are not.
The model describes a two level system coupled to a massless boson field in $d$ dimensions. The Hamiltonian is: $$H_D(m,d,f,\alpha)=m\,(1 -P)\otimes {\bf 1} + \alpha^{-1}\,
{\bf 1}\otimes E +
\sqrt{\alpha}\,\sigma_+\otimes a^{\dagger}(f) +
\sqrt{\alpha}\,\sigma_-\otimes a(f),\label{eq1}$$acting on the Hilbert space $\kern.1em{\raise.47ex\hbox{ $
\scriptscriptstyle|$}}\kern-.40em{\rm C}^2\otimes {\cal F}$ with ${\cal
F}$ being the symmetric Fock space over $L^2(R^d,d^dk)$. Here $$P=\left(\begin{array}{ll} 1&0\\
0&0\end{array}\right),\quad
\sigma_+=\left(\begin{array}{ll} 0&1\\ 0&0\end{array}\right),
\quad
\sigma_-=\left(\begin{array}{ll} 0&0\\ 1&0\end{array}\right)\quad
E=\int{|k|\, a^{\dagger}(k)a(k)d^dk}.$$ $m>0$ is the gap in the quantum Hamiltonian (without photons). $a(f)$ and $a^\dagger(f)$ are the usual creation and annihilation operators on ${\cal
F}$ obeying the canonical commutation relations $$[a(f),a^{\dagger}(g)]=\langle f\vert g\rangle.$$ We denote by $|0\rangle$ the field vacuum and by $\Omega$ the projection on the vacuum.
It may be worthwhile to explain where the various powers of $\alpha$ in $H$ come from. For the radiation field the $\alpha^{-1}$ comes from $\hbar\omega =\hbar c |k|$ which explains why the field energy comes with a large coupling constant. The $\sqrt{\alpha}$ has one inverse power of $c$ from minimal coupling, $\frac{e}{2mc} (p\cdot A+A\cdot p)$. Half a power of $\alpha$ comes from the standard formula for the vector potential $${A}({x}):=
\int d^3k\,\sqrt{\frac{2\pi c}{|k|}}\,\Big(e^{-i{k}\cdot
{x}}\,a^{\dagger}({k}) +e^{i{k}\cdot{x}} a({k})\Big).$$ Compare e.g. [@dicke].
With reasonable atomic eigenfunctions, $f(k)$, Eq. (\[fk\]) has fast decay at infinity and the model is ultraviolet regular. In the infrared limit $f(k)$ behaves like $$f(k)\to-i\,\sqrt{\frac{2\pi}{|k|}}
\int\,\Big(\psi_1^*(x)\,(\nabla\psi_2)(x)\,
-(\nabla\psi_1)^*(x)\,\psi_2(x)\Big)\,d^dx.$$ In particular we see that for small $k$ $$f(k)=K\,\sqrt{\frac{1}{ |k|}}.\label{K}$$ The square root singularity is a characteristic infrared divergence of QED, and it has consequences for the adiabatic theorem as we shall see. Note that with $f$ having a square root singularity the model makes sense (as an operator) provide $d>1$, for otherwise $a^\dagger(f)$ is ill defined since $f$ is not in $L^2$.
An important parameter in the model is $${\cal E}= \left\langle f \left\vert
\frac{1}{|k|}\right\vert f\right\rangle.$$ Bearing in mind the square root singularity of $f$ we see that $${\cal E}\sim \int \frac{d^dk}{|k|^2},$$ is finite for all $d>2$.
Spectral Properties
-------------------
What makes the Dicke model simple is that it has a constant of motion [@hs]. If we let $N=\int a^\dagger(k)
\, a(k) d^dk$ be the photon number operator, then ${\cal N}$ commutes with $H_D$ where $${\cal N}=\left(\begin{array}{ll} N&0 \\
0&N+1\end{array}\right)={\bf 1}\otimes N + P\otimes {\bf 1}.$$ The spectrum of ${\cal N}$ is the non-negative integers. The spectral properties of $H_D(m,d,f,\alpha)$ can be studied by restricting to subspaces of ${\cal N}$.
#### ${\cal N}=0$
: The kernel of ${\cal N}$ is one dimensional and is associate with the projection $$P=\left(\begin{array}{ll}\Omega&0 \\
0&0\end{array}\right).$$ $\Omega$ is the projection on the field vacuum. It is easy to see that $P\,H_D(m,d,f,\alpha)P=0$, so the model always has a state at zero energy. This state may or may not be the ground state. It is the ground state if $\alpha^2 {\cal E} <m$ [@hs].
#### ${\cal N}=1$
: The space is basically ${\cal H}$ of the Friedrichs model. The correspondence of vectors in the two spaces is $$\left(\begin{array}{c}a^\dagger(g)\\ \beta
\end{array}\right)\,|0\rangle\leftrightarrow \left(\begin{array}{c}g\\
\beta
\end{array}\right).$$ The Hamiltonian action in the Friedrichs model language is: $$H_D(m,d,f,\alpha)\leftrightarrow\left(\begin{array}{ll}\frac{|k|}{\alpha}&|
\sqrt{\alpha}\,f\rangle\\ \langle \sqrt{\alpha}\, f|&m\end{array}\right)\,.$$ It is a standard fact about the Friedrichs model [@friedrichs; @friedrichs1] that provided $$\alpha^2 {\cal E} <m,\label{standard}$$ the model has no bound state, and the spectrum is $[0,\infty)$ and is absolutely continuous. Since $f$ has square root singularity at the origin, (and has fast decay at infinity), this condition holds for $d\ge 3$ if $\alpha $ (or $f$) is not too large. In three dimensions, provided that the level spacing $m>>\alpha^2 $ in atomic units, (about $10^{-3}$ eV), the inequality holds. In two dimensions the left hand side is log divergent, and the spectrum in the ${\cal N}=1$ sector has a bound state at negative energy. This state lies below the bound state of the ${\cal N}=0$ sector. We do not consider this situation and henceforth stick to $d\ge 3$.
#### ${\cal N}\ge 2$
: It is known [@hs] that the bottom of the spectrum in all these sectors is at zero if (\[standard\]) holds.
Adiabatic Rotations
-------------------
Suppose, that the two level system of the Dicke model describes e.g. two Zeeman split energy levels of an atom in constant external magnetic field $B$ pointing in the z direction. Rotations about the z axis do not change the orientation of the magnetic field, and commute with $\cal N$ and are uninteresting. Rotations about the x axis change the orientation of the magnetic field and are implemented by $$V(s) =\exp\, (i\,s\, \sigma_x)\otimes {\bf 1}.\label{v}$$ Such rotations do not commute with ${\cal N}$. Indeed, $$=\left(\begin{array}{rl}0&{ 1}\\
-{1}&0\end{array}\right)\otimes {\bf 1}=J\otimes {\bf 1},\quad \sigma=
\sigma_x\otimes {\bf 1}.$$
The Adiabatic Theorem
---------------------
[**[Theorem IV.1]{}**]{} Let $H_D(s;m,d,f,\alpha)=V(s) H_D(m,d,f,\alpha )V^\dagger(s), \
s\in[0,1]$ be the family of time dependent Dicke models with $f$ square integrable, with square root singularity at $k=0$; $m>\alpha^2\,\langle f|\frac{1}{|k|}|f\rangle$; $d\ge 3$ and $V(s)=\exp \, (i\,s\,\sigma)$ as in Eq. (\[v\]). Then, $U_A$, the adiabatic evolution associated with the ground state of $H_D(s;m,d,f,\alpha )$, and $U_\tau$, the Schrödinger evolution are close in the sense that $$\Vert (U_A(s)- U_\tau(s))P(0)\vert \le
C\left\{\begin{array}{lr}\frac{1}{\tau}&\mbox{if $d>3$}\\
\frac{\sqrt{\log \tau}}{\tau}& d=3.\end{array}\right.$$ The time scale is determined by $m-\alpha^2 {\cal E}$ and coincides with the gap without photons, $m$, up to a correction by the Lamb shift, $\alpha^2 {\cal E}$. Proof: From Corollary 2.1 we find $K=\sigma\otimes\Omega$. We will first show that a solution of the commutator equation, Eq. (\[commute\]), for $d>0$, is $$X=\frac{i\,X_1-X_2(g)}{m-\alpha^2\,{\cal E}},\quad
Y=0,\label{x}$$ where $${X}_1 =J\otimes \Omega,
\quad X_2(g) = P\otimes(
a^{\dagger}(g)\,\Omega + {\rm h.c}), \quad g= i\,\alpha^{\frac{3}{2}}\,
\frac{f}{|k|}.\label{g}$$ Note that the gap of the two level system $m$ is renormalized to ${m+i\,\alpha\,\langle f|g\rangle}$, which is just the Lamb shift (See appendix). This is a small correction, of order $\alpha^2$.
A useful formula we shall need is $$E\, a^\dagger(g) \Omega =a^\dagger(|k|g)\Omega.$$
Let us compute the commutators of ${X}_1$, ${X}_2$ with $H$: $$\begin{aligned}
[H,{X_1}] &=&
\left[\left( \begin{array}{cc} \frac{E}{\alpha}&
\sqrt{\alpha}\,{a^\dagger(f)}
\\ \sqrt{\alpha}\, {a(f)}&m + \frac{E}{\alpha} \end{array} \right),
\left( \begin{array}{cc} {0}&-\Omega \\ \Omega & {0} \end{array}
\right)\right] \nonumber \\ &=&m\,\sigma\otimes \Omega+\sqrt{\alpha}\,
P\otimes
\Big(a^\dagger(f) \Omega+\Omega a(f)\Big).\end{aligned}$$ For the second commutator $$\begin{aligned}
[H,{X_2}] &=& \left[\left( \begin{array}{cc}
\frac{E}{\alpha}& {\sqrt{\alpha}\,a^\dagger(f)}
\\ \sqrt{\alpha}\, {a(f)}& m + \frac{E}{\alpha} \end{array}
\right),\left ( \begin{array}{cc} { a^{\dagger}(g)\,\Omega + {\rm h.c}}&
{0} \\ {0}& 0 \end{array} \right)\right] \nonumber \\
&=&\frac{1}{\alpha}\, P\otimes (a^{\dagger}(|k| g)\Omega - \Omega a(|k|
g)+\sqrt{\alpha}\,\left(\begin{array}{lr}0&-\langle g|f
\rangle\\ \langle f|g \rangle&0\end{array}\right)\otimes
\Omega.\label{second}\end{aligned}$$ So, if we take $g$ of Eq. (\[g\]) then $$[H,i\,{X_1}-{X_2}]=i\,\left(m-\alpha^2\,{\cal E}\right)
\sigma\otimes\Omega.$$ We see that we can formally solve for the commutator equation, Eq. (\[commute\]) provided ${\cal E}$ is finite.
This is, however, not the only condition. $X$ is a bounded operator in the Hilbert space provided $g\in L^2$, for otherwise $a^\dagger(g)$ is ill defined: $$\int{\frac{|f|^2}{|k|^2}d^dk}\sim \int{\frac{1}{|k|^3}d^dk}<\infty.$$ The integral is finite if $d\ge 4$ but is logarithmically divergent in $d=3$. For $d=3$ we need to squeeze $X$ back to the bounded operators. We do that by allowing for $Y\neq 0$.
Let $\chi_\varepsilon$ be the characteristic function of a ball of radius $\varepsilon$ and $\chi_\varepsilon^c=1-\chi_\varepsilon$ and let $g_\varepsilon^c =\chi_\varepsilon^c g$ and $g_\varepsilon
=\chi_\varepsilon g$. Let us take $X_2(g_\varepsilon^c)$, which is well defined and its norm is $O(\alpha^{\frac{3}{2}}\sqrt{\vert\log\varepsilon\vert} )$. For $X$ we take, as before, $$X=\frac{i\,X_1-X_2(g^c_\varepsilon)}{m+i\sqrt{\alpha}\langle
f|g^c_\varepsilon\rangle }.$$ From this $$\Vert X\Vert = O\left(\frac{1
+\alpha^{\frac{3}{2}}|\log\varepsilon|^{1/2}}{|m-\alpha^2 {\cal E} |}.\right)$$ For $Y$ we take $$\begin{aligned}
(m+i\, \sqrt{\alpha} \langle f| g^c_\varepsilon
\rangle)\,Y&=&[H,X_2(g)-X_2(g_\varepsilon^c)]=[H,X_2(g_\varepsilon)]
\\ &=&\frac{1}{\alpha}\, P\otimes \Big(a^{\dagger}(|k| g_e)\Omega -
\Omega a(|k| g_e)\Big)+\sqrt{\alpha}\,\left(\begin{array}{lr}0&-\langle
g_e|f
\rangle\\ \langle f|g_e \rangle&0\end{array}\right)\otimes\Omega,\nonumber\end{aligned}$$ and we used the computation of the commutator Eq. (\[second\]). With $f$ having a square root singularity, $$\Vert Y\Vert = O\left(\frac{\sqrt{\alpha}
\,\varepsilon +\alpha^2\,\varepsilon}{|m-\alpha^2{\cal E} |}\right) .$$ This puts us in the frame of theorem II.1, except for the minor modification the log appears with a square root. Chasing the square root establishes the main result.$\hfill\vbox{\hrule height 0.6pt\hbox{\vrule width 0.6pt height 1.8ex \kern
1.8ex \vrule width 0.6pt} \hrule height 0.6pt}$
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to I.M. Sigal, S. Graffi, A. Ori and C. Brif for useful discussions. This work was partially supported by a grant from the Israel Academy of Sciences, the Deutsche Forschungsgemeinschaft, and by the Fund for Promotion of Research at the Technion.
Resonance and Lamb Shift of the Dicke Model
===========================================
The ${\cal N}=1$ sector of the Dicke model has a resonance that serves to define the Lamb shift. The resonance is a solutions of the analytically extended eigenvalue equation, Eq. (\[res\]), see [@friedrichs; @howland; @ms]. The real part of the shift is, by definition, the Lamb shift of the model, while the imaginary shift is the life time. For $d\ge 3$, the Lamb shift is dominant and the life time is a higher order in $\alpha$. For the application to the adiabatic theorem we need only the dominant contribution, i.e. only the Lamb shift. Computing the Lamb shift is easy. Computing the life time is harder. For the sake of completeness we compute both, even though we only need one.
The eigenvalue equation is $$E-m=\alpha^2\, G(\alpha E)\label{res},$$ where $G(e)$ is defined as the analytic continuation from the upper half plane of $$\ G(e)=\int_{{\rm I\kern-.2em R}^d}\,\frac{|f|^2}{e-{|k|}}
\, d^d k,\quad \Im\, e\ge 0.$$ By taking the imaginary part, it is easy to see that Eq (\[res\]) has no solution in the upper half plane. To solve the equation in the lower half plane one needs an explicit expression, at least for small $\alpha$, and $e$ near $\alpha m$, of this analytic continuation. Then, we can solve Eq. (\[res\]) by iteration, and to lowest order we have $$E_r\approx m+\alpha^2\, G(\alpha m).\label{iterate}$$ Clearly $G(\alpha m)\to - {\cal E}$, in the limit $\alpha\to 0$, so to leading order $$E_r\approx m-\alpha^2\, {\cal E}.$$ To this order, one does not see the imaginary part of the resonance energy. $\alpha^2\,{\cal E}$ is, by definition, [@dicke], the Lamb shift of the model. It may be worthwhile to point out that the Lamb shift for the Hydrogen atom, [@bethe], is actually of [*higher*]{} order, namely, $\alpha^3
\log(\alpha^{-1})$. Since the Lamb shift of Hydrogen also involves an ultraviolet regularization, while the present model is ultraviolet regular, it is not surprising that the order of the two is different. What is surprising is that the order of Hydrogen is higher rather than lower.
Estimating the life time is, as we noted, irrelevant to the adiabatic theorem. So a reader will loose little by skipping the rest of this Appendix. However, for the benefit of the reader who is interested in how the computation of the life time goes, it is given below.
We shall show below that for $d\ge 3$, and $|e-\alpha m|<\alpha m$, the analytic continuation of $G(e)$ to the lower half plane, and to the next relevant order in $\alpha$, is given by $$G(e)=-{\cal E}-i\pi \,K\Omega^d e^{d-2}\,,\quad \Im\, e\le
0,\label{analytic}$$ where $K$ is as in Eq. (\[K\]), and $\Omega^d$ is the surface area of the unit ball in d dimensions. From Eq. (\[iterate\]), and taking into account Eq. (\[K\]), we get for the Lamb shift and the life-time: $$\begin{aligned}
E_r&\approx& m -\alpha^2 {\cal E} -i
\alpha^2 \pi \,K\,\Omega^d (m\alpha)^{d-2} \nonumber \\ &=& m -\alpha^2
{\cal E} -i
\alpha^2 \pi \Omega^d (m\alpha)^{d-1} \left\vert f(\alpha m)\right\vert^2.\end{aligned}$$ The life time is higher order in $\alpha$ than the Lamb shift, and is of order $\alpha^d$. For $d=3$ this is, indeed, the order of the life time of atomic levels that decay by dipole transition. For small $\alpha$ the Lamb shift dominates the life time, both in the Dicke model and in Hydrogen.
It remains to show that the analytic continuation of $G(e)$ to the lower half plane in a neighborhood of $m\alpha$, is indeed given by Eq. (\[analytic\]). This can be done as follows: Let $B_r$ be a ball of radius $r=2 m\alpha$. Then, in the upper half plane $$\ G(e)=\left(\int_{B_r}+\int_{B_r^c}\right) \frac{|f|^2}{e-{|k|}}
\, d^d k=G_r(e)+G^c_r(e).$$ Clearly, $G^c_r(e)$ extends analytically to a half circle in the lower half plane $|e-\alpha m|<\alpha m$. In the limit of $\alpha
\to 0$, by continuity, $$G^c_r(0)\to -{\cal E}.$$ This is the dominant piece, and it is real.
Consider the analytic continuation of $G_r(e)$ for $|e-\alpha m|\le \alpha
m$. Since, for small argument $f(k)$ is given by Eq. (\[K\]), one has (in the upper half plane) $$G_r(e)=K\Omega^d\int_0^{2m\alpha} \frac{k^{d-2}}{e-k}\,
dk= K\Omega^d\int_\gamma \frac{k^{d-2}}{e-k}\, dk,$$ where $\gamma$ is the obvious semi-circle in the complex $k$ plane and $\Omega^d$ the surface area of the unit ball in $d$ dimensions. The right hand side is analytic in $e$ in the lower half plane provided $|e-\alpha m|<\alpha m$, and so gives the requisite analytic continuation. Since $e$ is small, and of order $\alpha$, to leading order, we have $$\begin{aligned}
G_r(e)&=&K\Omega^d\int_\gamma
\frac{(k-e +e)^{d-2}}{e-k}\,dk\nonumber \\ &=&-K\Omega^d\,
\sum_{j=0}^{d-2} \left(\begin{array}{c} d-2\\
j\end{array}\right)\,\,e^{d-j-2}\int_\gamma (k-e)^{j-1}\,dk
\nonumber \\ &\approx&-K\Omega^d\, \,e^{d-2}\int_\gamma
\frac{dk}{k-e}\nonumber \\ &=&-\, K\Omega^d\, e^{d-2}
\left(i\,\pi\,+\log\left(\frac{2\alpha
m}{e}\right)\,+\,O(\alpha\log\alpha)\right)\end{aligned}$$ and the error term in approximation that we did not compute is real and being sub-dominant to ${\cal E}$ is irrelevant.
V. I. Arnold, [*Mathematical methods of classical mechanics*]{} ( Graduate texts in mathematics, Springer [**60**]{}, 1978). J. E. Avron, R. Seiler and L. G. Yaffe, Comm. Math. Phys. [**110**]{}, 33, (1987), (Erratum: Comm. Math. Phys. [**153**]{}, 649, (1993). V. Bach, J. Fröhlich and I. M. Sigal, Lett. Math. Phys. [**34**]{}, 183, (1995). V. Bach, J. Fröhlich and I. M. Sigal, Quantum electrodynamics of confined non-relativistic particles, Adv. in Math., to appear. V. Bach, J. Fröhlich and I. M. Sigal, Renormalization group analysis of spectral problems in quantum field theory, to appear. V. Bach, J. Fröhlich, I. M. Sigal and A. Sofer, Positive commutators and spectrum of non-relativistic QED, to appear. M.V. Berry, Proc. Roy. Soc. Lond. A [**392**]{}, 45, (1984); The quantum phase: Five years after, in [*Geometric phases in physics*]{} (A. Shapere and F. Wilczek, Eds., World Scientific, 1989). H. Bethe, Phys. Rev. [**72**]{}, 339, (1947). M. Born, [*The Mechanics of the Atom*]{}, (Ungar, 1960). M. Born and V. Fock, Z. Phys. [**51**]{}, 165, (1928). Davis and H. Spohn, J. Stat. Phys. [**19**]{}, 511, (1978). J. Dereziński and C. Gérard, Asymptotic completeness in quantum field theory. Massive Pauli-Fierz Hamiltonians, to appear. R.H. Dicke, Phys. Rev. [**93**]{}, 99, (1954). K. O. Friedrichs, Comm. Pure Appl. Math. [**1**]{}, 361, (1948). K. O. Friedrichs, [*Perturbations of spectra in Hilbert space*]{}, (AMS Providence, 1965). J. S. Howland, J. Math. Anal. Appl. [**50**]{}, no. 2, 415, (1975). H. Huebner and H. Spohn, Ann. Inst. H. Poincare Phys.Theor. [**62**]{}, no. 3, 289, (1995). T. Kato, Phys. Soc. Jap. [**5**]{}, 435, (1958). M. Klein and R. Seiler, Comm. Math. Phys. [**128**]{}, 141, (1990). C. Leonardi, F. Persico and G. Vetri, Rivista Del NuovoCimento [**9**]{}, ser. 3, no. 4, 1, (1986). R. Minlos and H. Spohn, Amer. Math. Soc. Transl. (2) [**177**]{}, 159, (1996). H. Narnhofer and W. Thirring, Phys. Rev. A [**26**]{}, 3646, (1982).
[^1]: $\{\cdot,\cdot\}$ stands for anticommutator
[^2]: for $X$ we also need that its derivative is bounded
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the structural and mechanical properties of jammed ellipse packings, and find that the nature of the jamming transition in these systems is fundamentally different from that for spherical particles. Ellipse packings are generically hypostatic with more degrees of freedom than constraints. The spectra of low energy excitations possess two gaps and three distinct branches over a range of aspect ratios. In the zero compression limit, the energy of the modes in the lowest branch increases [*quartically*]{} with deformation amplitude, and the density of states possesses a $\delta$-function at zero frequency. We identify scaling relations that collapse the low-frequency part of the spectra for different aspect ratios. Finally, we find that the degree of hypostaticity is determined by the number of quartic modes of the packing.'
author:
- 'Mitch Mailman$^{1}$, Carl F. Schreck$^{2}$, Corey S. O’Hern$^{2}$ and Bulbul Chakraborty$^{1}$\'
title: 'Jamming in Systems Composed of Frictionless Ellipse-Shaped Particles'
---
A decade ago, Liu and Nagel proposed that jamming transitions in glassy, granular, and other nonequilibrium amorphous systems can be described by the jamming phase diagram [@LiuNagel], and that a ‘fixed point’ (Point J) in the jamming phase diagram controls slow dynamics in these systems even far from Point J [@OHern2003]. For model disordered systems composed of frictionless, spherical grains interacting via purely repulsive, short-range potentials, the packings at Point J are *isostatic* [@Makse2000; @OHern2002], where the number of degrees of freedom exactly matches the number of constraints. It has been shown that isostatic packings of spherical grains have interesting mechanical properties; for example, they possess an abundance of spatially extended “floppy modes” of excitation [@Wyart1; @Wyart2; @somfai] and non-elastic response to applied deformations [@moukarzel].
However, there have been relatively few theoretical studies of jamming in systems with aspherical particles, despite the fact that most physical particulate media have grains with anisotropic shapes. The introduction of aspherical particles in equilibrium systems gives rise to completely new phases of matter and critical behavior as evidenced in the field of liquid crystals. What is the nature of jamming transitions in nonequilibrium systems when the grains are aspherical? For example, is there still a special point J in the jamming phase diagram, where packings are isostatic, that controls slow dynamics? In this letter, we begin to address these important open questions by investigating the structural and mechanical properties of static packings of frictionless anisotropic particles.
For a static packing of $N$ grains in $d$ spatial dimensions, with $d-1$ rotational and $d$ translational degrees of freedom per grain, total force and torque balance on each grain can be satisfied only if the total number of contacts satisfies $N_c \ge N_I \equiv
N\left(2d-1\right)$ in the large system limit. Isostatic packings satisfy $N_c = N_I$, while hypostatic packings possess $N_c < N_I$. In contrast to spherical particle packings, static packings of ellipsoidal particles, studied previously by Donev, [*et al.*]{} [@donev1], are generically hypostatic and possess higher densities without translational and orientational order. Experiments on packings of ellipsoidal $M\&M$ candies have also verified these results[@donev2]. These previous studies raise several fundamental questions about static packings of ellipsoidal particles; for example, why are they hypostatic and what is the nature of their low-energy excitations? In this letter, we study the low-energy, vibrational excitations of 2D static packings of ellipses using a numerical packing-generation algorithm in which soft ellipses interact via purely repulsive potentials at zero temperature. Our analysis demonstrates the existence of two gaps in the vibrational spectrum over a range of aspect ratios. The energy of the modes below the first gap increases [*quartically*]{} with deformation amplitude along the soft directions, and the number of these quartic modes determines the degree of hypostaticity of the packings.
#### Compression packing-generation protocol
We generated an extensive set of static packings of ellipses over a range of system sizes from $N=120$ to $1920$, in which particles are ‘just touching’, using a numerical packing-generation protocol similar to that employed to create static packings of spherical particles [@Corey_method; @makse_method]. We will refer to this protocol as the ‘compression method’. In this method, soft, purely repulsive ellipses are first randomly deposited in a square cell with periodic boundary conditions at a low packing fraction ($\phi=0.5$). The configurations are successively compressed in small steps $(\Delta
\phi = 10^{-4})$ and then relaxed using conjugate energy minimization after each step. Near the jamming packing fraction $\phi_J$, where the particles are just touching, the configurations are expanded or compressed by decreasing amounts until the system has vanishingly small total potential energy per particle $V_{\rm tol} < V < 2 V_{\rm
tol}$. $V_{\rm tol} = 10^{-12}$ for most simulations, $V = \sum_{i>j}
V(r_{ij})$ summed over all ellipse pairs, $$V(r_{ij})=\begin{cases}
\left(1-r_{ij}/\sigma_{ij}\right)^{2} & r_{ij} \leq \sigma_{ij}\\
0 & r_{ij} > \sigma_{ij},
\end{cases}
\label{eq:pair_pot}$$ and $r_{ij}$ is the separation between the centers of mass of ellipses $i$ and $j$. The separations and orientations ${\hat u}_i$ and ${\hat u}_j$ of the long axes of ellipses $i$ and $j$ determine the Perram and Wertheim overlap parameter [@Geo_Overlap_Der; @Geo_Overlap_Der_More; @BP; @Cleaver_Bidisperse] $$\begin{aligned}
\sigma_{ij} &=& \min_\lambda
\frac{\sigma_{0}(\lambda)}{\sqrt{1-\frac{\chi(\lambda)}{2} \sum_{\pm}
\frac{\beta(\lambda)\hat{r}_{ij}\cdot\hat{u}_i\pm\beta^{-1}(\lambda)
\hat{r}_{ij}\cdot\hat{u}_j}
{1\pm\chi(\lambda)\hat{u}_i\cdot\hat{u}_j}} } ,
\label{eq:sigma2}\end{aligned}$$ where $\sigma_0$, $\beta$, and $\chi$ depend on $\lambda$ and the major (minor) axis, $a_i$ ($b_i$), of the $i$th ellipse [@foot]. To determine $\sigma_{ij}$ for each pair, minimization of the parameter $0 < \lambda < 1$ must be performed. We simulate bidisperse mixtures to inhibit translational and orientational order: one-third of the particles are large with the major axis $1.4$ times that of the small particles [@donev1]. Using this procedure, we generated an ensemble of at least $100$ ellipse packings, each characterized by the jamming packing fraction $\phi_J$, over a range of aspect ratios from $\alpha=1$ to $2$. We calculated the global bond-orientational and nematic order parameters [@frenkel], and found no significant ordering with order parameters $\sim 1/\sqrt{N}$ for all $\alpha$.
#### Vibrational Spectra
To determine mechanical properties of ellipse packings, we calculate the dynamical matrix $$M_{mn}=\frac{\partial^2
V} {\partial \xi_{m}\partial \xi_{n}},
\label{dyn_matrix}$$ where $\xi_m=\{x_m,y_m,a_m \theta_m\}$, $x_m$ and $y_m$ are center of mass coordinates of ellipse $m$, $\theta_m$ is the angle between ${\hat u}_m$ and the $x$-axis, and $m,n=1,\cdots,N$ [@barrat]. When Eq. \[dyn\_matrix\] is evaluated for an ellipse packing and diagonalized using periodic boundary conditions, in principle one obtains $(2d-1)N' -d$ nontrivial vibrational eigenmodes, where $N' = N - N_r$ and $N_r$ is the number of ‘rattlers’ with fewer than $d+1$ contacts. If we assume that all ellipses have the same mass, the square roots of the eigenvalues of the dynamical matrix give the normal mode frequencies $\omega_i$ indexed by $i$. We denote the normalized eigenvector corresponding to $\omega_i$ by ${\hat e}_i = \{e_{xi}^{j=1},e_{yi}^{j=1},e_{\theta
i}^{j=1},\cdots,e_{xi}^{j=N'},e_{yi}^{j=N'},e_{\theta i}^{j=N'}\}$ with the constraint that ${\hat e}^2_i = 1$. Below, we show the relative contributions of translational and orientational components of the eigenvectors, for example, the translational contribution from mode $i$ is $T_i = \sum_{j=1}^{N'}\lbrace (e_{xi}^j)^2 + (e_{yi}^j)^2
\rbrace$ and $R_i = 1 - T_i$.
Over a range of aspect ratios, the spectrum $\omega_i$, sorted in order of increasing frequency, possesses three distinct regimes ([*cf*]{} Fig. \[fig:plot1\]): (1) modes with indexes $i - 2 <
i^*(\alpha)$ below the low-frequency gap, (2) modes with index $i^*
\leq i-2 \leq i_t = (d-1)N$, where for $\alpha \le \alpha_t$, there is a second gap at index $i_t$, and (3) modes with index $i-2 > i_t$. (We are explicitly not including the two trivial modes corresponding to translational invariance.) In the inset to Fig. \[fig:plot1\], we show that we are able to choose aspect ratio dependent scaling factors $\omega^*$ and $i^*$ that collapse the low-frequency part of the spectra including the first gap. We find that $\omega^*$ is roughly linear with $\alpha-1$, while $i^*$ possesses two different scaling regimes for $\alpha-1 \ll 1$ and $\alpha > 1$. As demonstrated in the inset to Fig. \[fig:plot2\], modes in regions (1) and (2) are mainly rotational, whereas those in region (3) correspond to mainly translational motion.
We find that our ellipse packings at finite, but small overcompression possess $(2d-1)N-d$ [*nonzero, positive*]{} eigenvalues [@foot2]. To rationalize this result with the fact that hard ellipse packings are generically hypostatic [@donev1], we perturbed our ellipse packings along each of the eigendirections of the dynamical matrix over a range of overcompression. If $\vec{\xi}_0$ characterizes the centers of mass and orientations of the original static ellipse packing, the perturbed configuration obtained after a shift by $\delta$ along eigenmode $i$ and relaxation to the nearest local energy minimum is $\vec{\xi}_i = \vec{\xi}_{0} + \delta {\hat e}_i$. In Fig. \[fig:plot2\], we plot the change in potential energy, $\Delta V_i \equiv V(\vec{\xi}_i)-V(\vec{\xi}_{0})$, of ellipse packings with $N=120$ at $\alpha=1.5$ arising from a perturbation along mode $i$ as a function of amplitude $\delta$ and for two values of overcompression $V_{\rm tol}$. As shown in Fig. \[fig:plot2\], for modes with indexes in regions (2) and (3) of the frequency spectrum $\Delta V_i \propto \delta^{2}$ for all $\delta$ independent of $V_{\rm tol}$. In contrast, there is range $\delta > \delta_c$ over which modes in region (1) display [*quartic*]{} dependence on $\delta$, $\Delta V_i \propto \delta^{4}$, whereas $\Delta V_i \propto
\delta^{2}$ for $\delta < \delta_c$. Since $\delta_c \sim V_{\rm
tol}^{1/4}$ for modes in region (1), quartic behavior will persist over the entire range of $\delta$ in the zero compression limit. Thus, ‘just-touching’ ellipse packings are stabilized by quartic rather than quadratic terms in the expansion of the total potential energy around the reference packing[@donev1].
The density of vibrational modes, $D(\omega)$, obtained from the spectrum, $\omega_{i}$, shown in Fig. \[fig:DOS\], exhibits several key differences from that for disk packings [@OHern2003]. First, the plateau at small $\omega$ characteristic of nearly-isostatic disk packings is replaced by a narrow peak at low frequencies. This peak is composed of modes in region (2) of the spectrum that display collective, primarily rotational motions. For $\alpha < \alpha_t$ (with $\alpha_t \approx 1.2$ for $N=120$), this peak is clearly separated by a gap from the broad, high-frequency regime. Note that on a logarithmic scale one would also see a peak that corresponds to modes in region (1), and is separated by a gap from the modes in region (2). As discussed before, this peak shifts to lower $\omega$, and narrows to a $\delta$-function at zero frequency in the zero-compression limit. Second, for large aspect ratios ($\alpha >
1.5$ for $N=120$), $D(\omega)$ has significantly more structure than that for disk packings for $\omega > 10$. These new features will be investigated in more detail in future work [@future]. We have verified that the vibrational properties described here also hold for packings in which ellipses interact via the Gay-Berne potential [@GB] with a simpler form for the overlap function than in Eq. \[eq:sigma2\].
#### Hypostaticity
Our analysis of the structural and mechanical properties of ellipse packings has led to several novel observations: (1) The quartic modes that exist for hard ellipse packings develop a quadratic contribution at finite overcompression, and therefore compression [*stabilizes*]{} ellipse packings[@donev1]; (2) The contribution of these quartic modes to the density of vibrational modes decreases as the aspect ratio increases; and (3) Ellipse packings are generically hypostatic, as shown in Fig. \[fig:plot4\], with smaller average contact number $\langle z \rangle$ than that predicted from isostatic counting arguments, i.e. $\langle z \rangle_{\rm iso} = 6$ for ellipses with $\alpha>1$. Hypostatic packings have fewer contacts than required to satisfy all force and torque balance conditions, and thus in some directions of configuration space these packings are only quartically, not quadratically stable. We find that quartic modes represent collective, primarily rotational motions of the ellipses, which do not lead to cage breaking and particle rearrangements. Thus, we expect that if the isostatic counting argument is reformulated so that the quartic modes are not constrained, $\langle z \rangle$ will correspond to the minimum number of contacts necessary to constrain the quadratic modes. Thus, ellipse packings are isostatic with respect to only the [*quadratic*]{} modes.
When each internal degree of freedom in an ellipse packing is stabilized, the isostatic conjecture gives $N \langle z
\rangle/2=3N-1$. If it is not necessary to constrain the quartic modes, this equation can be rewritten as $$\frac{N \langle z \rangle}{2}= 3N - i^*(\alpha),
\label{revised}$$ where $N_4 = i^{*}(\alpha)-1$ is the number of quartic modes in region 1 of the frequency spectrum. Thus, by measuring the number of quartic modes, we are able to determine $\langle z \rangle (\alpha)$ as shown in Fig. \[fig:plot4\]. In the inset we show that $N/i^{*}(\alpha)
-N/(N+1)$ has two power-law regimes: scaling as $\sqrt{\alpha-1}$ ($\alpha-1$) for small (large) $\alpha-1$. In these limits, we obtain $$\langle z \rangle (\alpha) = \langle z \rangle (1) + 2
\frac{A_n (\alpha -1)^n}{1 + A_n (\alpha-1)^n},
\label{eq:Model_Kfit}$$ where $n=0.5$ ($1$) for small (large) $\alpha-1$ and $A_{0.5}$ and $A_1$ are positive constants. In Fig. \[fig:plot4\], we use a function that interpolates between these power-laws and allows us to fit $\langle z \rangle$ for the compression method over the entire range of $\alpha$. These arguments imply that $\langle z \rangle = \langle
z \rangle_{\rm iso}$ in the $\alpha \rightarrow \infty$ limit, yet this is still an open question. Previous studies have predicted that $\langle z \rangle(\alpha) - \langle z \rangle (1)$ scale as $\sqrt{\alpha-1}$ based on the behavior of the pair distribution function $g(r)$ near contact for spherical particle packings [@donev1]. In contrast, our numerical results demonstrate hypostaticity in ellipse packings originates from the existence of quartic modes.
#### Annealed packings
The ellipse packings discussed up to this point were generated using the compression method at fixed aspect ratio. Using this method, we obtained similar $\langle z \rangle(\alpha)$ to that obtained previously for hard ellipse packings [@donev1]. Since ellipse packings are hypostatic, it is in principle possible to obtain packings with higher $\langle z \rangle$ than found in Fig. \[fig:plot4\] without increasing the translational or orientational order. To test this, we developed an ‘annealing’ method, which creates static packings by incrementally increasing the aspect ratio from $\alpha=1$. We initially create bidisperse disk packings. Each disk is then assigned the same aspect ratio $1+\Delta
\alpha$ with the direction of the long axis chosen randomly. A new ellipse packing is formed from this initial state using the compression method described above. The ellipses of the new packing are elongated again along their defined major axes, and the protocol is repeated until a packing with the desired aspect ratio is reached. Using this method, ellipse packings can be generated that are denser and possess contact numbers much closer to $z_{\rm iso}=6$ as shown in Fig. \[fig:plot4\]. The annealed packings still exhibit a low-frequency gap, and the location, $i^{*}$, of this gap can be used to predict $\langle z \rangle$ through Eq. \[revised\]. The variation of $i^{*}$ with aspect ratio, however, differs from the ‘compressed’ packings. The annealed packings tend to have higher $\langle z \rangle$, fewer quartic modes, and a plateau in $\langle z
\rangle$ at large $\alpha$.
In conclusion, we find that aspherical grains qualitatively change the nature of jamming at point J. Model systems with spherical grains appear to be the exceptional case: they are isostatic, and all nontrivial modes of excitation increase quadratically with deformation amplitude. In contrast, ellipse packings, which are more relevant for real granular media, possess quartic modes characterized by collective rotational motions. Thus, Landau theories for ellipse packings have a 4th order term as the lowest nonvanishing contribution to the generalized free energy. These results will likely stimulate further work on statistical descriptions of packings of aspherical particles.
Support from NSF grant numbers DMR-0549762 (BC $\&$ MM), DMR-0448838 (CSO) and CDI-0835742 (CS) is acknowledged. During the workshop “Dynamical heterogeneities in glasses, colloids, and granular media” in Leiden in August 2008, we learned of independent, parallel results by Zeravcic, et al. [@zeravcic] on 3D ellipsoids, with similar conclusions. We acknowledge discussions and suggestions from Z. Zeravcic and S. Nagel at the workshop, and also thank M. Bi, G. Lois, T. Witten, and N. Xu for helpful comments.
A. J. Liu and S. R. Nagel, Nature [**396**]{}, 21 (1998).
C. S. O’Hern, L. E. Silbert, A. J. Liu, and S. R. Nagel, Phys. Rev. E [**68**]{}, 011306 (2003).
H. A. Makse, D. L. Johnson, and L. M. Schwartz, Phys. Rev. Lett. [**84**]{}, 4160 (2000).
C. S. O’Hern, S. A. Langer, A. J. Liu, and S. R. Nagel, Phys. Rev. Lett. [**88**]{}, 075507 (2002).
M. Wyart, S. R. Nagel, and T. A. Witten, Europhys. Lett. [**72**]{}, 486 (2005).
M. Wyart, L. E. Silbert, S. R. Nagel, and T. A. Witten, Phys. Rev. E [**72**]{}, 051306 (2005).
E. Somfai, M. Van Hecke, W. G. Ellenbroek, K. Shundyak, and W. Van Saarloos, Phys. Rev. E [**75**]{}, 020301(R) (2007).
C. F. Moukarzel, Phys. Rev. Lett. [**81**]{}, 1634 (1998); R. Blumenfeld, Phys. Rev. Lett. [**93**]{}, 108301 (2004).
A. Donev, R. Connelly, F. H. Stillinger, and S. Torquato, Phys. Rev. E [**75**]{}, 051304 (2007).
A. Donev, I Cisse, D. Sachs, E. A. Variano, F. H. Stillinger, R. Connelly, and S. Torquato, Science [**303**]{}, 990 (2004).
G.-J. Gao, J. Blawzdziewicz, and C. S. O’Hern, Phys. Rev. E [**74**]{}, 061304 (2006).
H. P. Zhang and H. A. Makse, Phys. Rev. E [**72**]{}, 011301 (2005).
J.W. Perram and M.S. Wertheim, J. Comp. Phys. [**58**]{}, 409 (1985).
J.W. Perram, J. Rasmussen, E. Praestgaard, J.L. Lebowitz, Phys. Rev. E [**54**]{}, 6565 (1996).
B. J. Berne and P. Pechukas, J. Chem. Phys. [**46**]{}, 4213 (1972).
D. J. Cleaver, C. M. Care, M. P. Allen, and M. P. Neal, Phys. Rev. E [**54**]{}, 559 (1996).
The parameters $\sigma_0$, $\beta$, and $\chi$ in Eq. \[eq:sigma2\] are defined as $\sigma_{0}=\sqrt{\frac{b_{i}^{2}}
{\lambda}+\frac{b_{j}^{2}}{1-\lambda}}$, $\beta=\left(\frac{
\left(a_{i}^{2}-b_{i}^{2}\right)
\left(a_{j}^{2}+\frac{1-\lambda}{\lambda} b_{i}^{2}\right)}{
\left(a_{j}^{2}-b_{j}^{2}\right)
\left(a_{i}^{2}+\frac{\lambda}{1-\lambda}
b_{j}^{2}\right)}\right)^{1/4}$, and $\chi=\left(\frac{
\left(a_{i}^{2}-b_{i}^{2}\right) \left(a_{j}^{2}-b_{j}^{2}\right)}{
\left(a_{j}^{2}+\frac{1-\lambda}{\lambda} b_{i}^{2}\right)
\left(a_{i}^{2}+\frac{\lambda}{1-\lambda}
b_{j}^{2}\right)}\right)^{1/2}$.
J. A. Cuesta and D. Frenkel, Phys. Rev. A [**42**]{}, 2126 (1990).
A. Tanguy, J. P. Wittmer, F. Leonforte, and J.-L. Barrat, Phys. Rev. B [**66**]{}, 174205 (2002).
In the following discussion, we have accounted for rattler ellipses and set $N=N'$.
M. Mailman, C. Schreck, B. Chakraborty, and C. S. O’Hern (unpublished).
J. G. Gay and B. J. Berne, J. Chem. Phys. [**74**]{}, 3316 (1981).
Z. Zeravcic, N. Xu, et al. (unpublished).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Aurélien Drezet
- Andreas Hohenau
- 'Joachim R. Krenn'
title: 'Comment on “Far-field microscopy with a nanometer-scale resolution based on the in-plane image magnification by surface plasmon polaritons”'
---
Recently, Smolyaninov *et al.* \[1\] reported the realization of a 2D sub-wavelength resolution microscope using surface plasmon polaritons (SPP) propagating on a gold film bounded by the glass substratum and a glycerin droplet \[1-3\]. The main idea of \[1\] is that SPPs should, for a laser wavelength $\lambda_{0}$ of 500 nm, have a very short wavelength of $\lambda_{\rm SPP}\simeq$ 70 nm which implies an ultimate resolution of $\lambda_{\rm SPP}/2\simeq
35$ nm. We disagree on the conclusions of \[1\] because the SPP wavelength is too large and their propagation length too small to enable the described in-plane imaging process. Indeed, the claims in \[1\] are based on the assumption that the SPP modes confined on a gold metal film are described by an effective Drude model (see Fig. 1 of \[1,2\]) requiring precise compensation of the imaginary parts in the denominator of Eq. 1 of \[1\] (Eq. 1 supposes $\lambda_{\rm SPP}\ll \lambda_{0}$ and $\epsilon_{gold}\simeq-\epsilon_{d}$). However their analysis here is questionable since in a real metal (see \[4\]) losses modify fundamentally the ideal SPP dispersion relations considered in \[1\]. In order to show that we calculated numerically the dispersion relation for SPPs confined on a 50 nm thick film as used in \[1,2,3\]. Fig. 1 clearly reveals two bound SPP modes which are highly damped in the optical region $\lambda_0\simeq 500$ nm. As visible in Fig. 1a the SPP wavelength ($=2\pi/\textrm{Re}\{K_{\rm SPP}\}$) at such $\lambda_0$ takes values centered at 250 nm to 300 nm while the propagation lengths ($L_{\rm SPP}=1/(2\textrm{Im}\{K_{\rm SPP}\})$) reach values between 40 and 250 nm (see Fig. 1b). These values agree with similar theoretical \[5,6\] and experimental \[7\] analysis but disagree completely with the interpretation presented in \[1\]. Therefore the explanation of the image formation in Fig. 2 of \[1\] cannot be considered as correct. Furthermore one can easily see that with such a wavelength and propagation length SPPs can not be used to resolve nanoholes separated by 40 nm as claimed in \[1\] (see Figs. 3,4 of \[1\]). Reducing the film thickness could not solve the problem. Indeed, $\lambda_{\rm SPP}$ and $L_{\rm SPP}$ (of one given SPP mode) decrease in parallel monotonously (e. g., $\lambda_{\rm
SPP}=150$ nm and $L_{\rm SPP}=20$ nm for 20 nm thick films) \[6\]. Such SPPs are clearly non-propagative ($L_{\rm SPP}<\lambda_{\rm
SPP}$) and primarily represent a channel for ohmic losses and scattering. There is no hope in such conditions to justify high resolution by an hypothetical “mode coupling” mechanism \[1,2\] called to transfer the local information from SPPs to guided waves in the glycerin as $L_{\rm SPP}$ is simply too small. The additional high-resolution imaging mechanisms based on the properties of SPP Bloch waves with short wavelengths suggested in \[2, 8\] do not appear more convincing since i) it can not justify the result presented in Fig. 2 of \[1\] where only few holes are involved, and ii) it should imply an unjustified increase of $L_{\rm SPP}$ by 2 or 3 orders of magnitude necessary for the SPP to travel through the hole arrays of Figs. 3,4 of \[1\]. Outside the array, the SPPs still follow the dispersion relation of the unperturbed film and therefore have to be excluded from the image formation, due to their low propagation length.
[99]{} I. I. Smolyaninov, J. Elliott, A. Zayats, C. C. Davis, Phys. Rev. Lett. **94**, 057401 (2005). I. I. Smolyaninov, J. Opt. A**7**, S165 (2005). I. I. Smolyaninov, C. C. Davis, J. Elliott, A. Zayats, Opt. Lett. **30**, 382 (2005). P. B. Johnson and R. W. Christy, Phys. Rev. **B6** 4370 (1972). A. Hohenau *et al. *, Opt. Lett. **30**, 893 (2005). J. J. Burke, G. I. Stegeman, T. Tamir, Phys. Rev. **B33** 5186 (1986). J. T. van Wijngaarden *et al.*, Appl. Phys. Lett. **88**, 221111 (2006). I. I. Smolyaninov *et al.*, Phys. Rev. **B72**, 085442 (2005).
{width="8cm"}
figure 1
Caption for figure1:
(color online) Dispersion relation for bound SPPs propagating on a 50 nm thick gold film (medium 1) surrounded by glass (medium 2) and glycerin (medium 0). (a) shows the dispersion for the real part of the SPP wavevectors while (b) shows the imaginary part. The red and blues curves correspond to the two bound modes (see inset in (b) for the tangential magnetic field profiles). The green and magenta curves show the dispersion for SPPs bound at single interfaces gold/glass and gold/glycerin, respectively. The horizontal black lines show the range of frequency used in \[1,2,3\]. The light lines for glass (continuous) and glycerin (dashed) are included in (a).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'N. Astudillo-Defru'
- 'R. Cloutier'
- 'S. X. Wang'
- 'J. Teske'
- 'R. Brahm'
- 'C. Hellier'
- 'G. Ricker'
- 'R. Vanderspek'
- 'D. Latham'
- 'S. Seager'
- 'J. N. Winn'
- 'J. M. Jenkins'
- 'K. A. Collins'
- 'K. G. Stassun'
- 'C. Ziegler'
- 'J. M. Almenara'
- 'D. R. Anderson'
- 'E. Artigau'
- 'X. Bonfils'
- 'F. Bouchy'
- 'C. Briceño'
- 'R. P. Butler'
- 'D. Charbonneau'
- 'D. M. Conti'
- 'J. Crane'
- 'I. J .M. Crossfield'
- 'M. Davies'
- 'X. Delfosse'
- 'R. F. Díaz'
- 'R. Doyon'
- 'D. Dragomir'
- 'J. D. Eastman'
- 'N. Espinoza'
- 'Z. Essack'
- 'F. Feng'
- 'P. Figueira'
- 'T. Forveille'
- 'T. Gan'
- 'A. Glidden'
- 'N. Guerrero'
- 'R. Hart'
- 'Th. Henning'
- 'E. P. Horch'
- 'G. Isopi'
- 'J. S. Jenkins'
- 'A. Jordán'
- 'J. F. Kielkopf'
- 'N. Law'
- 'C. Lovis'
- 'F. Mallia'
- 'A. W. Mann'
- 'J. R. de Medeiros'
- 'C. Melo'
- 'R. E. Mennickent'
- 'L. Mignon'
- 'F. Murgas'
- 'D. A. Nusdeo'
- 'F. Pepe'
- 'H. M. Relles'
- 'M. Rose'
- 'N. C. Santos'
- 'D. Ségransan'
- 'S. Shectman'
- 'A. Shporer'
- 'J. C. Smith P. Torres'
- 'S. Udry'
- 'J. Villasenor'
- 'J .G. Winters'
- 'G. Zhou'
bibliography:
- 'TOI134.bib'
title: 'A hot terrestrial planet orbiting the bright M dwarf L 168-9 unveiled by *TESS* [^1] ^,^[^2] ^,^[^3] '
---
Introduction
============
The best planets for detailed characterization are transiting planets, first and foremost because they allow for the possibility of unambiguous mass measurement [e.g., HD209458b @Charbonneau2000; @Mazeh2000; @Henry2000]. From the Doppler effect we can determine the minimum mass of the planet ($m \sin i$), and from the transit light curve we can determine the planetary radius and the orbital inclination ($i$), thus yielding a measurement of the planet’s mass. Moreover, from this combination we can calculate the planet’s mean density, and shed light on its internal structure by comparison with models containing different amount of iron, silicates, water, hydrogen, and helium. Furthermore, transiting planets are unique because of the feasibility to characterize the upper atmosphere by spectroscopy during transits and occultations of a significant number of planets. The forthcoming *James Webb Space Telescope* [JWST, @Gardner2006] and the *Extremely Large Telescope* [ELT, @deZeeuw2014] will have unprecedented capabilities for detailed studies of the atmospheres of terrestrial planets, and the interpretation of the results will require an accurate mass measurement (e.g., @Batalha:2019).
The Transiting Exoplanet Survey Satellite (TESS) [@Ricker2015] started scientific operations in July 2018, aiming to detect transiting planets around bright and nearby stars – bright enough for Doppler mass measurement to be feasible. For this task, TESS surveys about 85% of the sky during the Prime Mission. The survey covering the southern ecliptic hemisphere is now complete, and the northern survey is underway. Each hemisphere is divided into 13 rectangular sectors of $96\degree \times 24\degree$ each. Each sector is continuously observed for an interval of 27-days, with a cadence of 2 minutes for several hundred thousand pre-selected stars deemed best suited for planet searching. Additionally, during the TESS Prime Mission, the Full Frame Images – the full set of all science and collateral pixels across all CCDs of a given camera – are available with a cadence of 30 minutes. M dwarfs are of special interest because the transit and radial-velocity signals of a given type of planet are larger for these low-mass stars than they are for Sun-like stars. In addition, compared to hotter stars M dwarfs present better conditions for the detection of planets orbiting the circumstellar habitable zone: less time consuming, larger Doppler signals, and an increased transit probability. @Sullivan2015 anticipated that from 556 small ($<$2R$_\oplus$) transiting planets discovered by TESS, 23% of them will be detected orbiting bright (K$_S<$9) stars, and that 75% of small planets will be found around M dwarfs.
This paper reports the discovery of a small planet orbiting the star L 168-9 (TOI-134), based on TESS data. The host star is a bright M dwarf. An intense precise radial-velocity campaign with HARPS and the Planet Finder Spectrograph (PFS) revealed the terrestrial nature of the newly detected world. This work is presented as follows: Section \[sec:L168-9\] describes the host star properties. Sections \[sec:phot\] and \[sec:rv\] describe the photometric and radial-velocity observations. Section \[sec:analysis\] presents an analysis of all the data, including the study of stellar activity. Finally, Section \[sec:conclusion\] places L 168-9 b within the larger context of the sample of detected planets.
L 168-9 {#sec:L168-9}
=======
[p[0.2]{}l c c c]{} Parameter & Units & Value & Reference\
R.A. & \[J2000\] & $23^h 20^m 07.52^s$ & Gaia2018\
Decl. & \[J2000\] & $-60\degree 03^\prime 54.64^{\prime\prime}$ & Gaia2018\
Spectral type & & M1V & Ga2014\
B & \[mag\] & 12.45$\pm$0.19 & Ho2000\
V & \[mag\] & 11.02$\pm$0.06 & Ho2000\
B$_A$ & \[mag\] & 12.460$\pm$0.025 & He2016\
V$_A$ & \[mag\] & 11.005$\pm$0.018 & He2016\
g$_A$ & \[mag\] & 11.752$\pm$0.032 & He2016\
r$_A$ & \[mag\] & 10.416$\pm$0.028 & He2016\
i$_A$ & \[mag\] & 9.675 & He2016\
W$_1$ & \[mag\] & 6.928$\pm$0.060 & Cu2013\
W$_2$ & \[mag\] & 6.984$\pm$0.020 & Cu2013\
W$_3$ & \[mag\] & 6.906$\pm$0.016 & Cu2013\
W$_4$ & \[mag\] & 6.897$\pm$0.074 & Cu2013\
T && 9.2298$\pm$0.0073 & St2018\
J & &7.941$\pm$0.019 & Cu2003\
H & &7.320$\pm$0.053 & Cu2003\
K$_s$ & &7.082$\pm$0.031 & Cu2003\
B$_p$ & &11.2811$\pm$0.0016 & Gaia2018\
G & & 10.2316$\pm$0.0008 & Gaia2018\
R$_p$ & &9.2523$\pm$0.0011 & Gaia2018\
$\pi$ & \[mas\] & 39.762$\pm$0.038 & Gaia2018\
Distance & \[pc\] & 25.150 $\pm$ 0.024 & Gaia2018\
$\mu_\alpha$ & \[mas/yr\] & -319.96$\pm$0.10 & Gaia2018\
$\mu_\delta$ & \[mas/yr\] & -127.78$\pm$0.12 & Gaia2018\
$dv_r/dt$ &\[m/s/yr\] & 0.06865$\pm$0.00011 & this work\
M$_s$ & \[M$_\odot$\] & 0.62$\pm$0.03 & Ma2019\
R$_s$ & \[R$_\odot$\] & 0.600 $\pm$0.022 & Sect. \[subsec:stellarparam\]\
T$_{\rm eff}$ & \[K\] & 3800$\pm$70 & Sect. \[subsec:stellarparam\]\
L$_s$ & \[L$_{\odot}$\] & 0.0673$\pm$0.0024 & Sect. \[subsec:stellarparam\]\
${log(g)}$ & \[g/cm$^{-3}$\] &4.04$\pm$0.49 & Sect. \[subsec:stellarparam\]\
$[\mathrm{Fe/H}]$ & & 0.04$\pm$0.17 & Ne2014\
${log(R^\prime_{HK})}$ & & -4.562$\pm$0.043 & As2017A\
P$_{\text{rot}}$ & \[days\] & $29.8\pm1.3$ & Sect. \[subsec:rotation\]\
Gaia2018 – @Gaia2018; Ga2014 – @Gaidos2014; Ho2000 – @Tycho-2; H22016 – @apass; Cu2013 – @wise; St2018 – @StassunTIC:2018; Cu2003 – @Cutri2003; Ma2019 – @Mann:2019; Ne2014 – @Neves2014; As2017a – @Astudillo2017a
L 168-9, also known as CD-60 8051, HIP 115211, 2MASS J23200751-6003545, with the entry 234994474 of the TESS Input Catalog (TIC) or 134 of the Tess Object of Interest (TOI) list, is a red dwarf of spectral type M1V. It appears in the southern sky, and resides at a distance of 25.150$\pm$0.024 $pc$ from the Sun [@Gaidos2014; @Gaia2018]. Table \[tab:stellarprop\] lists the key parameters of the star, namely its position, visual and near-infrared apparent magnitudes, parallax, proper motion, secular acceleration, and its essential physical properties.
Derived stellar properties {#subsec:stellarparam}
--------------------------
We performed an analysis of the broadband spectral energy distribution (SED) together with the [*Gaia*]{} DR2 parallax in order to determine an empirical measurement of the stellar radius, following the procedures described by @Stassun:2016 [@Stassun:2017; @Stassun:2018]. We took the $B_T V_T$ magnitudes from [*Tycho-2*]{}, the $BVgri$ magnitudes from APASS, the $JHK_S$ magnitudes from [*2MASS*]{}, the W1–W4 magnitudes from [*WISE*]{}, and the $G$ magnitude from [*Gaia*]{}. Together, the available photometry spans the full stellar SED over the wavelength range 0.35–22 $\mu$m (see Figure \[fig:sed\]).
We performed a fit using NextGen stellar atmosphere models [@Hauschildt1999], with the effective temperature ($T_{\rm eff}$) and surface gravity ($\log g$) constrained on the ranges reported in the TESS Input Catalog [@StassunTIC:2018], while the metallicity \[Fe/H\] was fixed to a typical M-dwarf metallicity of -0.5. We fixed the extinction ($A_V$) to be zero, considering proximity of the star, the degrees-of-freedom of the fit is 10. The resulting fit (Figure \[fig:sed\]) has a $\chi^2$ of 42.3 ($\chi^2_{red}$=4.2), with $T_{\rm eff} = 3800 \pm 70$ K. The relatively high $\chi^2$ is likely due to systematics, as the stellar atmosphere model is not perfect. We artificially increased the observational uncertainty estimates until $\chi^2_{red}=1$ was achieved. Integrating the (unreddened) model SED gives the bolometric flux at Earth of $F_{\rm bol} = 3.41 \pm 0.12 \times 10^{-9}$ erg s$^{-1}$ cm$^{-2}$. Taking the $F_{\rm bol}$ and $T_{\rm eff}$ together with the [*Gaia*]{} DR2 parallax, adjusted by $+0.08$ mas to account for the systematic offset reported by @StassunTorres:2018, gives the stellar radius as $R = 0.600 \pm 0.022$ R$_\odot$. Finally, estimating the stellar mass from the empirical relations of @Mann:2019 gives $M = 0.62 \pm 0.03 M_\odot$. With these values of the mass and radius, the stellar mean density is $\rho = 4.04 \pm 0.49$ g cm$^{-3}$. We also tested to fit the SED using BT-Settl theoretical grid of stellar model [@Allard2014], where we obtained a consistent result.
We searched for infrared (IR) excess in [*WISE*]{} data using the Virtual Observatory SED Analyser [VOSA, @Bayo:2008], which could point for the presence of debris disks. For that we computed the excess significance parameter [$\chi_\lambda$, @Beichman:2006; @Moor:2006], where $\chi_\lambda \ge 3$ represents a robust detection of IR excess. We obtained $\chi_\lambda = 0.70$ in the W$_3$ band, ruling out the presence of a debris disk around L 168-9.
Observations
============
The first hint of a planetary companion orbiting L 168-9 came from analyzing TESS data. After the Data Validation Report was released to the community, a follow-up campaign started with several instruments and by different teams to check on whether the transit-like signal seen by TESS originated from a planet, as opposed to a stellar binary or other source. The follow-up observations included supplementary time series photometry aiming to detect additional transits, seeing-limited and high-resolution imaging to analyze the possibility that the signal comes from a star on a nearby sightline, and precise radial-velocity monitoring to measure the companion’s mass.
Photometry {#sec:phot}
----------
### TESS
TESS observed Sector 1 from the 25th of July to the 22nd of August 2018 [^4], a 27.4-day interval that is typical of each sector. L 168-9 is listed in the Cool Dwarf Catalog that gathered the known properties of as many dwarf stars as possible with $V-J>2.7$ and effective temperatures lower than 4000 K [@Muirhead2018]. The predicted TESS-band apparent magnitudes are also given in this catalog. L 168-9 was chosen for 2-min time sampling as part of the TESS Candidate Target List [CTL, @StassunTIC:2018], consisting of a subset of the TESS Input Catalog (TIC) identified as high-priority stars in the search for small transiting planets. Time series observations of L 168-9 were made with CCD 2 of Camera 2.
The TESS Science Processing Operations Center (SPOC) at the NASA Ames Research Center performed the basic calibration, reduction, and de-trending of the time series, and also performed the search for transit-like signals [@Jenkins2016]. The light curves were derived by the SPOC pipeline and consist of a time series based on Simple Aperture Photometry (SAP), as well as a corrected time series based on Pre-search Data Conditioning [PDC, @Smith2012; @Stumpe2014] referred to as PDCSAP (as detailed by Tenenbaum and Jenkins 2018[^5]). This work made use of the PDCSAP time series available on the Mikulski Archive for Space Telescopes (MAST[^6]). The TESS photometry is presented in the upper panel in Figure \[fig:tess\_phot\].
A Data Validation Report for L 168-9 was released to the community as part of the MIT TESS Alerts[^7]. The report includes several validation tests to assess the probability that the signal is a false positive: eclipsing-binary discrimination tests, a statistical bootstrap test, a ghost diagnostic test, and difference-image centroid offset tests. These are described by @Twicken2018. All the tests were passed successfully. The formal false-alarm probability of the planet candidate was reported to be 5.85$\times 10^{-37}$.
The TESS time series covers 19 transits of what was originally deemed a planet candidate (L 168-9b or TOI-134.01), and reported on the TESS exoplanet Follow-up Observing Program (TFOP) website[^8]. According to the Data Validation Report the orbital period is P $=1.401461\pm0.000137$ days and the transit depth is $\Delta F/F_0=566\pm38\ ppm$, which translates into a planetary radius of R$_p=1.58\pm0.36$ R$_\oplus$ (Sect. \[subsec:stellarparam\] describes how we determined the stellar radius, which is the same value as the used in the Validation Report). The time of mid-transit at an arbitrarily chosen reference epoch is (BJD) T$_c=2458326.0332\pm0.0015$.
### LCOGT, MKO, and SSO T17 {#subsubsec:groundphot}
We acquired ground-based time series photometric follow-up of L 168-9 and the nearby field stars as part of the TESS Follow-up Observing Program (TFOP) to attempt to rule out nearby eclipsing binaries (NEBs) in all stars that are bright enough to cause the TESS detection and that could be blended in the TESS aperture. We used the [TESS Transit Finder]{}[^9], which is a customized version of the [Tapir]{} software package [@Jensen:2013], to schedule our transit observations.
We observed one full transit simultaneously using three 1-meter telescopes at the Las Cumbres Observatory Global Telescope (LCOGT) [@Brown2013] South Africa Astronomical Observatory node on 21 September 2018 in $i'$-band. The Sinistro detectors consist of 4K$\times$4K 15-$\mu m$ pixels with an image scale of $0\farcs389$ pixel$^{-1}$, resulting in a field-of-view of $26\farcm5 \times 26\farcm5$. The images were calibrated by the standard LCOGT BANZAI pipeline.
We observed a full transit from the Mount Kent Observatory (MKO) 0.7-meter telescope near Toowoomba, Australia on 23 September 2018 in $r'$-band. The Apogee U16 detector consists of 4K$\times$4K 9-$\mu m$ pixels with an image scale of $0\farcs41$ pixel$^{-1}$, resulting in a field-of-view of $27\arcmin\times27\arcmin$. The images were calibrated using [AstroImageJ]{} ([AIJ]{}) software package [@Collins:2017].
We observed a full transit from the Siding Spring Observatory, Australia, iTelescope T17 0.43-meter telescope on 27 September 2018 with no filter. The FLI ProLine PL4710 detector consists of 1K$\times$1K pixels with an image scale of $0\farcs92$ pixel$^{-1}$, resulting in a field-of-view of $15\farcm5\times15\farcm5$. The images were calibrated using [AstroImageJ]{}.
We used the [AstroImageJ]{} to extract differential light curves of L 168-9 and all known stars within $2\farcm5$ of the target star that are bright enough to have possibly produced the shallow TESS detection, which includes 11 neighbors brighter than TESS-band = 17.9 mag. This allows an extra 0.5 in delta magnitude relative to L 168-9 to account for any inaccuracies in the TESS band reported magnitudes in the TICv8. The L 168-9 light curves in all five photometric data sets show no significant detection of the shallow TESS detected event, as expected from our lower precision ground-based photometry. Considering a combination of all five photometric data sets, we exclude all 11 known neighbors that are close enough and bright enough to L 168-9 to have possibly caused the TESS detection as potential sources of the TESS detection.
### WASP
WASP-South, located in Sutherland, South Africa, is the southern station of the *Wide Angle Search for Planets* [WASP, @Pollacco2006]. It consists of an array of 8 cameras each backed by a 2048x2048 CCD. Observations in 2010 and 2011 (season A) used 200mm, f/1.8 lenses with a broadband filter spanning $400-700$ nm and a plate scale of $13.7\arcsec$/pixel. Then, from 2012 to 2014 (season B), WASP-South used 85mm, f/1.2 lenses with a Sloan r’ filter and a plate scale of $32\arcsec$/pixel. The array rastered a number of fields each clear night at typically 10-min cadence.
L 168-9 was monitored for four consecutive years, from to May 20, 2010 to December 12, 2014; typically covering 150 days in each year. In one campaign two cameras with overlapping fields observed the star, giving a total of 27300 data points; in another campaign, L 168-9 was observed by three cameras with overlapping fields, totalling 170000 data points. The photometry has a dispersion of 0.027 $\delta$mag and average uncertainty of 0.024 $\delta$mag, presenting clear signs of variability, as shown below in Sec. \[subsec:rotation\].
High-resolution Imaging {#sec:image}
-----------------------
The relatively large 21pixels of TESS can lead to photometric contamination from nearby sources. These must be accounted for to rule out astrophysical false positives, such as background eclipsing binaries, and to correct the estimated planetary radius, initially derived from the diluted transit in a blended light curve [@ziegler18]. Without this correction, the interpreted planet radius can be underestimated [e.g., @ciardi2015; @Teske2018].
### SOAR {#subsubsec:soar}
![SOAR speckle results of L 168-9. Black points represent the I-band contrast obtained at a given separation of the star. The solid black line shows the $5 \sigma$ detection limit curve.[]{data-label="fig:soar"}](figures/TIC234994474I-cz20180925_2.pdf)
We searched for close companions to L 168-9 with speckle imaging on the 4.1-m Southern Astrophysical Research telescope [SOAR, @tokovinin18] installed in Cerro Pachón, Chile, on 2018 September 25 UT using the I-band ($\lambda_{cen}=824\ nm$, full width at half maximum$=170\ nm$) centered approximately on the TESS passband. Further details of the TESS SOAR survey are published in [@Ziegler2019].
Figure \[fig:soar\] shows the 5$\sigma$ detection sensitivity. No nearby stars ($\rho < 1 \farcs 6$) to L 168-9 were detected within the sensitivity limits of SOAR.
### Gemini-South {#subsubsec:gemini}
![DSSI/Gemini-S detection limit curves of L 168-9 in the 692 nm (top) and 880 nm (bottom) filters. Squares and points on left panel represent local maxima and local minima, respectively. []{data-label="fig:speckle"}](figures/TOI134_f1ab.png)
Observations of L 168-9 were conducted with the Differential Speckle Survey Instrument (DSSI; @Horch2009) on Gemini South, Chile, on UT 28 October 2018 under program GS-2018B-LP-101 (PI: I. Crossfield). The usual 692 nm and 880 nm filters on DSSI were used, and three sequences of 60 ms/frame$\times$1000 frames were taken. The total time on target, including readout overhead, was six minutes. [@Howell2011], [@Horch2011a], and [@Horch2012] detail the speckle observing and data reduction procedures. The detection limit curves are shown in Figure \[fig:speckle\]. While the 692 nm image was taken at too low of gain, leading to a shallow detection limit, the 880 nm detection limit curve ($b$) in Figure \[fig:speckle\] indicates that L 168-9 lacks any companions of $\Delta m \sim$5.0 mag beyond 0.1 and any companions of $\Delta m \sim$5.5 mag beyond 0.2. Gemini-South and SOAR result (Sect. \[subsubsec:soar\]) mean the transit signal is likely to be associated with L 168-9.
Thus, we conclude that there is no significant contamination of the TESS photometric aperture that would bias the determination of the planet radius. The Sinistro data (Sect. \[subsubsec:groundphot\]) rule out surrounding stars as potential sources of the TESS detection, so we can assume the planet orbits L 168-9. Given the limits placed on nearby companions by the Gemini-South data, if there were a close companion it would have to be fainter by $\sim 5$ magnitudes than the primary star, meaning the planet radius correction factor would be at most $$R_{p,corr} = R_{p}\times \sqrt{1 + 10^{-0.4\Delta m}} = R_{p}\times1.005.$$
This is smaller than the derived planet radius uncertainty (Table \[tab:results\]).
Radial velocity {#sec:rv}
---------------
### HARPS
The *High Accuracy Radial velocity Planet Searcher* [HARPS @Mayor2003] is an echelle spectrograph mounted on the 3.6m telescope at La Silla Observatory, Chile. The light is spread over two CCDs (pixel size 15 $\mu m$) by a science fiber and a calibration fiber. The calibration fiber can be illuminated with the calibration lamp for the best radial velocity precision, or it can be placed on sky for moderate precision. HARPS is stabilized in pressure and temperature, and has a resolving power of 115[,]{}000. The achievable precision in radial velocity of better than 1 $m/s$.
We began monitoring L 168-9 with HARPS on September 29, 2018, soon after the TESS alert. We elected not to use the simultaneous wavelength calibration (i.e. the on-sky calibration fiber) to ensure that the bluer spectral regions would not be contaminated by the calibration lamp, that provides a much stronger flux for any instrumental setup.. The exposure time was set to 900 s for ESO programs 198.C-0838 and 1102.C-0339, and to 1[,]{}200 s for ESO program 0101.C-0510, translating in a median signal-to-noise ratio per spectral pixel of 51 and 70 at 650 nm, respectively. A single spectrum on October 1, 2018, had an exposure time of 609 s for an unknown reason. A total of 47 HARPS spectra were collected, ending with observations on December 19, 2018.
HARPS spectra were acquired in roughly three packs of data separated in time by about 40 days. This sampling is reflected in the window function presented in Figure \[fig:GLSP\]. Two archival spectra of L 168-9 are available at the ESO database. However, a radial velocity offset was introduced on May 2015 because the vacuum vessel was opened during a fiber upgrade [@LoCurto2015]. As this offset is not yet well characterized for M dwarfs, we decided to disregard those two points (from July 2008 and June 2009) in our subsequent analysis.
The HARPS Data Reduction Software [@LovisPepe2007] computes radial velocities by a cross-correlation function technique [@Baranne:1996 e.g.,]. Nevertheless we derived radial velocities by a different approach to exploit as much as possible the Doppler information of spectra [e.g., @Anglada2012]. We performed a maximum likelihood analysis between a stellar template and each individual spectrum following the procedure presented in @Astudillo-Defru2017b. The stellar template corresponds to a true stellar spectrum of the star, enhanced in signal-to-noise. It was made from the median of all the spectra, after shifting them into a common barycentric frame. The resulting template is Doppler shifted by a range of trial radial velocities to construct the maximum likelihood function, from which we derived the HARPS radial velocity used in the subsequent analysis. The obtained radial velocities – listed in Table \[tab:rvHARPS\] – present a dispersion of 4.01 $m/s$ and a median photon uncertainty of 1.71 $m/s$. Figure \[fig:phasedRV\] shows the radial velocities folded to the orbital period.
### PFS
The Planet Finder Spectrograph is an iodine-calibrated, environmentally controlled high resolution PRV spectrograph [@Crane:2006; @Crane:2008; @Crane:2010]. Since first light in October 2009, PFS has been running a long-term survey program to search for planets around nearby stars (e.g., @Teske:2016). In January 2018, PFS was upgraded with a new large format CCD with $9\mu m$ pixels and switched to a narrower slit for its regular operation mode to boost the resolution from 80[,]{}000 to 130[,]{}000. The PFS spectra are reduced and analyzed with a custom IDL pipeline that is capable of delivering RVs with $<$1 m/s precision [@Butler:1996].
We followed up L 168-9 with PFS on the 6.5 m Magellan II Clay telescope at Las Campanas Observatory in Chile from October 13–26, and then on December 16 and 21 in 2018. Observations were conducted on 15 nights, with multiple exposures per night. There were a total of 76 exposures of 20 minutes each. We typically took 2–6 exposures per night over a range of timescales, to increase the total SNR per epoch and also to average out the stellar and instrumental jitter. Each exposure had a typical SNR of 28 per pixel near the peak of the blaze function, or 56 per resolution element. The radial velocity dispersion is 4.61 m/s and the median RV uncertainty per exposure is about 1.8 m/s. Five consecutive 20-minute iodine-free exposures were obtained to allow for the construction of a stellar spectral template in order to extract the RVs. These were bracketed with spectra of rapidly rotating B stars taken through the iodine cell, for reconstruction of the spectral line spread function and wavelength calibration for the template observations.
The PFS observations of L 168-9 presented here are part of the Magellan TESS Survey (MTS) that will follow up $\sim$30 super-Earths and sub-Neptunes discovered by TESS in the next three years using PFS (Teske et al. in prep.). The goal of MTS is to conduct a statistically robust survey to understand the formation and evolution of super-Earths and sub-Neptunes. The observation schedules of all MTS targets, including L 168-9, can be found on the ExoFOP-TESS website.[^10]
Analysis {#sec:analysis}
========
Photometric rotation period {#subsec:rotation}
---------------------------
![*Left column*: Generalized Lomb-Scargle periodograms of the HARPS and PFS RV time series, window functions, and the S-index, $H\alpha$, $H\beta$, $H\gamma$, and sodium doublet activity indicators. The vertical dotted lines highlight the locations of the L 168-9 b orbital period, the stellar rotation period, and its first two harmonics. *Right column*: the false alarm probabilities computed from bootstrapping with replacement.[]{data-label="fig:GLSP"}](figures/periodogramsv2.png){width="0.98\hsize"}
Knowledge of the stellar rotation period helps to disentangle spurious RV signals arising from rotation and true RV signals due to orbital motion [e.g., @Queloz2001; @Cloutier2017b]. L 168-9 was photometrically monitored by WASP between May 2010 and December 2014 within five observing seasons each lasting approximately 200 days in duration. The photometric precision, observing cadence, and baselines within the WASP fields are sufficient to detect quasi-periodic (QP) photometric variations of L 168-9 due to active regions on the stellar surface rotating in and out of view at the stellar rotation period [$P_{\text{rot}}$]{}[.]{} The WASP photometry of L 168-9 is shown in Fig. \[fig:WASPphot\] along with the generalized Lomb-Scargle periodogram [GLSP; @Zechmeister2009] of the photometry in each of the five WASP observing seasons. It is clear from the GLSPs that a strong periodicity exists within the data whose timescale is often $\sim 30$ days except for within the second WASP season wherein the dominant periodicity appears at the first harmonic of [$P_{\text{rot}}$]{} $\sim 15$ days.
Given periodicities significantly detected in the WASP photometry, we proceeded to measure the photometric rotation period of L 168-9 [$P_{\text{rot}}$]{} with each WASP camera and in each WASP field[^11] in which L 168-9 was observed. As the photometric variations appear to vary nearly periodically, we modeled the photometry with a Gaussian process (GP) regression model and adopted a QP covariance kernel (see Eq. \[eq:covariance\]) [@Angus2018]. The covariance function’s periodic timescale was a free parameter [$P_{\text{rot}}$]{} for which the posterior probability density function (PDF) was sampled using a Markov Chain Monte Carlo (MCMC) method (see Appendix \[sec:GP\]). We modeled each binned WASP light curve with a GP. We adopted a bin size of 1 day to reduce the computation time. In preliminary analyses, we also tested bin sizes of 0.25, 0.5, and 2 days and found that, probably due to the very large number of points, the recovered values of [$P_{\text{rot}}$]{} were not very sensitive to this choice.
After sampling the posterior PDFs of the GP hyperparameters, we arrived at point estimates of each parameter value based on the maximum a-posteriori values and 68 percent confidence intervals. The resulting mean GP model of the data from each WASP observing season is depicted in Fig. \[fig:WASPphot\] along with the corresponding $1\sigma$ confidence interval. Over the five observing seasons, the measured (median) rotation period of L 168-9 was [$P_{\text{rot}}$]{} $=29.8\pm 1.3$ days.
In principle, the WASP signal could have arisen from any star within the 48" extraction aperture. However, L 168-9 is by far the brightest star in the aperture. Another concern with any photometric signal with a period near 30 days is whether it was affected by moonlight. To check on this possibility, we searched for modulations in the WASP data of several stars of similar brightness within the surrounding 10 arcmin field, but did not find any 30-d signals similar to the one that was seen for L 168-9. In any case, the star location is far from the ecliptic, and moonlight contamination is not expected. Furthermore, the modulation was sometimes seen at the 15-d first harmonic, which would not be expected for moonlight. We can therefore be confident that the 30-d periodicity in the WASP data arises from L 168-9. The $R^\prime_{HK}$ from HARPS spectra supports the obtained photometric rotation period, as $log(R^\prime_{HK})=-4.562\pm0.043$ (active star) translates into [$P_{\text{rot}}$]{} $=22\pm 2$ days using the $R^\prime_{HK}$vs.[$P_{\text{rot}}$]{} relationship from [@Astudillo2017a].
Radial Velocity Periodogram Analysis {#subsec:kep}
------------------------------------
A first identification of significant periodicities in the HARPS and PFS RV time series is required in order to develop an accurate model of the observed RV variations. In a manner similar to our analysis of the WASP photometry, we computed the GLSP of the following HARPS and PFS spectroscopic time series: the RVs, the window functions (WF), and the S-index, $H\alpha$, $H\beta$, $H\gamma$, and the sodium doublet NaD activity indicators. [@Astudillo-Defru2017b] details how these spectroscopic activity indicators were derived. Each GLSP is shown in Fig. \[fig:GLSP\] along with a false alarm probability (FAP) that was computed via bootstrapping with replacement using $10^4$ iterations and normalizing each periodogram by its standard deviation.
Each of the GLSPs of the HARPS and PFS RV time series is dominated by noise and aliases arising from the respective WF. For example, the HARPS WF contains a forest of peaks with comparable FAP from $\sim 8$ days and extending out towards long periodicities. Similarly the GLSP of the PFS WF reveals a series of broad peaks for periodicities $\gtrsim 1$ day. These features, particularly those from the PFS WF, have clear manifestations in the GLSPs of their respective RV and activity indicator time series, thereby complicating the robust identification of periodocities in the data. However, strong peaks at the orbital period of L 168-9 b ($\sim 1.4$ days) are discernible in both RV time series at FAP $\sim 0.4\%$ and $\sim 0.3\%$ with HARPS and PFS respectively. This periodicity is not apparent in any of the ancillary activity indicators time series as expected for a signal originating from an orbiting planet.
In addition to the signal from L 168-9 b, the HARPS RVs exhibit some power close to [$P_{\text{rot}}$]{} and its second harmonic [$P_{\text{rot}}$]{}$/3$. Although the FAPs of these periodicities over the full frequency domain are large, they each appear locally as strong periodicities as most of the power in the HARPS RVs exists at $\lesssim 5$ days. The GLSP of the PFS RVs is much more difficult to interpret at periodicities in the vicinity of [$P_{\text{rot}}$]{} and its first and second harmonics due to strong aliases from the PFS WF. Due to these effects it is difficult to discern from the available PFS activity indices (i.e. S-index and $H\alpha$) whether or not a coherent activity signal is seen with PFS. Although each of the HARPS and PFS RV time series are significantly affected by sampling aliases, we do see evidence for L 168-9 b and rotationally modulated stellar activity in the RVs we endeavor to mitigate with our adopted model discussed in Sect. \[subsec:model\].
Radial velocity + transit model {#subsec:model}
-------------------------------
Guided by the periodicities in the HARPS and PFS RV time series, we proceeded to fit a model to the RVs including the effects of both stellar activity and the planet. Following numerous successful applications on both Sun-like [e.g., @Haywood2014; @Grunblatt2015; @Faria2016; @Lopezmorales2016; @Mortier2016] and M dwarf stars [e.g., @Astudillo-Defru2017c; @Bonfils2018; @Cloutier2019a; @Ment2019], we adopted a QP kernel for the GP as a non-parametric model of the physical processes resulting in stellar activity. When used to model RV stellar activity, the QP covariance kernel is often interpreted as modelling the rotational component of stellar activity from active regions on the rotating stellar surface whose lifetimes typically exceed many rotation cycles on M dwarfs [@Giles2017] plus the evolutionary time scale of the active regions. The corresponding GP hyperparameters are described in detail in Appendix \[sec:GP\] and include each spectrograph’s covariance amplitudes $a_{\text{HARPS}}$, $a_{\text{PFS}}$, the common exponential timescale $\lambda_{\text{RV}}$, the common coherence parameter $\Gamma_{\text{RV}}$, and the common periodic timescale $P_{\text{RV}}$ equal to the stellar rotation period [$P_{\text{rot}}$]{}[.]{}
The planetary component attributed to the transiting planet L 168-9 b was fitted to the de-trended light curve with a [@MandelAgol2002] planetary transit model. The de-trended TESS light curve was produced by adjusting a QP GP systematic model to the photometry alone and with all the transits previously removed. The best QP GP model was subtracted to the entire TESS data set. The planetary component is modelled by a Keplerian solution parameterized by the planet’s orbital period $P_b$, time of mid-transit $T_0$, RV semi-amplitude $K$, and the orbital parameters $h=\sqrt{e}\cos{\omega}$ and $k=\sqrt{e}\sin{\omega}$ where $e$ and $\omega$ are the planet’s orbital eccentricity and argument of periastron respectively. In addition, our RV model contains each spectrograph’s zero point velocity $\gamma_{\text{HARPS}}$, $\gamma_{\text{PFS}}$ and an additive scalar jitter $s_{\text{HARPS}}$, $s_{\text{PFS}}$ is account for any residual jitter that, unlike the stellar activity signal, is not temporally correlated. The complete RV model therefore contains fourteen model parameters.
To ensure self-consistent planet solutions between the available TESS transit data and the RV observations, we simultaneously fitted the de-trended light curve and the RVs. The common planetary parameters between these two data sets are $P_b$, $T_0$, $h$, and $k$. The additional model parameters required to model the TESS transit light curve included an additive scalar jitter $s_{\text{TESS}}$, the baseline flux $\gamma_{\text{TESS}}$, the scaled semi-major axis $a/R_s$, the planet-star radius ratio $r_p/R_s$, the orbital inclination $i$, and the nearly-uncorrelated parameters $q_1$ and $q_2$ which are related to the quadratic limb darkening coefficients $u_1$ and $u_2$ via $$\begin{aligned}
q_1 &= (u_1+u_2)^2 \\
q_2 &= \frac{u_1}{2(u_1+u_2)}.
\label{eq:LDCs}\end{aligned}$$
[@Kipping2013]. Thus we required eleven model parameters to describe the TESS transit light curve and a total of twenty-one model parameters of the joint RV + light curve data set: $\boldsymbol{\Theta}=\{a_{\text{HARPS}}, a_{\text{PFS}}, \lambda_{\text{RV}},
\Gamma_{\text{RV}}, P_{\text{RV}}, s_{\text{TESS}}, s_{\text{HARPS}}, s_{\text{PFS}}, \gamma_{\text{TESS}}, \gamma_{\text{HARPS}}$, $\gamma_{\text{PFS}},
P_b, T_0, K, h, k, a/R_s, r_p/R_s, i, q_1, q_2 \}$.
We sampled the posterior PDF of this 21-dimensional parameter space using an MCMC sampler. Details on the sampler and the adopted prior distributions on each model parameter are given in Appendix \[sec:GP\] and Table \[tab:priors\].
Results {#sec:results}
=======
![Phase folded radial velocity acquired HARPS (blue points) and PFS (red points) where the best GP model was subtracted. The black curve represents the maximum a-posteriori model adjusted to the data set.[]{data-label="fig:phasedRV"}](figures/RVphased.png){width="1.0\hsize"}
In our analysis of the light curve and radial velocity time series of L 168-9, we evaluated the model presented in Sect. \[subsec:model\] on the separate RV data sets obtained with HARPS and PFS as well as the combined time series. Subtracting the model to the 46 HARPS radial velocity points reduces the dispersion to 3.37 m/s (equivalent to $\chi^2_{red}=3.5$), while the dispersion of the 76 PFS residual points gives 4.05 m/s (translating into $\chi^2_{red}=1.5$). The dispersion obtained from the 122 HARPS+PFS residual points is 3.80 m/s ($\chi^2_{red}=2.2$).
From point estimates of the model parameters $\boldsymbol{\Theta}$ from our joint RV plus transit analysis with each of these input data sets we retrieved that L 168-9 b has a radius of 1.39$\pm$0.09 R$_\oplus$ and a mass of 4.60$\pm$0.56 M$_\oplus$, translating into a bulk mean density of $9.6^{+2.4}_{-1.8}$ g cm$^{-3}$. The planet is orbiting at 0.02091$\pm$0.00024 AU from the parent star, therefore the hot terrestrial planet has an equilibrium temperature between 668 K and 965 K assuming a Venus-like and zero bond albedo, respectively. Results for the entire set of parameters are reported in Table \[tab:results\]. Explicitly, we report the maximum a-posteriori value of each parameter along with its 16$^{th}$ and 84$^{th}$ percentiles, corresponding to a $1 \sigma$ confidence interval. We check for consistency of our joint analysis by performing the analysis for each instrument independently. Table \[tab:results\] details the results from this test. We note that HARPS and PFS results are in agreement within their uncertainties, translating into a robust detection of the planetary signal in radial velocity data.
Figure \[fig:tess\_phot\] show the TESS photometry and adjusted transit model and Figure \[fig:phasedRV\] the phase folded radial velocity with the model that best fits the data.
[lccc]{} Measured transit model parameters & & TESS &\
Baseline flux, $\gamma_{\text{TESS}}$ &\
Orbital period, $P_b$ \[days\] &\
Time of mid-transit, $T_0$ \[BJD-2,457,000\] &\
Scaled semi-major axis, $a/R_s$ &\
Planet-star radius ratio, $r_p/R_s$ &\
Orbital inclination, $i$ \[deg\] &\
Linear limb darkening coefficient, $q_1$ &\
Quadratic limb darkening coefficient, $q_2$ &\
TESS additive jitter, $s_{\text{TESS}}$ &\
Radial velocity GP hyperparameters & HARPS+PFS & HARPS & PFS\
ln HARPS covariance amplitude, $\ln{(a_{\text{HARPS}}/\text{m s}^{-1})}$ & $3.27^{+1.38}_{-1.03}$ & $3.24^{+1.24}_{-1.19}$ & -\
ln PFS covariance amplitude, $\ln{(a_{\text{PFS}}/\text{m s}^{-1})}$ & $3.63^{+1.24}_{-1.08}$ & - & $4.01^{+1.23}_{-0.97}$\
ln RV exponential timescale, $\ln{(\lambda_{\text{RV}}/\text{day})}$ & $11.90^{+2.81}_{-1.99}$ & $12.25^{+3.16}_{-2.18}$ & $13.3^{+2.02}_{-1.44}$\
ln RV coherence, $\ln{(\Gamma_{\text{RV}})}$ & $-0.09^{+0.20}_{-0.24}$ & $-0.42^{+0.41}_{-0.55}$ & $-0.22^{+0.39}_{-0.44}$\
ln RV periodic timescale, $\ln{(P_{\text{RV}}/\text{day})}$ & $3.47^{+0.02}_{-0.03}$ & $3.47^{+0.03}_{-0.03}$ & $3.48^{+0.04}_{-0.03}$\
HARPS additive jitter, $s_{\text{HARPS}}$ \[[m s$^{-1}$]{}[\]]{} & $0.10^{+0.20}_{-0.09}$ & $0.86^{+0.98}_{-0.86}$ & -\
PFS additive jitter, $s_{\text{PFS}}$ \[[m s$^{-1}$]{}[\]]{} & $2.82\pm 0.42$ & - & $2.78\pm 0.30$\
Measured RV model parameters & HARPS+PFS & HARPS & PFS\
HARPS zero point velocity, $\gamma_{\text{HARPS}}$ \[km s$^{-1}$\] & $29.7687\pm 0.0013$ & $29.7692\pm 0.0013$ & -\
PFS zero point velocity, $\gamma_{\text{PFS}}$ \[km s$^{-1}$\] & $0.00077 \pm 0.00186$ & - & $0.00063\pm 0.00142$\
Semi-amplitude, $K$ \[[m s$^{-1}$]{}[\]]{} & $3.66^{+0.47}_{-0.46}$ & $3.74^{+0.54}_{-0.59}$ & $3.26^{+0.52}_{-0.70}$\
$h=\sqrt{e}\cos{\omega}$ & $-0.10^{+0.17}_{-0.12}$ & $0.04^{+0.15}_{-0.18}$ & $-0.01^{+0.18}_{-0.21}$\
$k=\sqrt{e}\sin{\omega}$ & $0.00^{+0.17}_{-0.16}$ & $-0.09^{+0.22}_{-0.23}$ & $0.01^{+0.23}_{-0.30}$\
Derived L 168-9 b parameters & HARPS+PFS+TESS & HARPS+TESS & PFS+TESS\
Semi-major axis, $a$ \[AU\] & $0.02091\pm 0.00024$ &&\
Equilibrium temperature, $T_{\text{eq}}$ \[K\] &&&\
Zero bond albedo & $965\pm 20$ &&\
Venus-like bond albedo = 0.77 & $668\pm 14$ &&\
Planetary radius, $R_p$ \[R$_{\oplus}$\] & $1.39\pm 0.09$ &&\
Planetary mass, $M_p$ \[M$_{\oplus}$\] & $4.60\pm 0.56$ & $4.74^{+0.71}_{-0.75}$ & $4.08^{+0.70}_{-0.90}$\
Planetary bulk density, $\rho_p$ \[g cm$^{-3}$\] & $9.6^{+2.4}_{-1.8}$ & $10.0^{+2.5}_{-1.9}$ & $8.7^{+2.3}_{-2.1}$\
Planetary surface gravity, $g_p$ \[m s$^{-2}$\] & $23.9^{+4.5}_{-3.9}$ & $24.4^{+4.8}_{-4.2}$ & $21.4^{+4.4}_{-4.8}$\
Planetary escape velocity, $v_{\text{esc}}$ \[km s$^{-1}$\] & $20.5^{+1.4}_{-1.4}$ & $20.8^{+1.6}_{-1.7}$ & $19.4^{+1.6}_{-2.2}$\
Orbital eccentricity, $e^\dagger$ & $< 0.21$ & $< 0.25$ & $< 0.26$\
95% confidence interval.
Discussion & Conclusions {#sec:conclusion}
========================
L 168-9 b adds to the family of small ($<2R_\oplus$) transiting planets around bright ($J<$ 8 mag) stars with mass measurements and contributes to the completion of the TESS Level One Science Requirement to detect and measure the masses of 50 small planets. In particular, L 168-9 b is one of fourteen[^12] likely rocky planets without primordial hydrogen-helium envelopes – that is, with a radius $<1.8R_\oplus$ – for which the mass has been measured with an uncertainty smaller than 33%. Thus, our result represents progress toward the understanding of the transition between super-Earths and mini-Neptunes previously reported in the radii of planets [e.g. @Fulton2017; @Cloutier2019c] but, here, including the information on mass.
Figure \[fig:MR\] shows the mass - radius diagram centered in the sub-Earth to mini-Neptune regime. With about twice the Earth average density, L 168-9 b bulk density is compatible with a terrestrial planet with an iron core (50%) surrounded by a mantle of silicates (50%). In this diagram the detected planet is located in an interesting place: for masses lower than that of L 168-9 b the great majority of planets are consistent with a 50% Fe–50% MgSiO3 or 100% MgSiO3 bulk composition, while for higher planetary masses there is a great diversity of density. Being one of the densest planets for masses greater than 4M$_\oplus$, L 168-9 b can help to define the mass limits of the rocky planets population.
![The mass-radius diagram showing L 168-9 b (red circle) in the context of known exoplanets. The transparency of each point is proportional to its associated mass uncertainty. Error bars correspond to $1 \sigma$ uncertainties. Different models for the bulk composition are plotted where the legend details the fraction of iron, silicates, and/or water for each color-coded curve.[]{data-label="fig:MR"}](figures/MR.pdf)
Good targets for atmospheric characterization with transmission/emission spectroscopy are those transiting nearby, bright stars ($J<$ 8 mag). There are currently 11 small planets detected transiting a bright star, according to the NASA Exoplanet Archive[^13], two of them orbit M dwarfs. Overall, there would not be a large number of small, transiting planets around nearby, bright M dwarfs. There are roughly 200 M dwarfs within the 25 pc solar neighborhood with $J<$ 8 mag, about 2/3 of which are single stars [@Winters:2015; @Winters:2019]. Considering the occurrence rate of small planets with orbital period smaller than 10 days from [@Dressing:2015] and combining with the transit probability of such planets (about a couple % to $\sim$20%), there would be a few to up to about twenty such planets.
The measured properties of L 168-9 b and its host star make it a promising target for the atmospheric characterization of a terrestrial planet via either transmission or emission spectroscopy measurements with JWST [@Morley2017; @Kempton2018] and/or thermal phase curve analysis to infer the absence of a thick atmosphere [e.g., @Seager2009]. Its transmission and emission spectroscopy metrics from [@Kempton2018] are reported in Table \[tab:atmosphere\] and compared to other confirmed transiting terrestrial planets with known masses that are of interest for atmospheric characterization. Based on this assessment, L 168-9 b is an excellent candidate for emission spectroscopy or for detecting the planetary day-side phase curve as recently done for the similar planet, LHS 3844 b [@Kreidberg2019].
[lcccccccccccccc]{} Star ID & R$_{p}$ & M$_{p}$ & P & a& T$_{eff}$ & T$_{eq}$ & T$_{day}$ & J & K$_{\rm s}$ & R$_\star$ & M$_\star$ & TSM & ESM & Ref.\
units & \[R$_\oplus$\] & \[M$_\oplus$\] & \[days\] & \[AU\]& \[K\] & \[K\] & \[K\] & \[mag\] & \[mag\] & \[R$_\odot$\] & \[M$_\odot$\] & & &\
L 168-9 b & 1.39 & 4.39 & 1.401 & 0.021 & 3743 & 963.01 & 1059.31& 7.941 & 7.0819 & 0.60 & 0.62 & 8.025 & 9.692&\
LHS 3844 b & 1.30 & – & 0.463 & 0.006 & 3036 & 805.90 & 886.49 & 10.046& 9.145 & 0.19 & 0.15 & – & 29.004 & Kr2019\
GJ 1132 b & 1.13 & 1.66 & 1.629 & 0.015 & 3270 & 590.61 & 649.67 & 9.245 & 8.322 & 0.21 & 0.18 & 31.166 & 9.872 & Bo2018\
L 98-59 c & 1.35 & 2.17 & 3.690 & 0.032 & 3412 & 515.31 & 566.84 & 7.933 & 7.101 & 0.31 & 0.31 & 29.168 & 6.696 & Cl2019b\
LTT 1445A b & 1.38 & 2.20 & 5.359 & 0.038 & 3335 & 433.34 & 476.68 & 7.29 & 6.5 & 0.28 & 0.26 & 44.976 & 6.382 & Wi2019\
TRAPPIST-1 b & 1.09 & 1.02 & 1.511 & 0.011 & 2559 & 402.38 & 442.62 & 11.4 & 10.3 & 0.12 & 0.08 & 36.914 & 4.007 & Gi2017\
LHS 1140 c & 1.28 & 1.81 & 3.778 & 0.027 & 3216 & 436.43 & 480.07 & 9.612 & 8.821 & 0.21 & 0.18 & 25.225 & 3.401 & Me2019\
Reference notes: Kr2019 – [@Kreidberg2019]; Bo2018 – @Bonfils2018; Cl2019b – @Cloutier2019b; Wi2019 – @Winters2019; Gi2017 – @Gillon2017; Me2019 – @Ment2019;
N. A.-D. acknowledges the support of FONDECYT project 3180063. J.K.T. acknowledges that support for this work was provided by NASA through Hubble Fellowship grant HST-HF2-51399.001 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS5-26555. R.B. acknowledges support from FONDECYT Post-doctoral Fellowship Project 3180246, and from the Millennium Institute of Astrophysics (MAS). X.B. and J.M-A. acknowledge funding from the European Research Council under the ERC Grant Agreement n. 337591-ExTrA. X.D.; X.B.; T.F.; et L.M. acknowledge the support by the French National Research Agency in the framework of the Investissements d’Avenir program (ANR-15-IDEX-02), through the funding of the “Origin of Life” project of the Univ. Grenoble-Alpes.” LM acknowedge the support of the Labex OSUG@2020 (Investissements d’avenir – ANR10 LABX56). JRM acknowledges CAPES, CNPq and FAPERN brazilian agencies. This work was supported by FCT/MCTES through national funds and by FEDER - Fundo Europeu de Desenvolvimento Regional through COMPETE2020 - Programa Operacional Competitividade e Internacionalização by these grants: UID/FIS/04434/2019; PTDC/FIS-AST/32113/2017 & POCI-01-0145-FEDER-032113; PTDC/FIS-AST/28953/2017 & POCI-01-0145-FEDER-028953. T.H. acknowledges support from the European Research Council under the Horizon 2020 Framework Program via the ERC Advanced Grant Origins 83 24 28. REM acknowledges support by the BASAL Centro de Astrofísica y Tecnologías Afines (CATA) and FONDECYT 1190621. A.J. acknowledges support from FONDECYT project 1171208 and by the Ministry for the Economy, Development, and Tourism’s Programa Iniciativa Científica Milenio through grant IC120009, awarded to the Millennium Institute of Astrophysics (MAS). JGW is supported by a grant from the John Templeton Foundation. The opinions expressed here are those of the authors and do not necessarily reflect the views of the John Templeton Foundation. Work J.N.W. was partly funded by the Heising-Simons Foundation. The authors would like to acknowledge Zachary Hartman for his help conducting the Gemini-South/DSSI observations. Some of the work here is based on observations obtained at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), National Research Council (Canada), CONICYT (Chile), Ministerio de Ciencia, Tecnología e Innovación Productiva (Argentina), Ministério da Ciência, Tecnologia e Inovação (Brazil), and Korea Astronomy and Space Science Institute (Republic of Korea). This work makes use of observations from the LCOGT network. Funding for the TESS mission is provided by NASA’s Science Mission directorate. We acknowledge the use of public TESS Alert data from pipelines at the TESS Science Office and at the TESS Science Processing Operations Center. This research has made use of the Exoplanet Follow-up Observation Program website, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program. Resources supporting this work were provided by the NASA High-End Computing (HEC) Program through the NASA Advanced Supercomputing (NAS) Division at Ames Research Center for the production of the SPOC data products. This paper includes data collected by the TESS mission, which are publicly available from the Mikulski Archive for Space Telescopes (MAST). This research has made use of the NASA Exoplanet Archive, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program.
Gaussian process model {#sec:GP}
======================
Gaussian process (GP) regression is widely used in the exoplanet community as a non-parametric Bayesian approach to model temporally correlated stellar activity signals in RV data. These rotationally-modulated activity signals prohibit the accurate measurement of planetary parameters and often produce in planetary false positives.
Here we model the RV activity signals of L 168-9 as a stochastic process whose temporal evolution is well-described by a quasi-periodic covariance kernel. A GP with a quasi-periodic covariance kernel $k(t_i,t_j)$ is included in our joint model describing the data RV and transit data and takes the following form:
$$\label{eq:covariance}
k(t_i,t_j) = a^2 \exp{\left[ -\frac{(t_i-t_j)^2}{2\lambda^2} -\Gamma^2 \sin^2{\left( \frac{\pi |t_i-t_j|}{P_{\text{GP}}} \right)} \right]}$$
and is described by the covariance amplitude $a$, the exponential evolutionary timescale $\lambda$, the coherence $\Gamma$, and the periodic timescale $P_{\text{GP}}$.
As usual in the Bayesian context, prior probability density functions (PDF) for the hyperparameters $\{ a,\ \lambda,\ \Gamma,\ P_{\text{GP}} \}$ are listed in Table \[tab:priors\] for the multiple GPs applied our data analysis. The posterior PDF are sampled through a Markov chain Monte Carlo [MCMC, @GoodmanAndWeare2010]; in particular we used the `emcee` ensemble sampler [@Foremanmackey2013]. The sampling of the joint posterior is made with the Gaussian ln likelihood function given by
$$\ln{\mathcal{L}} = -\frac{1}{2} \left( y^{\text{T}}\cdot K \cdot y + \ln{\text{det}K} + N\ln{2\pi} \right)$$
where $y$ is the vector of $N$ measurement taken at times $t=\{t_1,t_2,\dots,t_N \}$ and the $N \times N$ covariance matrix $K$ is given by
$$K_{ij} = k(t_i,t_j) + \delta_{ij}[\sigma(t_i)^2 + s^2].$$
being $\delta_{ij}$ the Kronecker delta that adds the measurement uncertainties $\sigma$ to the diagonal elements of $K$ and includes an additive jitter factor $s$.
[lc]{} Parameter & Prior\
\
Baseline flux, $\gamma_{\text{TESS}}$ & $\mathcal{U}(-0.1,0.1)$\
Limb darkening coefficient, $u_1$ &$\mathcal{U}(0.1,0.4)$\
Limb darkening coefficient, $u_2$ & $\mathcal{U}(0.25,0.55)$\
TESS additive jitter, $s_{\text{TESS}}$ & $\mathcal{J}(10^{-2},1)$\
&\
\
HARPS zero point velocity, $\gamma_{\text{0,HARPS}}$ \[[m s$^{-1}$]{}[\]]{} & $\mathcal{U}(-1,1)$\
PFS zero point velocity, $\gamma_{\text{0,PFS}}$ \[[m s$^{-1}$]{}[\]]{} & $\mathcal{U}(-1,1)$\
ln HARPS covariance amplitude, $\ln{(a_{\text{HARPS}}/\text{m s}^{-1})}$ & $\mathcal{U}(-5,10)$\
ln PFS covariance amplitude, $\ln{(a_{\text{PFS}}/\text{m s}^{-1})}$ & $\mathcal{U}(-5,10)$\
ln RV exponential time scale, $\ln{(\lambda_{\text{RV}}/\text{day})}$ & From H$\alpha$ training\
ln RV coherence, $\ln{(\Gamma_{\text{RV}})}$ & From H$\alpha$ training\
ln RV periodic timescale, $\ln{(P_{\text{RV}}/\text{day})}$ & From H$\alpha$ training\
HARPS additive jitter, $s_{\text{HARPS}}$ \[[m s$^{-1}$]{}[\]]{} & $\mathcal{J}(10^{-2},10)$\
PFS additive jitter, $s_{\text{PFS}}$ \[[m s$^{-1}$]{}[\]]{} & $\mathcal{J}(10^{-2},10)$\
\
Orbital period, $P_b$ \[days\] & $\mathcal{U}(1.370,1.405)$\
Time of mid-transit, $T_{0,b}$ \[BJD - 2,457,000\] & $\mathcal{U}(1338, 1342)$\
Scaled semi-major axis, $a/R_s$ & $\mathcal{G}(7.542, 0.27)$\
Planet-star radius ratio, $r_p/R_s$ & $\mathcal{U}(0,0.1)$\
Orbital inclination, $i$ \[deg\] & $\mathcal{U}(75,105)$\
Semi-amplitude, $K_b$ \[[m s$^{-1}$]{}[\]]{} & $\mathcal{J}(10^{-2},10)$\
$h_b=\sqrt{e_b}\cos{\omega_b}$ & $\mathcal{U}(-1,1)$\
$k_b=\sqrt{e_b}\sin{\omega_b}$ & $\mathcal{U}(-1,1)$\
Spectroscopic Data
==================
[@lllllllllllll]{}
BJD & RV & $\sigma_{RV}$ & H$\alpha$ & $\sigma_{H \alpha}$ & H$\beta$ & $\sigma_{H \beta}$ & H$\gamma$ & $\sigma_{H \gamma}$ & NaD & $\sigma_{NaD}$ & S & $\sigma_{S}$\
& $[ms^{-1}]$ & $[ms^{-1}]$ &&&&&&&&&&\
4664.954608 & 29781.65 & 3.48 & 0.06224 & 0.00025 & 0.04679 & 0.00052 & 0.10087 & 0.00131 & 0.01232 & 0.00020 & 1.616 & 0.068\
4991.927461 & 29786.04 & 2.95 & 0.06074 & 0.00021 & 0.04632 & 0.00044 & 0.10385 & 0.00114 & 0.01172 & 0.00016 & 2.167 & 0.073\
8367.523380 & 29769.97 & 1.37 & 0.05962 & 0.00011 & 0.04606 & 0.00019 & 0.10582 & 0.00052 & 0.01166 & 0.00007 & 1.892 & 0.018\
8367.551447 & 29773.09 & 1.19 & 0.05928 & 0.00010 & 0.04534 & 0.00016 & 0.10573 & 0.00045 & 0.01158 & 0.00006 & 1.826 & 0.014\
8367.623921 & 29772.57 & 1.20 & 0.05944 & 0.00010 & 0.04562 & 0.00017 & 0.10639 & 0.00046 & 0.01164 & 0.00006 & 1.868 & 0.014\
[@lllll]{}
BJD & RV & $\sigma_{RV}$ & H$\alpha$ & S\
& $[ms^{-1}]$ & $[ms^{-1}]$ &&\
8409.51799 & 8.80 & 1.74 & 1.293 & 0.05807\
8409.53206 & 4.10 & 1.74 & 1.334 & 0.05813\
8409.60491 & -3.11 & 1.71 & 1.285 & 0.05879\
8409.61960 & 4.80 & 1.75 & 1.340 & 0.05901\
8409.67495 & 8.23 & 2.09 & 1.311 & 0.05976\
[^1]: Partially based on observations made with the HARPS instrument on the ESO 3.6 m telescope under the program IDs 198.C-0838(A), 0101.C-0510(C), and 1102.C-0339(A) at Cerro La Silla (Chile).
[^2]: Data (Tables XXX) are only available at the CDS via anonymous ftp to cdsarc.u-strasbg.fr (130.79.128.5) or via http://cdsarc.u-strasbg.fr/viz-bin/qcat?J/A+A/XXX/XXX
[^3]: This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile.
[^4]: The Sector 1 pointing direction was $RA(J2000):+352.68\degree,\
Dec(J2000):-64.85\degree,\ Roll:-137.85\degree$.
[^5]: <https://archive.stsci.edu/missions/tess/doc/EXP-TESS-ARC-ICD-TM-0014.pdf>
[^6]: <https://mast.stsci.edu/portal/Mashup/Clients/Mast/Portal.html>
[^7]: <https://tev.mit.edu/toi/alerts/>
[^8]: <https://exofop.ipac.caltech.edu/tess/>
[^9]: <https://astro.swarthmore.edu/telescope/tess-secure/find_tess_transits.cgi>
[^10]: <https://exofop.ipac.caltech.edu/tess/>
[^11]: At times, L 168-9 appeared within the fields-of-view of multiple WASP cameras.
[^12]: L 98-59 bc [@Cloutier2019b], GJ 357 b [@Luque2019], HD 15337 b [@Dumusque2019], HD 213885 b [@Espinoza2019], GJ 9827 b [@Rice2019], K2-265 b [@Lam2018], K2-141 b [@Barragan2018], K2-229 b [@Santerne2018], HD 3167 b [@Gandolfi2017], K2-106 b [@Guenther2017], TRAPPIST-1 fh [@Wang2017], HD 219134 b [@Motalebi2015]
[^13]: <https://exoplanetarchive.ipac.caltech.edu/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We compare among themselves two different methods for the derivation of results following from the requirement of polynomial boundedness of tree-level chiral amplitudes. It is shown that the results of the algebraic approach are valid also in the framework of the analytical one. This means that the system of Sum Rules and Bootstrap equations previously obtained with the help of the latter approach can be analyzed in terms of reducible representations of the unbroken Chiral group with the known structure of the mass matrix.'
---
-2.0cm -3.0cm
SPbU-IP-96-33
[**Chiral Symmetry in algebraic and analytic approaches**]{}\
\
St.Petersburg State University, St.Petersburg, Russia\
and\
Institute for Theor. Physics III, Univ. of Erlangen-Nuernberg, Erlangen, Germany\
M. Dillig\
Institute for Theor. Physics III, Univ. of Erlangen-Nuernberg, Erlangen, Germany\
A. Vereshagin\
St.Petersburg State University, St.Petersburg, Russia\
Novadays it is commonly realised that the Spontaneously Broken Chiral Symmetry (SBChS) plays an essential role in the hadronization mechanism provided by the large-distance forces in OCD. The smallness of the pion mass (compared to the typical hadronic scale of 1 Gev) made it possible to develop the Chiral Perturbation Theory (ChPT) [@1; @2] (for recent reviews see also [@3; @4; @5]) allowing one to calculate in a systematic manner the corrections to the zero-order (“bare”) amplitude given by a sum of corresponding tree graphs. Thus, if the bare amplitudes are known, we can analyze the low energy processes with the arbitrary high degree of accuracy.
Unfortunately, the bare (tree-level) amplitudes themselves contain a lot of free parameters since they depend of infinite number of basically unknown coupling constants describing the off-shell hadron interactions and point-like vertices. Just because of this reason the current situation looks unsatisfactory: we have an excellent tool – ChPT – for the computation of corrections to the quantities which we know not so much about.
A remarkable step in the understanding of the bare couplings was done by S.Weinberg [@6], first suggested to use the asymptotic boundedness requirements for the chiral amplitudes describing the forward scattering of massless pions on arbitrary targets (the list of the corresponding refs. can be found, e.g., in [@7] ; the paper [@8] should be added to this list). It was shown that the SBChS – despite of its dynamical origin – manifests itself in a customary algebraic manner. This means that it is possible to classify the states with a given helicity according to linear reducible representations of unbroken Chiral group, the mass matrix being constructed as a sum of chiral scalar and 4-th component of chiral vector.
A variety of interesting results derived by many authors in the framework of Weinberg’s scheme clearly demonstrates both the importance and the fruitfulness of the asymptotic requirements.
There are known two technically different approaches to the derivation of results from those requirements.
Weinberg’s original approach [@6] (later on called algebraic) consisting of the group theoretical analysis of mass (squared) and coupling matrices, shows both attractive features and serious limits. One of the most attractive features of this method is that it allows one to work from the very beginning with purely algebraic structures. However, it is applicable only for the case of scattering of pions; its feasibility is limited on a small number of partial waves taken into account. Moreover (as stressed by Weinberg himself [@6]), from the mathematical point of view this method is only valid if the number of resonances is finite.
A different approach (first suggested in [@9]) has been recently formulated in its final form in [@7]. Later on we call this approach analytic. It takes advantage of the analytic structure of a tree-level amplitude, this structure being defined by the requirements of meromorphy and polynomial boundedness. In this method there is no necessity to require that pion is massless. Also, it is not restricted to forward processes only: the nonforward ones can be analyzed on the same ground. The method does not contain any limitations on the number of resonances or partial waves included. Moreover, it follows from the results of ref. [@7] that, to avoid self-contradictions in the case if a partial wave with $l \geq 2$ (or, the same, the resonance with a spin $J \geq 2$) is taken into account, one has to include also the full tower of higher spin resonances with couplings and masses strongly restricted by a certain infinite set of self-consistency (bootstrap) equations. So, compared to the algebraic approach, the analytic one looks more general, since it contains the algebraic method as a particular case. Moreover, it contains also the built-in mechanism guaranteing the absence of mutually contradicting results in those cases when the algebraic method is applied for the simultaneous analysis of direct- and cross-channel processes.
However, as a serious drawback, the problem of algebraization of the system of bootstrap equations – which is clearly absent in the algebraic approach – looks rather complicated in the analytical framework. Thus, it is interesting to ask the question if it is possible to get Weinberg’s results from the analysis of the corresponding part of a system of bootstrap equations derived with a help of the analytic method.
Below we give the positive answer to this question along with the general outline of the corresponding proof. A more detailed consideration will be published elsewhere.
Let us first recall some key points of the analytic approach (for more details see [@7]). This approach is based on the following postulates:\
1. The tree-level amplitude of a given binary process is a meromorphic function in a space of 3 dependent complex Mandelstam variables $s, t, u$ ($s+t+u=\sum_{i=1}^4 {m_i}^2)$.\
2. In accordance with the crossing symmetry requirement, this amplitude describes all three channels of a process under consideration.\
3. The principal part of the amplitude is completely determined by the formally written tree-level expression following from the “naive” (unitary) Feynman rules (no ghosts, no tachions!).\
4. As a function of every complex variable $x$ (considered as the CMS energy squared in a given channel) this amplitude is polynomially bounded at zero value of the corresponding momentum transfer, the degree of a polynomial being dictated by the value of relevant Regge intercept (or, equivalently, by experiment).
The main technical tool used in the analytic approach is provided by the Cauchy method. This method allows one to construct the power fraction expansion for the meromorphic function $f(z)$ of one complex variable $z$ with a given principal part, this expansion being coordinated with the asymptotic condition of the form $$\begin{aligned}
\max_{z \in C_n}\left|\frac{f(z)}{z^{N+1}}\right|
\stackrel{n \to \infty}{\longrightarrow} 0 \;,
\label{1}\end{aligned}$$ where $C_n$ denotes a smooth contour (for definiteness, a circle) which contains inside of it the first $n$ poles ( $|p_{i+1}|> |p_i| , p_i$ is the $i$-th pole position, $r_i$ – the corresponding residue; $p_i$ being enumerated in the order of increasing modulo) and does not contain any other poles. The minimal integer number $N$ in (\[1\]) shows the degree of the bounding polynomial.
With the condition (\[1\]) the Cauchy method gives: $$\begin{aligned}
f(z)= \sum_{p=0}^N f^{p}(0)\frac{z^p}{p!}+\sum_{i=1}^{\infty}
\left( \frac{r_i}{z-p_i} - {\Pi}_i (z) \right) .
\label{2}\end{aligned}$$ Here ${\Pi}_i (z)$ stands for the so-called correcting polynomial: it is nothing but the sum of the first $N$ terms of the power series expansion of $r_i / (z-p_i)$ around the point $z=0$. The necessity in correcting polynomials is caused by the convergency condition. In fact, the eq. (\[2\]) provides a special case of the general theorem by Mittag-Leffler.
In a framework of the analytic method one works with the amplitude written in the form (\[2\]). In the case of forward process all the parameters appearing in (\[2\]) (namely, $ f^{(0)}(0), r_i , p_i $ ) are just constants; otherwise, they should be considered as the functions of the momentum transfer. The form (\[2\]) is convenient because it does not contain any unwanted terms breaking the allowed asymptotic regime, thus from the very beginning there is no need to control the correctness of the asymptotics.
This is in strong contrast to the algebraic method, where results are derived from the condition of cancellation of just those “unwanted” terms which increase too rapidly with energy but appear in the scattering amplitude if the Lagrangian is written in terms of covariant derivatives (see [@6]). In this approach one calculates the Laurent series expansion coefficient which breaks the allowed asymptotic regime and then requires of it to be zero.
At the first glance two approaches described above look quite different. Indeed, since the “naive” Laurent series expansion is only applied for the study of asymptotics if the number of poles is finite, it looks impossible to use it for the analysis of cross-conjugated processes where the infinite set of resonances is required to provide the correct asymptotic behavior in both channels simultaneously (see [@7]). This, in turn, could mean that the results of early papers [@10], based on the calculation of the Laurent coefficients, have no ground. At the same time, it is possible to show that the Sum Rules given in [@10] can be also derived with a help of the form (\[2\]). So, it looks that the conclusions obtained in a framework of the algebraic method remain also valid with respect to the results of the analytic approach.
To check (and prove) this idea [@11] let us consider the case $N=0$ in (\[1\]) and (\[2\]) (just in order to simplify the corresponding formulas). In this case from the integral form of the asymptotic condition (\[1\]) $$\begin{aligned}
\oint\limits_{C_n}^{}\left| \frac{f(z)}{z^2}\,dz \right|
\stackrel{n \to \infty}{\longrightarrow} 0 ,
\label{3}\end{aligned}$$ it follows that $$\begin{aligned}
\sum_{i=n+1}^{\infty} \frac{r_i}{p_i^2}
\stackrel{n \to \infty}{\longrightarrow} 0 .
\label{4}\end{aligned}$$
Next, let us consider the true Laurent series expansion $$\begin{aligned}
f(z) = \sum_{- \infty}^{\infty} C_p^{(n)}z^p,\;\; \;\;\; (z\in R_n) ,
\label{5}\end{aligned}$$ where $R_n$ stands for the ring $|z| \in (|p_n|, |p_{n+1}|)$ and calculate the lowest “unwanted” coefficient $C_{+1}^{(n)}$. It is easy to get $$\begin{aligned}
C_{+1}^{(n)} = - \sum_{i=n+1}^{\infty}\frac{r_i}{p_i^2} .
\label{6}\end{aligned}$$ To study the behavior of the amplitude $f(z)$ at large values of $|z|$ one can use the expansion (\[5\]) and take a limit $ n \to \infty $. In this case – as it follows immediately from the comparison of (\[6\]) with (\[4\]) – one obtains the result $$\begin{aligned}
C_{+1}^{(n)} \stackrel{n \to \infty}{\longrightarrow} 0 ,
\label{7}\end{aligned}$$ showing that the lowest “unwanted” Laurent coefficient vanishes (along with all higher ones $C_i ^{(n)}, i=+2,+3,...$). This result justifies Weinberg’s formal method for the case of infinite number of resonances. It explains also the complete coincidence of the systems of Sum Rules derived in [@10] (from the formal manipulation with Laurent series expansions) with those following directly from the convergent partial fraction expansion (\[2\]).
Summarizing, we conclude with the statement that the main result following from the algebraic method is also valid in the framework of the analytic one: in the chiral limit $ m_{\pi} = 0 $ the direct channel resonances with a given helicity fall into reducible representations of unbroken chiral group $ SU_2 \times SU_2 $, the mass matrix being constructed as a sum of a chiral scalar and the 4-th component of chiral vector.
This work was supported in part by DAAD (visiting Grant for V.V.), RFBR (Grant 96-02- 18017) and GRACENAS (Grant 95-06.3-13).
[11]{} S.Weinberg. Physica, 96A (1979) 327. J.Gasser, H.Leutwyler. Ann. of Physics, 158 (1984) 142. Ulf-G. Meissner. Rep. Progr. Phys., 56 (1993) 903. A.Pich. Rep. Progr. Phys., 58 (1995) 563. J.Bijnens. Phys. Reports. 265 (1996) 371. S.Weinberg. Phys. Rev., 177 (1969) 2604; Phys. Rev. Letters, 65 (1990) 1177. V.Vereshagin. Preprint SPbU-IP-96-20, 1996; hep-ph/9606464. S.R.Beane. Preprint DUKE-TH-95-98, 1995; hep-ph/9512228. V.Vereshagin. Nucl. Phys., B55 (1973) 621. A.Bolokhov, A.Manashov, M.Polyakov, V.Vereshagin. Phys. Letters, B303 (1993) 220; Phys. Rev., D48 (1993) 3090. In this point we follow the argumentation suggested by A.Vereshagin in his Bachelor Thesis (unpublished).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Constantin Loizides
date: 'Received: / Revised version: '
title: |
High transverse momentum suppression and surface effects\
in and collisions within the PQM model
---
\[cl:intro\]Introduction
========================
One of the very early and very exciting findings from the experiments running at the Relativistic Heavy Ion Collider (RHIC) was the observation of apparent jet suppression. At the top RHIC energy, ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$, the mid-rapidity yield of high transverse momentum leading particles in collisions is about a factor of five lower than expected from the measurements in collisions at the same energy [@cl:phenixRAA; @cl:starRAA]. Similarly, jet-like correlations on the azimuthally-opposite (‘away’) side of a high-${\ensuremath{p_{\mathrm{t}}}}$ trigger particle are suppressed by a factor of four to five, while the near-side correlation strength is almost unchanged [@cl:starIAA]. The observed suppression persists also at lower center-of-mass energies [@cl:na57RCP], as well as in collisions at 62.4 and 200 GeV [@cl:phobosRAA]. The absence of these effects in collisions at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$ [@cl:dAu] strongly supports the picture of partonic energy loss, where energetic partons, produced in initial hard scattering processes, lose energy as a consequence of the final-state interaction with the dense partonic matter created in Åcollisions. The dominant contribution to the energy loss is believed to originate from medium-induced gluon radiation (see [Ref. [@cl:eloss]]{} and references therein). Recent calculations point out, however, that the collisional contribution to the energy loss might not be negligible [@cl:collloss]. Still, strong interest in these probes arises mainly from the fact that modification of their properties due to interaction with the medium provides access to fundamental properties of the created matter, such as its density and nature [@cl:modelsdata].
A quite simple model that includes final-state gluon radiation is the Parton Quenching Model (PQM) [@cl:PQM]. It combines the pQCD framework for the probabilistic calculation of parton energy loss in extended partonic matter of given size and density [@cl:carlosurs] with a realistic description of the collision overlap geometry in a static medium. High-energy partons (and parton pairs) are treated on an event-by-event basis using Monte Carlo techniques. The model has one single free parameter that sets the scale of the medium transport coefficient $\hat{q}$, the average transverse momentum squared transferred to the hard parton per unit path length, and, thus, the scale of the energy loss.
In these proceedings, we will, after a short introduction to the PQM model, compare the suppression phenomena introduced above to the calculations obtained with PQM. In particular, we will discuss to what extent light hadronic probes are sensitive to the dense matter formed at RHIC.
\[cl:pqmmodel\]The PQM model
============================
The Monte Carlo calculation of the unquenched and quenched transverse momentum spectra in PQM consists of four main steps: 1) Determination of a parton type and its transverse momentum according to PYTHIA (LO) parton distribution functions. 2) Determination of its parton-production point in the transverse plane according to the nuclear density profile (Glauber) and evaluation of path length and transport coefficient seen by the produced parton using two line integrals weighted with $\rho_{\rm coll}$. 3) Calculation of the energy loss using constrained quenching weights to extrapolate from the eikonal approach used in the BDMPS-SW framework to finite parton energies. Two types of constraints are constructed to estimate the systematic uncertainty of the approach: In the reweighted case the energy-loss distribution is simply normalized with-in the kinematically allowed regime, whereas in the non-reweighted case the fraction of the distribution larger than the energy of the parton contributes to the probability of maximum energy loss. 4) Finally, independent fragmentation is applied to the quenched and original parton. Back-to-back parton pairs initially consist of a pair of partons with the same ${\ensuremath{p_{\mathrm{t}}}}$ at the same production point, but with opposite-side emission angle.
Currently, the model is restricted to mid-rapidity, and, for simplicity, we ignore initial-state effects. The single parameter of the model ($k$) is fixed to set the scale of the energy loss by fitting to data from 0–10% central collisions at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$. Once the scale is fixed we implicitly vary the medium density by its dependence on the centrality as given by Glauber.
Details on the quenching procedure and its application to high-${\ensuremath{p_{\mathrm{t}}}}$ data can be found in [Ref. [@cl:PQM]]{}.
\[cl:raasuppression\]Suppression of leading particles
=====================================================
Suppression of leading particles is usually quantified via the nuclear modification factor, $$\label{cl:eq:raa}
R_{\rm AA}({\ensuremath{p_{\mathrm{t}}}},\eta) \equiv
\frac{1}{{\ensuremath{\left< {\ensuremath{N_{\rm coll}}}\right>}}} \times
\frac{{\rm d}^2 N_{\rm AA}/{\rm d}{\ensuremath{p_{\mathrm{t}}}}{\rm d}\eta}
{{\rm d}^2 N_{\rm pp}/{\rm d}{\ensuremath{p_{\mathrm{t}}}}{\rm d}\eta} \,,$$ the ratio of the yield of light hadrons in Å over proton–proton yield scaled by the number of binary collisions in a given centrality class. The ratio is normalized so that, if no final-state effects were present, it would be close to one. Indeed, this scaling is observed in the measurement of direct photon yield in collisions at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$ [@cl:photons].
![\[cl:fig1\] ${R_{\rm AA}}({\ensuremath{p_{\mathrm{t}}}})$ in at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$ for different centralities. Data are PHENIX charged hadrons and $\pi^0$ [@cl:phenixRAA] and STAR charged hadrons [@cl:starRAA] with combined statistical and ${\ensuremath{p_{\mathrm{t}}}}$-dependent systematic errors (bars on the data points) and ${\ensuremath{p_{\mathrm{t}}}}$-independent systematic errors (bars at ). The gray band is the original PQM calculation from [Ref. [@cl:PQM]]{} using reweighted and non-reweighted quenching weights. ](cRAAvsPt-K5000000-all-centralities){width="49.00000%"}
However, in collisions at the same energy, ${R_{\rm AA}}$ for light hadrons at mid-rapidity is found to decrease from peripheral (${R_{\rm AA}}\simeq 1$) to central events (${R_{\rm AA}}\simeq 0.2$), for ${\ensuremath{p_{\mathrm{t}}}}{\,{\buildrel > \over {_\sim}}\,}5~{\mbox{${\rm GeV}$}}$ (see [Fig. \[cl:fig1\]]{}). In this high-${\ensuremath{p_{\mathrm{t}}}}$ region, ${R_{\rm AA}}$ is independent of the particle type and rather independent of ${\ensuremath{p_{\mathrm{t}}}}$. In [Fig. \[cl:fig1\]]{}, the data is compared to the original calculation of PQM, where the single parameter of the model was adjusted to match the suppression measured in 0–10% central collisions at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$. For central collisions, this corresponds to an average transport coefficient of about ${\ensuremath{\left< \hat{q} \right>}}=14~{\mbox{${\rm GeV}$}}^2/{\mbox{${\rm fm}$}}$, which decreases with decreasing centrality to essentially zero for the most peripheral collisions. The average is taken over all produced hard partons and given in the equivalent static scenario [@cl:carlosurs].
{width="99.00000%"}
{width="99.00000%"}
{width="99.00000%"}
{width="99.00000%"}
In [Fig. \[cl:fig2\]]{}, we show ${R_{\rm AA}}({\ensuremath{p_{\mathrm{t}}}})$ for neutral pions in 0–10% central collisions at for preliminary PHENIX data [@cl:phenixPrelimRAA], which extends beyond the previous data by almost a factor of two in ${\ensuremath{p_{\mathrm{t}}}}$. The new data is compared to the original PQM calculation for ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=14~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ (simply extended to larger ${\ensuremath{p_{\mathrm{t}}}}$) in the reweighted and non-reweighted approximation of the quenching weights for finite (small) parton energies. For a given medium density, the reweighted values generally produce a weaker energy loss than the non-reweighted ones. However, in the reweighted case, only a rather small fraction of the full quenching weights is used. This results in the unphysical behaviour that, beyond a certain medium density, the probability of ‘zero energy loss’ increases for increasing densities. Therefore, we show in addition, results for lower densities, for ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=4$ and $7~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$, in the non-reweighted case. We furthermore compare the realistic calculations with calculations where we either fix the geometry ($L=6~{\mbox{${\rm fm}$}}$ and ${\ensuremath{\hat{q}}}=1~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$) or the relative energy loss ($\Delta E/E=0.25$). While the fixed geometry case leads to a more strongly rising ${R_{\rm AA}}$ with increasing ${\ensuremath{p_{\mathrm{t}}}}$, the fixed relative energy loss approximation over a very wide range of ${\ensuremath{p_{\mathrm{t}}}}$ describes the data.
Similar conclusions can be made for the comparison with data from interactions at RHIC. In [Fig. \[cl:fig3\]]{}, we show ${R_{\rm AA}}({\ensuremath{p_{\mathrm{t}}}})$ for neutral pions in 0–10% central collisions at for preliminary PHENIX data [@cl:phenixPrelimRAA]. The data are compared to the PQM prediction (made before QM’05 [@cl:preQMprediction]) for ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=9~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ for the reweighted and non-reweighted case. The value of ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=9~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ relies on the proportionality of the transport coefficient to the initial volume-density of gluons [@cl:baier] and on the predictions of the saturation model [@cl:ekrt], as outlined in [Ref. [@cl:PQM]]{}. The prediction seems to slightly overestimate the suppression in the smaller system. We therefore also present results for lower densities, for ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=3$ and $5.5~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$, in the non-reweighted case. As before, the calculations including the nuclear geometry are compared to calculations with either fixed geometry ($L=4~{\mbox{${\rm fm}$}}$ and ${\ensuremath{\hat{q}}}=1~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$) or relative energy loss ($\Delta E/E=0.15$).
In [Fig. \[cl:fig4\]]{}, we show calculations of the centrality dependence of the nuclear modification factor for ${\ensuremath{p_{\mathrm{t}}}}>7~{\mbox{${\rm GeV}$}}$ in and collisions at compared to the preliminary PHENIX data [@cl:phenixPrelimRAA]. The data seem to favour values for the transport coefficient of ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=7-14$ for and $3-5.5~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ for collisions.
\[cl:iaasuppression\]Suppression of jet-like correlations
=========================================================
Within the PQM framework, we consider the suppression of back-to-back jet-like correlations by simulation of back-to-back pairs of partons in a simple LO parton model (no intrinsic $k_{\rm t}$). The magnitude of the suppression is usually quantified by the factor ${I^{\rm away}_{\rm AA}}= D^{\rm away}_{\rm AA}/D^{\rm away}_{\rm pp}$, where the di-hadron correlation strength, $D^{\rm away}_{\rm pp(AA)}$, for an associated hadron, $h_2$, with $p_{\rm t,2}$ in the opposite azimuthal direction from a trigger hadron, $h_1$, with $p_{\rm t,1}$, is integrated over the considered trigger- and associated- ${\ensuremath{p_{\mathrm{t}}}}$ intervals [@cl:wang], $$\begin{aligned}
\label{cl:eq:daapp}
D^{\rm away}_{\rm pp(AA)} &=&
\int_{p^{\rm trig}_{\rm t,min}}^{p^{\rm trig}_{\rm t,max}}{{\mathrm}{d}}p_{\rm t,1}
\int_{p^{\rm assoc}_{\rm t,min}}^{p^{\rm assoc}_{\rm t,max}}{{\mathrm}{d}}p_{\rm t,2}
\\\nonumber
& &\int_{\rm{away\, side}}{{\mathrm}{d}}\Delta\phi\,
\frac{{{\mathrm}{d}}^3\sigma_{\rm pp(AA)}^{h_1h_2}/{{\mathrm}{d}}p_{\rm t,1}{{\mathrm}{d}}p_{\rm t,2}
{{\mathrm}{d}}\Delta\phi}
{{{\mathrm}{d}}\sigma_{\rm pp(AA)}^{h_1}/{{\mathrm}{d}}p_{\rm t,1}}\,,\end{aligned}$$
As was found for the nuclear modification factor in collisions at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$, ${I^{\rm away}_{\rm AA}}$ for $4<{\ensuremath{p^{\rm trig}_{\rm t}}}\le6~{\mbox{${\rm GeV}$}}$ and $2~{\mbox{${\rm GeV}$}}\le{\ensuremath{p^{\rm assoc}_{\rm t}}}\le{\ensuremath{p^{\rm trig}_{\rm t}}}$ is found to decrease with increasing centrality, down to about $0.2$–$0.3$ for the most central events (see [Fig. \[cl:fig5\]]{}). In the figure, we also show ${I^{\rm away}_{\rm AA}}$ data taken from [Ref. [@cl:starDirectJets]]{} for higher ${\ensuremath{p_{\mathrm{t}}}}$ cuts of $8<{\ensuremath{p^{\rm trig}_{\rm t}}}\le15~{\mbox{${\rm GeV}$}}$, where we normalize the away-side yields measured in to the yields measured in collisions. Both data sets are compared to PQM calculations (using as the reference in both cases) for the non-reweighted case with ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=4$, $7$ and $14~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ for collisions. The data for the higher trigger cuts favour smaller medium densities, while the uncertainties on the data for the lower cuts make it difficult to draw strong conclusions. For completeness, we also present predictions for the system with ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=3$, $5.5$ and $9~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ (only high trigger cuts).
\[cl:sensitivity\]Sensitivity of light hadronic probes
======================================================
Another complication arises from the fact that probes based on leading-particle analyzes are affected by several biases. The biases result primarily from the steeply-falling underlying production cross sections and the emission from regions close to the surface. The latter effect dominates for large medium densities. This is illustrated in [Fig. \[cl:fig6\]]{}, which shows the behaviour of ${R_{\rm AA}}$ (at $10~{\mbox{${\rm GeV}$}}$) and ${I^{\rm away}_{\rm AA}}$ (for trigger cuts of $8<{\ensuremath{p^{\rm trig}_{\rm t}}}\le15~{\mbox{${\rm GeV}$}}$ and $6~{\mbox{${\rm GeV}$}}\le{\ensuremath{p^{\rm assoc}_{\rm t}}}\le{\ensuremath{p^{\rm trig}_{\rm t}}}$) as a function of the medium density, expressed as ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}$, in and collisions at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$. Clearly, beyond a certain medium density, the numerical values for the ratios saturate at a non-zero value. The fundamental reason is that the probability of no medium-induced gluon radiation, $P(\Delta E=0)$, for a medium of finite size and finite density is not zero. The ‘no radiation’ contribution to the spectra is dramatically enhanced when realistic nuclear path-length distributions and density profiles are taken into account, since $P(\Delta E=0)$ is decreasing as a function of $\hat{q}\,L^3$, giving significant weight to partons that ‘feel’ lower values of $\hat{q}\,L^3$. We have reported in [Ref. [@cl:PQM]]{} that the ${R_{\rm AA}}$ and ${I^{\rm away}_{\rm AA}}$ data (for low trigger cuts) can be described by taking into account only $P(\Delta E=0)$ and $1-P(\Delta E=0)$ (see also [@cl:drees]).
Recently, direct measurements of dijets in heavy-ion collisions have been performed by the STAR collaboration [@cl:starDirectJets]. It is found that, while the relative yields of the hadron-triggered fragmentation function relative to , the integrand in [eq. (\[cl:eq:daapp\])]{}, are suppressed, the shape is not modified, even in the most central collisions at $200~{\mbox{${\rm GeV}$}}$. Such a scenario naturally follows from the trigger bias since, due to the cuts on near-side and away-side particle ${\ensuremath{p_{\mathrm{t}}}}$, those dijet pairs that escaped the collision region without losing a significant fraction of their initial energy are preferentially selected.
![\[cl:fig6\] ${R_{\rm AA}}$ at $10~{\mbox{${\rm GeV}$}}$ and ${I^{\rm away}_{\rm AA}}$ for trigger cuts of $8<{\ensuremath{p^{\rm trig}_{\rm t}}}\le15~{\mbox{${\rm GeV}$}}$ and $6~{\mbox{${\rm GeV}$}}\le{\ensuremath{p^{\rm assoc}_{\rm t}}}\le{\ensuremath{p^{\rm trig}_{\rm t}}}$ as a function of ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}$ in 0–5% central and collisions at ${\sqrt{s_{\scriptscriptstyle{{\rm NN}}}}}=200~{\mbox{${\rm GeV}$}}$. The calculations are in the non-reweighted approximation. ](cRaaAndIaavsQhat){width="49.00000%"}
{width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"} {width="42.50000%"}
For the PQM framework, the situation is illustrated in [Fig. \[cl:fig7\]]{}, which shows production points and emission direction for surviving back-to-back pairs in the transverse plane for 0–10% central collisions. Each parton pair shown yields a hadron pair within $8<{\ensuremath{p^{\rm trig}_{\rm t}}}\le15~{\mbox{${\rm GeV}$}}$ and $6~{\mbox{${\rm GeV}$}}\le{\ensuremath{p^{\rm assoc}_{\rm t}}}\le{\ensuremath{p^{\rm trig}_{\rm t}}}$. The chosen set of transport coefficients are the same as previously used to describe the ${I^{\rm away}_{\rm AA}}$ data, ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=4$, $7$ and $14~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$, in the non-reweighted case. In panels of [Fig. \[cl:fig7\]]{}, the center of any line indicates the production point of a parton pair, and the two partons emerge along the line in opposite direction. The color of each line either indicates the medium density (left) or the relative energy loss (right panels) of the away-side parton, which is defined as the parton fragmenting into the associated hadron. The left panels of [Fig. \[cl:fig7\]]{} clearly illustrates that the largest medium densities are encountered by away-side partons that pass through the central zone of the collision, which is expected for a static scenario. The core is surrounded by dijets that escape tangentially with respect to the central region [Ref. [@cl:qm05]]{}. In the case of an expanding medium, the interpretation of the underlying trigger bias remains the same: surviving dijets are always selected from regions (and times) of the collision evolution where they suffered the least amount of interactions with the medium. The right panels of [Fig. \[cl:fig7\]]{}, which color-code the relative energy loss of the away-side parton, illustrate this point from a slightly different perspective. It is apparent that in all three cases, the energy loss is close to zero for a large fraction of the dijets, even for the core region. This is due to the probability of zero-energy loss for the away-side parton, $P(\Delta E^{\rm away}=0)$, which for the core is already about 0.05, even for the ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=14~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ case. As shown in [Fig. \[cl:fig8a\]]{}, $P(\Delta E^{\rm away}=0)$ quickly reaches values between 0.1 and 0.3 for the $d{\,{\buildrel < \over {_\sim}}\,}4~{\mbox{${\rm fm}$}}$, where $d=\sqrt{x_0^2+y_0^2}$ is the distance of the production point to the collision center. A large fraction of these parton pairs are initially produced, see [Fig. \[cl:fig8b\]]{} (dashed line, no medium). The same figure illustrates the shift of the production point distribution to larger distances with increasing medium densities, when energy loss is included. The mean values change from $3.2~{\mbox{${\rm fm}$}}$ in vacuum to $3.5$, $3.7$ and $4.1~{\mbox{${\rm fm}$}}$ for ${\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}=4$, $7$ and $14~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$, respectively. In [Fig. \[cl:fig8c\]]{}, we finally present the average $\Delta E/E$ versus $d$ for near-side and away-side partons. For distances of $d{\,{\buildrel > \over {_\sim}}\,}4~{\mbox{${\rm fm}$}}$, that contribute significantly to the near- and away-side yields, $0.1{\,{\buildrel < \over {_\sim}}\,}\Delta E/E{\,{\buildrel < \over {_\sim}}\,}0.25$ for the three densities, and the average relative energy loss is nearly the same on both sides.
\[cl:concl\]Summary
===================
Jet quenching effects at the top RHIC energy are discussed within the Parton Quenching Model that includes a probabilistic treatment of the BDMPS quenching weights and a Glauber-based implementation of the collision geometry. The available high-${\ensuremath{p_{\mathrm{t}}}}$ data for ${R_{\rm AA}}$, ${I^{\rm away}_{\rm AA}}$ and their centrality dependence constrain the extracted medium density to about $3{\,{\buildrel < \over {_\sim}}\,}{\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}{\,{\buildrel < \over {_\sim}}\,}9~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ for central and $4{\,{\buildrel < \over {_\sim}}\,}{\ensuremath{{\ensuremath{\left< \hat{q} \right>}}}}{\,{\buildrel < \over {_\sim}}\,}14~{{\mbox{${\rm GeV}$}}^{2}/{\mbox{${\rm fm}$}}}$ for central collisions. Our analysis suggests that particle production in central collisions is ‘surface’ dominated, not only for single hadrons, but also for dijets. The properties of surviving dijets therefore are very similar to the vacuum properties.
S.S. Adler [*et al.*]{} \[PHENIX\], [Phys. Rev. Lett.]{} [**91**]{} (2003) 072301, [Phys. Rev.]{} [**C69**]{} (2004) 034910.
J. Adams [*et al.*]{} \[STAR\], [Phys. Rev. Lett.]{} [**91**]{} (2003) 172302.
C. Adler [*et al.*]{} \[STAR\], [Phys. Rev. Lett.]{} [**90**]{} (2003) 082302, [Phys. Rev. Lett.]{} [**95**]{} (2005) 152301.
F. Antinori [*et al.*]{} \[NA57\], [Phys. Lett. B]{} [**623**]{} (2005) 17; C. Hohne [*et al.*]{} \[NA49\], [[](http://www.arxiv.org/nucl-ex/0510049)]{}.
B. Alver [*et al.*]{} \[PHOBOS\], [Phys. Rev. Lett.]{} [**96**]{}, 212301 (2006). B.B. Back [*et al.*]{} \[PHOBOS\], [Phys. Rev. Lett.]{} [**91**]{} (2003) 072302; J. Adams [*et al.*]{} \[STAR\], [ibidem]{} 072304; S.S. Adler [*et al.*]{} \[PHENIX\], [ibidem]{} 072303; I. Arsene [*et al.*]{} \[BRAHMS\], [ibidem]{} 072305.
C.A. Salgado, [[](http://www.arxiv.org/hep-ph/0510062)]{}; X.N. Wang, [[](http://www.arxiv.org/nucl-th/0511001)]{}; I. Vitev, [[](http://www.arxiv.org/hep-ph/0603010)]{}.
A. Peshier, these proceedings, [[](http://www.arxiv.org/hep-ph/0607275)]{}.
D. d’Enterria, [Eur. Phys. J. C]{} [**43**]{}, 295 (2005). A. Dainese, C. Loizides and G. Paić, [Eur. Phys. J.]{} [**C38**]{} (2005) 461. C.A. Salgado and U.A. Wiedemann, [Phys. Rev.]{} [**D68**]{} (2003) 014008.
S. Adler [*et al.*]{} \[PHENIX\], [Phys. Rev. Lett.]{} [**94**]{}, 232301 (2005).
M. Shimomura \[PHENIX\], [[](http://www.arxiv.org/nucl-ex/0510023)]{}.
C. Loizides, QM’05 poster (not published), see P. M. Jacobs and M. van Leeuwen, [[](http://www.arxiv.org/nucl-ex/0511013)]{}.
R. Baier, [Nucl. Phys.]{} [**A715**]{} (2003) 209.
K.J. Eskola [*et al.*]{}, [Nucl. Phys.]{} [**B570**]{} (2000) 379. E. Wang and X.N. Wang, [Phys. Rev. Lett.]{} [**89**]{} (2002) 162301.
A. Drees, H. Feng and J. Jia, [Phys. Rev. C]{} [**71**]{}, 034909 (2005).
A. Dainese, C. Loizides and G. Paic, [[](http://www.arxiv.org/hep-ph/0511045)]{}. J. Adams [*et al.*]{} \[STAR\], [[](http://www.arxiv.org/nucl-ex/0604018)]{}.
K.J. Eskola [et al.]{}, [Nucl. Phys.]{} [**A747**]{} (2005) 511.
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Department of Physics, Hiroshima University\
Higashi-Hiroshima 739, Japan\
E-mail: [email protected]
author:
- 'Cong-Feng Qiao[^1]'
title: 'THE STATUS OF CHARMONIUM PRODUCTION IN PHOTON-PHOTON COLLISIONS'
---
=cmr8
1.5pt \#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
A Brief Overview of Onium Production Theory and Models
======================================================
Quarkonium is a bound state of heavy quark and its antiquark mediated by the strong interaction (QCD). The quarkonia can be classified and labeled in the conventional spectroscopic way $n$ $^{2S + 1}L_J (J^{PC})$, where $n$ is the radial quantum number; $P$ is the parity and $C$ is the charge conjugation; S, L, and J are total intrinsic spin, orbital angular momentum, and total angular momentum, respectively.
Since the discovery of $J/\psi$ in 1974, quarkonium physics has become one of the most fruitful areas in high energy physics. With the development of experiment in the past two decades a great advance has been made in understanding of the nature of quarkonium production and decays. At the time being, the wisdoms on quarkonium production and decays can be categorized as following theory and models:
[**$\bullet$ Color-Singlet Model [@t.a.degrand]**]{}
In color-singlet model, quarkonia are interpreted as non-relativistic bound states of $Q\bar{Q}$ pair. It assumes that the heavy quark pair produced in high energy collision can bind to form a given quarkonium state if the $Q\bar{Q}$ is created with exactly the same quantum numbers possessed by the bound state. And, the quarkonium production and decay amplitudes are supposed to be factorized into short- and long-distance sectors. The former can be calculated by using perturbative QCD; the latter, referring to non-perturbative effect, can be absorbed into a wave function factor; i.e., $$d\sigma(\psi + {\sc{x}}) = d\sigma (c\bar{c_1}(^3S_1) + \sc{x})
|R_\psi(0)|^2.$$ The wave function can be either determined phenomenologically from experimental measurements of quarkonium leptonic decays or calculated from bound state potential model.
[**$\bullet$ Color Evaporation Model [@h.fritzsch]**]{}
An alternative prescription for quarkonium production is the so-called color-evaporation(duality) model. In this approach, the probability of a $Q\bar{Q}$ pair with invariant mass between $2m_c$ and $D\bar{D}$ threshold $2m_D$(in case of Charmonium) evolving into a quarkonium state will be taken nearly independent of its color and spin states. e.g., for $J/\psi$ production, the cross section can be written as $$\sigma(J/\psi) = \hat{\sigma} (c\bar{c}(4m_c^2<s<4m_D^2)) f_{J/\psi},$$ where $\hat{\sigma}$ is the cross section for producing a $c\bar{c}$ pair with invariant mass below $D\bar{D}$ threshold, $f_{J/\psi}$ is a phenomenological parameter. The cross section $\hat{\sigma}(c\bar{c})$ is spin-summed and can be in both color-singlet and -octet configurations. This model has the flaw of incapable of describing the variation of production ratios for different states, though with some phenomenological success.
[**$\bullet$ NRQCD Factorization Theory [@g.t.bodwin]**]{}
Non-relativistic QCD(NRQCD) provides a rigorous QCD analysis of the production and decays of heavy quarkonium, which enables one to make perturbative corrections to all orders in $\alpha_s$, and relativistic corrections as well. The key point of this novel theory is that the inventors noticed that in quarkonium production and decays several typical energy scales are well-separated, $$(M_Q v^2)^2 \ll (M_Q v)^2 \ll M_Q^2.$$ With this hierarchy, the NRQCD effective Lagrangian can be expressed as $${\cal L}_{\rm NRQCD} = {\cal L}_{light} + {\cal L}_{heavy} +
\delta{\cal L},$$ where $ {\cal L}_{light} + {\cal L}_{heavy}$ describes ordinary QCD coupled to a Schrödinger field theory for the heavy quarks and antiquarks. The relativistic effects of full QCD are reproduced through the correction term $\delta{\cal L}$ in the Lagrangian.
In NRQCD formalism, the inclusive production cross-section of heavy quarkonium is argued taking a factorized form $$d\sigma(H + X) = \sum d\hat{\sigma}(c\bar{c} + X)<{\cal{O}}_H> .$$ Here, $d\hat{\sigma}(c\bar{c} + X)$ is the hard part calculable using perturbative QCD, $<{\cal{O}}_H> $ is the non-perturbative sector which can be expressed as vacuum matrix elements of NRQCD four quark operators. In the NRQCD Lagrangian, four-fermion operators can couple to both color-neutral states and colored states, which makes the NRQCD distinctively different from the color-singlet hypothesis in describing the quarkonium production mechanism, i.e., the Octet mechanism may play a role in quarkonium production as well.
Quarkonium Production in Photon-Photon Collisions
=================================================
The NRQCD has a list of merits in describing heavy quarkonium production and decays, especially in properly regulating the singularities appeared in color-singlet model and explaining the Tevatron large-$p_T$ $\psi'$ surplus problem based the color-octet model [@com]. However, it is not the end of story. There still lacks of direct evidence for the octet scenario at currently running colliders. Not to mention the large-$p_T$ $\psi'$ polarization “disaster” it encounters recently [@rothstein]. The point is that in principle NRQCD should be a correct theory in heavy quark limit, but in practice whether it can be applied to the Charmonium system is not clear.
People believe that non-hadronic collisions may give more clear signals and predictions by experiment and theory respectively than hadronic ones, and think that the study of quarkonium production at linear colliders may be helpful in clarifying the Onium production mechanism. In recent years several new concepts on linear colliders aimed at providing collisions at the center-of-mass energy from hundreds GeV to multi-TeV with high luminosity were proposed and the feasibilities pretested, such as JLC at KEK, TESLA at DESY and CLIC at CERN, etc. Theoretically, high-intensity photon beams may be obtained by the Compton backscattering of laser light off the linac electron beams, and photon-photon collision with approximately the same luminosity as that of the $e^+$ $e^-$ beams. Such a photon linear collider can have high energy up to TeV order. During the past more than twenty years researches on quarknoium production at $e^+$ $e^-$ colliders at various energies were carried out in detail [@ee1]. However, studies of photon-photon scattering are very limited and have just begun [@gg1](here, we focus on the direct photon-photon collision, for the resolved case see, e.g., ref. 6), though the preliminary results were obtained from LEP II data.
In $\gamma\gamma$ scattering, at leading order in $\alpha$ the $J/\psi$ is produced via the process $$\begin{aligned}
\label{eq:leading1}
\gamma + \gamma \rightarrow J/\psi + \gamma\;.\end{aligned}$$ However, since on the scale of heavy quark mass, the strong coupling constant is not too small, the process $$\begin{aligned}
\label{eq:leading2}
\gamma + \gamma \rightarrow J/\psi^{(8)} + g\end{aligned}$$ may compete with the pure electromagnetic process (\[eq:leading1\]) through the Color-Octet mechanism [@gg1]. Here, $J/\psi^{(8)}$ denotes those states evolved from the Color-Octet states.
When going up one order in $\alpha_s$, one may still expect to obtain the same order of magnitude in the $J/\psi$ production rate, because in this case $J/\psi$ may be produced in color-singlet and therefore will compensate for $\alpha_s$ suppression from the non-perturbative sector relative to the octet process (\[eq:leading2\]). This argument was confirmed recently by the calculation of double quarkonium, the $J/\psi$, production in direct photon-photon collision[@c.f.qiao], which is a sub-category of the inclusive $J/\psi$ production process at order $\alpha^2\alpha^2_s$. That is the process $$\begin{aligned}
\label{eq:leading3}
\gamma(k_1) + \gamma(k_2) \rightarrow J/\psi(P) + J/\psi(P')\;. \end{aligned}$$
-4cm
The results for processes (1) – (3) are shown in Figure 1. Since projected linear colliders with a luminosity of hundreds of fb$^{-1}$ per year, and the integrated total cross-sections increase with the colliding energy decreasing, we may have hundreds of events being observed in one year at colliding energies 500 GeV or less. It is obvious that the color-octet process is the smallest one over the entire energy scope of LEP II to next generation of linear colliders in the three processes being concerned, though they are in the same order.
In addition it was proved in Ref. 8 that the cross-section for single $J/\psi$ production via only the fragmentation mechanism would be about one order larger than that of processes (1) – (3) at 500 GeV. That means that the process at order $\alpha^2\alpha_s^2$ for single $J/\psi$ inclusive production should be the dominant one in the photon-photon collision, which unfortunately is still not studied in theory.
Concluding Remarks
==================
We have given a brief review of the previously investigated works appeared in literature on $J/\psi$ production at photon colliders. It is found that at a moderate energy of next generation photon colliders, there could be hundreds of events to be detected per year with high projected luminosity. Since the production rates of $J/\psi$ via the color-octet mechanism, the electromagnetic process, and the double production are almost the same in the full scope of the colliding energy, to differentiate the color-octet mechanism from the color-singlet one in these processes, experimentally one should detect not only the $J/\psi$ but also other final states. Furthermore, without distinguishing the final states in experiment we can not make any conclusion on $J/\psi$ production mechanisms from photon-photon collision, since the dominant production process, the inclusive one at order $\alpha^2\alpha^2_s$, is left un-investigated in theory.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by a Grant-in-Aid aid of JSPS committee.
[99]{} T.A. DeGrand and D. Toussiant, Phys. Lett. B89(1980)256; M. Wise Phys. Lett. B89(1980)229; J.H. Kühn, S. Nussinov and R.Z. Rückle, Phys. C5(1080)117; C.H. Chang, Nucl. Phys. B172(1980)425. H. Fritzsch, Phys. Lett. B67(1977)217. G.T. Bodwin, E. Braaten, and G.P. Lepage, Phys. Rev. D51(1995)1125; D55(1997)5853(E). G.T. Bodwin, E. Braaten, and G.P. Lepage, Phys. Rev. D[**46**]{} (1992) R3703; E. Braaten and S. Fleming, Phys. Rev. Lett. [**74**]{} (1995) 3327. For recent review see: I. Rothstein, Hep-ph/9911276. J.H. Kühn and H. Schneider, Phys. Rev. D24 (1981) 2996; Z. Phys. C11 (1981) 263; W.Y. Keung, Phys. Rev. D23 (1981) 2072; Feng Yuan, Cong-Feng Qiao, Kuang-Ta Chao, Phys. Rev. D56 (1997)321; Phys. Rev. D56 (1997) 1663; E. Braaten and Y.-Q. Chen, Phys. Rev. Lett. 76 (1996) 730; Chao-Hsi Chang, Cong-Feng Qiao, Jian-Xiong Wang, Phys. Rev. D56 (1997) R1363; Phys. Rev. D57 (1998) 4035; G.A. Schuler and M. Vänttinen, phys. Rev. D58 (1998) 017502; C. Glenn Boyd, Adam K. Leibovich, and I.Z. Rothstein, Phys. Rev. D59 (1999) 054016; R.M. Godbole, D. Indumathi, and M. Krämer, hep-ph/0101333. J.P. Ma, B.H.J. McKellar and C.B. Paranavitabe, Phys. Rev. D57 (1998) 606; G. Japaridze and A. Tkabladze, Phys. Lett. B433 (1998) 139; M. Klasen, B.A. Kniehl, L. Mihaila, and M. Steinhauser, hep-ph/0104044. Cong-Feng Qiao, Phys. Rev. D64 (2001) 077503.
[^1]: JSPS Research Fellow
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Baker and Norine introduced a graph-theoretic analogue of the Riemann-Roch theory. A central notion in this theory is the rank of a divisor. In this paper we prove that computing the rank of a divisor on a graph is -hard, even for simple graphs.
The determination of the rank of a divisor can be translated to a question about a chip-firing game on the same underlying graph. We prove the -hardness of this question by relating chip-firing on directed and undirected graphs.
address:
- 'Department of Analysis, Eötvös Loránd University, Pázmány Péter sétány 1/C, Budapest H-1117, Hungary'
- 'Department of Computer Science, Eötvös Loránd University, Pázmány Péter sétány 1/C, Budapest H-1117, Hungary'
author:
- Viktor Kiss
- Lilla Tóthmérész
bibliography:
- 'gon\_rank.bib'
title: 'Chip-firing games on Eulerian digraphs and -hardness of computing the rank of a divisor on a graph'
---
chip-firing game ,Riemann-Roch theory ,computational complexity ,Eulerian graph 05C57 ,05C45 ,14H55
Introduction
============
The Riemann-Roch theory for graphs was introduced by Baker and Norine in 2007 as the discrete analogue of the Riemann-Roch theory for Riemann surfaces [@BN-Riem-Roch]. They defined the notions *divisor, linear equivalence* and *rank* also in this combinatorial setting, and showed that the analogue of basic theorems as for example the Riemann-Roch theorem, remains true. Theorems like Baker’s specialization lemma [@specializ_lemma] establish a connection between the rank of a divisor on a graph and on a curve, which enables a rich interaction of the discrete and continuous theories.
A central notion in the Riemann-Roch theory is the rank of a divisor. The question whether the rank can be computed in polynomial time has been posed in several papers [@hladky; @Manjunath; @pot_theory], originally attributed to H. Lenstra.
Let us say a few words about previous work concerning the computation of the rank. Hladký, Král’ and Norine [@hladky] gave a finite algorithm for computing the rank of a divisor on a metric graph. Manjunath [@Manjunath] gave an algorithm for computing the rank of a divisor on a graph (possibly with multiple edges), that runs in polynomial time if the number of vertices of the graph is a constant. It can be decided in polynomial time, whether the rank of a divisor on a graph is at least $c$, where $c$ is a constant [@pot_theory]. Computing the rank of a divisor on a complete graph can be done in polynomial time [@cori]. For divisors of degree greater than $2g-2$ (where $g$ is the genus of the graph), the rank can be computed in polynomial time [@Manjunath]. On the other hand, there is a generalized model in which deciding whether the rank of a divisor is at least zero is already -hard [@lattice].
Our main goal in this paper is to show that computing the rank of a divisor on a graph is -hard, even for simple graphs. This result implies also the -hardness of computing the rank of a divisor on a tropical curve by [@Luo Theorem 1.6]. We also show that deciding whether the rank of a divisor on a graph is at most $k$ is in .
Our method is the following: We translate the question of computing the rank of a divisor to a question about the chip-firing game of Björner, Lovász and Shor using the duality between these frameworks discovered by Baker and Norine [@BN-Riem-Roch]. We get that the following question is computationally equivalent to the determination of the rank: Given an initial chip-distribution on an (undirected) graph $G$, what is the minimum number of extra chips we need to put on this distribution to make the game non-terminating.
We first prove the -hardness of computing the minimum number of chips that enables a non-terminating game on a simple Eulerian digraph by showing that it equals to the number of arcs in a minimum cardinality feedback arc set. This result is mentioned in a note added in proof of [@BL92], where only the larger or equal part is proved. Recently, Perrot and Pham [@Perrot] solved an analogous question in the abelian sandpile model, which is a closely related variant of the chip-firing game. Our result follows by applying their method to the chip-firing game.
Then we show that the second question (concerning chip-firing games on directed graphs) can be reduced to the first one (concerning undirected graphs). In order to do so, to any Eulerian digraph and initial chip-distribution, we assign an undirected graph with a chip-distribution such that in the short run, chip-firing on the undirected graph imitates chip-firing on the digraph.
Preliminaries
=============
Basic notations
---------------
Throughout this paper, *graph* means a connected undirected graph that can have multiple edges but no loops. A graph is *simple* if it does not have multiple edges. A graph is usually denoted by $G$. The vertex set and the edge set of a graph $G$ are denoted by $V(G)$ and $E(G)$, respectively. The degree of a vertex $v$ is denoted by $d(v)$, the multiplicity of the edge $(u,v)$ by $d(u,v)$. The *Laplacian matrix* of a graph $G$ means the following matrix $L$: $$L(i,j) = \left\{\begin{array}{cl} -d(v_i) & \text{if } i=j \\
d(v_i, v_j) & \text{if } i\neq j.
\end{array} \right.$$
*Digraph* means a (weakly) connected directed graph that can have multiple edges but no loops. We usually denote a digraph by $D$. The vertex set and edge set are denoted by $V(D)$ and $E(D)$, respectively. For a vertex $v$ the indegree and the outdegree of $v$ are denoted by $d^-(v)$ and $d^+(v)$, respectively. A digraph $D$ is *Eulerian* if $d^+(v) = d^-(v)$ for each vertex $v \in V(D)$. The *head* of the directed edge $(u, v)\in E(D)$ is $v$, and the *tail* of the edge is $u$. The multiplicity of the directed edge $(u,v)$ is denoted by $\overrightarrow{d}(u, v)$. A digraph is *simple* if $\overrightarrow{d}(u, v) \le 1$ for each pair of different vertices $u, v \in V(D)$.
The Laplacian matrix of a digraph $D$ means the following matrix $L$: $$L(i,j) = \left\{\begin{array}{cl} -d^+(v_i) & \text{if } i=j \\
\overrightarrow{d}(v_j, v_i) & \text{if } i\neq j.
\end{array} \right.$$
An important notion concerning digraphs is the feedback arc set. It also plays a crucial role in this paper.
A *feedback arc set* of a digraph $D$ is a set of edges $F \subseteq E(D)$ such that the digraph $D'=(V(D), E(D) \setminus F)$ is acyclic. We denote $$\minfas(D) = \min\{|F| : F \subseteq E(D) \text{ is a feedback
arc set}\}.$$
Let $G$ be a graph. An *orientation* of $G$ is a directed graph $D$ obtained from $G$ by directing each edge. We identify the vertices of $G$ with the corresponding vertices of $D$. We denote the indegree and the outdegree of a vertex $v \in V(G)$ in the orientation $D$ by $d^-_D(v)$ and $d^+_D(v)$, respectively.
For a graph $G$ let us denote by $\mathbf{0}_G$ the vector with each coordinate equal to $0$, and by $\mathbf{1}_G$ the vector with each coordinate equal to $1$, where the coordinates are indexed by the vertices of $G$. For a vertex $v$ of $G$ we denote the characteristic vector of $v$ by $\mathbf{1}_v$. We use the same notations for digraphs.
Riemann-Roch theory on graphs
-----------------------------
In this section we give some basic definitions of the Riemann-Roch theory on graphs. The basic objects are called *divisors*. For a graph $G$, $\Div(G)$ is the free abelian group on the set of vertices of $G$. An element $f\in \Div(G)$ is called *divisor*. We either think of a divisor $f\in \Div(G)$ as a function $f:V(G)\to \mathbb{Z}$, or as a vector $f \in \mathbb{Z}^{|V(G)|}$, where the coordinates are indexed by the vertices of the graph.
The *degree* of a divisor is the following: $$\deg(f)=\sum_{v\in V(G)} f(v).$$
The following equivalence relation on $\Div(G)$ is called *linear equivalence*: For $f, g \in \Div(G)$, $f\sim g$ if there exists a $z\in \mathbb{Z}^{|V(G)|}$ such that $g = f + Lz$.
A divisor $f\in\Div(G)$ is *effective*, if $f(v)\geq 0$ for each $v\in V(G)$.
For a divisor $f\in \Div(G)$, the *rank* of $f$ is $$\begin{split}
\rank(f) = \min\{\deg(g) - 1 : \text{$g \in \Div(G)$, $g$ is effective, } \\
\text{$\nexists h \in \Div(G)$ such that $h \sim f-g$ and $h$ is effective}\}.
\end{split}$$ When we wish to emphasize the underlying graph, we write $\rank_G(f)$ instead of $\rank(f)$.
Chip-firing
-----------
It was noted already by Baker and Norine [@BN-Riem-Roch], that there is a duality between divisors on graphs and the objects of the chip-firing game, as defined by Björner, Lovász and Shor [@BLS91]. Using this duality we can translate some questions in divisor theory to questions in chip-firing. We would like to use the latter language in the article, so let us include here a short introduction to chip-firing.
Often the term “chip-firing game” is used also in the setting of Baker and Norine, but for clarity, we only use this term for the game of [@BLS91].
The theory of chip-firing games was developed both for graphs (see [@BLS91]) and digraphs (see [@BL92]). Although the notions of the Riemann-Roch theory on graphs are in duality with notions concerning the chip-firing game on undirected graphs, later on, we also need chip-firing on digraphs.
The basic idea is that on each vertex of a graph (or digraph), there is a certain amount of chips. If a vertex has at least as many chips as its degree (in the directed case: as its outdegree), then it can be fired. In the undirected case, this means, that the vertex passes a chip to its neighbors along each edge incident to it, and so the number of chips on itself decreases by its degree. In the directed case, the fired vertex passes a chip along each outgoing edge. In fact, if we think of an undirected graph as a special digraph where we replace each edge by a pair of oppositely directed edges, then the two definitions coincide for these graphs.
Now we give the exact definitions.
Let $H$ be an undirected graph or a digraph. A *chip-distribution*, or *distribution*, is a function $x:V(H) \to \mathbb{Z}^+\cup \{0\}$. We sometimes say: vertex $v$ has $x(v)$ chips. We use the notation $|x|$ for the number of chips in the distribution $x$, i.e., $|x| = \sum_{v \in V(H)} x(v)$. We denote the set of chip-distributions on $H$ by $\Chip(H)$.
*Firing* a vertex $v$ means taking the new chip-distribution $x + L\mathbf{1}_v$ instead of $x$. Note that the Laplacian matrix $L$ is different in the undirected and in the directed case, and that in both cases $|x+L\mathbf{1}_v|=|x|$ so a firing preserves the number of chips. A vertex $v \in V(H)$ is *active* (with respect to $x$) if after firing it, it still has a nonnegative number of chips, i.e., in the undirected case if $x(v)\geq d(v)$, while in the directed case if $x(v) \ge d^+(v)$. The firing of a vertex $v \in V(H)$ is *legal*, if $v$ was active before the firing. A *legal game* is a sequence of distributions in which every distribution is obtained from the previous one by a legal firing. A game terminates if there is no active vertex with respect to the last distribution.
The following theorem was proved by Björner, Lovász and Shor for undirected graphs and by Björner and Lovász for digraphs. (Originally the theorem for the undirected case was proved for simple graphs, but the proof works also for graphs with multiple edges.)
\[thm::vegesseg\_kommutativ\] Let $H$ be a graph or a digraph and let $x \in \Chip(H)$ be a distribution. Then starting from $x$, either every legal game can be continued indefinitely, or every legal game terminates after the same number of moves with the same final distribution. Moreover, the number of times a given node is fired is the same in every legal game.
Let us call a chip-distribution $x\in \Chip(H)$ *terminating*, if every legal chip-firing game played starting from $x$ terminates, and call it *non-terminating*, if every legal chip-firing game played starting from $x$ can be continued indefinitely. According to Theorem \[thm::vegesseg\_kommutativ\], a chip-distribution is either terminating or non-terminating.
It can easily be seen by the pigeonhole principle that if for a graph $G$ a distribution $x \in \Chip(G)$ has $|x| > 2|E(G)| - |V(G)|$ then $x$ is non-terminating (see [@BLS91]). And similarly, for a digraph $D$ if a distribution $x \in \Chip(D)$ has $|x| > |E(D)| - |V(D)|$ then $x$ is non-terminating (see [@BL92]).
From this it follows that the following quantity, which measures how far a given distribution is from being non-terminating, is well defined. For a distribution $x \in \Chip(H)$, let $$\dist(x) = \min\{|y| : y \in \Chip(H), x + y \text{ is non-terminating}\}.$$ We say that $\dist(x)$ is the *distance* of $x$ from non-terminating distributions.
Note that $\dist(\mathbf{0}_H)$ is exactly the minimum number of chips in a non-terminating distribution on the graph/digraph $H$.
Chip-firing and the Riemann-Roch theory
---------------------------------------
Now we describe the duality between divisors on graphs and chip-distributions discovered by Baker and Norine [@BN-Riem-Roch].
Let $G$ be a graph and let $K^+ = K^+_G$ be the chip-distribution with $K^+(v)=d(v) - 1$ for each $v \in V(G)$.
For a divisor $f\in \Div(G)$ with $f(v)\leq d(v)-1$ for each $v\in V(G)$, we have $(K^+ - f)(v)\geq 0$ for each $v\in V(G)$, therefore $K^+ - f\in \Chip(G)$. We call $K^+ - f$ the *dual pair* of $f$. Note that each chip-distribution is a dual pair of some divisor.
\[prop::dual\_pair\] For a divisor $f\in \Div(G)$ with $f(v)\leq d(v)-1$ for each $v\in V(G)$, there exists an effective divisor equivalent to $f$ if and only if $K^+ - f$ is a terminating distribution.
We could have defined the chip-firing game for not necessarily nonnegative distributions as well with the same rules (only active vertices can fire). In this case Theorem \[thm::vegesseg\_kommutativ\] would still hold, and we could have a dual pair for any divisor, but this is not necessary for our purposes.
The following is a straightforward consequence of Proposition \[prop::dual\_pair\].
\[prop::rank\_of\_dual\] Let $f \in \Div(G)$ be a divisor with $f(v)\leq d(v)-1$ for each $v\in V(G)$, and let $x \in \Chip(G)$ be its dual pair. Then $\rank(f) = \dist(x) - 1$.
-hardness results
-----------------
In Section \[sec::dist\_digraph\], based on recent results of Perrot and Pham [@Perrot], we prove the following.
\[thm::rang\_eulerben\_nehez\] Given a digraph $D$, computing $\dist(\mathbf{0}_D)$ is -hard, even for simple Eulerian digraphs.
Using this result we prove the main theorem of this article.
\[thm::chip\_rang\_NP-teljes\] For a distribution $x\in \Chip(G)$ on a graph $G$, computing $\dist(x)$ is -hard.
The proof can be found in Section \[sec::rank\_chip\_dist\_NP\_hard\]. As a corollary of the theorem and Proposition \[prop::rank\_of\_dual\], we get the following.
For a divisor $f\in \Div(G)$ on a graph $G$, computing $\rank(f)$ is -hard, even for a divisor $f$ with $f(v)\leq d(v)-1$ for every $v \in V(G)$.
In [@hladky], Hladký, Král’ and Norine prove the following statement:
\[prop::hladky\_elosztos\] Let $f$ be a divisor on a graph $G$. Let $G'$ be the simple graph obtained from $G$ by subdividing each edge of $G$ by an inner point and let $f'$ be the divisor on $G'$ that agrees with $f$ on the vertices of $G$ and has value $0$ on new points. Then $\rank_G(f)=\rank_{G'}(f')$.
By dualizing this statement, we get the following: For a distribution $x\in\Chip(G)$, if we get $x'\in \Chip(G')$ from $x$ so that we put $d(v) - 1 - 0 = 1$ chip on each new vertex, and on the vertices of $G$, $x'$ agrees with $x$, then $\dist_G(x) = \rank_G(K^+_G - x) + 1 = \rank_{G'}(K^+_{G'} - x') + 1 =
\dist_{G'}(x')$.
When proving Theorem \[thm::chip\_rang\_NP-teljes\], we show a somewhat stronger statement: By Remark \[rem::dist\_NP\_erosites\], computing $\dist$ is -hard even for graphs with $|E(G)|\leq 9|V(G)|^5$. For such a $G$, $|V(G')|\leq |V(G)|+|E(G)|\leq 10|V(G)|^5$. Hence $G'$ and $x'$ can be computed in polynomial time for such a graph $G$ and $x\in \Chip(G)$, giving the following corollary.
For a distribution $x\in \Chip(G)$ on a simple graph $G$, computing $\dist(x)$ is -hard.
Using Proposition \[prop::rank\_of\_dual\] again, we have the following.
For a divisor $f\in \Div(G)$ on a simple graph $G$, computing $\rank(f)$ is -hard.
Using a result of [@Luo], we get that computing the rank of divisors is also -hard for so called tropical curves. Informally, a metric graph is a graph, where each edge has a positive length, and we consider our graph to be a metric space (the inner points of the edges are also points of this metric space). Tropical curve is more general in that we also allow some edges incident with vertices of degree one to have infinite length. A divisor on a tropical curve is an integer-valued function on the curve with only finitely many nonzero values. The notions of the degree of a divisor, linear equivalence, effective divisor and the rank can be defined as well, see [@hladky].
A metric graph $\Gamma$ corresponds to the graph $G$, if $\Gamma$ is obtained from $G$ by assigning some positive length to each edge.
\[thm::metrikus\_rang\_rang\] Let $f$ be a divisor on a graph $G$, and $\Gamma$ be a metric graph corresponding to $G$. Then $\rank_G(f)=\rank_{\Gamma}(f)$.
As a metric graph is a special tropical curve, we get the following corollary:
For a tropical curve $\Gamma$, $f\in \Div(\Gamma)$, computing $\rank(f)$ is -hard.
From the positive side, we show the following:
Deciding whether for a given divisor $f$ on a graph $G$, and integer $k$, $\rank(f)\leq k$ is in .
For an input $(f,k)$ with $\rank(f)\leq k$, our witness is the divisor $g\geq 0$ such that $\rank(f-g)=-1$, and $\deg(g)\leq k+1$ (such a $g$ exists because $\rank(f)\leq k$).
First, we need to check that $g$ can be given so that it has size polynomial in the size of $(f,k)$. As $\deg(g)\leq k+1$, and $g\geq 0$, we have $g(v)\leq k+1$ for each vertex $v$. Therefore, the size of $g$ is at most $O(|V(G)|\cdot \log k)$.
On the other hand, it can be checked in polynomial time if $\rank(f-g)=-1$ [@pot_theory], and also whether $\deg(g)\leq k+1$.
By applying Proposition \[prop::rank\_of\_dual\], deciding whether for a given chip-distribution $x$, and integer $k$, $\dist(x)\leq k$ is also in .
Minimal non-terminating distributions on Eulerian digraphs {#sec::dist_digraph}
==========================================================
In this section we prove Theorem \[thm::rang\_eulerben\_nehez\], i.e., that computing the minimum number of chips in a non-terminating distribution is -hard for a simple Eulerian digraph $D$.
We use the method of Perrot and Pham. In the paper [@Perrot], they prove the -hardness of an analogous question in the abelian sandpile model, which is a closely related variant of the chip-firing game.
Using the ideas of [@Perrot], we first give a formula for the minimum number of chips in a non-terminating distribution on an Eulerian digraph. As a motivation, let us have a look at the analogous question on undirected graphs, which was solved by Björner, Lovász and Shor.
\[thm::also\_korlat\] Let $G$ be a graph. Then $\dist(\mathbf{0}_G)=|E(G)|$.
We sketch the proof as a motivation for the directed case.
First we prove the following useful lemma.
\[lem::chip\_dist\_over\_acyclic\] Let $D$ be an acyclic orientation of $G$ and let $x\in \Chip(G)$ be a distribution with $x(v) \ge d^-_D(v)$ for each $v \in V(G)$. Then $x$ is non-terminating.
Since the orientation is acyclic, there is a sink, i.e., a vertex $v_0 \in V(G)$ with $d(v_0) = d^-_D(v_0) \leq x(v_0)$. Hence $v_0$ is active with respect to $x$. Fire $v_0$ and denote the resulting distribution by $x'$. Reverse the direction of the edges incident to $v_0$ and denote the resulting directed graph by $D'$. It is easy to see that $D'$ is acyclic and $d^-_{D'}(v) \le x'(v)$ for each $v \in V(G)$. Hence we can repeat the above argument. This shows that the distribution $x$ is indeed non-terminating.
Now taking an acyclic orientation $D$ of $G$ and setting $x(v) = d^-_D(v)$ for each $v \in V(G)$ we have a distribution with $|x| = |E(G)|$ that is non-terminating from the lemma. This shows that $\dist(\mathbf{0}_G)\leq |E(G)|$.
For proving $\dist(\mathbf{0}_G)\geq |E(G)|$, take a non-terminating distribution $x \in \Chip(G)$. It is enough to show that $|x|\geq |E(G)|$. Since in a non-terminating game every vertex is fired infinitely often (see [@BLS91 Lemma 2.1]), after finitely many firings, every vertex of $G$ has been fired at least once. Let $x'$ be the distribution at such a moment. Then $|x| = |x'|$. Let $D$ be the orientation of $G$ that we get by directing each edge toward the vertex whose last firing occurred earlier. It is straightforward to check that $x'(v) \ge d^-_D(v)$ for each $v \in V(G)$. This fact implies that $|x| = |x'| \ge |E(G)|$, completing the proof.
Now let us consider Eulerian digraphs.
\[thm::dist = minfas\] Let $D$ be an Eulerian digraph. Then $\dist(\mathbf{0}_D) = \minfas(D)$.
This theorem is already stated in a note added in proof of [@BL92], but there only the direction $\dist(\mathbf{0}_D) \ge \minfas(D)$ is proved. We give a proof following ideas of Perrot and Pham [@Perrot]. The idea of the proof can be thought of as the generalization of the idea of the proof of Theorem \[thm::also\_korlat\]. For proving Theorem \[thm::dist = minfas\] we need a classical result about chip-firing on an Eulerian digraph:
\[prop::euler\_vegt\_mind\_vegt\_sokszor\_lo\] On an Eulerian digraph $D$ if a chip-distribution is non-terminating then in any legal game every vertex is fired infinitely often.
The key lemma is the following observation of Perrot and Pham; they proved it for the recurrent configurations of the abelian sandpile model, but the two models are very closely related.
\[lemma::euler\_ir\_felett\_vegtelen\] Let $F\subseteq E(D)$ be a minimum cardinality feedback arc set. Denote by $d^+_F(v)$ and $d^-_F(v)$ the outdegree and indegree of a vertex $v$ in the digraph $D_F=(V(D), F)$. Then a distribution $x\in \Chip(D)$ satisfying $$\label{eq::feedback_felett}
x(v)\geq d^-_F(v) \text{ for every $v\in V(D)$}$$ is non-terminating.
First we prove that if $F$ is a minimum cardinality feedback arc set then there exists a vertex $v \in V(D)$ such that among the edges incident to $v$, $F$ contains exactly the in-edges of $v$.
Let $A = E(D) \setminus F$. From the definition of feedback arc set, $D_A = (V(D), A)$ is an acyclic graph. Therefore, it has a source $v_0$. We claim that no out-edge of $v_0$ is in $F$. Indeed, if some out-edges of $v_0$ would be in $F$, removing them from $F$ would mean adding some out-edges to the source $v_0$ in $D_A$, which cannot create a cycle. So we could get a smaller feedback arc set.
The fact that $v_0$ is a source in $D_A$ means that all the in-edges of $v_0$ are in $F$. Hence from the edges incident to $v_0$, $F$ contains exactly the in-edges of $v_0$.
Now take such a vertex $v_0$. From , the choice of $v_0$ and the fact that $D$ is Eulerian, we have that $x(v_0)\geq d^-_F(v_0)=d^-(v_0)=d^+(v_0)$, therefore $v_0$ is active with respect to $x$. Fire $v_0$. Let $x'$ be the resulting distribution. We show, that we can modify the feedback arc set $F$, such that for the new feedback arc set $F'$ we have $x'(v)\geq d^-_{F'}(v)$ for every $v\in V(D)$.
Let $F'$ be the set of arcs obtained from $F$ by removing the in-edges of $v_0$ and adding the out-edges of $v_0$ (see Figure \[fig::feedback\]). Then $D_{A'} = (V(D), A')$ with $A'=E(D)\setminus F'$ is acyclic, since the new edges are all incident to a sink in $D_{A'}$. Moreover, since $D$ is Eulerian, and from the choice of $v_0$, we have $|F'|=|F|$, hence $F'$ is also a feedback arc set of minimum cardinality. It is straightforward to check that indeed $x'(v)\geq d^-_{F'}(v)$ for every $v \in V(D)$.
So we are again in the starting situation, which shows that $x$ is indeed non-terminating.
\(1) at (0, 1) ; (2) at (0, -1) ; (3) at (-1, 0) ; (4) at (1, 0) ; (1) edge \[dashed\] node (4) (3) edge node (1) (2) edge node (3) (4) edge node (2) (4) edge \[bend left=20\] node (3) (3) edge \[dashed,bend left=20\] node (4);
\(1) at (0, 1) ; (2) at (0, -1) ; (3) at (-1, 0) ; (4) at (1, 0) ; (1) edge node (4) (3) edge node (1) (2) edge node (3) (4) edge \[dashed\] node (2) (4) edge \[dashed,bend left=20\] node (3) (3) edge \[bend left=20\] node (4);
Now take a feedback arc set $F$ of minimum cardinality, and let $x(v)=d^-_F(v)$ for every $v\in V(D)$. Then $|x|=|F|=\minfas(D)$, and from the lemma, $x$ is non-terminating. This proves that $\dist(\mathbf{0}_D)\leq \minfas(D)$.
The direction $\dist(\mathbf{0}_D)\geq \minfas(D)$ is shown in the note added in proof of [@BL92] for a general digraph, however, as we need implications of its idea, we also include this part of the proof.
Take a non-terminating distribution $x$. It is enough to prove that $|x|\geq \minfas(D)$.
Let us play a chip-firing game with initial distribution $x$. Proposition \[prop::euler\_vegt\_mind\_vegt\_sokszor\_lo\] says that after finitely many steps, every vertex has fired. Play until such a moment, and let the distribution at that moment be $x'$.
Let $A$ be the following set of edges: $$A=\{(u,v) \in E(D): \textrm{ the last firing of $u$ preceeds the last firing of $v$}\}.$$ As every vertex has fired, $A$ is well defined. Let $v_1,v_2, \dots v_{|V(D)|}$ be the ordering of the vertices by the time of their last firing. Then $v_1,v_2, \dots v_{|V(D)|}$ is a topological order of $D_A=(V(D),A)$, so $D_A$ is acyclic, hence $F=E(D)\setminus A$ is a feedback arc set. We show that $x'(v)\geq d^-_F(v)$ for every $v\in V(D)$. For $1 \le i \le |V(D)|$ the vertex $v_i$ has $d^-_F(v_i)=\sum_{j>i}\overrightarrow{d}(v_j,v_i)$. After its last firing, $v_i$ had a nonnegative number of chips. Since then, it kept all chips it received. And as $v_{i+1}, \dots, v_{|V(D)|}$ all fired since the last firing of $v_i$, it received at least $\sum_{j>i}\overrightarrow{d}(v_j,v_i)=d^-_F(v_i)$ chips. So indeed, we have $x'(v_i)\geq d^-_F(v_i)$.
Therefore $|x|=|x'| \ge |F|\geq \minfas(D)$.
Note that in the above setting, starting from $x'$, then firing the vertices in the order $v_1,v_2,\dots, v_{|V(D)|}$ (once each) is a legal game. Indeed, we proved that $x'(v_i)\geq d^-_F(v_i)=\sum_{j>i}\overrightarrow{d}(v_j,v_i)$. After firing $v_1, \dots v_{i-1}$, the vertex $v_i$ receives $\sum_{j<i}\overrightarrow{d}(v_j,v_i)$ more chips, so it indeed becomes active ($d^+(v_i)=d^-(v_i)=\sum_{j\neq i}\overrightarrow{d}(v_j,v_i))$ as we did not allow loops).
We need this observation in the next section, so we state it as a proposition:
\[prop::euler\_ha mar mindenki lott\] In a chip-firing game on an Eulerian digraph $D$, if at some moment every vertex has already fired then there is an order of the vertices in which they can be legally fired once each, starting from that moment.
It is worth noting that on an Eulerian digraph if starting from an initial distribution $x$ we fired each vertex exactly once, then we get back to distribution $x$: each vertex $v$ gave and received $d^-(v)=d^+(v)$ chips.
Finally, we prove Theorem \[thm::rang\_eulerben\_nehez\].
Perrot and Pham proved that computing $\minfas(D)$ for a simple Eulerian digraph $D$ is -hard [@Perrot Theorem 2], by reducing it to the -hardness of computing $\minfas(D)$ for general digraphs. From this, and from Theorem \[thm::dist = minfas\], the statement follows.
The distance from non-terminating distributions is -hard on graphs {#sec::rank_chip_dist_NP_hard}
==================================================================
In this section we prove Theorem \[thm::chip\_rang\_NP-teljes\], the main theorem of this article. In our proof of the -hardness, we rely on the fact that a terminating chip-firing game on an Eulerian digraph $D$ terminates after at most $2|V(D)|^2 |E(D)| \Delta(D)$ steps (see [@BL92 Corollary 4.9]), where $\Delta(D)$ denotes the maximum of all the indegrees and the outdegrees of $D$, i.e., $\Delta(D) = \max_{v \in V(D)}\max\{d^-(v), d^+(v)\}$. With this in mind, we define the following transformation:
Let $\varphi$ be the following transformation, assigning an undirected graph $G=\varphi(D)$ to any digraph $D$:
Split each directed edge by an inner point, and substitute the tail segment by $M=8|V(D)|^2 |E(D)| \Delta(D)$ parallel edges. Then forget the orientations.
We maintain the effect of the transformation by a bijective function $\psi: (V(D)\cup E(D)) \to V(\varphi(D))$:
For a vertex $v\in V(D)$ let $\psi(v)$ be the corresponding vertex of $\varphi(D)$. For an edge $e \in E(D)$, let $\psi(e)$ be the vertex with which we have split $e$.
Then the degrees in $\varphi(D)$ are the following: $$d(v) = \left\{\begin{array}{cl} d^+\left(\psi^{-1}(v)\right)\cdot M + d^-(\psi^{-1}(v)) & \text{if } \psi^{-1}(v) \in V(D) \\
M + 1 & \text{if } \psi^{-1}(v) \in E(D).
\end{array} \right.$$
\(1) at (0, 1) [$v_1$]{}; (2) at (0, -1) [$v_2$]{}; (3) at (-1, 0) [$v_3$]{}; (4) at (1, 0) [$v_4$]{}; (1) edge node [$e_1$]{} (4) (3) edge node [$e_2$]{} (1) (2) edge node [$e_3$]{} (3) (4) edge node [$e_4$]{} (2) (4) edge \[bend left=20\] node [$e_5$]{} (3) (3) edge \[bend left=20\] node [$e_6$]{} (4);
\(1) at (0, 1) [$\psi(v_1)$]{}; (2) at (0, -1) [$\psi(v_2)$]{}; (3) at (-1, 0) [$\psi(v_3)$]{}; (4) at (1, 0) [$\psi(v_4)$]{}; (41) at (0.5, 0.5) [$\psi(e_1)$]{}; (13) at (-0.5, 0.5) [$\psi(e_2)$]{}; (32) at (-0.5, -0.5) [$\psi(e_3)$]{}; (24) at (0.5, -0.5) [$\psi(e_4)$]{}; (34) at (0, -0.25) [$\psi(e_5)$]{}; (43) at (0, 0.25) [$\psi(e_6)$]{};
\(4) edge (41) (41) edge \[bend right=10\] (1) (41) edge \[bend left=10\] (1) (41) edge (1) (1) edge (13) (13) edge \[bend right=10\] (3) (13) edge \[bend left=10\] (3) (13) edge (3) (3) edge (32) (32) edge \[bend right=10\] (2) (32) edge \[bend left=10\] (2) (32) edge (2) (2) edge (24) (24) edge \[bend right=10\] (4) (24) edge \[bend left=10\] (4) (24) edge (4) (3) edge (34) (34) edge \[bend right=10\] (4) (34) edge \[bend left=10\] (4) (34) edge (4) (4) edge (43) (43) edge \[bend right=10\] (3) (43) edge \[bend left=10\] (3) (43) edge (3);
Let us define a certain chip-distribution on the graph $\varphi(D)$:
Let $base_D \in \Chip(\varphi(D))$ on a vertex $v \in V(\varphi(D))$ be the following: $$\begin{split}
base_D(v) = \left\{\begin{array}{cl} d^+\left(\psi^{-1}(v)\right)\cdot M & \text{if } \psi^{-1}(v) \in V(D) \\
M/2 & \text{if } \psi^{-1}(v) \in E(D).
\end{array} \right.
\end{split}$$
The key lemma in our proof of Theorem \[thm::chip\_rang\_NP-teljes\] is the following:
For an Eulerian digraph $D$, $\dist_D(\mathbf{0}_D) = \dist_{\varphi(D)}(base_D)$.
Let $G = \varphi(D)$. First we show that $\dist_D(\mathbf{0}_D) \ge \dist_{G}(base_D)$.
Let $x \in \Chip(D)$ be a non-terminating chip-distribution such that $|x|$ is minimal. We can assume that there is an order of the vertices of $D$ such that from initial distribution $x$ we can fire the vertices in that order (once each). Otherwise, from Proposition \[prop::euler\_vegt\_mind\_vegt\_sokszor\_lo\] we can play a chip-firing game from $x$ until each vertex has fired. Denoting the distribution at that moment by $x'$, from Proposition \[prop::euler\_ha mar mindenki lott\] for $x'$ there is such an order. As firing does not change the number of chips in the game, $|x'|$ is still minimal, so we can substitute $x$ with $x'$.
Let $y\in\Chip(G)$ be the distribution “$x + base_D$”, i.e., for a vertex $v \in V(D)$ let $y(\psi(v)) = x(v) + base_D(\psi(v))$ and for an edge $e \in E(D)$ let $y(\psi(e)) = base_D(\psi(e))$. Since $y(w) \ge base_D(w)$ for each $w \in V(G)$ and $|y - base_D| = |x| = \dist_D(\mathbf{0}_D)$, it is enough to show that $y$ is non-terminating.
For that, it is enough to show that we can fire each vertex of $G$ exactly once in some order. Then each vertex $w \in V(G)$ gives and receives $d(w)$ chips, so we get back to the distribution $y$ and can repeat this period indefinitely.
To get such an order of the vertices of $G$, we will play the chip-firing game simultaneously on $D$ and $G$.
To firing a vertex $v$ in $D$, let the corresponding firings in $G$ be: Fire $\psi(v)$, then fire $\psi(e)$ for every out-edge $e$ of $v$ (in some order).
If a sequence of firings of length $k \leq M/2$ on $D$ with initial distribution $x$ is legal then the sequence of the corresponding firings on $G$ with initial distribution $y$ is also legal. Moreover, if we denote the resulting distribution on $D$ by $\tilde{x}$ and on $G$ by $\tilde{y}$ then $$\label{eq::vertex}
\tilde{y}(\psi(v)) = \tilde{x}(v) + d^+(v) \cdot M
\text{ for each $v \in V(D)$}$$ and $$\label{eq::edge}
M/2 - k \le \tilde{y}(\psi(e)) \le M/2 + k \text{ for each $e \in E(D)$.}$$
We show this by induction on $k$. For $k = 0$ this is trivial. Take a sequence of firings of length $k \leq M/2$ and assume that the claim holds for $k - 1$. Denote the distribution on $D$ after the first $k - 1$ firings by $x'$ and the corresponding distribution on $G$ by $y'$. Assume that the vertex $v$ is the last to be fired on $D$. Hence $v$ is active with respect to $x'$. Denote the distribution after firing $v$ by $x''$. Vertex $\psi(v)$ is active with respect to $y'$, since using of the induction hypothesis, the fact that $v$ is active with respect to $x'$ and that $D$ is Eulerian, we get that $y'(\psi(v)) = x'(v) + d^+(v) \cdot M \ge
d^+(v) + d^+(v) \cdot M =
d^-(v) + d^+(v) \cdot M = d(\psi(v))$. Fire $\psi(v)$. Now for each out-edge $e$ of $v$ the vertex $\psi(e)$ is active, since using of the induction hypothesis, it has at least $M + y'(\psi(e)) \ge M + M/2 - (k - 1) \geq M + 1 = d(\psi(e))$ chips. Fire these vertices in an arbitrary order. (Firing one leaves the others active.) Denote by $y''$ the resulting distribution. It is easy to check that the distributions $x''$ and $y''$ satisfy conditions and .
We have chosen the distribution $x$ such that we can fire the vertices of $D$ in some order (once each) with initial distribution $x$. This is a legal sequence of firings of length $|V(D)|<M/2$. According to the previous claim, the sequence of the corresponding firings on $G$ is also legal. Moreover, on $G$ we also fire each vertex exactly once. This finishes the proof of the direction $\dist_D(\mathbf{0}_D) \ge \dist_G(base_D)$.
Now we prove that $\dist_D(\mathbf{0}_D) \le \dist_G(base_D)$. For this, let $y \in \Chip(G)$ be a minimal non-terminating chip-distribution with $base_D(w) \le y(w)$ for each $w \in V(G)$. Let $x(v) = y(\psi(v)) - base_D(\psi(v))$ on each $v\in V(D)$. It is enough to show that $x$ is non-terminating.
First note that $\dist_D(\mathbf{0}_D) \le |E(D)| - |V(D)| + 1$, since having a chip-distribution with at least $|E(D)|-|V(D)|+1$ chips, at every stage of the game at least one of the vertices has the sufficient number of chips to fire. Consequently, using also the first part of the lemma, we have that $|y - base_D| = \dist_G(base_D) \le \dist_D(\mathbf{0}_D) \le
|E(D)|-|V(D)| + 1 \le \frac{1}{8}M$.
Now we play the game on $G$ and $D$ simultaneously from initial distributions $y$ and $x$, respectively, in the following way. Let a step be the following: Choose a vertex $v \in V(D)$ for which $\psi(v)$ can fire. On $G$ fire $\psi(v)$, then for every out-edge $e$ of $v$, fire $\psi(e)$. On $D$ fire $v$.
We show that for $\frac{3}{8}M \ge
2|V(D)|^2 |E(D)| \Delta(D) + 1$ steps we can play this legally on both graphs. Note first that for an edge $e$ of $D$, the change of the number of chips on $\psi(e)$ is at most one after each step. Hence at the beginning of a step a vertex of $G$ of the form $\psi(e)$ can have at most $M/2 + |y - base_D| + \frac{3}{8}M \leq
M/2 + \frac{1}{8}M + \frac{3}{8}M < M + 1 = d(\psi(e))$ chips, so it cannot be fired. It also follows from this that on every such vertex the number of chips is positive, since it is at least $M/2 - \frac{3}{8}M > 0$. But $y$ is a non-terminating distribution, hence at the beginning of a step we can find an active vertex, which therefore must be of the form $\psi(v)$ with $v \in V(D)$. After firing $\psi(v)$, $\psi(e)$ becomes active for every out-edge $e$ of $v$, since $\psi(e)$ had a positive number of chips at the beginning of the step, and received $M$ chips. Hence on $G$ we can play in the desired way for $\frac{3}{8}M$ steps.
For the initial distributions, we have $y(\psi(v)) = d^+(v) \cdot M + x(v)$ for each $v\in V(D)$, so a vertex $v \in V(D)$ is active with respect to $x$ if and only if $\psi(v)$ is active with respect to $y$. Let $x'$ be the distribution on $D$ and $y'$ the distribution on $G$ at the end of an arbitrary (but at most $\frac{3}{8}M^\text{th}$) step. Then it can be shown by induction that $y'(\psi(v)) = d^+(v) \cdot M + x'(v)$ for each $v \in V(D)$. So in each step we have that a vertex $v \in V(D)$ is active if and only if $\psi(v)$ is active.
Hence for $\frac{3}{8}M$ steps, the corresponding game on $D$ is also legal. This means that there is a chip-firing game of length at least $\frac{3}{8}M \ge 2|V(D)|^2 |E(D)| \Delta(D) + 1$ on $D$ with initial distribution $x$, which by [@BL92 Corollary 4.9] implies that the distribution $x$ is non-terminating. This finishes the proof.
For a general digraph, the construction of the proof imitates the following game: If a vertex $v$ fires, each of its out-neighbors $u$ receives $\overrightarrow{d}(vu)$ chips, but the number of chips on $v$ decreases by the in-degree of $v$. This modification of the chip-firing game has been studied by Asadi and Backman [@asadi].
The theorem follows from Theorem \[thm::rang\_eulerben\_nehez\] and the previous lemma.
\[rem::dist\_NP\_erosites\] For a simple Eulerian digraph $D$, one has $$|E(\varphi(D))|\leq |E(D)|\cdot 9|V(D)|^3|E(D)|\leq 9|V(\varphi(D))|^5,$$ therefore the computation of $\dist$ is -hard even for graphs with $|E(G)|\leq 9|V(G)|^5$.
Polynomial time computability in a special case
===============================================
In this section we consider undirected graphs, and observe that for chip-distributions that are in a sense “small”, computing the distance from non-terminating distributions can be done in polynomial time. Moreover, for these distributions, the distance from non-terminating distributions only depends on the number of edges of the graph and the number of chips in the distribution.
The corollaries of this observation for the case of divisors give a special case of the Riemann-Roch theorem.
Recall that Theorem \[thm::also\_korlat\] stated that $\dist(\mathbf{0}_G)=|E(G)|$ for any undirected graph $G$. We would like to generalize this statement for “small enough” distributions. We say that a distribution $x \in \Chip(G)$ is *under an acyclic orientation*, if there exists an acyclic orientation $D$ of $G$ such that $x(v) \le d^-_D(v)$ for each $v \in V(G)$.
\[prop::acikl\_alatt\] Let $G$ be a graph and let $x \in \Chip(G)$ be a distribution. If $x$ is under an acyclic orientation then $\dist(x) = |E(G)| - |x|$.
From Theorem \[thm::also\_korlat\], a non-terminating distribution has at least $|E(G)|$ chips, therefore $\dist(x) \ge |E(G)| - |x|$.
For the other direction, let $D$ be an acyclic orientation of $G$ with $x(v) \le d^-_D(v)$ for each $v \in V(G)$. Let $y$ be the distribution on $G$ corresponding to the indegrees of the orientation, i.e., $y(v) = d^-_D(v)$ for each $v \in V(G)$. Then, using Lemma \[lem::chip\_dist\_over\_acyclic\], $y$ is non-terminating, moreover $|y| = |E(G)|$ and $y(v) \ge x(v)$ for each $v \in V(G)$. Hence $\dist(x) \le |y - x| = |E(G)| - |x|$. This completes the proof of the proposition.
It can also be decided in polynomial time whether a distribution $x \in \Chip(G)$ is under an acyclic orientation. A greedy algorithm solves the problem.
From the previous proposition, using the duality between chip-distributions and divisors, we get a special case of the Riemann-Roch theorem for graphs.
Let us denote by $K$ the canonical divisor on a graph $G$, that is, $K(v)=d(v)-2$ for each vertex $v\in V(G)$.
Let $G$ be a graph, and let $f$ be a divisor on $G$. Then $$\rank(f)-\rank(K-f)=\deg(f)-|E(G)|+|V(G)|.$$
Now, from Proposition \[prop::rank\_of\_dual\] and Proposition \[prop::acikl\_alatt\] we have for $f=K^+ - x$ that $$\rank(f) = \dist(x) - 1 = |E(G)| - |x| - 1 =
\deg(f)- |E(G)| + |V(G)| -1,$$ if $x$ is under an acyclic orientation.
We claim that in this case, $\rank(K-f)=-1$. Indeed, $K-f=K-(K^+ - x)= x - \mathbf{1}$, so the dual of $K-f$ is $K^+ - x + \mathbf{1}$. The distribution $x$ is under an acyclic orientation, let $D$ be an orientation witnessing this, i.e., $x(v) \le d^-_D(v)$ for each vertex $v \in V(G)$. Then $(K^+ - x + \mathbf{1})(v) = d(v) - x(v) \ge d^+_D(v)$ for each vertex $v \in V(G)$, hence we can use Lemma \[lem::chip\_dist\_over\_acyclic\] for $K^+ - x + \mathbf{1}$ and the directed graph obtained from $D$ by reversing every edge. It follows that $K^+ - x + \mathbf{1}$ is non-terminating, hence for its dual, $\rank(K-f)=-1$ by Proposition \[prop::rank\_of\_dual\].
Therefore, we have $\rank(f) - \rank(K-f)= \deg(f)- |E(G)| + |V(G)|$, showing the Riemann-Roch theorem in this special case.
Acknowledgement {#acknowledgement .unnumbered}
===============
Research was supported by the MTA-ELTE Egerváry Research Group and by the Hungarian Scientific Research Fund - OTKA, K109240 (Lilla Tóthmérész), 104178 and 105645 (Viktor Kiss).
We would like to thank Bálint Hujter for introducing us to this topic, Erika Bérczi-Kovács and Kristóf Bérczi for suggesting us to try to reduce the feedback arc set problem to the computation of the rank. We would also like to thank Tamás Király, Zoltán Király, Márton Elekes and the anonymous referees for their useful comments about the manuscript.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider differential operators between sections of arbitrary powers of the determinant line bundle over a contact manifold. We extend the standard notions of the Heisenberg calculus: noncommutative symbolic calculus, the principal symbol, and the contact order to such differential operators. Our first main result is an intrinsically defined “subsymbol” of a differential operator, which is a differential invariant of degree one lower than that of the principal symbol. In particular, this subsymbol associates a contact vector field to an arbitrary second order linear differential operator. Our second main result is the construction of a filtration that strengthens the well-known contact order filtration of the Heisenberg calculus.'
address:
- |
Department of Mathematics\
University of North Texas\
Denton TX 76203, USA
- |
CNRS, Institut Camille Jordan\
Université Claude Bernard Lyon I\
21 Avenue Claude Bernard, 69622 Villeurbanne Cedex, France
author:
- 'Charles H. Conley'
- Valentin Ovsienko
title: Linear differential operators on contact manifolds
---
Introduction {#Intro}
============
The space $\D(M)$ of linear differential operators on a smooth manifold $M$ has a rich geometric structure. By the *geometry* of $\D(M)$, we understand its structure as a module over the group of all diffeomorphisms of $M$, and thereby also over its Lie algebra, the space $\vm$ of smooth vector fields on $M$. The most interesting geometric properties of $\D(M)$ are described by its invariants under the group of diffeomorphisms.
Additional structure on $M$ leads to a smaller group of diffeomorphisms, and therefore a richer set of invariants of $\D(M)$. Contact manifolds provide an important class of examples of geometric structures. In this paper we study the geometric properties of $\D(M)$ viewed as a module over the Lie algebra $\K(M)$ of all contact vector fields on $M$. This viewpoint fits into the general framework of *Heisenberg calculus*, see [@BG88; @EM98; @vE10], where the geometric structure is a codimension-1 distribution in $T(M)$.
Our first main result is the association of a contact vector field to an arbitrary second order linear differential operator in a contact-invariant manner. Although we do not carry out the investigation here, this could provide a means to associate topological invariants to second order operators. We generalize the result to differential operators of arbitrary order, associating to each a certain *tensor density* on $M$. This tensor density is independent of the symbol of the operator and may be thought of as a partial “subsymbol”.
By a tensor density, we mean a section of a power of the determinant line bundle. In fact, we state our results in the more general context of the spaces $\D_{\l,\mu}(M)$ of differential operators between such line bundles, rather than simply for differential operators on functions. An interesting feature is appearance of *contact resonances*, that is, of special powers of the determinant line bundle for which the geometric properties of differential operators are more complicated. These resonances were already observed in [@FMP08]. Let us mention that the usual case where $\l=\mu$ is non-resonant.
Our second main result is the existence of a filtration refining the usual filtration given by the Heisenberg calculus. Recall that differential operators on a contact manifold have a *contact order*, in which vector fields tangent to the distribution are of order 1, and contact vector fields are of order 2. We introduce a contact-invariant filtration on $\D_{\l,\mu}(M)$ for which, roughly speaking, tangential vector fields have order 1 and contact vector fields have order 3. However, this filtration is not compatible with composition. We prove its existence in the non-resonant case.
We remark that there is also an invariant double filtration on the space of differential forms on $M$ [@Ru94]. In some ways, the situation for contact manifolds appears to be analogous to that for foliated manifolds. This may at first be surprising, as contact distributions are completely non-integrable, but such analogies have been observed before [@ET98].
Although our results and their applications are essentially geometric, the proofs are algebraic. The considerations are local, so we may work in the Euclidean case, replacing $M$ by $\bR^m$, where $m=2\ell+1$. Moreover, we need only consider the Lie algebra of polynomial contact vector fields on $\bR^m$, which is the classical infinite dimensional Cartan algebra $\K_m$; see [@Fu86].
Our main theorems are proven using certain underlying structural results concerning the cohomology of $\K_m$ with coefficients in spaces of s between refined symbol modules. These results in turn are obtained using a quantization map which is equivariant with respect to the projective subalgebra $\ds_m$ of $\K_m$, a maximal subalgebra isomorphic to $\dsp_{2(\ell+1)}$. We prove the existence and uniqueness of this quantization map using the description of the s of $\ds_m$ given by the Harish-Chandra homomorphism. We also calculate the map explicitly.
An $\ds_m$-module is said to have an if the center $\dZ(\ds_m)$ of the of $\ds_m$ acts on it by scalars. The is then the resulting homomorphism from $\dZ(\ds_m)$ to $\bC$. If an $\ds_m$-module has a finite Jordan-Hölder composition series of modules with distinct s, then the module splits as the direct sum of its composition series modules.
Let us discuss at this point the role of s in other forms of quantization. The [*Casimir element*]{} is the best known and simplest element of $\dZ(\ds_m)$. It turns out that for the contact projective quantization studied in this paper, it is not sufficient to consider the eigenvalues of the Casimir element alone, because there are fine symbol modules with distinct s but identical Casimir eigenvalues. This is in contrast with the situation for projective quantization with respect to the full vector field Lie algebra $\vrm$, whose projective subalgebra is $\dsl_{m+1}$. The full principal symbol modules have s under the action of $\dsl_{m+1}$, and these s are distinct their Casimir eigenvalues are different [@Le00]. Therefore in this setting there is no need to consider s.
For conformal quantization, one replaces the projective subalgebra with the conformal subalgebra $\dgo_{p+1, q+1}$, a maximal subalgebra of $\Vec(\bR^{p+q})$. As was first observed in [@DLO99], in this setting the Casimir element of $\dgo_{p+1, q+1}$ is not sufficient to detect distinct s among these submodules. Complete results concerning the existence and uniqueness of conformal quantization for s between s have recently been obtained in [@Si09] and [@Mi11]. It would be interesting to determine to what extent s can be used to replicate them.
The crucial property that allows us to apply algebraic results in the geometric situation of an arbitrary contact manifold $M$ is the *uniqueness* of the $\dsp_{2(\ell+1)}$-equivariant quantization map. For example, the subsymbol is first defined locally in Darboux coordinates. Its uniqueness then implies that it is defined globally on $M$.
This paper is organized as follows. In Section \[Results\] we define the modules of tensor densities, s, and symbols, and formulate our main results. In Section \[K\] we fix local Darboux coordinates and review local properties of the contact Lie algebra. In Section \[S\] we study the modules of Section \[Results\] under the action of the projective subalgebra, using s to compute contact resonances. In Section \[PQ\] we prove the existence and uniqueness of the projective quantization, the natural projective equivalence from symbols to s. Section \[CRCs\] contains the proofs of two of our main results: the existence and uniqueness of the subsymbol and of the fine filtration. Section \[T\] gives the explicit formula for the projective quantization and the subsymbol and proves our third main result.
Main results {#Results}
============
Fix $m = 2\ell + 1$ odd, and let $M$ be a smooth $m$-dimensional manifold equipped with a contact distribution $\Xi$: a completely non-integrable distribution of codimension 1. As usual, locally we define the contact structure in terms of a contact form $\theta$ whose kernel is $\Xi$. The non-integrability of $\Xi$ is equivalent to the fact that $\theta\wedge(d\theta)^\ell$ is a local volume form. We define the subspace $\tm$ of $\vm$ to consist of the sections of $\Xi$, that is, the vector fields annihilated by $\theta$. We will refer to such vector fields as *tangential vector fields.*
Definitions {#Definitions}
-----------
We will use the following notation throughout this paper. For $X \in \vm$, we write $L(X)$ for the associated *Lie derivative.* The non-negative integers will be denoted by $\bN$, and the positive integers by $\bZ^+$. For $x \in \bR$, we use the floor notation for the greatest integer $\le x$ and the ceiling notation for the least integer $\ge x$: $$\lfloor x \rfloor := \sup \{ n \in \bZ: n \le x \}, \qquad
\lceil x \rceil := \inf \{ n \in \bZ: n \ge x \}.$$
Within $\vm$ we have the Lie subalgebra $\cm$ of *contact vector fields,* those which preserve $\Xi$. Contact vector fields are characterized locally as those whose Lie derivatives preserve the conformal class of $\theta$. More precisely, a vector field $X$ on $M$ is contact if $$\label{LXalpha}
L(X)\theta = {\ts \frac{1}{\ell+1}}\, \Div(X) \theta,$$ where $\Div$ is the divergence with respect to the volume form $\theta \wedge (d \theta)^\ell$. The complete non-integrability of $\Xi$ translates to $$\vm = \cm \oplus \tm.$$ This decomposition is invariant under the Lie action of $\cm$. Observe that $\cm$ is not invariant under multiplication by functions, and $\tm$ is not a Lie algebra. Let $$\pi:\Vec(M)\to\K(M)$$ be the projection along $\Tan(M)$.
We now make several definitions valid for arbitrary (not necessarily contact) manifolds.
1. For $\l \in \bC$, let $|\Lambda^m T^*(M)|^\l$ be the line bundle of homogeneous functions of degree $\l$ on the determinant bundle. The space $\F_\l(M)$ of [*tensor densities of degree $\lambda$*]{} consists of the smooth sections of $|\Lambda^m T^*(M)|^\l$ with *complex coefficients.* It is a module for $\vm$, and we write $L_\l(X)$ for the action of a vector field $X$ on it.
2. Let $\D_{\l, \mu}(M)$ be the space of s from $\F_\l(M)$ to $\F_\mu(M)$, and let $L_{\l, \mu}$ be the natural action of $\vm$ on it. For $k \in \bN$, let $\D^k_{\l, \mu}(M)$ be the subspace of operators of order $\le k$. The spaces $\D^k_{\l,\mu}(M)$ comprise the *order filtration* of $\D_{\l,\mu}(M)$ and are invariant under $\vm$.
3. We write $\delta$ for the difference between $\mu$ and $\l$: $$\delta := \mu -\l.$$
4. The space of *principal symbols* of degree $k$ is the quotient $$\S^k_\delta(M)\ :=\ \D^k_{\l,\mu}(M)/\D^{k-1}_{\l,\mu}(M).$$ It is well-known that its $\Vec(M)$-module structure depends only on $\delta$.
5. The *principal symbol* is the natural projection $$\sigma^k_{\l,\mu}: \D^k_{\l,\mu}(M) \to \S^k_\delta(M).$$
Let us give some natural examples of s. The simplest is $C^\infty(M)$, which is $\F_0(M)$. In the contact setting, the following facts are well-known.
- The adjoint action of the Lie algebra $\cm$ of contact s on itself is equivalent to $\F_{-\frac{1}{\ell+1}}(M)$. In other words, there is a $\K(M)$-equivalence $$X:\F_{-\frac{1}{\ell+1}}(M)\to\K(M),$$ associating a contact vector field $X_\varphi$ to each element $\varphi$ of $\F_{-\frac{1}{\ell+1}}(M)$. The tensor density $\varphi$ is called the *contact Hamiltonian* of $X_\varphi$. The notion of the contact Hamiltonian is independent of the choice of a contact form $\theta$. However, fixing $\theta$ one can (locally) identify tensor densities and functions and think of a contact Hamiltonian as of a function.
- The conformal class $C^\infty(M) \theta$ of the contact form $\theta$ is equivalent to $\F_{\frac{1}{\ell+1}}(M)$ as a $\cm$-module.
In fact, the second statement follows from (\[LXalpha\]), and the first follows from Lemma \[Lagrange\] below. Thus the $\K(M)$-modules of contact Hamiltonians and contact forms are dual over $C^\infty(M)$. We remark that the algebraic direct sum $\bigoplus_\l \F_\l(M)$ of all s is a Poisson algebra under the Lagrange bracket.
The space $\D_{\l, \mu}(M)$ generalizes $\D_{0,0}(M)$, which is the usual space of differential operators acting on functions. Geometric properties of $\D_{\l, \mu}(M)$ vary with the parameters, and the structure of $\D_{\l, \mu}(M)$ viewed as a $\K(M)$-module can be special for certain values of $\l$ and $\mu$. Let us stress the fact that differential operators between tensor densities appear naturally in many geometric situations. We mention for example the classical notion of the conformally invariant Laplace operator, also known as the Yamabe Laplacian, which is an element of $\D_{\frac{1}{2}-\frac{1}{m},\frac{1}{2}+\frac{1}{m}}(M)$. The case $\l+\mu=1$ is particularly special. This is the case where the notions of *symmetric* and *skew-symmetric* operators are well-defined. More generally, if $A\in\D_{\l, \mu}(M)$, then the adjoint operator $A^*$ belongs to $\D_{1-\mu,1- \l}(M)$.
We now recall the classical notion of the Heisenberg order of a differential operator on a contact manifold; see for example [@vE10] and references therein.
1. The space of differential operators of *Heisenberg order* $\leq{}d$ is $$\P^d_{\l,\mu}(M)\ :=\ \Span\bigl\{
T_c \circ L_\l(Y_1) \circ \cdots \circ L_\l(Y_{t})\,:\,
T_c \in \D^c_{\l, \mu}(M),\, Y_i\in\Tan(M),\,
2c+t\leq{}d
\bigr\}.$$ The spaces $\P^d_{\l,\mu}(M)$, comprise the *Heisenberg filtration* of $\D_{\l,\mu}(M)$. They are invariant under $\K(M)$.
2. The bifiltration $\D^{k,d}_{\l,\mu}(M):=\D^k_{\l,\mu}(M)\cap\P^d_{\l,\mu}(M)$ gives rise to the *fine symbol* modules: $$\FS^{k,d}_\delta(M)\ :=\
\D^{k,d}_{\l,\mu}(M) / \bigl(\D^{k-1, d}_{\l,\mu}(M)+\D^{k,d-1}_{\l,\mu}(M)\bigr).$$
3. The *fine symbol* is the corresponding projection $$\fs^{k,d}_{\l,\mu}: \D^{k,d}_{\l,\mu}(M) \to \FS^{k,d}_\delta(M).$$
In the simplest case $k=1$ and $\delta=0$, $\FS^{1,1}_0(M)$ is $\Tan(M)$ and $\FS^{1,2}_0(M)$ is $\K(M)$. Here $ \fs^{1,2}_{\l,\l}$ is nothing but the projection $\pi$ defined above. More generally, it follows from Proposition \[LXf\] below that $$\FS^{k,2k}_\delta(M) \cong \F_{\delta-\frac{k}{\ell+1}}(M).$$ Therefore $\fs^{k,2k}_{\l,\mu}$ may be regarded as a $\K(M)$-equivariant linear projection from $\D^k_{\l,\mu}(M)$ to $\F_{\delta-\frac{k}{\ell+1}}(M)$. This type of equivariant assignment of a tensor density to a is known as a *differential invariant.*
We say that $\delta$ is [*contact-resonant*]{} if it lies in the set $$\left\{\frac{1}{\ell+1} + \frac{n}{2(\ell+1)}
\,\Big|\,n\in\bN \right\}.$$
We will see in Section \[ICS\] that contact resonances arise from the representation theory of $\dsp_{2(\ell+1)}$.
The subsymbol
-------------
Our first main theorem gives a new contact differential invariant.
\[subsymbol\] If $\delta$ is not contact-resonant, then there exists a unique $\K(M)$-equivariant linear map $$\ss^k_{\l,\mu}: \D^k_{\l,\mu}(M) \to \FS^{k-1,\,2(k-1)}_\delta(M)$$ whose restriction to $\D^{k-1}_{\lambda,\mu}(M)$ is $\fs^{k-1,\,2(k-1)}_{\lambda,\mu}$.
We refer to $\ss^k_{\l,\mu}$ as the [*contact subsymbol.*]{} We will give an explicit formula for it in Proposition \[subsymbol formula\]. It may be regarded as a $\K(M)$-equivariant projection from $\D^k_{\lambda,\mu}(M)$ to $\F_{\delta - \frac{k-1}{\ell+1}}(M)$.
We remark that in the general self-adjoint case, where $\lambda + \mu = 1$ and $k$ is arbitrary, the existence of such a differential invariant is obvious. Indeed, for $T$ in $\D^k_{\lambda,\mu}(M)$, the operator $T - (-1)^k T^*$ is in $\D^{k-1}_{\lambda,\mu}(M)$, and so can be projected to $\FS^{k-1,\,2(k-1)}_\delta(M)$.
Since $\F_{-\frac{1}{\ell+1}}(M)$ is equivalent to $\cm$, the case that $k=2$ and $\mu=\l$ is of particular interest, as there the differential invariant given by the contact subsymbol may be viewed as a contact vector field. In other words, for all $\l\in\bC$, the subsymbol $\ss^2_{\lambda,\lambda}$ defines a $\cm$-equivariant projection from $\D^2_{\lambda,\lambda}(M)$ to $\K(M)$.
In order to give an intrinsically defined and manifestly contact-invariant formula for $\ss^2_{\lambda,\lambda}$, observe that any second order differential operator can be represented as a linear combination of compositions of vector fields. On contact manifolds, contact vector fields and tangential vector fields are intrinsically distinguished. Thus we are led to express an arbitrary second order operator on $\F_\l(M)$ as a linear combination of operators of the form $$\label{SecOp}
\begin{array}{rcl}
T&=&L_{\l}(X_{\varphi_1})\circ{}L_{\l}(X_{\varphi_2})+L_{\l}(X_{\varphi_3})\circ{}L_{\l}(Y_1)+
L_{\l}(Y_2)\circ{}L_{\l}(Y_3)\\[8pt]
&&+L_{\l}(X_{\varphi_4})+L_{\l}(Y_4)+f,
\end{array}$$ where the $\varphi_i$ are arbitrary contact Hamiltonians, the $Y_i$ are tangential vector fields, and $f$ is a function.
\[subsOrd2\] The subsymbol $\ss^2_{\lambda,\lambda}(T)$ is the contact vector field $$\ts
\oh\bigl[X_{\varphi_1},X_{\varphi_2}\bigr]-
\bigl(\frac{\ell+1}{\ell+2}\bigr)\bigl(\l-\oh\bigr)X_{L(Y_1)\varphi_3}+
\oh\pi\bigl[Y_2,Y_3\bigr]+X_{\varphi_4},$$ where $L(Y_1)\varphi_3$ denotes the natural action of $Y_1$ on the $-\frac{1}{\ell+1}$-density $\varphi_3$.
Let us comment on this formula. It only contains natural operations, so it is clearly contact-invariant. Conversely, equivariance with respect to $\K(M)$ (in fact the affine subalgebra suffices) implies that $\ss^2_{\lambda,\lambda}(T)$ has to be of the form $$\ts
c_{12}\bigl[X_{\varphi_1},X_{\varphi_2}\bigr]+
c_{13}X_{L(Y_1)\varphi_3}+
c_{23}\pi\bigl[Y_2,Y_3\bigr]+c_4X_{\varphi_4},$$ where the $c$’s are constants. The normalization condition on $\D^1_{\l,\l}$ gives $c_4=1$. Skew-symmetrizing the expression then yields $c_{12}=c_{23}=\oh$. Symmetrizing the expression implies that $c_{13}$ vanishes in the self-adjoint case $\l=\oh$, but its exact form must be deduced by computation.
The main content of the theorem is that the formula is actually well-defined. Indeed, the choice of the $\varphi_i$ and $Y_i$ in (\[SecOp\]) is not unique: one can write an operator as a linear combination of such expressions in many different ways. However, the formula is independent of the choice. Moreover, the uniqueness statement of Theorem \[subsymbol\] implies that, up to a scalar, this is not true for any other choice of the $c$’s.
The fine filtration
-------------------
In order to explain the significance of our next theorem, consider the following arrangement of the fine symbol modules (we have omitted $M$ and $\delta$ for clarity):
$$\begin{array}{llllllllll}
& & & & & & & & \FS^{6,12}&\\[10pt]
& & & & & && \FS^{5,10}\ & \FS^{6,11}&\\[10pt]
& & & & & \FS^{4,8}& & \FS^{5,9}\ & \FS^{6,10}&\\[10pt]
& & & & \FS^{3,6}\ & \FS^{4,7}&& \FS^{5,8}\ & \FS^{6,9}\ & \cdots \\[4pt]
\cline{3-3}\\[-8pt]
& & \FS^{2,4}&\!\vline\!\!\!\! & \FS^{3,5}\
& \FS^{4,6} & & \FS^{5,7}\ & \FS^{6,8}&\\[4pt]
\cline{5-6}\\[-8pt]
& \FS^{1,2}\ & \FS^{2,3}&& \FS^{3,4}\ & \FS^{4,5}
&\!\vline & \FS^{5,6}\ & \FS^{6,7}& \\[4pt]
\cline{8-9}\\[-8pt]
\FS^{0,0}\ & \FS^{1,1}\ & \FS^{2,2}& & \FS^{3,3}\
& \FS^{4,4}& & \FS^{5,5}\ & \FS^{6,6}&\!\!\vline
\end{array}$$
Observe that the graded module of $\S^k_\delta(M)$ defined by the bifiltration $\D^{k,d}_{\l,\mu}(M)$ is the “vertical” sum $$\gr\S^k_\delta(M)=\bigoplus_{k \le d \le 2k}\FS^{k,d}_\delta(M).$$ The graded module of $\P^d_{\l,\mu}(M)/\P^{d-1}_{\l,\mu}(M)$ is the “slope $-1$” sum $$\gr\bigl(\P^d_{\l,\mu}(M)/\P^{d-1}_{\l,\mu}(M)\bigr)=
\bigoplus_{\lceil\frac{d}{2}\rceil \le k \le d}\FS^{k,d}_\delta(M).$$
The content of our next theorem is that there exists a $\K(M)$-invariant filtration that strengthens the filtration $\P^d_{\l,\mu}(M)$. The graded modules of its subquotients are the “slope $-\frac{1}{2}$” sums.
\[bifilt\] Assume that $\delta$ is not contact-resonant. Then there is a unique $\K(M)$-invariant filtration of $\D_{\l,\mu}(M)$, $$\D^{(0)}_{\l,\mu}(M)
\subset\cdots\subset
\D^{(b)}_{\l,\mu}(M)\subset
\D^{(b+1)}_{\l,\mu}(M)
\subset\cdots,$$ such that the graded module of $\D^{(b)}_{\l,\mu}(M)$ is given by $$\mathrm{gr}\,\D^{(b)}_{\l,\mu}(M)=
\bigoplus_{2d-k\leq{}b}\,\FS^{k,d}_\delta(M).$$
For example, $\mathrm{gr}\bigl(\D^{(6)}_{\l,\mu}(M)/\D^{(5)}_{\l,\mu}(M)\bigr)
=
\FS^{6,6}_\delta(M)\oplus{}\FS^{4,5}_\delta(M)\oplus{}\FS^{2,4}_\delta(M),
$ as indicated by the boundaries in the diagram above. We will define $\D^{(b)}_{\l,\mu}(M)$ via the projective quantization: see Section \[C proof\].
Comments on non-existence and conjectures
-----------------------------------------
In this paper we are concerned with existence results rather than non-existence results, but we remark that in the contact-resonant case our theorems are false for most values of $\l$. For example, if $k = \delta (\ell + 1) - \ell$ in Theorem \[subsymbol\], then there exists no subsymbol except in the self-adjoint case $\lambda + \mu = 1$, where there exists a 1-parameter family of such maps. Non-existence results can be interpreted in terms of cohomological obstructions. We conjecture that the filtration of Theorem \[bifilt\] does not exist in the contact-resonant case. These questions will be addressed elsewhere.
A module is called *uniserial,* or *completely indecomposable,* if it has a unique maximal invariant filtration. Such modules may be thought of as opposite to completely reducible modules. We conjecture that for generic values of $\l$ and $\mu$, $\D^{(b)}_{\l,\mu}(M) / \D^{(b-1)}_{\l,\mu}(M)$ is uniserial, because we expect that in the Euclidean case, there is a non-trivial projectively relative 1-cohomology class of $\K_m$ linking $\FS^{k,d}_\delta$ to $\FS^{k-2,d+1}_\delta$: an analog of the Schwarzian derivative. This conjecture says essentially that the bifiltration $\D^k_{\l,\mu}(M)\cap\D^{(b)}_{\l,\mu}(M)$ is the best possible. In other words, if the conjecture is true then there is no $\K(M)$-invariant filtration of $\D_{\l,\mu}(M)$ whose elements are composed of all those fine symbol modules on or below the lines of any fixed slope shallower than $-\frac{1}{2}$ passing through the above diagram.
The Euclidean contact Lie algebra {#K}
=================================
Since all of the theorems in Section \[Results\] are local, their proofs essentially reduce to the case $M = \bR^m$. Therefore in this section we establish notation and state some well-known results for Euclidean contact manifolds: the proofs are straightforward and are usually omitted. All $\cirm$-modules of finite rank are equipped with their usual topologies as Frechet spaces, and by definition all $\Hom$ spaces between such modules include only continuous linear maps.
Darboux coordinates {#DarbSect}
-------------------
Fix coordinates $x_i$, $y_i$, and $z$ on $\bR^m$, where $m=2\ell+1$ and $1 \le i \le \ell$. Henceforth we use [*Einstein’s summation convention:*]{} unless stated otherwise, [*repeated indices are summed over from $1$ to $\ell$.*]{} Let $\theta$ and $\vol$ be the standard contact and volume forms on $\bR^m$: $$\theta := dz + \oh (x_i dy_i - y_i dx_i),
\qquad
\vol := {\ts \frac{1}{\ell!}} \th \theta \wedge (d \theta)^\ell
= dz \wedge {\ts \bigwedge_1^\ell} (dx_i \wedge dy_i).$$ The standard divergence operator $\Div: \vrm \to \cirm$ is defined by $$L(X)\, \vol = \Div(X) \vol.$$
Recall from Section \[Results\] the definitions of the Lie algebra of contact vector fields and the space of tangent vector fields: $$\begin{aligned}
\crm &:=& \bigl\{ X \in \vrm:
L(X)\, \theta = {\ts \frac{1}{\ell + 1}} \Div(X) \theta \bigr\}, \\[6pt]
\trm &:=& \bigl\{ X \in \vrm: \la \theta, X \ra = 0 \bigr\}.\end{aligned}$$ The space $\trm$ is a module over $\crm$, and so one has the $\crm$-invariant decomposition $$\vrm = \crm \oplus \trm.$$ It is important to keep in mind that while $\trm$ is closed under multiplication by smooth functions, $\crm$ is not. We now give explicit descriptions of both spaces.
Define the following [*Euler operators:*]{} $$\E_z := z \pz, \qquad \E_{xy} := x_i \pxi + y_i \pyi.$$ For $1 \le i \le \ell$, define the vector fields $$A_i := \pxi + \oh y_i \pz, \qquad B_i := -\pyi + \oh x_i \pz.$$ The following statements are classical.
- There is a linear bijection $X: \cirm \to \crm$ mapping $f$ to $X_f$, the unique contact vector field such that $\la \theta, X_f \ra = f$. It has the following explicit formulas: $$\label{Xf}
\begin{array}{rcl}
X_f &=& f\pz + B_i(f) A_i - A_i(f) B_i \\[6pt]
&=& \bigl( 1 - \oh \E_{xy} \bigr)(f) \pz
+ \oh \pz(f) \E_{xy} + \bigl(\pxi(f) \pyi - \pyi(f) \pxi\bigr).
\end{array}$$
- $\{ A_i, \th B_i: 1 \le i \le \ell \}$ is a basis of $\trm$ over $\cirm$, and $$[A_i, B_j] = \delta_{ij} \pz,
\qquad [\pz, A_i] = 0,
\qquad [\pz, B_i] = 0.$$
We remark that one can verify (\[LXalpha\]) directly in this setting by checking that $$\label{explicit LXalpha}
L(X_f) \,\theta = \pz(f)\,\theta, \qquad \Div(X_f) = (\ell + 1) \pz(f) .$$
The [*Lagrange bracket*]{} $\{f,g\}$ on $\cirm$ is defined by $$X_{\{f, g\}} := [X_f, X_g].$$ The following formulas for $\{f, g\}$ may be deduced from (\[Xf\]).
\[Lagrange\] The Lagrange bracket is given by $$\begin{aligned}
\{f, g\} &=& X_f(g) - g \pz(f) \ =\ f \pz(g) - X_g(f) \\[6pt]
&=& f \pz(g) - \pz(f) g + B_i(f) A_i(g) - A_i(f) B_i(g) \\[6pt]
&=& \bigl( 1 - \oh \E_{xy} \bigr)(f) \pz(g)
- \bigl( 1 - \oh \E_{xy} \bigr)(g) \pz(f)
+ \bigl(\pxi(f) \pyi(g) - \pyi(f) \pxi(g)\bigr).\end{aligned}$$
Tensor density modules {#TDMs}
----------------------
Recall from Section \[Results\] the tensor density module $\F_\lambda(\bR^m)$ of $\vrm$. As a vector space, $\F_\lambda(\bR^m)$ is nothing but $\cirm$. However, the action of $\vrm$ depends on $\lambda$ and is given by $$L_\lambda(X) (g) :=
X(g) + \lambda \Div(X)\,g,$$ where $g\in\cirm$ and $\lambda\in\mathbb{R}$. Since the volume form $\vol$ is global, we may regard $\F_\lambda(\bR^m)$ as $\vol^\lambda \cirm$, so that the action of $\vrm$ is identified with the usual Lie derivative: $$L(X) (\vol^\lambda g) = \vol^\lambda \,L_\lambda(X) (g).$$ We remark that the full family $\bigl\{ \F_\lambda(\bR^m): \lambda \in \bC \bigr\}$ of $\vrm$-modules can be understood algebraically as a non-trivial deformation of the module $\cirm$.
We will consider $\F_\lambda(\bR^m)$ as a module over the subalgebra of contact vector fields $\crm$. In light of (\[LXalpha\]) and the global contact form $\theta$, we may regard $\F_\lambda(\bR^m)$ in this context as either $\vol^\lambda \cirm$ or $\theta^{\lambda (\ell+1)} \cirm$. In particular, (\[explicit LXalpha\]) gives $$L(X_f) (\theta^{\lambda (\ell + 1)} g)
= \theta^{\lambda (\ell + 1)} L_\lambda(X_f) (g)
= \theta^{\lambda (\ell + 1)} \bigl(X_f(g)
+ \lambda (\ell + 1) \pz(f) g \bigr).$$
As mentioned in Section \[Results\], the adjoint action of $\crm$ on itself is equivalent to the module of $-\frac{1}{\ell + 1}$-densities. The following definition and lemma state this formally.
Henceforth we regard $X$ as the map $$X: \F_{-\frac{1}{\ell + 1}}(\bR^m) \to \crm,
\qquad X(\theta^{-1} f) := X_f.$$
\[X\] The map $X$ is a linear bijection and a $\crm$-equivalence.
By Lemma \[Lagrange\], $\{f,g\}=L_{-\frac{1}{\ell + 1}}(X_f) (g)$. Therefore $X$ intertwines the $\crm$-action on $\F_{-\frac{1}{\ell + 1}}(\bR^m)$ and the adjoint action. $\Box$
Differential operator modules {#DOMs}
-----------------------------
We now turn to the focus of the paper, the modules $\D_{\l,\mu}(\bR^m)$ of s from $\F_\l(\bR^m)$ to $\F_\mu(\bR^m)$. We can write the action $L_{\l,\mu}$ of $\vrm$ on these modules concretely as follows: $$L_{\lambda, \mu}(X) (T) := L_\mu(X) \circ T - T \circ L_\lambda(X).$$
The structure of the spaces $\D_{\l,\mu}(\bR^m)$ viewed as $\vrm$-modules has been thoroughly studied; see for example [@LMT96; @DO97; @GMO05; @Co09] and the references therein. We will be interested in these spaces viewed as $\crm$-modules, as which they are less rigid, because $\crm$ is smaller than $\vrm$. In particular, the $\D_{\l, \mu}(\bR^m)$ admit more $\crm$-invariant operations than they do $\vrm$-invariant operations, such as projections to tensor fields. We will understand such operations as differential invariants.
Symbol modules {#SMs}
--------------
In Section \[Results\] we defined the principal symbol modules $\S^k_\delta(\bR^m)$ and the fine symbol modules $\FS^{k,d}_\delta(\bR^m)$, $k \le d \le 2k$. We will write $\S_\delta(\bR^m)$ and $\FS_\delta(\bR^m)$ for the [*total symbol module*]{} and the [*total fine symbol module,*]{} respectively; the graded modules associated to the order filtration and the bifiltration of $\D_{\l,\mu}(\bR^m)$: $$\S_\delta(\bR^m)\ :=\
\bigoplus_{k=0}^\infty \S^k_\delta(\bR^m), \qquad
\FS_\delta(\bR^m)\ :=\
\bigoplus_{k=0}^\infty \bigoplus_{d=k}^{2k}\FS^{k,d}_\delta(\bR^m).$$
Let $L^\S_{\delta}$ be the natural action of $\vrm$ on $\S_\delta(\bR^m)$, and let $L^{\FS}_\delta$ be the natural action of $\crm$ on $\FS_\delta(\bR^m)$. Our next proposition gives formulas for $L^\S_\delta(X_f)$ and $L^{\FS}_\delta(X_f)$. In order to state it we must develop a variation of the usual symbol calculus which is adapted to the fine filtration. Let $\alpha_i$, $\beta_i$, and $\zeta$ be the symbols associated to the vector fields $A_i$, $B_i$, and $\pz$. More explicitly, $$\label{abxi}
\alpha_i=\xi_{x_i}+\oh\,y_i\xi_z,
\qquad
\beta_i=-\xi_{y_i}+\oh\,x_i\xi_z,
\qquad
\zeta=\xi_z,$$ where $\xi_{x_1}, \xi_{y_1}, \ldots, \xi_{x_\ell}, \xi_{y_\ell}, \xi_z$ are the coordinates on $T^*\bR^{m}$ dual to $x_1, y_1, \ldots, x_\ell, y_\ell, z$.
We shall abuse notation and use $\alpha_i$ and $\beta_i$ also to denote fine symbols. Thus $$\label{symbol bases} \begin{array}{rcl}
\S^k_\delta (\bR^m) &=& \Span_{\cirm}\bigl\{
\zeta^c \alpha^I \beta^J: |I| + |J| + c = k \bigr\}, \\[6pt]
\FS^{k,d}_\delta (\bR^m) &=& \Span_{\cirm}\bigl\{
\zeta^{d-k} \alpha^I \beta^J: |I| + |J| = 2k - d \bigr\},
\end{array}$$ where $I$ and $J$ are multi-indices: $I=(I_1,\ldots,I_\ell)$ and $J=(J_1,\ldots,J_\ell)$. Note that in local coordinates we do not write the shift $\omega^\delta$ in the tensor density degree explicitly.
Any from one symbol module to another, or from one fine symbol module to another, may be written as a linear combination over $\cirm$ of products of the operators $$A_i,\ B_i,\ \pz,\ \alpha_i,\ \beta_i,\ \zeta,\
\partial_{\alpha_i},\ \partial_{\beta_i},\ \partial_\zeta.$$ Such a combination is to be interpreted as follows. The operators $A_i$, $B_i$, and $\pz$ act solely on the $\cirm$ coefficients of the basis elements in (\[symbol bases\]), while the remaining operators act solely on the basis elements themselves.
Imitating the definitions of $\E_z$ and $\E_{xy}$, we set $$\E_\zeta := \zeta \partial_\zeta, \qquad \E_{\alpha \beta} :=
\alpha_i \partial_{\alpha_i} + \beta_i \partial_{\beta_i}.$$
\[LXf\]
1. The action $L_\delta^{\FS}$ of $\crm$ on $\FS_\delta(\bR^m)$ is $$\begin{array}{rcl}
L_\delta^{\FS}(X_f) &=& f \pz + \th B_i(f) A_i - \th A_i(f) B_i + \th \pz(f)
\bigl(\delta (\ell + 1) - \E_\zeta - \oh \E_{\alpha \beta} \bigr)
\\[6pt]
&&+ \th \oh (A_i B_i + B_i A_i)(f) (\beta_i \partial_{\beta_j}
- \alpha_i \partial_{\alpha_j})
+ \th A_i A_j(f) \beta_i \partial_{\alpha_j}
- \th B_i B_j(f) \alpha_i \partial_{\beta_j}.
\end{array}$$
2. The action $L_\delta^{\FS}$ of $\crm$ on $\S_\delta(\bR^m)$ is $$L_\delta^{\S}(X_f) =
L_\delta^{\FS}(X_f)+\bigl(\pz A_i(f) \beta_i
- \pz B_i(f) \alpha_i \bigr) \partial_\zeta.$$
Taking in to account the shift $\omega^\delta$ in tensor density degree, $L^\S$ and $L^{\FS}$ are derivations in an obvious sense. Therefore it is only necessary to check the formulas on the generators $\alpha_i$, $\beta_i$, and $\zeta$ and on functions $g$. Keep in mind that $A_i$ and $B_i$ do not commute, although $\alpha_i$ and $\beta_i$, being symbols, do. $\Box$
The difference between $L_\delta^{\S}$ and $L_\delta^{\FS}$ is due to the fact that $\FS_\delta(\bR^m)$ is the graded module of $\S_\delta(\bR^m)$. Observe that $L_\delta^{\S}-L_\delta^{\FS}$ maps $\FS^{k,d}_\delta(\bR^m)$ to $\FS^{k,d-1}_\delta(\bR^m)$.
Henceforth we will frequently drop the argument $\bR^m$ of the various tensor density, , and symbol modules, writing simply $\F_\lambda$, $\D_{\lambda, \mu}$, $\S^k_\delta$, $\FS^{k,d}_\delta$, and so on.
The projective subalgebra {#S}
=========================
Here we recall the [*projective subalgebra*]{} $\ds_m$ of $\crm$, which is isomorphic to $\dsp_{2(\ell+1)}$. Restriction of $\crm$-modules to $\ds_m$ will be central to our strategy throughout this paper, for two reasons. First, $\ds_m$ is a maximal polynomial subalgebra of $\crm$, and it turns out that for most values of their parameters, the s and the fine symbol modules are not only algebraically under $\crm$, they remain so under restriction to $\ds_m$. (By “algebraically ”, we mean in the polynomial category.) Second, $\ds_m$ is a finite dimensional semisimple Lie algebra, and so we can bring the theory of such algebras to bear on the restricted modules.
In fact, the restrictions to $\ds_m$ of the s and fine symbol modules are duals of $\dsp_{2\ell}$-relative Verma modules. Our approach to the construction of the projective quantization referred to in the introduction will be to observe that for generic $\delta$, the fine symbol modules composing $\D_{\l,\mu}$ have distinct infinitesimal characters under the action of $\ds_m$. This implies that there is a unique $\ds_m$-equivariant splitting of $\D_{\l,\mu}$ into the sum of its fine symbol modules. This splitting may be regarded as a projectively invariant total symbol. The projective quantization is, by definition, its inverse.
In addition to the projective subalgebra $\ds_m$, two other subalgebras of $\crm$ will be important to us: the [*affine subalgebra*]{} $\dt_m$ and its nilradical $\du_m$.
The projective subalgebra of the full vector field Lie algebra {#SlCoord}
--------------------------------------------------------------
We first recall the definitions of the analogous subalgebras of the full vector field Lie algebra. Let $u_1, \ldots, u_m$ be any coordinates on $\bR^m$. The full Euler operator is $$\E_u:=\sum_{i=1}^m u_i \partial_{u_i}.$$ Within $\Vec(\bR^m)$ we have the projective subalgebra $\da_m$, the affine subalgebra $\db_m$, and the constant coefficient subalgebra $\dc_m$: $$\begin{array}{rcl}
\da_m &:=& \Span_\bC \bigl\{ \partial_{u_i},
u_j \partial_{u_i}, u_j \E_u: 1 \le i, j \le m \bigr\}, \\[6pt]
\db_m &:=& \Span_\bC \bigl\{ \partial_{u_i},
u_j \partial_{u_i}: 1 \le i, j \le m \bigr\}, \\[6pt]
\dc_m &:=& \Span_\bC \bigl\{ \partial_{u_i}:
1 \le i \le m \bigr\}.
\end{array}$$
Clearly $\da_m \supset \db_m \supset \dc_m$. In fact, $\da_m$ is isomorphic to $\dsl_{m+1}$, and $\db_m$ is a maximal parabolic subalgebra of $\da_m$ with Levi factor $\dgl_m$ and nilradical $\dc_m$. The center of the Levi factor is $\bC \E_u$.
There is a standard conceptual proof of $\da_m \cong \dsl_{m+1}$ which we briefly sketch. Let $u_0, \ldots, u_m$ be coordinates on $\bR^{m+1}$, and note that $\Span_\bC \{ u_j \partial_{u_i}: 0 \le i, j \le m \}$ is a subalgebra of $\Vec (\bR^{m+1})$ isomorphic to $\dgl_{m+1}$. Restricting the action of $\Vec (\bR^{m+1})$ to functions on $\bP^m$ defines the canonical projection from this copy of $\dgl_{m+1}$ to $\dsl_{m+1}$. Regarding $\bP^m$ locally as the hyperplane defined by $u_0 = 1$ and identifying this hyperplane with $\bR^m$ yields the isomorphism.
The projective subalgebra of the contact Lie algebra
----------------------------------------------------
We now define the projective and affine subalgebras of $\crm$: they are simply the intersections of the projective and affine subalgebras of $\vrm$ with $\crm$.
The [*projective subalgebra*]{} $\ds_m$ and the [*affine subalgebra*]{} $\dt_m$ of $\crm$ are $$\ds_m := \da_m \cap \crm, \qquad
\dt_m := \db_m \cap \crm.$$ We define also the following subspaces of $\crm$: $$\begin{aligned}
\du_m &:=& \Span_\bC \bigl\{ X_1, X_{x_i}, X_{y_i}:
1\le i \le \ell \bigr\}, \\[6pt]
\dl_m^{\rms} &:=& \Span_\bC \bigl\{ X_{x_i y_j}, X_{x_i x_j}, X_{y_i y_j}:
1\le i, j \le \ell \bigr\}, \\[6pt]
\dl_m &:=& \dl_m^{\rms} \oplus \bC X_z.\end{aligned}$$
Observe that $X_z = \E_z + \oh \E_{xy}$. This is the natural Euler operator in $\crm$. The space $\dl_m$ is the $0$-weight space of its adjoint action, and $\du_m$ is the sum of its $-\oh$-eigenspace $\Span_\bC \{X_{x_i},\th X_{y_i}:\th 1 \le i \le \ell \}$ and its $-1$-eigenspace $\bC X_1$.
Next we give explicit descriptions of $\ds_m$ and $\dt_m$ and prove that $\ds_m$ is symplectic. We also show that $\du_m$, $\dl_m^{\rms}$, and $\dl_m$ are subalgebras of $\dt_m$, and that $\du_m$ is the contact-analog of the constant coefficient algebra $\dc_m$.
\[sp iso\]
1. The space $\dl_m^{\rms}$ is a subalgebra of $\ds_m$ isomorphic to $\dsp_{2\ell}$.
2. The space $\dl_m$ is a Levi subalgebra of $\ds_m$, with semisimple part $\dl_m^{\rms}$ and center $\bC X_z$.
3. The space $\du_m$ is a Heisenberg Lie algebra with center $\bC X_1$.
4. The affine subalgebra $\dt_m$ is the semidirect sum $\dl_m \oplus_{\rms} \du_m$. It is a maximal parabolic subalgebra of $\ds_m$ with nilradical $\du_m$.
5. The projective subalgebra $\ds_m$ is isomorphic to $\dsp_{2(\ell + 1)}$, and $$\ds_m = X \bigl(\theta^{-1} \bigl\{
\mbox{\rm Polynomials of degree~$\le 2$ on $\bR^m$} \bigr\}\bigr).$$
Most of this can be left to the reader. To prove (i), use (\[Xf\]) to obtain $$X_{x_i y_j} = y_j \partial_{y_i} - x_i \partial_{x_j}, \quad
X_{x_i x_j} = x_i \partial_{y_j} + x_j \partial_{y_i}, \quad
X_{y_i y_j} = -y_i \partial_{x_j} - y_j \partial_{x_i}.$$ Then note that the natural action of $\dl_m^{\rms}$ on $\Span_{\bC} \{ x_i,\th y_i:\th 1 \le i \le \ell \}$ preserves the skew-symmetric form defined by $\la x_i, x_j \ra = 0$, $\la y_i, y_j \ra = 0$, and $\la x_i, y_j \ra = \delta_{ij}$. (Alternately, observe that the Lagrange bracket defines a non-degenerate $\dl_m$-invariant skew-symmetric form on the $-\oh$-eigenspace of $X_z$.) Parts (ii), (iii), and (iv) now follow by computation.
For a direct proof of the displayed equation in (v), first check that $\E_{xy} = x_i A_i - y_i B_i$. Then verify that the vector fields $$A_i,\quad B_i,\quad x_j A_i - y_i B_j,\quad
x_j B_i - x_i B_j,\quad y_j A_i - y_i A_j,\quad z A_i + \oh y_i \E_{xy},\quad
z B_i + \oh x_i \E_{xy},$$ $1 \le i, j \le \ell$, span a $(2\ell + 3) \ell$-dimensional subspace of $\da_m \cap \trm$.
Next, use Lemma \[X\] to verify that the right side of the display in (v) is a $(2\ell + 3) (\ell + 1)$-dimensional subspace of $\da_m \cap \crm$. Since $\da_m$ is $(2\ell + 3) (2\ell + 1)$-dimensional, (v) is proven.
The fact that $\ds_m$ is a copy of $\dsp_{2(\ell + 1)}$ can now be proven using (ii) and an adaptation of the above argument proving that $\da_m$ is a copy of $\dsl_{m+1}$. $\Box$
Let us establish some notation for later use: we will write $\bC^{2\ell}$ for the basic module of $\dl_m^{\rms} \cong \dsp_{2\ell}$, and $\Sym^r\bC^{2\ell}$ for its $r^\thup$ symmetric power. It is well-known that $\Sym^r\bC^{2\ell}$ is self-dual and for all $r$.
The following lemma defines a Cartan subalgebra $\dh_m$ of $\ds_m$ and gives the associated root system and Weyl group. Its proof may be found in any text on Lie theory; see [@Va84].
\[roots\]
1. The set $\{ 2X_z,\th X_{x_1 y_1}, \ldots, X_{x_\ell y_\ell} \}$ is a basis of a Cartan subalgebra $\dh_m$ of $\ds_m$. Under the Cartan-Killing form this basis is orthogonal and all of its elements are of length $2 \sqrt{\ell+2}$. Let $\{ e_0, e_1, \ldots, e_\ell \}$ be the dual basis of $\dh_m^*$.
2. The adjoint action of $\dh_m$ on the polynomial contact vector fields is semisimple: $X(x^I y^J z^c)$ is of $\dh_m$-weight $(2c + |I| + |J| - 2) e_0 - (I_i - J_i) e_i$.
3. The roots of $\ds_m$ are all of the non-zero $\pm e_i \pm e_j$. The Weyl group $W(\ds_m)$ is $S_{\ell+1} \ltimes \bZ_2^{\ell+1}$, acting in the usual way on the basis $\{ e_0, \ldots, e_\ell \}$.
4. The order $0 < e_\ell < \cdots < e_0$ gives the following simple root system $\Pi^+(\ds_m)$ and positive root system $\De^+(\ds_m)$ of $\ds_m$: $$\Pi^+(\ds_m) :=
\bigl\{ e_{i-1} - e_i,\th 2e_\ell \bigr\}_{i=1}^\ell, \quad
\De^+(\ds_m) :=
\bigl\{ e_i - e_j \}_{i < j} \cup \{ e_i + e_j \bigr\}_{i,j}.$$ The half-sum of the positive roots is $\rho(\ds_m) = \sum_0^\ell (\ell+1-i)e_i$. The dominant weights are those $\sum_0^\ell \gamma_i e_i$ with $0 \le \gamma_\ell \le \cdots \le \gamma_0$. The negative root vectors are $X_1$, $X_{x_i}$, $X_{y_i}$, $X_{x_i x_j}$, and those $X_{x_i y_j}$ with $i < j$.
5. The algebra $\dh_m$ is also a Cartan subalgebra of the Levi subalgebra $\dl_m$ of $\ds_m$. The roots of $\dl_m$ are the non-zero $\pm e_i \pm e_j$ with $i, j > 0$. It inherits the simple root system $$\Pi^+(\dl_m) = \bigl\{ e_{i-1} - e_i,\th 2e_\ell \bigr\}_{i=2}^\ell.$$
6. The subalgebra $\dh_m^{\rms}:=\Span\{X_{x_1 y_1}, \ldots, X_{x_\ell y_\ell} \}$ of $\dh_m$ is a Cartan subalgebra of $\dl_m^{\rms}$. The roots of $\dl_m^{\rms}$ are the same as those of $\dl_m$.
The $\ds_m$-structure of the symbol modules {#PSA on symbols}
-------------------------------------------
In this subsection we analyze the action of $\ds_m$ on the principal and fine symbol modules. Our first lemma gives the restriction to $\ds_m$ of the actions $L^\S_\delta$ and $L^{\FS}_\delta$ on $\S_\delta$ and $\FS_\delta$. It is a corollary of Proposition \[LXf\]. In order to state it concisely, we define the [*total weight operator*]{} $W$ to act on both $\S_\delta$ and $\FS_\delta$ by $$W|_{\S_\delta} = W|_{\FS_\delta} := (\E_z + \oh \E_{xy}) -
(\E_\zeta + \oh \E_{\alpha \beta}) + \delta (\ell + 1).$$
\[LXs\]
1. The restrictions of $L^\S_\delta$ and $L^{\FS}_\delta$ to the affine subalgebra $\dt_m$ coincide.
2. Their restriction to the nilradical $\du_m$ of $\dt_m$ is the identity map $X_f \mapsto X_f$: $$X_1\ \mapsto\ \pz, \quad
X_{x_i}\ \mapsto\ x_i \pz - B_i, \quad
X_{y_i}\ \mapsto\ y_i \pz - A_i.$$
3. Their restriction to the Levi factor $\dl_m$ of $\dt_m$ is given by $$\begin{aligned}
X_z &\mapsto& W, \\[6pt]
X_{x_i y_j} &\mapsto& x_i y_j \pz - (x_i A_j + y_j B_i) +
(\alpha_j \partial_{\alpha_i} - \beta_i \partial_{\beta_j}), \\[6pt]
X_{x_i x_j} &\mapsto& x_i x_j \pz - (x_i B_j + x_j B_i) +
(\beta_i \partial_{\alpha_j} + \beta_j \partial_{\alpha_i}), \\[6pt]
X_{y_i y_j} &\mapsto& y_i y_j \pz - (y_i A_j + y_j A_i) -
(\alpha_i \partial_{\beta_j} + \alpha_j \partial_{\beta_i}).\end{aligned}$$
4. The action of the rest of $\ds_m$ under $L^{\FS}_\delta$ is given by $$\begin{aligned}
L^{\FS}_\delta (X_{x_i z}) &=& x_i W - z B_i
- \oh (x_r \alpha_r - y_r \beta_r) \partial_{\alpha_i}
+ \oh \beta_i (x_s \partial_{\beta_s}
+ y_s \partial_{\alpha_s}), \\[6pt]
L^{\FS}_\delta (X_{y_i z}) &=& y_i W - z A_i
+ \oh (x_r \alpha_r - y_r \beta_r) \partial_{\beta_i}
+ \oh \alpha_i (x_s \partial_{\beta_s}
+ y_s \partial_{\alpha_s}), \\[6pt]
L^{\FS}_\delta (X_{z^2}) &=& 2z W - z^2 \pz - \oh (x_r \alpha_r - y_r \beta_r)
(x_s \partial_{\beta_s} + y_s \partial_{\alpha_s}).\end{aligned}$$
5. The action of the rest of $\ds_m$ under $L^\S_\delta$ is given by $$\begin{aligned}
L^\S_\delta (X_{x_i z}) &=& L^{\FS}_\delta (X_{x_i z}) + \beta_i \partial_\zeta, \\[6pt]
L^\S_\delta (X_{y_i z}) &=& L^{\FS}_\delta (X_{y_i z}) + \alpha_i \partial_\zeta, \\[6pt]
L^\S_\delta (X_{z^2}) &=& L^{\FS}_\delta (X_{z^2}) - (x_r \alpha_r - y_r \beta_r) \partial_\zeta.\end{aligned}$$
Note that the operators $x_r \alpha_r - y_r \beta_r$ and $x_s \partial_{\beta_s} + y_s \partial_{\alpha_s}$ occurring in (iv) are $\dl^{\rms}_m$-invariants of total weights $0$ and $1$, respectively.
We now define a space which will regularly play an important role in our arguments: $$(\FS^{k,d}_\delta)^{\du_m} :=
\bigl\{ P \in \FS^{k,d}_\delta:
L^{\FS}_\delta(\du_m) P = 0 \bigr\},$$ the subspace of $\FS_\delta^{k,d}$ invariant under the Heisenberg algebra $\du_m$. Since $\dt_m$ normalizes $\du_m$, $(\FS^{k,d}_\delta)^{\du_m}$ is a $\dt_m$-module on which $\du_m$ acts trivially.
\[FS ltrs\]
1. $(\FS^{k,d}_\delta)^{\du_m} =
\Span_\bC\{ \zeta^{d-k} \alpha^I \beta^J: |I| + |J| = 2k - d \}$.
2. Under $W$, $(\FS^{k,d}_\delta)^{\du_m}$ has total weight $\delta (\ell+1) -\oh d$.
3. Under $\dl^{\rms}_m \cong \dsp_{2\ell}$, $(\FS^{k,d}_\delta)^{\du_m}$ is equivalent to $\Sym^{2k-d} \bC^{2\ell}$.
Part (i) is clear from the action of $\du_m$ given in Lemma \[LXs\]. Part (ii) is straightforward, and Part (iii) follows from the fact that as a module of $\dl_m^{\rms}$, $\Span_\bC\{\alpha_i,\beta_i\}_i$ is equivalent to the irreducible module $\bC^{2\ell}$ of $\dsp_{2\ell}$. $\Box$
Thus the $\du_m$-invariant fine symbols are precisely the constant fine symbols. In addition to the spaces of constant symbols, we will encounter the spaces of [*polynomial fine symbols,*]{} which are modules of the Lie algebra of the [*polynomial contact s.*]{} Under the action of $\ds_m$, these modules turn out to be the restricted duals of $\dl_m$-relative Verma modules.
To make this precise, let us write $\Poly(\bR^m)$ for the polynomials in $\cirm$. We will denote the polynomial subspaces of $\crm$, $\F_\lambda(\bR^m)$, $\D_{\lambda, \mu}(\bR^m)$, $\S_\delta(\bR^m)$, and $\FS_\delta(\bR^m)$ by writing $\rmp$ in place of $\bR^m$. Thus for example $\K(\rmp)$ is the classical Cartan algebra $\K_m$ discussed in the introduction, and $$\FS^{k,d}_\delta (\rmp) := \Span_{\prm} \bigl\{
\zeta^{d-k} \alpha^I \beta^J: |I| + |J| = 2k - d \bigr\}.$$ Note that $\K_m$ is a dense subalgebra of $\crm$ containing $\ds_m$, and all of the above polynomial subspaces are dense $\K_m$-submodules of their smooth counterparts.
We will need the fact that smooth globally defined eigenfunctions of Euler operators are polynomials.
\[EuLem\]
1. Let $f\in{}C^\infty(\bR^m)$ be an eigenfunction of the full Euler operator $\E_u=\sum_{i=1}^m{}u_i\partial_i$ with eigenvalue $\l$. Then $\l\in\bN$ and $f$ is a homogeneous polynomial in $u$ of degree $\l$.
2. Let $f\in{}C^\infty(\bR^m)$ be an eigenfunction of the contact Euler operator $\E_z+\oh\E_{xy}$ with eigenvalue $\l$. Then $\l\in\oh\bN$ and $f$ is a homogeneous polynomial in $(x,y,z)$ of degree $\l$, in the sense that $x_i$ and $y_i$ have degree $\oh$ and $z$ has degree $1$.
The first statement is classical, and the second follows from the first by the change of coordinates $(x,y,z)\mapsto(x,y,z^2)$. $\Box$
Recall from Lemma \[roots\] the Cartan subalgebra $\dh_m$ of $\ds_m$. Given any $\dh_m$-module $V$ and any $\nu \in \dh_m^*$, we use the standard notation $V_\nu$ for the $\nu$-weight space of $V$. By the [*restricted dual*]{} of $V$, we mean the direct sum of the duals of its weight spaces. A [*lowest weight vector*]{} in an $\ds_m$-module is a weight vector annihilated by the negative root vectors; see Lemma \[roots\](iv). Recall also that total weights are eigenvalues of the operator $W$.
\[FSpolyOne\]
1. $\FS^{k,d}_\delta(\rmp)$ is the span of those elements of $\FS^{k,d}_\delta$ with well-defined total weight.
2. $(\FS^{k,d}_\delta)^{\du_m}$ is the lowest total weight space of $\FS^{k,d}_\delta(\rmp)$.
3. $\FS^{k,d}_\delta(\rmp)$ has lowest $\dh_m$-weight $$\nu^{k,d}_\delta = \bigl[ 2\delta (\ell+1) - d \bigr] e_0
- \bigl[ 2k-d \bigr] e_1.$$ Its lowest $\dh_m$-weight space is the line spanned by $\zeta^{d-k} \beta_1^{2k-d}$.
Lemma \[EuLem\] implies (i), and (ii) is due to the fact that polynomials have non-negative total weights. For (iii), recall from Lemma \[roots\] the basis $\{e_0, \ldots, e_\ell\}$ of the weight space $\dh_m^*$ and its order $0 < e_\ell < \cdots < e_0$. The lowest $\dh_m$-weight space is contained in the lowest total weight space $(\FS^{k,d}_\delta)^{\du_m}$. Since the $\dh_m$-weights of $\zeta$, $\alpha_i$, and $\beta_i$ are, respectively, $-2e_0$, $-e_0 + e_i$, and $-e_0 - e_i$, (iii) follows from Lemma \[FS ltrs\]. $\Box$
We now recall the basic notions of Verma modules. Let us denote by $\du^+_m$ and $\dt^+_m$ the subalgebras of $\ds_m$ opposite to $\du_m$ and $\dt_m$, respectively: $$\du^+_m=\Span_\bC\bigl\{X_{x_iz},\,X_{y_iz},\,X_{z^2}\,:\,1\leq{}i\leq\ell\bigr\},
\qquad
\dt_m^+ = \dl_m \oplus_{\rms} \du^+_m.$$ Given an irreducible module $V$ of $\dl_m$, the associated relative Verma module of $\ds_m$ is $\dU(\ds_m) \otimes_{\dt_m^+} V$, where $\du_m^+$ acts trivially on $V$.
\[FS poly\] As an $\ds_m$-module, $\FS^{k,d}_\delta (\rmp)$ is equivalent to the restricted dual of the relative Verma module $\dU(\ds_m)\otimes_{\dt_m^+} \Sym^{2k-d} \bC^{2\ell}$.
By Lemmas \[LXs\] and \[FS ltrs\], any non-zero $\dt_m$-invariant subspace of $\FS^{k,d}_\delta (\rmp)$ contains $(\FS^{k,d}_\delta)^{\du_m}$. Taking duals, we see that the dual of $(\FS^{k,d}_\delta)^{\du_m}$ generates the restricted dual $\FS^{k,d}_\delta (\rmp)^*$ under the action of $\dt_m$. It follows that $\FS^{k,d}_\delta (\rmp)^*$ is a quotient of $\dU(\ds_m)\otimes_{\dt_m^+}
\bigl((\FS_\delta^{k,d})^{\du_m}\bigr)^*$. The reader may easily check that the two have the same total weight space dimensions, so they are equivalent. The result now follows from Lemma \[FS ltrs\](iii) and the fact that $\bC^{2\ell}$ is self-dual. $\Box$
The $\ds_m$-structure of $\K_m$
-------------------------------
As a particular case of the $\ds_m$-module structures of spaces of fine symbols, we investigate the $\ds_m$-structure of the algebra $\K_m$ itself, which is by definition $\FS^{1,2}_0$.
\[K is irr\] The quotient $\K_m/\ds_m$ is under $\ds_m$. Its is $X_{x_1^3}$, which has weight $e_0-3e_1$.
The reader may use Lemma \[LXs\] to check that the lowest weight vectors of $\K_m/\ds_m$ under $\dl_m$ are precisely all the elements of the form $X_{x_1^i z^c}$ with $i+c \ge 3$. The same lemma shows that it is possible to move from any one of these s to any other using the elements $X_1$, $X_{x_1}$, $X_{x_1 z}$, and $X_{z^2}$ of $\ds_m$. The weight of $X_{x^3_1}$ is given by Lemma \[roots\](ii). $\Box$
It will be important to understand the space of $\du_m$-invariants in $\K_m/\ds_m$. The following result is immediate from Proposition \[K is irr\].
\[WeightK\]
1. $(\K_m/\ds_m)^{\du_m}=\Span\{X_{x^Iy^J}:|I|+|J|=3\}$.
2. $(\K_m/\ds_m)^{\du_m}$ has total weight $\oh$.
3. $(\K_m/\ds_m)^{\du_m}$ is equivalent to $\Sym^3\bC^{2\ell}$ under the action of $\dl_m^{\rms}$.
Infinitesimal characters {#ICS}
------------------------
We now turn to the s of the fine symbol modules. A module of a complex semisimple Lie algebra $\dg$ is said to have an [**]{} if the center $\dZ(\dg)$ of the universal enveloping algebra $\dU(\dg)$ acts on it by scalars. In this case, the is the resulting homomorphism from $\dZ(\dg)$ to $\bC$. By Schur’s lemma, all modules have s.
Suppose that a Cartan subalgebra and a positive root system of $\dg$ are fixed. Let $\rho$ be the half-sum of the positive roots. If $V$ and $V'$ are two lowest weight modules of $\dg$ with lowest weights $\nu$ and $\nu'$, respectively, and both modules have s, then it is a consequence of the Harish-Chandra homomorphism that their s are the same $\nu - \rho$ and $\nu' - \rho$ lie in the same orbit of the Weyl group of $\dg$.
\[FS eqvs\] Under the action of $\ds_m$, the fine symbol modules have s. Fix two fine symbol modules $\FS^{k,d}_\delta$ and $\FS^{k',d'}_\delta$ such that either $k' < k$, or $k' = k$ and $d' < d$. They have the same s at least one of the following four conditions holds: $$\begin{array}{lrclrcl}
\mbox{\em (i)\ \ } &
k' &=& k, &
d' &=& 2\delta(\ell+1) - 1+2k - d. \\[6pt]
\mbox{\em (ii)\ \ } &
k' &=& 2(\delta - 1) (\ell + 1) + 1 - k, &
d' &=& 2(\delta - 1) (\ell + 1) + 1 - 2k+d. \\[6pt]
\mbox{\em (iii)\ \ } &
k' &=& (2\delta-1) (\ell+1) +k- d, &
d' &=& 2(2\delta-1) (\ell+1) - d. \\[6pt]
\mbox{\em (iv)\ \ } &
k' &=& d-k - \ell, &
d' &=& 2(\delta-1) (\ell+1) +1 - 2k+d.
\end{array}$$ Cases (i), (ii), (iii), and (iv) cannot occur unless $2\delta (\ell+1)$ is in $$2 + \bN, \quad 2(\ell + 1) + \bN,
\quad \ell + 2 + \bN, \quad 2\ell + 1 + \bN,$$ respectively. Moreover, if $\ell = 0$, then Case (iv) cannot occur unless $2\delta \in 2 + \bN$. Therefore if $\delta$ is not contact-resonant (see Section \[Definitions\]), then all of the fine symbol modules of $\D_{\lambda,\mu}$ have distinct s.
It is well-known that Verma modules have s, and so their restricted duals do also. Therefore by Lemma \[FS poly\], $\dZ(\ds_m)$ acts by scalars on $\FS^{k,d}_\delta(\rmp)$. By a density argument, it acts by the same scalars on $\FS^{k,d}_\delta(\bR^m)$. Thus the fine symbol modules have s.
Recall from Lemma \[FSpolyOne\] the lowest weight $\nu^{k,d}_\delta$ of $\FS^{k,d}_\delta(\rmp)$, and from Lemma \[roots\] the half-sum $\rho(\ds_m)$ of the positive roots: $$\rho(\ds_m) = \sum_{i=0}^\ell (\ell+1-i) e_i.$$ Recall also that the Weyl group $W(\ds_m) = S_{\ell+1} \ltimes \bZ_2^{\ell+1}$ acts by permutations and sign changes on the $e_i$.
As stated above, the Harish-Chandra homomorphism shows that the s of $\FS^{k,d}_\delta$ and $\FS^{k',d'}_\delta$ are the same there is an element $w$ of $W(\ds_m)$ such that $$w \bigl(\nu^{k,d}_\delta - \rho(\ds_m)\bigr) =
\nu^{k',d'}_\delta - \rho(\ds_m).$$ By Lemma \[FSpolyOne\], such a $w$ exists the two sets $$\begin{aligned}
&& \bigl\{ |(2\delta-1)(\ell+1)-d |,\ \ell + 2k - d,\
\ell - 1,\ \ell - 2,\ \ldots, 2,\ 1 \bigr\}, \\[6pt]
&& \bigl\{ |(2\delta-1)(\ell+1)-d' |,\ \ell + 2k' - d',\
\ell - 1,\ \ell - 2,\ \ldots, 2,\ 1 \bigr\}\end{aligned}$$ are equal. Since $\ell + 2k - d > \ell - 1 > \cdots > 1 > 0$, this can occur only in the following ways.
First, $|(2\delta-1)(\ell+1)-d|$ and $|(2\delta-1)(\ell+1)-d') |$ can be equal. In this case $\ell + 2k - d$ and $\ell + 2k' - d'$ must also be equal. Since we are assuming that $(k,d)$ and $(k',d')$ are distinct, this leads to Case (iii).
Second, we could have $$|(2\delta-1)(\ell+1)-d| = \ell + 2k' - d', \qquad
|(2\delta-1)(\ell+1)-d' | = \ell + 2k - d.$$ This can occur in three ways, depending on the signs of the arguments of the absolute values. If both are negative we arrive at Case (i), if both are positive we are in Case (ii), and if they are different we obtain Case (iv).
Finally, if $|(2\delta-1)(\ell+1)-d |$ is equal to one of $\ell - 1, \ldots, 1$, say $i$, then again $\ell + 2k - d$ and $\ell + 2k' - d'$ must be equal and $|(2\delta-1)(\ell+1)-d|$ must also be $i$. Therefore here we are still in Case (iii), albeit with a different $w$. $\Box$
Projective quantization {#PQ}
=======================
Recall from Section \[SMs\] the total symbol modules $\S_\delta$ and the total fine symbol modules $\FS_\delta$. In this section we study quantizations of $\FS_\delta$ invariant under the projective subalgebra $\ds_m$ of $\crm$. We begin in Section \[S PQ\] with a review of quantizations of $\S_\delta$ invariant under the projective subalgebra $\da_m$ of $\vrm$, as these quantizations are a component of the fine projective quantizations.
Projective quantization of symbols {#S PQ}
----------------------------------
A [*quantization of*]{} $\S_\delta$ is defined to be a linear bijection $Q$ from $\S_\delta$ to $\D_{\lambda,\mu}$ which [*preserves degree*]{} and is the [*identity on symbols.*]{} By this we mean that it carries $\S^k_\delta$ into $\D^k_{\lambda,\mu}$, and its restriction to $\S^k_\delta$ is a right-inverse of the principal symbol map $\sigma^k_{\lambda, \mu}$: $$\sigma^k_{\lambda, \mu} \circ Q|_{\S^k_\delta}:
\S^k_\delta \to \S^k_\delta \mbox{\rm\ \ is the identity map.}$$ Suppose that $\dg$ is any Lie subalgebra of $\vrm$. A quantization is said to be a [*$\dg$-equivariant quantization*]{} (or simply a [*$\dg$-quantization*]{}) if it intertwines the two $\dg$-actions $L^\S_\delta|_\dg$ and $L_{\lambda,\mu}|_\dg$.
The more vector fields a quantization is invariant with respect to, the more useful it is. Recall from Section \[SlCoord\] the projective subalgebra $\da_m$ and the affine subalgebra $\db_m$ of $\vrm$. At one extreme, one might ask for a $\vrm$-quantization of $\S_\delta$. However, there is no such map: $\S_\delta$ and $\D_{\lambda,\mu}$ are not $\vrm$-equivalent for any $(\lambda, \mu)$. At the opposite extreme, $\db_m$-quantizations are easy to find, not unique, and not very useful. The critical intermediate case is afforded by $\da_m$, because it is a simple finite dimensional maximal subalgebra of $\vrm$.
We say that $\delta$ is [*projectively resonant*]{} if it lies in the set $$\left\{1 + \frac{n}{m+1}\;\Big|\; n\in\bN\right\}.$$
The following theorem was proven in [@CMZ97] for $m=1$, in [@LO99] for arbitrary $m$ at $p=0$, and in general in [@Le00].
\[PQa\] For $\delta$ not projectively resonant, there exists a unique $\da_m$-quantization $$\PQ^{\da_m}_{\lambda,\mu}: \S_\delta \to \D_{\lambda,\mu}.$$
Theorem \[PQa\] may be proven using only the eigenvalues of the Casimir operator; the full s are not needed. This is the approach taken in [@Le00]. The explicit formula for the projective quantization $\PQ^{\da_m}_{\lambda,\mu}$ was given in [@CMZ97] for $m=1$, in [@LO99] for arbitrary $m$ at $p=0$, and in general in [@DO01]; see Section \[SlExp\] below.
Projective quantization of fine symbols {#FS PQ}
---------------------------------------
By analogy with quantizations of $\S_\delta$, we define a [*quantization of*]{} $\FS_\delta$, sometimes called a [*fine quantization,*]{} to be a linear bijection $Q$ from $\FS_\delta$ to $\D_{\lambda,\mu}$, carrying $\FS^{k,d}_\delta$ into $\D^{k,d}_{\lambda,\mu}$, which is the identity on fine symbols in the sense that its restriction to $\FS^{k,d}_\delta$ is a right-inverse for the fine symbol map $\fs^{k,d}_{\lambda,\mu}$: $$\fs^{k,d}_{\lambda, \mu} \circ\, Q|_{\FS^{k,d}_\delta}:
\FS^{k,d}_\delta \to \FS^{k,d}_\delta
\mbox{\rm\ \ is the identity map.}$$ For any Lie subalgebra $\dg$ of $\crm$, we say that a fine quantization is a [*fine $\dg$-equivariant quantization*]{} if it intertwines $L^{\FS}_\delta|_\dg$ and $L_{\lambda,\mu}|_\dg$.
The picture for quantizations of $\FS_\delta$ is similar to that for quantizations of $\S_\delta$: there is no fine $\crm$-quantization, there are many fine $\dt_m$-quantizations, and for most $\delta$ there is a unique fine $\ds_m$-quantization. The following theorem makes this precise. It may be proven by combining the results of [@DO01] and [@FMP08]. We understand it as a corollary of Proposition \[FS eqvs\]. As remarked before that proposition, it cannot be proven using the Casimir operator of $\ds_m$ alone; the full s of the fine symbol modules are required.
\[PQs\] For $\delta$ not contact-resonant (see Section \[Definitions\]), there exists a unique fine $\ds_m$-quantization $$\PQ^{\ds_m}_{\lambda,\mu}: \FS_\delta \to \D_{\lambda,\mu}.$$
Write $\chi^{k,d}_\delta$ for the $\ds_m$- of $\FS^{k,d}_\delta$, and $(\D_{\lambda,\mu})_{\chi^{k,d}_\delta}$ for the subspace of $\D_{\lambda,\mu}$ on which the center $\dZ(\ds_m)$ of $\dU(\ds_m)$ acts by $\chi^{k,d}_\delta$. The $\chi^{k,d}_\delta$ are distinct by Proposition \[FS eqvs\], so we have the $\ds_m$-decomposition $$\D_{\lambda,\mu} = \bigoplus_{k,d}
(\D_{\lambda,\mu})_{\chi^{k,d}_\delta}.$$
The fine symbol map $\fs^{k,d}_{\lambda,\mu}$ restricts to the unique fine symbol-preserving $\ds_m$-equivalence from $(\D_{\lambda,\mu})_{\chi^{k,d}_\delta}$ to $\FS^{k,d}_\delta$. The fine $\ds_m$-quantization $\PQ^{\ds_m}_{\lambda,\mu}$ is the direct sum of the inverses of these restrictions. $\Box$
The explicit formula for $\PQ^{\ds_m}_{\lambda,\mu}$ will be given in Section \[SpExp\].
Lowest weight calculations {#CRCs}
==========================
In this section we prove Theorem \[subsymbol\] and Theorem \[bifilt\]. The proofs rely on lowest weight calculations in modules of homomorphisms between the fine symbol spaces.
The structure of $\Hom_{\du_m}(\FS^{k,d}_\delta, \FS^{k',d'}_\delta)$
---------------------------------------------------------------------
We begin with a description of the total weight spaces of $\Hom_{\du_m}(\FS^{k,d}_\delta, \FS^{k',d'}_\delta)$ which will be needed in both proofs. Recall that for $\nu\in\dh_m^*$, $V_\nu$ denotes the $\nu$-weight space of any $\dh_m$-module $V$. We will also use the following notation: for $w\in\bC$, $V_{(w)}$ denotes the $w$-total weight space of $V$. Note that if $\nu = \sum_0^\ell \nu_i e_i$, then $V_\nu \subseteq V_{(\frac{\nu_0}{2})}$.
We will abbreviate $\bC[\pz, A_1, B_1 \ldots, A_\ell, B_\ell]$ by $\bC[\pz, A, B]$.
\[Homu\]
1. The total weight space $\End_{\du_m}\bigl(\cirm\bigr)_{(w)}$ is zero unless $w\in -\oh\bN$, when it is $\bC[\pz, A, B]_{(w)}$.
2. The total weight space $\Hom_{\du_m}(\FS^{k,d}_\delta, \FS^{k',d'}_\delta)_{(w)}$ is zero unless $w \in \oh(d-d') - \oh \bN$, when it is $$\begin{aligned}
\Span_\bC\bigl\{ \zeta^{d'-k'} \alpha^{I'} \beta^{J'}
\partial_\zeta^{d-k}\; \partial_\alpha^I\; \partial_\beta^J&:&
|I|+|J| = 2k-d,\ |I'|+|J'| = 2k'-d' \bigr\} \\[6pt]
& \otimes & \bC[\partial_z,A,B]_{\left(w-\frac{1}{2}(d-d') \right)}.\end{aligned}$$
Let $T:C^\infty(\bR^m)\to{}C^\infty(\bR^m)$ be of total weight $w$. Use Lemma \[EuLem\] to see that $T$ maps polynomials to polynomials, and verify that $\End\bigl(\Poly(\bR^m) \bigr)$ is $\bC[x,y,z]\,[[\partial_z,A,B]]$. Hence $$\End\bigl(\Poly(\bR^m)\bigr)_{(w)}=\bC[x,y,z]\,[\partial_z,A,B]_{(w)}.$$ Since $\partial_z$, $A$, and $B$ commute with $\du_m$, the $\du_m$-equivariant maps in this space are precisely those independent of $x$, $y$, and $z$. This proves (i).
Under $\du_m$, $\FS^{k,d}_\delta$ is equivalent to $(\FS^{k,d}_\delta)^{\du_m}\otimes{}C^\infty(\bR^m)$, so $\Hom_{\du_m}\bigl(\FS^{k,d}_\delta,\FS^{k',d'}_\delta\bigr)$ is $$\Hom_{\bC}\bigl((\FS^{k,d}_\delta)^{\du_m},(\FS^{k',d'}_\delta)^{\du_m}\bigr)\otimes
\End_{\du_m} \bigl(C^\infty(\bR^m) \bigr).$$ Thus (ii) follows from (i) and the fact that $\Hom_{\bC} \bigl((\FS^{k,d}_\delta)^{\du_m},(\FS^{k',d'}_\delta)^{\du_m} \bigr)$ is $$\Span_\bC\bigl\{ \zeta^{d'-k'} \alpha^{I'} \beta^{J'}
\partial_\zeta^{d-k}\; \partial_\alpha^I\; \partial_\beta^J\;:\;
|I|+|J| = 2k-d,\ |I'|+|J'| = 2k'-d' \bigr\} .
\qquad
\Box$$
We will also need the structure of $\Hom_{\du_m} \bigl(\FS^{k,d}_\delta, \FS^{k',d'}_\delta \bigr)_{(w)}$ as a module of the subalgebra $\dl_m^{\rms}$, which recall is isomorphic to $\dsp_{2\ell}$.
\[SymLem\] Suppose that $w$ is in $\oh(d-d') - \oh \bN$. Then under the action of $\dl_m^{\rms}$, $$\Hom_{\du_m} \bigl(\FS^{k,d}_\delta, \FS^{k',d'}_\delta \bigr)_{(w)}
\th \cong \th
\Sym^{2k-d}\bC^{2\ell}\,\otimes\,\Sym^{2k'-d'}\bC^{2\ell}\,\otimes\,
\biggl(\th
\bigoplus_{r\geq0}\Sym^{d-d'-2r-2w}\bC^{2\ell}
\biggr),$$ where the direct sum is only over non-negative exponents.
Consider the explicit basis of $\Hom_{\du_m}\bigl(\FS^{k,d}_\delta, \FS^{k',d'}_\delta\bigr)_{(w)}$ given in Lemma \[Homu\](ii). We have the following $\dl_m^{\rms}$-equivalences: $$\begin{array}{rcl}
\Span_\bC
\bigl\{ \alpha^{I'} \beta^{J'}\;:\;
|I'|+|J'| = 2k'-d' \bigr\} &\cong& \Sym^{2k'-d'}\bC^{2\ell},\\[6pt]
\Span_\bC\bigl\{
\partial_\alpha^I\; \partial_\beta^J\;:\;
|I|+|J| = 2k-d\bigr\} &\cong& \Sym^{2k-d}\bC^{2\ell},
\end{array}$$ and for $v \in -\oh\bN$, $\bC[\partial_z,A,B]_{(v)} \cong
\bigoplus_{r\geq0}\Sym^{-2v-2r}\bC^{2\ell}$. Since $\zeta$, $\partial_\zeta$, and $\partial_z$ are $\dl_m^{\rms}$-invariant, the result follows. $\Box$
Proof of Theorem \[subsymbol\]
------------------------------
Assume temporarily that $M$ is $\bR^m$, equipped with the standard contact structure. Since $\delta$ is not contact-resonant, Theorem \[PQs\] shows that there is a unique $\ds_m$-equivariant map from $\D^k_{\l,\mu}$ to $\FS^{k-1,\,2(k-1)}_\delta$, namely $$\label{ssEQ}
\ss^k_{\l,\mu} := \pi_{k-1,\,2(k-1)} \circ (\PQ^{\ds_m}_{\lambda,\mu})^{-1},$$ where $\pi_{j,d}$ denotes the canonical projection from $\FS_\delta$ to $\FS^{j,d}_\delta$. Our task is to prove that this map is in fact $\crm$-equivariant.
We use the projective quantization to pull the action $L_{\lambda,\mu}$ of $\crm$ on $\D_{\l,\mu}$ back to an action $\L_{\lambda, \mu}$ of $\crm$ on $\FS_\delta$: $$\L_{\l,\mu}(X_f)\ :=\ (\PQ^{\ds_m}_{\lambda,\mu})^{-1} \circ
L_{\l,\mu}(X_f) \circ \PQ^{\ds_m}_{\lambda,\mu}.$$ The statement that $\ss^k_{\l,\mu}$ is $\crm$-equivariant is equivalent to the statement that $$\pi_{k-1,\,2(k-1)}: \bigoplus_{0\leq{}j\leq{}k} \bigoplus_{j\leq{}d\leq{}2j}
\FS^{j,d}_\delta \to\FS^{k-1,\,2(k-1)}_\delta$$ intertwines the $\crm$-actions $ \L_{\l,\mu}$ and $L^{\FS}_\delta$.
We may regard $ \L_{\l,\mu}$ as a block matrix with entries $$\L_{\l,\mu}^{(j,d),(j',d')}(X_f):
\FS^{j,d}_\delta \to \FS^{j',d'}_\delta.$$ This matrix is triangular with respect to the dictionary order on $(j,d)$, and the diagonal entries are simply the usual actions on fine symbols: $\L_{\l,\mu}^{(j,d),(j,d)}$ is $L^{\FS}_\delta$ restricted to $\FS^{j,d}_\delta$. As a result of the $\ds_m$-equivariance of $\PQ^{\ds_m}_{\l,\mu}$, the off-diagonal entries $\L_{\l,\mu}^{(j,d),(j',d')}$ are $\ds_m$-relative 1-cochains of $\crm$: they vanish on $\ds_m$ and are $\ds_m$-equivariant maps from $\crm/\ds_m$ to $\Hom(\FS^{j,d}_\delta, \FS^{j',d'}_\delta)$.
One finds that $\pi_{k-1,\,2(k-1)}\circ{}\L_{\l,\mu}$ restricted to $\bigoplus_{0\leq{}j\leq{}k}\bigoplus_{j\leq{}d\leq{}2j}\FS^{j,d}_\delta$ is $$L_\delta^{\FS}\circ\pi_{k-1,\,2(k-1)}+
\sum_{d=k}^{2k}\L_{\l,\mu}^{(k,d),(k-1,\,2(k-1))}\circ\pi_{k,d}.$$ Therefore it suffices to prove that entries $\L_{\l,\mu}^{(k,d),(k-1,\,2(k-1))}$ are zero for all $d$. The following lemma is the key to the situation.
\[NoExist\] Let $V$ be a module of $\crm$ such that the space $V_{(\frac{1}{2})}^{\du_m}$ of $\du_m$-invariants in $V$ of total weight $\oh$ contains no copies of $\Sym^3\bC^{2\ell}$ under the action of $\dl_m^{\rms}$. Then the space $C^1\bigl(\K(\bR^m), \ds_m; V\bigr)$ of $\ds_m$-relative 1-cochains of $\crm$ with values in $V$ is zero.
Apply Corollary \[WeightK\]. $\Box$
In order to apply this lemma, we must prove that $\Hom_{\du_m} \bigl(\FS^{k,d}_\delta, \FS^{k-1,\,2(k-1)}_\delta \bigr)_{(\frac{1}{2})}$ contains no copies of $\Sym^3\bC^{2\ell}$ under $\dl_m^{\rms}$ for $k \le d \le 2k$. When $d=2k$, by Lemma \[Homu\] we obtain $$\Hom_{\du_m}\bigl(\FS^{k,2k}_\delta, \FS^{k-1,\,2(k-1)}_\delta\bigr)_{(\frac{1}{2})}=
\zeta^{k-1}\partial_\zeta^k\,\Span\{A,B\},$$ which is a copy of $\bC^{2\ell}$. When $d=2k-1$, we obtain $$\Hom_{\du_m} \bigl(\FS^{k,2k-1}_\delta, \FS^{k-1,\,2(k-1)}_\delta \bigr)_{(\frac{1}{2})}=
\zeta^{k-1}\partial_\zeta^{k-1}\,\Span\{\partial_\alpha,\partial_\beta\},$$ which is again a copy of $\bC^{2\ell}$. Finally, when $d<2k-1$, we obtain zero.
This completes the proof of Theorem \[subsymbol\] when $M=\bR^m$. If $M$ is an arbitrary contact manifold, Darboux’s theorem implies the existence of an atlas of local charts on $M$ that are diffeomorphic to the standard contact structure on $\bR^m$. We have just proved that for every chart $U$, there is a unique (locally defined) map $\ss^k_{\l,\mu}(U)$ equivariant with respect to $\K(U)$. If $U'$ is another chart, uniqueness implies that the maps $\ss^k_{\l,\mu}(U)$ and $\ss^k_{\l,\mu}(U')$ coincide on $U\cap{}U'$. Therefore $\ss^k_{\l,\mu}(M)$ is well-defined on all of $M$ and obviously commutes with $\K(M)$.
Proof of Theorem \[bifilt\] {#C proof}
---------------------------
As in the proof of Theorem \[subsymbol\], take $M=\bR^m$ and consider the action $\L_{\lambda,\mu}$ of $\crm$ on $\FS_\delta$. We first prove that certain of its matrix elements $\L_{\l,\mu}^{(k,d),(k',d')}$ vanish.
\[k-1\] In both of the following cases, the space of $\ds_m$-relative 1-cochains of $\crm$ with coefficients in $\Hom \bigl(\FS^{k,d}_\delta, \FS^{k',d'}_\delta \bigr)$ is zero:
1. $k' = k-1$ and $d' \ge d$.
2. $k' < k-1$ and $d' \ge d - (k-k') + 2$.
By Lemma \[NoExist\], we must prove that under $\dl_m^{\rms}$ the space $\Hom_{\du_m}(\FS^{k,d}_\delta, \FS^{k',d'}_\delta)_{(\frac{1}{2})}$ contains no copies of $\Sym^3\bC^{2\ell}$ in either Case (i) or Case (ii). By Lemma \[Homu\], this space is zero unless $d-d' \in \bZ^+$. In particular, it is zero in Case (i). In Case (ii) with $d-d' \in \bZ^+$, Lemma \[SymLem\] shows that under $\dl_m^{\rms}$ it is equivalent to $$\label{PRV}
\Sym^{2k-d} \bC^{2\ell}\, \otimes\,\Sym^{2k'-d'} \bC^{2\ell}\, \otimes\,
\biggl(\th \bigoplus_{r\geq0}\Sym^{d-d'-2r-1}\bC^{2\ell} \biggr).$$
It is well-known (see, [@Va84]) that the largest irreducible component of $\Sym^r\bC^{2\ell} \otimes{} \Sym^s\bC^{2\ell}$ is $\Sym^{r+s}\bC^{2\ell}$, and for $r\geq{}s$, its smallest irreducible component is its Parthasarathy - Ranga Rao - Varadarajan submodule, a copy of $\Sym^{r-s}\bC^{2\ell}$. Therefore the maximal component of $$\Sym^{2k'-d'}\bC^{2\ell}\,\otimes\,
\biggl(\th \bigoplus_{r\geq0}\Sym^{d-d'-2r-1}\bC^{2\ell} \biggr)$$ is $\Sym^{d-2d'+2k'-1}\bC^{2\ell}$. In Case (ii) we have $k'-d'<k-d$, so the smallest irreducible component of (\[PRV\]) is $\Sym^{2(d'-d+k-k')+1}\bC^{2\ell}$, which is larger than $\Sym^{3}\bC^{2\ell}$. $\Box$
This lemma implies that $\L_{\l,\mu}^{(k,d),(k',d')}=0$ under the conditions of Cases (i) and (ii). Therefore the space $\D^{(b)}_{\l,\mu}(\bR^m)$ defined by $$\D^{(b)}_{\l,\mu}(\bR^m) \th:=\th
\PQ^{\ds_m}_{\l,\mu} \biggl(\th \bigoplus_{2d-k\leq{}b}\FS^{k,d}_\delta \biggr)$$ is invariant under the action $L_{\l,\mu}$ of $\K(\bR^m)$. Since under $\ds_m$ there is unique copy of $\FS^{k,d}_\delta$ in $\D_{\l,\mu}$ for all $(k,d)$, $\D^{(b)}_{\l,\mu}$ is the unique subspace of $\D_{\l,\mu}$ whose graded module is as in Theorem \[bifilt\]. This completes the proof for $M = \bR^m$. For $M$ arbitrary, local existence and uniqueness allows us to conclude global existence and uniqueness of $\D^{(b)}_{\l,\mu}(M)$ as in the proof of Theorem \[subsymbol\].
Explicit formulas {#T}
=================
The affine invariants
---------------------
We begin by proving that the following maps between fine symbol modules are equivariant with respect to the affine subalgebra $\dt_m$:
The [*contact divergence*]{} is the map $$\DivC := \pz \partial_\zeta:
\FS^{k, d}_\delta \to \FS^{k-1, d-2}_\delta.$$ The [*tangential divergence*]{} is the map $$\DivT := A_r \partial_{\alpha_r} + B_r \partial_{\beta_r}:
\FS^{k, d}_\delta \to \FS^{k-1, d-1}_\delta.$$ Finally, define $$\De := (\alpha_r B_r - \beta_r A_r) \partial_\zeta:
\FS^{k, d}_\delta \to \FS^{k, d-1}_\delta.$$
\[divergences\]
1. $\DivC$, $\DivT$, and $\De$ are all $\dt_m$-equivariant.
2. The contact divergence $\DivC$ commutes with both $\DivT$ and $\De$, and $$[\DivT, \De] = (\ell + \E_{\alpha \beta}) \DivC.$$
3. Regarded as a map from $\S^k_\delta$ to $\S^{k-1}_\delta$, $\DivT + \DivC$ is the full divergence $\Div$.
The statement follows from Lemma \[LXs\] and short computations. $\Box$
It follows from the first fundamental theorem of invariant theory for $\dsp_{2\ell}$ that the associative algebra generated by the operators $$\DivC,
\qquad
\DivT,
\qquad
\De,
\qquad
\E_{\alpha \beta},
\qquad
\E_{\zeta}$$ coincides with the algebra $\End_{\dt_m}(\FS_\delta)$ of all affine invariants: see [@FMP07].
The $\dsl_{m+1}$-equivariant quantization of $\S_\delta$ {#SlExp}
--------------------------------------------------------
Continuing the discussion of Section \[S PQ\], we now give an explicit formula for the quantization map $\PQ^{\da_m}_{\lambda,\mu}$ of Theorem \[PQa\]. This formula is part of the explicit formula for the fine projective quantization in the contact setting.
We begin with the standard symbol calculus on $\S_\delta$. As in Section \[SlCoord\], fix any coordinates $u_1, \ldots, u_m$ on $\bR^m$. Write $\xi_i$ for the symbol of the $\partial_{u_i}$. Then $$\S_\delta = \Span_{\cirm} \bigl\{ \xi^I: I \in \bN^m \bigr\},
\qquad
\D_{\lambda,\mu} = \Span_{\cirm} \bigl\{ \partial_u^I: I \in \bN^m \bigr\}.$$ As noted in Section \[SMs\], these spaces carry the $\vrm$-actions $L^\S_\delta$ and $L_{\lambda,\mu}$, respectively.
One of the simplest quantizations is the [*normal order quantization:*]{} $$\NQ: \S_\delta \to \D_{\lambda,\mu}, \qquad
\NQ \bigl( f_I(u) \xi^I \bigr)
:= f_I(u) \partial_u^I.$$ In the literature, $\S_\delta$ and $\D_{\lambda,\mu}$ are frequently identified via $\NQ$, which then does not appear explicitly in the formulas. Although $\NQ$ is not an $\da_m$-equivariant quantization, it does turn out to be a $\db_m$-quantization. The first step in computing $\PQ^{\da_m}_{\lambda,\mu}$ is to find explicit formulas for $L^\S_\delta$ and $L_{\lambda,\mu}$. In fact, one computes the pull-back of $L_{\lambda,\mu}$ to $\S_\delta$ via $\NQ$, that is, $$L_{\lambda,\mu}^{\NQ^{-1}} (X)\ :=\
\NQ^{-1} \circ L_{\lambda,\mu}(X) \circ \NQ.$$ Towards this end, note that any on $\S_\delta$ may be written as a $\cirm$-linear combination of monomials $\partial_u^I \xi^J \partial_\xi^K$.
Given any vector field $X = \sum_1^m X_i \partial_{u_i}$, it is straightforward to obtain $$L_{\lambda,\mu}^{\NQ^{-1}} (X) = X + \delta \Div(X)
- \sum_{|I| > 0} \ts\frac{1}{I!}
\Bigl( \partial_u^I \bigl[ \lambda \Div(X)
+ {\ts\sum_j X_j \xi_j} \bigr]\Bigr) \partial_\xi^I.$$ The action $L^\S_\delta$ is simply the part of $L_{\lambda,\mu}^{\NQ^{-1}}$ which preserves $\xi$-degree: $$L^\S_\delta(X) = X + \delta \Div(X)
- \sum_{i,j=1}^m (\partial_{u_i} X_j) \xi_j \partial_{\xi_i}.$$ Observe that these two formulas are the same $X \in \db_m$. Thus as claimed, $\NQ$ is a $\db_m$-quantization.
As usual, let $\E_\xi$ denote the $\xi$-Euler operator. The full divergence operator is $$\Div = \sum _{1\leq{}i\leq{}m} \partial_{u_i} \partial_{\xi_i}
: \S^k_\delta \to \S^{k-1}_\delta.$$ Bear in mind that $\E_\xi$ and $\Div$ do not commute: $[\E_\xi, \Div] = -\Div$. The theorem is as follows.
\[PQa formula\] [@LO99; @DO01] For $\delta$ non-resonant, $$\PQ^{\da_m}_{\lambda,\mu} = \NQ \circ
\sum_{s=0}^\infty {\ts \frac{1}{s!}} \Div^s
\circ {\E_\xi + \lambda(m+1) - 1 \choose s}
{2\E_\xi - \delta(m+1) + m - 1 \choose s}^{-1}.$$
It is worth mentioning that this formula can be understood as a (non-commutative) hypergeometric function: see [@DO01].
The $\dsp_{2(\ell+1)}$-equivariant quantization of $\FS_\delta$ {#SpExp}
---------------------------------------------------------------
We now proceed to derive an explicit formula for $\PQ^{\ds_m}_{\lambda,\mu}$. We begin by defining a map $\SPQ^{\ds_m}_\delta$ from $\FS_\delta$ to $\S_\delta$: $$\label{SPQ formula}
\SPQ^{\ds_m}_\delta := \sum_{s=0}^\infty { \frac{2^s}{(s!)^2}}
\Delta^s \circ {2\E_\zeta - 2\delta (\ell + 1) \choose s}^{-1}.$$ This formula is well-defined provided that $\delta$ is not contact-resonant. Recall from (\[symbol bases\]) that we are abusing notation and using the same bases for $\FS_\delta$ and $\S_\delta$ (this is analogous to regarding the normal order quantization $\NQ$ as the identity). Therefore we may and do regard $\SPQ^{\ds_m}_\delta$ as a map from $\FS_\delta$ to $\S_\delta$.
\[SPThExp\] The fine $\ds_m$-equivariant quantization of Theorem \[PQs\] is $$\PQ_{\l,\mu}^{\ds_m}=\PQ_{\l,\mu}^{\da_m}\circ\SPQ^{\ds_m}_\delta.$$
Observe that it suffices to prove that $\SPQ^{\ds_m}_\delta$ intertwines the restrictions of the actions $L^{\FS}_\delta$ and $L^\S_\delta$ to $\ds_m$. By Lemma \[divergences\], $\SPQ^{\ds_m}_\delta$ intertwines the $\dt_m$-actions. Restricted to $\FS^{k,d}_\delta$, it is of the form $\sum_0^\infty \Delta^s C_s$ for some constants $C_s$. One obtains (\[SPQ formula\]) by deriving a recursion relation for these constants. Since $X_{x_i z}$ generates $\ds_m$ under $\dt_m$ for any $i$, we need only impose the condition $$L^\S_\delta (X_{x_i z}) \circ \SPQ^{\ds_m}_\delta =
\SPQ^{\ds_m}_\delta \circ L^{\FS}_\delta (X_{x_i z}).$$
By Lemma \[LXs\], $L^\S_\delta (X_{x_i z}) - L^{\FS}_\delta(X_{x_i z})$ is $\beta_i \partial_\zeta$, which commutes with $\Delta$. Therefore we find that $\SPQ^{\ds_m}_\delta\circ \beta_i \partial_\zeta$ and $\bigl[\SPQ^{\ds_m}_\delta, L^{\FS}_\delta(X_{x_i z}) \bigr]$ must be equal, $$0 = \ts \sum_s \bigl( \Delta^s \circ \beta_i \partial_\zeta
- \bigl[\De^s, L^{\FS}_\delta(X_{x_i z}) \bigr] \bigr) C_s.$$ Using the same lemma, deduce the following commutator: $$[L^{\FS}_\delta(X_{x_i z}), \Delta^s]= -s \Delta^{s-1} \beta_i \partial_\zeta
\bigl( \E_\zeta - \delta(\ell+1) - \oh(s-1) \bigr),$$ This gives the recursion relation $$C_{s-1} = \oh s \bigl( 2c - 2\delta (\ell+1) - (s-1) \bigr) C_s.$$ Since $C_0 = 1$, the theorem follows. $\Box$
The subsymbol {#SubsExp}
-------------
In order to give an explicit local formula for $\ss^k_{\l,\mu}$, let us fix a system of Darboux coordinates as in Section \[DarbSect\]. An arbitrary differential operator $T$ of order $\leq{}k$ may be expressed as $$T=\sum_{c+|I|+|J|\leq{}k}T_{c,I,J}\,\partial_z^c\partial_x^I\partial_y^J,$$ where the $T_{c,I,J}$ are smooth functions. As usual, we replace $\partial_z,\partial_x,\partial_y$ by their symbols $\xi_z,\xi_x,\xi_y$, respectively. This amounts to replacing $T$ by $$\NQ^{-1}(T)=
\sum_{c+|I|+|J|\leq{}k}T_{c,I,J}\,\xi_z^c\xi_x^I\xi_y^J.$$
The formula has two ingredients: the full divergence $\Div$, and the projection $\pi_{k-1,2(k-1)}:\S^{k-1}_\delta\to\FS^{k-1,2(k-1)}_\delta$. The full divergence $\Div(\NQ^{-1}(T))$ is $$\sum_{c+|I|+|J|\leq{}k}\biggl(
c\partial_z(T_{c,I,J})\,\xi_z^{c-1}\xi_x^I\xi_y^J+
\sum_{s=1}^\ell{}\Bigl(
I_s\partial_{x_s}(T_{c,I,J})\,\xi_z^c\xi_x^{I-e_s}\xi_y^J+
J_s\partial_{y_s}(T_{c,I,J})\,\xi_z^c\xi_x^{I}\xi_y^{J-e_s}
\Bigr)\biggr).$$ In the $(\alpha, \beta, \zeta)$-coordinates on $\S^{k-1}_\delta$ given in Section \[SMs\], the projection $\pi_{k-1,2(k-1)}$ simply gives the $\zeta^{k-1}$ term. In Darboux coordinates, $$\pi_{k-1,2(k-1)}\bigl(T_{c,I,J}\,\xi_z^{k-1-|I|-|J|}\xi_x^I\xi_y^J\bigr)=
(-1)^{|I|}\bigl(\oh\bigr)^{|I|+|J|}y^{I}x^{J}
T_{c,I,J}\zeta^{k-1},$$ because by (\[abxi\]), $\xi_{x_i}=\alpha_i-\oh{}y_i\zeta$, $\xi_{y_i}=-\beta_i+\oh{}x_i\zeta$, and $\xi_z=\zeta$.
\[subsymbol formula\] In Darboux coordinates, the subsymbol $\ss^k_{\l,\mu}(T)$ is given by $$\ss^k_{\l,\mu}(T) =\pi_{k-1,2(k-1)}\circ\Bigl(
1-\frac{(k-1)+2\l(\ell+1)}{2(k-1)-2(\delta-1)(\ell+1)}\Div
\Bigr)\circ{}\NQ^{-1}(T).$$
Using (\[ssEQ\]) and Theorem \[SPThExp\], we obtain $\ss^k_{\l,\mu}(T) =\pi_{k-1,2(k-1)}\circ
\bigl(\SPQ^{\ds_m}_\delta\bigr)^{-1}
\circ
\bigl(\PQ_{\l, \mu}^{\da_m}\bigr)^{-1}$. Since $\FS^{k-1,2(k-1)}_\delta$ is the fine symbol module in $\S^{k-1}_\delta$ of the highest contact order, (\[SPQ formula\]) gives $$\pi_{k-1,2(k-1)}\circ
\bigl(\SPQ^{\ds_m}_\delta\bigr)^{-1}=\pi_{k-1,2(k-1)}.$$ To calculate the $\S^{k-1}_\delta$-component of $\bigl(\PQ_{\l, \mu}^{\da_m}\bigr)^{-1}$, we use the formula of Theorem \[PQa formula\]. $\Box$
Note that the formula for $\ss^k_{\l,\mu}$ is well-defined for all but one contact-resonant value of $\delta$, the value $\delta=\frac{\ell+k}{\ell+1}$. By continuity, it retains $\K_m$-equivariance whenever it is well defined.
Proof of Theorem \[subsOrd2\] {#thmB}
-----------------------------
Let $T$ be a second order operator from $\F_\l$ to $\F_\l$. Since $\FS^{1,2}_0$ is equivalent to $\F_{-\frac{1}{\ell+1}}$, the subsymbol $\ss^2_{\l,\l}(T)$ may be written as a contact Hamiltonian. If the operator $T$ is of the form $$\begin{array}{rcl}
T&=&T_{2,0,0}\,\partial_z^2+T_{1,i,0}\,\partial_z\partial_{x_i}
+T_{1,0,i}\,\partial_z\partial_{y_i}+T_{0,ij,0}\,\partial_{x_i}\partial_{x_j}+
T_{0,i,j}\,\partial_{x_i}\partial_{y_j}+T_{0,0,ij}\,\partial_{y_i}\partial_{y_j}\\[4pt]
&&+T_{1,0,0}\,\partial_z+T_{0,i,0}\,\partial_{x_i}+T_{0,0,i}\,\partial_{y_i}+T_{0,0,0},
\end{array}$$ then the formula of Proposition \[subsymbol formula\] reads $$\begin{array}{rcl}
\ss^2_{\l,\l}(T) &=&
-\frac{1+2\l(\ell+1)}{\ell+2}
\Bigl(
\partial_z(T_{2,0,0}-\oh{}y_iT_{1,i,0}+\oh{}x_iT_{1,0,i})\\[6pt]
&&\hskip1.6cm
+\partial_{x_i}(T_{1,i,0}-\oh{}y_jT_{0,ij,0}+\oh{}x_jT_{0,i,j})\\[6pt]
&&\hskip1.6cm
+\partial_{y_i}(T_{1,0,i}+\oh{}x_jT_{0,0,ij}-\oh{}y_jT_{0,j,i})\Bigr)\\[6pt]
&&
+T_{1,0,0}-\oh{}y_iT_{0,i,0}+\oh{}x_iT_{0,0,i}.
\end{array}$$ One can verify directly that this formula coincides with that of Theorem \[subsOrd2\]. The most efficient approach is to prove first the comments following the statement of the theorem, after which it suffices to carry out the verification for the single operator $T = L_\lambda(X_z) \circ L_\lambda(A_1)$, because that determines $c_{13}$.
**Acknowledgments**. The first author was partially supported by Simons Foundation Collaboration Grant 207736. The second author was partially supported by Grant PICS05974, “PENTAFRIZ”, of the CNRS. The article was begun while both authors were supported by the Research in Pairs program of the Mathematisches Forschungsinstitut Oberwolfach, for whose support they are very grateful. They are also pleased to thank Dmitry Fuchs, Jérôme Germoni, Peter Littelmann, Sophie Morier-Genoud, Sergei Tabachnikov, and André Unterberger for enlightening discussions.
[CMZ97]{}
, Calculus on Heisenberg manifolds, [*Ann. Math. Studies*]{} [**119**]{}, Princeton University Press, Princeton, 1988.
, Automorphic pseudodifferential operators, in “Algebraic Aspects of Integrable Systems,” [*Progr. Nonlinear Differential Equations Appl.*]{} [26]{}, Birkhäuser, Boston, 1997, 17–47.
, Conformal symbols and the action of contact s over the superline, [*J. Reine Angew. Math.*]{} [**633**]{} (2009), 115–163.
, Conformally equivariant quantization: existence and uniqueness, [*Ann. Inst. Fourier (Grenoble)*]{} [49]{} (1999), no. 6, 1999–2029.
, Space of second-order linear differential operators as a module over the Lie algebra of vector fields, [*Adv. Math.*]{} [**132**]{} (1997), no. 2, 316–333.
, Projectively equivariant quantization and symbol calculus: noncommutative hypergeometric functions, [*Lett. Math. Phys.*]{} [**57**]{} (2001) 61–67.
, Confoliations, University Lecture Series, 13, American Mathematical Society, Providence, RI, 1998.
, Contact degree and the index of Fourier integral operators, [*Math. Res. Lett.*]{} [**5**]{} (1998), no. 3, 363–381.
, Dequantized differential operators between tensor densities as modules over the Lie algebra of contact vector fields, [arXiv:0710.0222.]{}
, Decomposition of symmetric tensor fields in the presence of a flat contact projective structure, [*J. Nonlin. Math. Phys.*]{} [**15**]{} (2008), no. 2, 252–269.
, Cohomology of Infinite Dimensional Lie Algebras, Plenum, New York, 1986.
, Symmetries of modules of differential operators, [*J. Nonlinear Math. Phys.*]{} [**12**]{} (2005), no. 3, 348–380.
, On the cohomology of ${\rm sl}(m+1,\Bbb R)$ acting on differential operators and ${\rm sl}(m+1,\Bbb R)$-equivariant symbol, [*Indag. Math. (N.S.)*]{} [11]{} (2000), no. 1, 95–114.
Comparison of some modules of the Lie algebra of vector fields, [*Indag. Math. (N.S.)*]{} [**7**]{} (1996), no. 4, 461–471.
, Projectively invariant symbol calculus, [*Lett. Math. Phys.*]{} [49]{} (1999), no. 3, 173–196.
, Conformally equivariant quantization – a complete classification, [*SIGMA*]{} [**8**]{} (2012), 022.
, Formes différentielles sur les variétés de contact, [*J. Differential Geom.*]{} [**39**]{} (1994), no. 2, 281–330.
, Conformally equivariant quantization – towards complete classification, [arXiv:0903.4798.]{}
, Noncommutative microlocal analysis I, [*Mem. Amer. Math. Soc.*]{} [**52**]{} (1984), no. 313.
, The Atiyah-Singer index formula for subelliptic operators on contact manifolds, Part I, [*Ann. Math. (2)*]{} [**171**]{} (2010), no. 3, 1647–1681.
Lie groups, Lie algebras, and their representations, Graduate Texts in Math., Springer-Verlag, New York, 1984.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, nilpotent $n$-Lie algebras of dimension $n+3$ as well as nilpotent $n$-Lie algebras of class $2$ and dimension $n+4$ are classified.'
address:
- '$^{1}$Farhangiyan University, Shahid Beheshti Campus, Mashhad, Iran.'
- '$^{2}$Department of Mathematics, Mashhad Branch, Islamic Azad University, Mashhad, Iran.'
- '$^{3}$ Department of Mathematics, Esfarayen University of Technology, Esfarayen, Iran.'
author:
- mehdi eshrati$^1$
- farshid saeedi$^2$
- hamid darabi$^3$
title: 'Low dimensional nilpotent $n$-Lie algebras'
---
Introduction and preliminaries
==============================
The classification of low dimensional Lie algebras is an early problem in the study of Lie algebras. Such classifications can be find in standard references of Lie algebras. A first step to classify $6$-dimensional Lie algebras was done by Umlauf [@ku] in 1891. In 1950, Mozorov [@vm] obtained the classification of nilpotent Lie algebras of dimension less than $6$ over arbitrary fields and those of dimensions $6$ over a field of characteristic zero. The classification of $6$-dimensional nilpotent Lie algebras is completed by Cicalo [@sc.wg.cs] in 2012. The $7$-dimensional nilpotent Lie algebras over the fields of real and complex numbers are classified in [@cs]. Also, the $8$-dimensional nilpotent Lie algebras of class $2$ with a $4$-dimensional center, those with a $2$-dimensional center, and those with a $4$-dimensional center over the field of complex numbers are classified in [@sx.br], [@br.lz] and [@yw.hc.yn], respectively. We know that a classification of Lie algebras with respect to the dimension of their multiplier is given in [@lb].
In 1985, Fillipov [@vf] defined an *$n$-Lie algebra* as an antisymmetric $n$-linear map on a vector space that satisfy the following Jacobi identity: $$[[x_1,x_2,\ldots,x_n],y_2,\ldots,y_n]=\sum_{i=1}^n[x_1,\ldots,x_{i-1},[x_i,y_2,\ldots,y_n],x_{i+1},\ldots,x_n]$$ for all $x_i,y_i\in L$, $1\leq i\leq n$ and $2\leq j\leq n$. Also, he has classified the $n$-Lie algebras of dimension $n+1$ over an algebraically closed field of characteristic zero. In 2008, Bai [@rb.xw.wx.ha] classified those $n$-Lie algebras of dimension $n+1$ whose underlying field has characteristic $2$. Also, in 2011, Bai et. al. [@rb.gs.yz] classified the $n$-Lie algebras of dimension $n+2$ over algebraically closed fields of characteristic zero.
Let $A_1,A_2,\ldots,A_n$ be subalgebras of $n$-Lie algebra $A$. Then the subalgebra of $A$ generated by all commutators $[x_1,\ldots,x_n]$, in which $x_i\in A_i$, is denoted by $[A_1,\ldots,A_n]$. The subalgebra $A^2=[A,\ldots,A]$ is called the *derived* $n$-Lie subalgebra of $A$. If $A^2=0$, we call $A$ an abelian $n$-Lie algebra. The *center* of the $n$-Lie algebra $A$ is defined as $$Z(A)=\{x\in A:[x,A,\ldots,A]=0\}.$$ Assume $Z_0(A)=0$, then $i$th center of $A$ is defined inductively as $Z_i(A)/Z_{i-1}(A)=Z(A/Z_{i-1}(A))$ for all $i\geq1$. In 1987, Kasimov [@sk] defines the notion of nilpotency of $n$-Lie algebras as follows:
An $n$-Lie algebra $A$ is called nilpotent if $A^s=0$ for some non-negative integer $s$, in which $A^{i+1}$ is defined inductively as $A^1=A$ and $A^{i+1}=[A^i,A,\ldots,A]$. The $n$-Lie algebra $A$ is nilpotent of class $c$ provided that $A^{c+1}=0$ and $A^c\neq0$. Similar results for the same groups are obtained for nilpotent $n$-Lie algebras in [@mw].
The goal of this paper is to classify $(n+3)$-dimensional nilpotent $n$-Lie algebras as well as $(n+4)$-dimensional nilpotent $n$-Lie algebras of class $2$ over an arbitrary field. This paper is organized as follows: Section 2 presents the preliminary results that will be used in the next sections. In section 3, we give a classification of all $(n+3)$-dimensional nilpotent $n$-Lie algebras over an arbitrary field. Also, in section 4, we shall classify all $(n+4)$-dimensional nilpotent $n$-Lie algebras of class $2$ over an arbitrary field.
Known results
=============
In this section, we shall present some known results, without proofs, that will be used later. Recall that an $n$-Lie algebra $A$ is called *Special Heisenberg* if $A^2=Z(A)$ and $\dim A^2=1$.
\[special Heisenberg\] Every Special Heisenberg $n$-Lie algebra is of dimension $mn+1$ for some positive integer $m$ and it has a presentation as: $$H(n,m)={\langlex,x_1,\ldots,x_{mn}:[x_{n(i-1)+1},x_{n(i-1)+2},\ldots,x_{ni}]=x,i=1,\ldots,m\rangle}.$$
\[dimA\^2=1\] Let $A$ be a nilpotent $n$-Lie algebra of dimension $d$ satisfying $\dim A^2=1$. Then $A\cong H(n,m)\oplus F(d-mn-1)$ for some $m\geq1$.
\[A\^2=Z(A)\] Let $A$ be a nilpotent $n$-Lie algebra of dimension $d=n+k$, for $3 \leq k\leq n+1$ such that $A^2=Z(A)$ and $\dim A^2=2$. Then $$A\cong {\langlee_1,\ldots,e_{n+k}:[e_{k-1},\ldots,e_{n+k-2}]=e_{n+k},[e_1,\ldots,e_n]=e_{n+k-1}\rangle}.$$ This $n$-Lie algebra is denoted by $A_{n,k}$.
\[d<=n+2\] Let $A$ be a non-abelian nilpotent $n$-Lie algebra of dimension $d\leq n+2$. Then $$A\cong H(n,1),\ H(n,1)\oplus F(1)\ \text{or}\ A_{n+2,3},$$ in which $$A_{n+2,3}={\langlee_1,\ldots,e_{n+2}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+2}\rangle}.$$
- Over a field $F$ of characteristic different from $2$, the list of the isomorphisms type of $6$-dimensional nilpotent Lie algebras is the following: $L_{5,k}\oplus F$ with $k\in\{1,\ldots,9\}$ ; $L_{6,k}$ with $k\in\{10,\ldots,18,20,23,25,\ldots,28\}$; $L_{6,k}({\varepsilon}_1)$ with $k\in\{19,21\}$ and ${\varepsilon}_1\in F^*/(\overset{*}{\mathop \sim})$; $L_{6,k}({\varepsilon}_2)$ with $k\in\{22,24\}$ and ${\varepsilon}_2\in F/(\overset{*}{\mathop \sim})$.
- Over a field $F$ of characteristic $2$, the isomorphism types of $6$-dimensional nilpotent Lie algebras are: $L_{5,k}\oplus F$ with $k\in\{1,\ldots,9\}$; $L_{6,k}$ with $k\in\{10,\ldots,18,20,23,25,\ldots,28\}$; $L_{6,k}({\varepsilon}_1)$ with $k\in\{19,21\}$ and ${\varepsilon}_1\in F^*/(\overset{*}{\mathop \sim})$; $L_{6,k}({\varepsilon}_2)$ with $k\in\{22,24\}$ and ${\varepsilon}_2\in F/(\overset{*+}{\mathop \sim})$; $L_{6,k}^{(2)}$ with $k\in\{1,2,5,6\}$; $L_{5,k}^{(2)}({\varepsilon}_3)$ with $k\in\{3,4\}$ and ${\varepsilon}_3\in F^*/(\overset{*+}{\mathop \sim})$; $L_{6,k}^{(2)}({\varepsilon}_4)$ with $k\in\{7,8\}$ and ${\varepsilon}_4\in\{0,\omega\}$.
Classification of $(n+3)$-dimensional nilpotent $n$-Lie algebras
================================================================
By Theorem \[d<=n+2\], the number of nilpotent $n$-Lie algebras of dimensions $n$, $n+1$ and $n+2$ are one, two and three up to isomorphism. In this section, we shall classify all $(n+3)$-dimensional nilpotent $n$-Lie algebras over an arbitrary field.
Let $A$ be a nilpotent $n$-Lie algebra of dimension $n+3$ with basis $\{e_1,\ldots,e_{n+3}\}$. If $e_{n+3}$ is a central element of $A$, then $A/{\langlee_{n+3}\rangle}$ is a nilpotent $n$-Lie algebra of dimension $n+2$. We discuss on the abelian-ness of $A/{\langlee_{n+3}\rangle}$.
If $A/{\langlee_{n+3}\rangle}$ is abelian, then brackets in $A$ can be written as $$[e_1,\ldots,\hat{e}_i,\ldots,\hat{e}_j,\ldots,e_{n+2}]=\theta_{ij}e_{n+3},\quad1\leq i<j\leq n+2.$$ If $\theta_{ij}=0$ for all $1\leq i<j\leq n+2$, then $A$ is an abelian $n$-Lie algebra, which we denote it by $A_{n+3,1}$. If $\theta_{ij}$ are not all equal to zero, then $A$ is non-abelian satisfying $\dim A^2=1$. Hence, by Theorem \[dimA\^2=1\], $A\cong H(n,m)\oplus F(n+3-nm-1)$.
In case $n>2$, we must have $m=1$ and hence $$A={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+3}\rangle}\cong H(n,1)\oplus F(2),$$ which we denote it by $A_{n+3,2}$.
Also, in case $n=2$, we have two $n$-Lie algebras for $A$, namely $H(2,2)$ and $H(2,1)\oplus F(2)$.
Now, assume that $A/{\langlee_{n+3}\rangle}$ is a non-abelian $n$-Lie algebra. By Theorem \[d<=n+2\], we have two possibilities for $A/{\langlee_{n+3}\rangle}$:
Case 1: $A/{\langlee_{n+3}\rangle}\cong{\langle{\overline}{e}_1,\ldots,{\overline}{e}_{n+2}:[{\overline}{e}_1,\ldots,{\overline}{e}_n]={\overline}{e}_{n+1}\rangle}$. Then the brackets in $A$ can be written as $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1}+{\alpha}e_{n+3},&\\
[e_1,\ldots,\hat{e}_i,\ldots,\hat{e}_j,\ldots,e_n,e_{n+1},e_{n+2}]=\theta_{ij}e_{n+3},&1\leq i<j\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_i e_{n+3},&1\leq i\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+2}]={\gamma}_ie_{n+3},&1\leq i\leq n.
\end{cases}$$
Regarding a suitable change of basis, one can assume that ${\alpha}=0$, and the Jacobi identities give us $$\theta_{ij}=0,\quad1\leq i<j\leq n.$$ Hence $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1},&\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+3},&1\leq i\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+2}]={\gamma}_ie_{n+3},&1\leq i\leq n.
\end{cases}$$ The above brackets show that the dimension of the center of $A$ is at most $3$. We discuss on the dimension of the center of $A$.
\(i) $\dim Z(A)=1$. Then, we must have ${\beta}_i,{\gamma}_j\neq0$ for some $i$ and $j$. Without loss of generality assume that ${\beta}_1,{\gamma}_1\neq0$. Applying the following transformations $$e'_1=e_1+\sum_{i=2}^n(-1)^{i-1}\frac{{\beta}_i}{{\beta}_1}e_i,\quad e'_j=e_j,\quad2\leq j\leq n+2,\quad e'_{n+3}={\beta}_1e_{n+3}$$ we obtain $$\begin{cases}
[e'_1,\ldots,e'_n]=e'_{n+1},&\\
[e'_2,\ldots,e'_{n+1}]=e'_{n+3},&\\
[e'_2,\ldots,e'_n,e'_{n+2}]=\frac{{\gamma}_1}{{\beta}_1}e'_{n+3},&\\
[e'_1,\ldots,\hat{e}'_i,\ldots,e'_n,e'_{n+2}]=\frac{1}{{\beta}_1}{\left({\gamma}_i-\frac{{\beta}_i{\gamma}_1}{{\beta}_1}\right)}e'_{n+3},&2\leq i\leq n.
\end{cases}$$ Next, by applying the transformations $$\begin{aligned}
e''_1&=e'_1+\sum_{i=2}^n(-1)^{i-1}{\left(\frac{{\gamma}_i}{{\gamma}_1}-\frac{{\beta}_i}{{\beta}_1}\right)}e'_i,\\
e''_j&=e'_j,\quad2\leq j\leq n+1,\quad e''_{n+2}=\frac{{\beta}_1}{{\gamma}_1}e'_{n+2},\quad e''_{n+3}=e'_{n+3}\end{aligned}$$ it yields $$[e''_1,\ldots,e''_n]=e''_{n+1},\quad[e''_2,\ldots,e''_{n+1}]=e''_{n+3},\quad[e''_2,\ldots,e''_n,e''_{n+2}]=e''_{n+3}.$$ Hence, we conclude that $$A={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+3},[e_2,\ldots,e_n,e_{n+2}]=e_{n+3}\rangle},$$ which we denote it by $A_{n+3,3}$.
\(ii) $\dim Z(A)=2$. We have two possibilities:
(ii-a) If $Z(A)={\langlee_{n+2},e_{n+3}\rangle}$, then at least one of the ${\beta}_i$ is non-zero while all ${\gamma}_i$ are zero. Without loss of generality assume that ${\beta}_1\neq0$. Then $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1},&\\
[e_2,\ldots,e_{n+1}]={\beta}_1e_{n+3},&\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+3},&2\leq i\leq n.
\end{cases}$$ Applying the following transformations $$e'_1=e_1+\sum_{i=2}^n(-1)^{i-1}\frac{{\beta}_i}{{\beta}_1}e_i,\quad e'_j=e_j,\quad2\leq j\leq n+2,\quad e'_{n+3}={\beta}_1e_{n+3}$$ we obtain $$\begin{cases}
[e'_1,\ldots,e'_n]=e'_{n+1},&\\
[e'_2,\ldots,e'_{n+1}]=e'_{n+3}.
\end{cases}$$ Hence $$A={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+3}\rangle},$$ which is isomorphic to $A_{n+2,3}\oplus F(1)$ and it is denoted by $A_{n+3,4}$.
(ii-b) If $Z(A)={\langlee_{n+1},e_{n+3}\rangle}$, then $A^2=Z(A)$ and Theorem \[A\^2=Z(A)\] yields $$A={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_n,e_{n+2}]=e_{n+3}\rangle},$$ which is denoted by $A_{n+3,5}$.
\(iii) $\dim Z(A)=3$. Then $Z(A)={\langlee_{n+1},e_{n+2},e_{n+3}\rangle}$ and so that ${\beta}_i={\gamma}_i=0$ for all $1\leq i\leq n$. The only non-zero bracket is $[e_1,\ldots,e_n]=e_{n+1}$, which gives rise to the algebra $H(n,1)\oplus F(2)=A_{n+3,2}$.
Case 2: $A/{\langlee_{n+3}\rangle}\cong{\langle{\overline}{e}_1,\ldots,{\overline}{e}_{n+2}:[{\overline}{e}_1,\ldots,{\overline}{e}_n]={\overline}{e}_{n+1},[{\overline}{e}_2,\ldots,{\overline}{e}_{n+1}]={\overline}{e}_{n+2}\rangle}$. Then the brackets in $A$ are as follows: $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1}+{\alpha}e_{n+3},&\\
[e_2,\ldots,e_{n+1}]=e_{n+2}+{\beta}e_{n+3},&\\
[e_1,\ldots,\hat{e}_i,\ldots,\hat{e}_j,\ldots,e_n,e_{n+1},e_{n+2}]=\theta_{ij}e_{n+3},&1\leq i<j\leq n,\\
[e_2,\ldots,e_n,e_{n+2}]=\gamma e_{n+3},&\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+2}]={\alpha}_ie_{n+3},&2\leq i\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+3},&2\leq i\leq n.
\end{cases}$$ With a suitable change of basis, one can assume that ${\alpha}={\beta}=0$. Moreover, from the Jacobi identities, it follows that $$\theta_{ij}=0,\quad1\leq i<j\leq n,\quad{\alpha}_i=0,\quad2\leq i\leq n,$$ so $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1},&\\
[e_2,\ldots,e_{n+1}]=e_{n+2},&\\
[e_2,\ldots,e_n,e_{n+2}]=\gamma e_{n+3},&\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+3},&2\leq i\leq n.
\end{cases}$$ The above relations show that the dimension of $Z(A)$ is at most $2$. We have two possibilities:
\(i) $\dim Z(A)=1$. Then ${\gamma}\neq0$ and $e'_{n+3}={\gamma}e_{n+3},\quad e'_j=e_j,\quad 1\leq j\leq n+2$ yields $$\begin{cases}
[e'_1,\ldots,e'_n]=e'_{n+1},&\\
[e'_2,\ldots,e'_{n+1}]=e'_{n+2},&\\
[e'_2,\ldots,e'_n,e'_{n+2}]=e'_{n+3},&\\
[e'_1,\ldots,\hat{e'}_i,\ldots,e'_n,e'_{n+1}]=\frac{{\beta}_i}{{\gamma}}e'_{n+3},&2\leq i\leq n.
\end{cases}$$ If ${\beta}_i=0$ for all $2\leq i\leq n$, then $$A={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+2},[e_2,\ldots,e_n,e_{n+2}]=e_{n+3}\rangle},$$ which is denoted by $A_{n+3,6}$. On the other hand, if ${\beta}_i\neq0$ for some $2\leq i\leq n$, say ${\beta}_2\neq0$, then, by applying the following transformations, $$\begin{aligned}
e'_1&=\frac{{\beta}_2}{{\gamma}}e_1,\quad e'_2={\left(\frac{{\beta}_2}{{\gamma}}\right)}^2 e_2+\sum_{i=3}^n(-1)^i\frac{{\beta}_2{\beta}_i}{{\gamma}}e_i,\quad e'_i=e_i,\quad3\leq i\leq n-1,\\
e'_n&={\left(\frac{{\gamma}}{{\beta}_2}\right)}^2e_n,e'_i=\frac{{\beta}_2}{{\gamma}}e_i,\quad n+1\leq i\leq n+2,\quad e'_{n+3}=\frac{{\beta}_2}{{\gamma}}e_{n+3},\end{aligned}$$ we observe that $$\begin{gathered}
A=\langle e_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+2},\\
[e_2,\ldots,e_n,e_{n+2}]=[e_1,e_3,\ldots,e_{n+1}]=e_{n+3}\rangle,\end{gathered}$$ which we denote it by $A_{n+3,7}$.
\(ii) $\dim Z(A)=2$. Then ${\gamma}=0$ and the brackets in $A$ reduce to $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1},&\\
[e_2,\ldots,e_{n+1}]=e_{n+2},&\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+3},&2\leq i\leq n.
\end{cases}$$ If ${\beta}_i=0$ for all $2\leq i\leq n$, then $$A={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+2}\rangle}.$$ One can easily see that this algebra is isomorphic to $A_{n+3,4}$, while if ${\beta}_i\neq0$ for some $2\leq i\leq n$, say ${\beta}_2\neq0$, then the following transformations $$e'_1=e_1,\quad e'_2=e_2+\sum_{i=3}^n(-1)^i\frac{{\beta}_i}{{\beta}_2}e_i,\quad e'_j=e_j,\quad3\leq j\leq n+2,\quad e'_{n+3}={\beta}_2e_{n+3}$$ show that $$\begin{gathered}
A=\langle e_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},\\
[e_2,\ldots,e_{n+1}]=e_{n+2},[e_1,e_3,\ldots,e_{n+1}]=e_{n+3}\rangle.\end{gathered}$$ This algebra is denoted by $A_{n+3,8}$.
The above results summary as:
\[d=n+3\] The only nilpotent $n$-Lie algebras of dimension $n+3$ with $n>2$ are $A_{n+3,i}$ with $i\in\{1,\ldots,8\}$. For $n=2$ we have instead the algebra $H(2,2)$.
Classification of $(n+4)$-dimensional nilpotent $n$-Lie algebras of class $2$
=============================================================================
Nilpotent $n$-Lie algebras of class $2$ appear in some problems of geometry like commutative Riemannian manifolds. Also, classifying nilpotent Lie algebras of class $2$ has been an important problem in Lie algebras. In [@br.lz] nilpotent Lie algebras of class $2$ and dimension $8$ with $2$-dimensional center over the field of complex numbers are classified. In this section, we intend to classify $(n+4)$-dimensional nilpotent $n$-Lie algebras of class $2$. Regarding Theorem \[d=n+3\], the following result obtains immediately.
\[d=n+3 of class 2\] The only $(n+3)$-dimensional nilpotent $n$-Lie algebras of class $2$ are
- $A_{n+3,2}={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+3}\rangle}\cong H(n,1)\oplus F(2)$;
- $A_{n+3,5}={\langlee_1,\ldots,e_{n+3}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_n,e_{n+2}]=e_{n+3}\rangle}$;
- $H(2,2)$.
Suppose $A$ is an $(n+4)$-dimensional nilpotent $n$-Lie algebra of class $2$ with basis $\{e_1,\ldots,e_{n+4}\}$. If $\dim A^2=1$, Theorem \[dimA\^2=1\] yields $A\cong H(n,m)\oplus F(n+4-nm-1)$. Hence $A$ is isomorphic to one of the following algebras $$H(n,1)\oplus F(3),\quad H(2,2)\oplus F(1)\ \text{or}\ H(3,2).$$ Now, we assume that $\dim A^2\geq2$ and ${\langlee_{n+3},e_{n+4}\rangle}\subseteq A^2$. Then $A/{\langlee_{n+4}\rangle}$ is an $(n+3)$-dimensional nilpotent $n$-Lie algebra of class $2$ so that, by Theorem \[d=n+3 of class 2\], $A/{\langlee_{n+4}\rangle}$ is of the following forms:
Case 1: $A/{\langlee_{n+4}\rangle}\cong{\langle{\overline}{e}_1,\ldots,{\overline}{e}_{n+3}:[{\overline}{e}_1,\ldots,{\overline}{e}_n]={\overline}{e}_{n+3}\rangle}$. Then the brackets in $A$ can be written as: $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+3}+{\alpha}e_{n+4},&\\
[e_1,\ldots,\hat{e}_i,\ldots,\hat{e}_j,\ldots,e_n,e_{n+1},e_{n+2}]=\theta_{ij}e_{n+4},&1\leq i<j\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+4},&1\leq i\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+2}]={\gamma}_ie_{n+4},&1\leq i\leq n.
\end{cases}$$ By a suitable change of basis, we may assume that ${\alpha}=0$, hence $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+3},&\\
[e_1,\ldots,\hat{e}_i,\ldots,\hat{e}_j,\ldots,e_n,e_{n+1},e_{n+2}]=\theta_{ij}e_{n+4},&1\leq i<j\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+4},&1\leq i\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+2}]={\gamma}_ie_{n+4},&1\leq i\leq n.
\end{cases}$$ Since $\dim A^2\geq2$, it is obvious that $\dim Z(A)\leq3$. First, assume that $\dim Z(A)=3$. Then, without loss of generality, we may assume that $Z(A)={\langlee_{n+2},e_{n+3},e_{n+4}\rangle}$. Hence ${\gamma}_i$ and $\theta_{ij}$ are all zero while ${\beta}_i\neq0$ for some $i$. We may assume that ${\beta}_1\neq0$. Then $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+3},&\\
[e_2,\ldots,,e_n,e_{n+1}]={\beta}_1e_{n+4},&\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+4},&2\leq i\leq n.
\end{cases}$$ Using the following transformations $$e'_1=e_1+\sum_{i=1}^n(-1)^{i-1}\frac{{\beta}_i}{{\beta}_1}e_i,\quad e'_j=e_j,\quad2\leq j\leq n+3,\quad e'_{n+4}={\beta}_1e_{n+4},$$ it follows that $$A={\langlee_1,\ldots,e_{n+4}:[e_1,\ldots,e_n]=e_{n+3},[e_2,\ldots,e_{n+1}]=e_{n+4}\rangle}.$$ This algebra is denoted by $A_{n+4,1}$.
Next, assume that $\dim Z(A)=2$. Then $Z(A)=A^2$. If $n\geq3$, then, by Theorem \[A\^2=Z(A)\], we have $$A={\langlee_1,\ldots,e_{n+4}:[e_1,\ldots,e_n]=e_{n+3},[e_3,\ldots,e_{n+2}]=e_{n+4}\rangle},$$ which is denoted by $A_{n+3,2}$. From [@sc.wg.cs] in case $n=2$, the only $6$-dimensional Lie algebra satisfying $A^2=Z(A)$ and $\dim Z(A)=2$ is $$A\cong L_{6,22}({\varepsilon})={\langlee_1,\ldots,e_6:[e_1,e_2]=e_5,[e_1,e_3]=e_6,[e_2,e_4]={\varepsilon}e_6,[e_3,e_4]=e_5\rangle}.$$ Note that this algebra does not satisfy $A/{\langlee_6\rangle}\cong H(2,1)\oplus F(2)$.
Case 2: $A/{\langlee_{n+4}\rangle}\cong{\langle{\overline}{e}_1,\ldots,{\overline}{e}_{n+3}:[{\overline}{e}_1,\ldots,{\overline}{e}_n]={\overline}{e}_{n+1},[{\overline}{e}_2,\ldots,{\overline}{e}_n,{\overline}{e}_{n+2}]={\overline}{e}_{n+3}\rangle}$. Then the brackets in $A$ can be written as $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1}+{\alpha}e_{n+4},&\\
[e_2,\ldots,e_n,e_{n+2}]=e_{n+3}+{\beta}e_{n+4},&\\
[e_1,\ldots,\hat{e}_i,\ldots,\hat{e}_j,\ldots,e_n,e_{n+1},e_{n+2}]=\theta_{ij}e_{n+4},&1\leq i<j\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+1}]={\beta}_ie_{n+4},&1\leq i\leq n,\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+2}]={\gamma}_ie_{n+4},&2\leq i\leq n.
\end{cases}$$ With a suitable change of basis, one can assume that ${\alpha}={\beta}=0$. A simple verification shows $Z(A)={\langlee_{n+1},e_{n+3},e_{n+4}\rangle}$ such that $$\begin{cases}
[e_1,\ldots,e_n]=e_{n+1},&\\
[e_2,\ldots,e_n,e_{n+2}]=e_{n+3},&\\
[e_1,\ldots,\hat{e}_i,\ldots,e_n,e_{n+2}]={\gamma}_ie_{n+4},&2\leq i\leq n.
\end{cases}$$ If ${\gamma}_i=0$ for all $2\leq i\leq n$, then $$A={\langlee_1,\ldots,e_{n+4}:[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_n,e_{n+2}]=e_{n+3}\rangle}.$$ One can easily see that this algebra is isomorphic to $A_{n+4,1}$. On the other hand, if ${\gamma}_i\neq0$ for some $2\leq i\leq n$, say ${\gamma}_2\neq0$, then we may apply the following transformations $$e'_2=e_2+\sum_{j=3}^n(-1)^{j-1}\frac{{\gamma}_j}{{\gamma}_2}e_j,\quad e'_i=e_i,\quad i=1,3,\ldots,n+3,\quad e'_{n+4}={\gamma}_2e_{n+4}.$$ Then $$\begin{gathered}
A=\langle e_1,\ldots,e_{n+4}:[e_1,\ldots,e_n]=e_{n+1},\\
[e_2,\ldots,e_n,e_{n+2}]=e_{n+3},[e_1,e_3,\ldots,e_n,e_{n+2}]=e_{n+4}\rangle,\end{gathered}$$ which we denote it by $A_{n+4,3}$.
Case 3: $A/{\langlee_{n+4}\rangle}\cong H(2,2)$. From [@sc.wg.cs], we observe that $$A\cong L_{6,22}({\varepsilon})={\langlee_1,\ldots,e_6:[e_1,e_2]=e_5,[e_1,e_3]=e_6,[e_2,e_4]={\varepsilon}e_6,[e_3,e_4]=e_5\rangle}.$$
The only $(n+4)$-dimensional nilpotent $n$-Lie algebras of class $2$ are $$H(n,1)\oplus F(3),\ H(2,2)\oplus F(1),\ H(3,2),\ A_{n+4,1},\ A_{n+4,2},\ A_{n+4,3}\ \text{and}\ L_{6,22}({\varepsilon}).$$
In table I, we have illustrated all $n$-Lie algebras obtained in this paper.
Table I
Nilpotent $n$-Lie algebra Non-zero multiplications
--------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------
$A_{n+3,1}$ –
$A_{n+3,2}$ $[e_1,\ldots,e_n]=e_{n+3}$
$A_{n+3,3}$ $\begin{array}{c}[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+3},\\{[e_2,\ldots,e_n,e_{n+2}]}=e_{n+3}\end{array}$
$A_{n+3,4}$ $[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+3}$
$A_{n+3,5}$ $[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_n,e_{n+2}]=e_{n+3}$
$A_{n+3,6}$ $\begin{array}{c}[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+2},\\{[e_2,\ldots,e_n,e_{n+2}]}=e_{n+3}\end{array}$
$A_{n+3,7}$ $\begin{array}{c}[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+2},\\{[e_2,\ldots,e_n,e_{n+2}]}=[e_1,e_3,\ldots,e_{n+1}]=e_{n+3}\end{array}$
$A_{n+3,8}$ $\begin{array}{c}[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_{n+1}]=e_{n+2},\\{[e_1,e_3,\ldots,e_{n+1}]}=e_{n+3}\end{array}$
$A_{n+4,1}$ $[e_1,\ldots,e_n]=e_{n+3},[e_2,\ldots,e_{n+1}]=e_{n+4}$
$A_{n+4,2}$ $[e_1,\ldots,e_n]=e_{n+3},[e_3,\ldots,e_{n+2}]=e_{n+4}$ ($n\geq3$)
$A_{n+4,3}$ $\begin{array}{c}[e_1,\ldots,e_n]=e_{n+1},[e_2,\ldots,e_n,e_{n+2}]=e_{n+3},\\{[e_1,e_3,\ldots,e_n,e_{n+2}]}=e_{n+4}\end{array}$
[0]{} R. Bai, G. Song and Y. Zhang, On classification of $n$-Lie algebras, *Front. Math. China* **6**(4) (2011), 581–606.
R. Bai, X. Wang, W. Xiao and H. An, Structure of low dimensional $n$-Lie algebras over a field of characteristic $2$, *Linear Alg. Appl.* **428** (2008), 1912–1920.
L. Bosko, On Schur multipliers of Lie algebras and groups of maximal class, *Int. J. Algebra Comput.* **20** (2010), 807–821.
S. Cicalo, W. A. de Graaf and Csaba Schneider, Six-dimensional nilpotent Lie algebras, *Linear Algebra Appl.* **436** (2012), 163–189.
H. Darabi, F. Saeedi and M. Eshrati, A characterization of finite dimensional nilpotent Fillipov algebras, *J. Geom. Phys.* **101** (2016), 100–107.
M. Eshrati, F. Saeedi and H. Darabi, On the multiplier of nilpotent $n$-Lie algebras, *J. Algebra* **450** (2016), 162–172.
V. T. Fillipov, $n$-Lie algebras, *Sib. Math. Zh.* **26**(6) (1985), 126–140.
S. M. Kasymov, Theory of $n$-Lie algebras, *Algebra Logika* **26**(3) (1987), 277–297.
V. V. Morozov, Classification des algebras de Lie nilpotent de dimension $6$, *Izv. Vyssh. Ucheb. Zar* **4** (1958), 161–171.
B. Ren and L. Zhu, Classification of $2$-step nilpotent Lie algebras of dimension $8$ with $2$-dimensional center, *Comm. Algebra* **39**(6) (2011), 2068–2081.
C. Seeley, $7$-dimensional nilpotent Lie algebras, *Trans. Amer. Math. Soc.* **335**(2) (1993), 479–496.
K. A. Umlauf, *Ber die Zusammensetzung der endlichen continuierlichen transformationsgruppen insbesondere der gruppen vom range null*, Ph.D. Thesis, University of Leipzig, Germany, 1891.
Y. Wang, H. Chen and Y. Niu, Classification of $8$-dimensional nilpotent Lie algebras with $4$-dimensional center, *J. Math. Res. Appl.* **33**(6) (2013), 699–707.
M. P. Williams, Nilpotent $n$-Lie algebras, *Comm. Algebra* **37** (2009), 1843–1849.
S. Xia and B. Ren, Classification of $2$-step nilpotent Lie algebras of dimension $8$ with $4$ dimensional center, *J. University of Science and Technology of Suzhou* **27** (2010), 19–23.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $\mathbb{F}_q$ be a finite field with $q$ elements and denote by $\theta : \mathbb{F}_q\to\mathbb{F}_q$ an automorphism of $\mathbb{F}_q$. In this note, we deal with skew constacyclic codes, i.e. linear codes of $\mathbb{F}_q^n$ which are invariant under the action of a semi-linear map $T:\mathbb{F}_q^n\to\mathbb{F}_q^n$, defined by $T(a_0,...,a_{n-2}, a_{n-1}):=(\alpha a_{n-1},a_0,...,a_{n-2})$ for some $\alpha\in\mathbb{F}_q-\{0\}$ and $n\geq 2$. In particular, we study some algebraic and geometric properties of their dual codes and we give some Magma Programs as applications of the main theoretical results.'
address: ' Departamento de Matemática, Universidad de Concepción, Casilla 160-C, Concepción, Chile'
author:
- 'Alexis E. Almendras Valdebenito and Andrea Luigi Tironi'
---
[^1]
Introduction {#introduction .unnumbered}
============
Let $\mathbb{F}_q$ be a field with $q$ elements. A linear code $\mathcal{C}$ of length $n$ and dimension $k$, called an $[n,k]_q$-code, is a $k$-dimensional subspace of $\mathbb{F}_q^n$. Moreover, an $[n,k]_q$-code $\mathcal{C}$ with minimum Hamming distance $d:=d(\mathcal{C})$ is denoted as an $[n,k,d]_q$-code. A fundamental problem in Coding Theory is that of optimizing one of the parameters $n, k, d$ for given the other two. If $d_q(n,k)$ denotes the largest value of $d$ for which an $[n,k,d]_q$-code exists, we call an $[n,k,d_q(n,k)]_q$-code simply an optimal code. It is well known that a large number of new linear codes achieving the best known bounds $d_q(n,k)$, in particular over small fields, have been constructed as cyclic, constacyclic or quasi-cyclic codes, including the corresponding ones in the non-commutative case (e.g., see [@B0] and [@M2]). For this reason, the main purpose of this note is to study some algebraic and geometric properties of skew constacyclic codes (Definition \[skew GCC\]) and their duals, the latest ones being strongly related to the above minimum Hamming distance $d$ which is useful in error-correcting codes and for some decoding algorithms. Finally, two lower bounds for the distance of these kind of codes are given, together with some MAGMA programs (Programs 1 to 5) as an application of the main theoretical results.
After some notions and basic remarks, in Section \[Sec2\] we recall some known properties of skew constacyclic codes and as a consequence of these facts, we give some geometric properties about the parity check matrices of skew constacyclic codes (Theorems \[GMaruta:2\] and \[GMaruta:3\]), whose columns are composed of orbits of points in projective spaces via semi-linear maps. Furthermore, inspired by [@M2] and [@L], we firstly generalize at the end of Section \[Sec2\] a result about $1$-generator QT codes to the non-commutative case (Theorem \[GMaruta:QT\]) and secondly, by using a factorization algorithm of A. Leroy in the commutative case, in Section \[Sec3\] we give two lower bounds for the minimum Hamming distance $d$ of skew $\theta$-module codes, in particular for skew constacyclic codes (Theorems \[corchete:2\] and \[corchete:log\]). Finally, as an application of the above results, we write five MAGMA Programs [@magma] and for small fields we construct a table with some $1$-generator skew QT codes (see Definition \[skew QT codes\]) which reach the known best minimum Hamming distance.
Properties of skew constacyclic codes and their duals {#Sec2}
=====================================================
Along all this note, we will use the following notation.
Let $\F_q$ be a finite field with $q$ elements, where $q=p^n$ for some prime $p$. Define $\F_q^\ast:=\F_q-\{ 0\}$ and take $\alpha\in\F_q^\ast$. Let $\theta$ be an automorphism of $\F_q$ and let us recall here the definition of the main object which we will treat in this note.
\[skew GCC\] A linear code $\mathscr{C}\subseteq \mathbb{F}_q^n$ is called a *skew $(\alpha ,\theta )$-cyclic code* (simply a *skew $\alpha$-cyclic code*, or a $\theta$-constacyclic code) if $\mathscr{C}$ is invariant under the semi-linear map $$\phi_{\alpha,\theta} : (c_0,c_1\dots,c_{n-1})\mapsto (\alpha \theta(c_{n-1}),\theta(c_0),\dots,\theta(c_{n-2})).$$ Moreover, for some fixed $\theta$, we call a linear code $\mathscr{C}\subseteq \mathbb{F}_q^n$ a *skew constacyclic code* (or, a *skew pseudo-cyclic code*) if $\mathscr{C}$ is a skew $\alpha$-cyclic code for some $\alpha\in \F_q^\ast$.
The above definition can be reinterpreted in an algebraic way by the following
\[boucher:A\] Let $\alpha\in\F_q^\ast$. A linear code $\mathscr{C}\subseteq \mathbb{F}_q^n$ is a skew $(\alpha,\theta )$-cyclic code if and only if $\mathscr{C}$ is invariant under the semi-linear map $\T:=\Theta \circ A$, that is, $(\vec{c})\T\in \mathscr{C}$ for all $\vec{c}\in\mathscr{C}$, where $\Theta\colon \F_q^n\to\F_q^n$ is defined by $( (c_0,\dots,c_{n-1}))\Theta:=(\theta(c_0),\dots,\theta(c_{n-1}))$ and $A$ is the $n\times n$ matrix given by $$A:=\left( \begin{array}{ c | c c c } 0 & 1 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & 1 \\ \hline \alpha & 0 & \cdots & 0 \end{array}\right).$$
Let $\mathscr{C}\subseteq \mathbb{F}_q^n$ be a skew $(\alpha ,\theta )$-cyclic code. Then $$\mathscr{C}\star (\Theta\circ A )\subseteq \mathscr{C},$$ where $\mathscr{C}\star (\Theta\circ A):=\{ (\vec{c})\Theta A\mid \vec{c}\in\mathscr{C} \}$. When $\theta=Id$, a skew $(\alpha, Id)$-cyclic code is simply a constacyclic code.
From now on, we mainly show some algebraic and geometric properties for the dual code of a skew constacyclic codes.
First of all, with the purpose of giving an algebraic structure to skew $\alpha$-cyclic codes, write $R:=\F_q[x;\theta]$ and consider the following one-to-one correspondence: $$\label{cod:pi3}
\begin{aligned}
\pi \colon \phantom{aaaaat}\F^n_q\phantom{aaaaa} &\longrightarrow \phantom{aaa} R/R(x^n-\alpha)\\
(a_0,a_1,\dots,a_{n-1})&\longmapsto a_0+a_1x+\cdots a_{n-1}x^{n-1}\quad .\\
\end{aligned}$$ Note that $\pi$ is an $\F_q$-linear isomorphism of vectorial spaces over $\F_q$. So, we can identify $\F^n_q$ with $R/R(x^n-\alpha)$ and any vector $\vec{a}=(a_0,a_1,\dots,a_{n-1})\in \F^n_q$ with the polynomial class $\pi(\vec{a}):=\sum_{i=0}^{n-1}a_ix^i\in R/R(x^n-\alpha)$.
Put $m:=|\langle \theta \rangle|$. If either $m\nmid n$ or $\alpha\notin\F_q^\theta$, we know that $R/R(x^n-\alpha)$ is not a ring and we cannot argue about its ideals, as in the commutative case. For instance, when $\alpha =1$ the condition $m\mid n$ is assumed (e.g., see [@B0] and [@B2]) and one can construct a one-to-one correspondence between skew cyclic codes and the ideals of $R/R(x^n-1)$. On the other hand, the set $R/R(x^n-\alpha)$ could be considered a left $\F_q$-module or a left $\F_q[x;\theta]$-module.
The next two results give an equivalent definition of skew constacyclic codes and some of their well-known properties.
A nonempty subset $\mathscr{C}\subset\F^n_q$ is a skew $\alpha$-cyclic code if and only if $\pi(\mathscr{C})$ is a left $R$-submodule of the left $R$-module $R/R(x^n-\alpha)$.
\[GMaruta:1\] Let $\pi(\mathscr{C})$ be a left $R$-submodule of $R/R(x^n-\alpha)$ with $R:=\F_q[x;\theta]$, i.e. $\mathscr{C}$ is a skew $\alpha$-cyclic code of $\F_q^n$. Then there exists a unique monic polynomial of minimal degree in $R$ such that
(a) $\pi (\mathscr{C})=Rg(x)$, i.e. $g(x)$ is the generator polynomial of $\pi (\mathscr{C})$;
(b) $g(x)$ is a right divisor of $x^n-\alpha$;
(c) Every $c(x)\in\pi (\mathscr{C})$ can be written uniquely as $c(x)=f(x)g(x)\in R/R(x^n-\alpha)$, where $f(x)\in R$ has degree less than or equal to $n-\text{deg }g(x)$. Moreover, the dimension of $\mathscr{C}$ is equal to $n-\text{deg }g(x)$;
(d) If $g(x):=\sum\limits_{x=0}^k g_ix^i$, then $\mathscr{C}$ has a generator matrix $G$ given by $$\begin{aligned}
G&=\begin{pmatrix} g(x)\\xg(x)\\ \vdots \\x^{n-k-1}g(x) \end{pmatrix}\\ &=
\begin{pmatrix}
g_0 & g_1 & \cdots & g_{k} & 0 & 0 & \cdots & 0 \\
0 & \theta(g_0) & \theta(g_1) & \cdots & \theta(g_{k}) & 0 & \cdots & 0 \\
\vdots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \vdots \\
0 & 0 & 0 & \theta^{n-k-1}(g_0) & \theta^{n-k-1}(g_1) & \cdots & \cdots & \theta^{n-k-1}(g_{k}) \\
\end{pmatrix}. \end{aligned}$$
For skew a-cyclic codes of $\mathbb{F}_q^n$ with $\theta (x):=x^{p^t}$, where $q=p^r$ for some prime $p$, $r\in\mathbb{Z}_{\geq 2}$ and an integer $t$ such that $1\leq t\leq r-1$, type the following Program $1$ before the Programs $2$ to $5$:
\[Prog-1\]
p:=...; r:=...; t:=...; F<w>:=GF(p^r);
R<X>:=TwistedPolynomials(F:q:=p^t);
The next Program $2$ constructs by the command +SD(n,a)+ all the skew a-cyclic codes of $\mathbb{F}_q^n$ with $\theta (x):=x^{p^t}$:
\[Prog0\]
SD:=function(n,a)
P:=[0]; for i in [1..n-2] do
P:=P cat [0]; end for; T:= [-a] cat P cat [1]; f:=R!T;
V:=VectorSpace(F,n); dd:=[]; E:=[x : x in F | x ne 0];
S:=CartesianProduct(E,CartesianPower(F,n-1));
for s in S do
ll:=[s[1]] cat [p : p in s[2]];
if LeadingCoefficient(R!ll) eq 1 then
q,r:=Quotrem(f,R!ll); if r eq R![0] then
R!ll; dd := dd cat [R!ll]; end if; end if; end for;
return dd; end function;
In $\mathbb{F}_4^{14}$ with $\theta (x):=x^{2}$, the command +SD(14,1)+ gives $603$ different nontrivial right divisors of $X^{14}-1$, i.e. $603$ different nontrivial skew cyclic codes of $\mathbb{F}_4^{14}$ instead of $25$ different nontrivial cyclic codes of $\mathbb{F}_4^{14}$ in the commutative case.
The following technical result will be useful to prove Theorem \[GMaruta:1e\].
\[Tironi\] Let $\mathscr{C}\subseteq \F_q^n$ be a linear code and let $\mathcal{T}\colon \F_q^n\to\F_q^n$ be a semi-linear map such that $\mathscr{C}\star \mathcal{T}\subseteq
\mathscr{C}$.
(a) If $\mathcal{T}$ is invertible, then $\mathscr{C}\star
\mathcal{T}= \mathscr{C}$. Furthermore, $\mathscr{C}\star
\T=\mathscr{C}$ if and only if $\mathscr{C}\star
(\T)^{-1}=\mathscr{C}$;
(b) if $\T$ is as in [*Proposition \[boucher:A\]*]{}, then $\mathscr{C}\star
\mathcal{T}=\mathscr{C}\star (\T)^{-1}=\mathscr{C}$; moreover, the dual code $\mathscr{C}^\perp$ is a linear code invariant under $\T'=\Theta^{-1}\circ ({}^t \! A)_{\theta^{-1}}$ with $\mathscr{C}^{\perp}\star
\T'=\mathscr{C}^{\perp}\star (\T')^{-1}=\mathscr{C}^{\perp}$, where $M_{\theta^{-1}}:=[\theta^{-1}(m_{ij})]$ for any matrix $M=[m_{ij}]$.
$(a)$ Since $\T$ is invertible, the semi-linear map $\T$ is injective. Hence $|\mathscr{C}\star \T |=|\mathscr{C}|$. From $\mathscr{C}\star \T\subseteq \mathscr{C}$ we deduce that $\mathscr{C}\star \T=\mathscr{C}$.
$(b)$ Note that $\T$ is invertible since $\alpha\in\F_q^\ast$. Then we can conclude by [@TT1 Proposition 25].
Now, we are able to show in an easy and direct way that, similarly to the commutative case, the dual code of a skew $\alpha$-cyclic code is a skew $\alpha^{-1}$-cyclic code. This result was first presented and proven in very different forms in [@B2 Theorem 8], [@B1 Theorem 1] and [@FG Theorem 6.1].
\[GMaruta:1e\] Let $\mathscr{C}\subseteq\F_q^n$ be a linear code and take $\alpha\in\F_q^\ast$. Then $\mathscr{C}$ is a skew $\alpha$-cyclic code if and only if its dual code $\mathscr{C}^\perp$ is a skew $\alpha^{-1}$-cyclic code.
Suppose that $\mathscr{C}$ is a skew $\alpha$-cyclic code invariant under the semi-linear map $\Theta\circ A$, that is, $\mathscr{C}\star (\Theta\circ A)\subseteq \mathscr{C}$. By Lemma \[Tironi\](b), we have $$\mathscr{C}^\perp \star \left(\Theta^{-1}\circ ({}^t \! A)_{\theta^{-1}} \right)= \mathscr{C}^\perp.$$ Since $\Theta\circ M_\theta = M\circ \Theta$ for any matrix $M$, by Lemma \[Tironi\](a) we get $$\begin{aligned}
\mathscr{C}^\perp \star \left(\Theta^{-1}\circ ({}^t \! A)_{\theta^{-1}} \right)= \mathscr{C}^\perp &\Longleftrightarrow \mathscr{C}^\perp \star \left(\Theta^{-1}\circ ({}^t \! A)_{\theta^{-1}} \right)^{-1}= \mathscr{C}^\perp\\
&\Longleftrightarrow \mathscr{C}^\perp \star \left( (({}^t \! A)^{-1})_{\theta^{-1}} \circ \Theta \right)= \mathscr{C}^\perp\\
&\Longleftrightarrow \mathscr{C}^\perp \star \left( \Theta\circ
({}^t \! A)^{-1} \right)= \mathscr{C}^\perp ,
\end{aligned}$$ where $$({}^t \! A)^{-1}=\left( \begin{array}{ c | c c c } 0 & 1 & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & 1 \\ \hline
\alpha^{-1} & 0 & \cdots & 0 \end{array}\right).$$ Thus we can conclude that $\mathscr{C}^\perp$ is a skew $\alpha^{-1}$-cyclic code.
Finally, having in mind that $(\mathscr{C}^\perp)^\perp
=\mathscr{C}$, the converse of the statement can be immediately obtained.
Let $\C$ be a skew $\alpha$-cyclic code. If $\mathscr{C} = \mathscr{C}^\perp$, then $n$ is even and $\alpha=\pm 1$.
First of all, observe that $n=\dim\mathscr{C}+\dim\mathscr{C}^{\perp}=2\dim\mathscr{C}$. Finally, let $g(x)$ be the generator polynomial of $\mathscr{C}$. By Theorem \[GMaruta:1e\] we have $$X^n-\alpha=h_1(x)g(x)\qquad \mathrm{and} \qquad X^n-\alpha^{-1}=h_2(x)g(x)\ ,$$ for some $h_1(x),h_2(x)\in R$. So we get $\deg \left[ \left( h_1(x)-h_2(x)\right) g(x)\right]=0$ and since $\deg g(x)>0$, we conclude that $h_2(x)=h_1(x)$ and this shows that $\alpha = \alpha^{-1}$, i.e. $\alpha^2=1$.
With the next result one can write directly the generator polynomial of the dual code of any skew constacyclic code.
\[asterisco\] Let $\C$ be a skew $\alpha$-cyclic code generated by the polynomial $g(x)=g_0+g_1x+\cdots+g_{n-k}x^{n-k}$. Then the dual code $\mathscr{C}^\perp$ is generated by $$h(x):=\theta^k(\hbar_0^{-1})\left[\sum_{i=0}^k \theta^i\left( \hbar_{k-i}\right)x^i\right]\ ,$$ where $\hbar(x):=\hbar_0+\hbar_1x+\cdots+\hbar_kx^k$ is such that $x^n-\theta^{-k}(\alpha)=g(x)\hbar(x).$
Since $\mathscr{C}$ is a skew $\alpha$-cyclic code, by Theorem \[GMaruta:1e\] we know that the dual code $\mathscr{C}^\perp$ is a skew $\alpha^{-1}$-cyclic code. Thus from Theorem \[GMaruta:1\] it follows that there exists a unique monic polynomial $h(x)$ of minimal degree in $R:=\F_q[x;\theta]$ such that $\mathscr{C}^\perp=Rh(x)$. Therefore by [@B2 Theorem 8] we deduce that there exist $\hbar(x)\in R$ and $c\in\mathbb{F}_q^*$ such that $$X^n-c=g(x)\hbar(x)$$ and $h(x)=\theta^k(\hbar_0^{-1})g^\perp(x)$, where $g^\perp(x):=\sum_{i=0}^k \theta^i\left( \hbar_{k-i}\right)x^i$ and $\hbar(x)=\hbar_0+\hbar_1x+\cdots+\hbar_kx^k$. Since $g(x)\in R$ is monic and $X^n-\alpha =t(x)g(x)$ for some monic $t(x)\in R$, by [@B1 Lemma 2] we obtain that $$X^n-\theta^{-k}(\alpha)=g(x)s(x)$$ for some $s(x)\in R$. Hence $c-\theta^{-k}(\alpha)=g(x)(s(x)-\hbar(x))$ and since $\deg g(x)\geq 1$, we conclude that $c=\theta^{-k}(\alpha)$.
The following MAGMA Program $3$ defines a function +DualCode(n,a,g)+ which gives all the main informations about the dual code of a skew a-cyclic codes of length n with generator polynomial g :
\[Prog1\]
DualCode:=function(n,a,g)
k:=n-Degree(R!g); C:=a^(p^(k*(r-t))); P:=[0];
for i in [1..n-2] do
P:=P cat [0]; end for;
T:= [-C] cat P cat [1]; f:=R!T; g1:=R!g; E:=[x : x in F | x ne 0];
S:=CartesianProduct(E,CartesianPower(F,n-Degree(g1)));
for ss in S do
ll:=[ss[1]] cat [p : p in ss[2]];
if g1*R!ll eq f then
h:=R!ll; end if; end for;
d:=Degree(h); h1:=Matrix(F,1,d+1,[Eltseq(h)[i]: i in {1..d+1}]);
h2:=ReverseColumns(h1);
h3:=Matrix(F,1,d+1,[h2[1][i]^((p^t)^(i-1)): i in {1..d+1}]);
h4:=(1/h3[1][d+1])*h3; h5:=R![h4[1][i]: i in {1..d+1}];
k:=Degree(f)-Degree(h5); V:=VectorSpace(F,n);
H:=Matrix(F,k,n,[V!(HorizontalJoin(Matrix(1, j+Degree(h5)+1,
Eltseq((R![0,1])^j*h5)), ZeroMatrix(F, 1, n-j-Degree(h5)-1))):
j in {0..k-1}]); k2:=Degree(f)-Degree(g1);
G:=Matrix(F,k2,n,[V!(HorizontalJoin(Matrix(1, j+Degree(g1)+1,
Eltseq((R![0,1])^j*g1)), ZeroMatrix(F, 1, n-j-Degree(g1)-1))):
j in {0..k2-1}]); L:=LinearCode(G); LD:=LinearCode(H);
print "==========================================="; print " ";
print "Code Skew", a,"-cyclic of type: ", n, k2, MinimumWeight(L);
print " "; print "Polynomial generator:"; print g1; print " ";
print "Generator Matrix:"; print G; print " ";
print "-------------------------------------------"; print " ";
print "Dual Code Skew", 1/a,"-cyclic of type: ", n, k,
MinimumWeight(LD); print " "; print "Check Polynomial:";
print h5; print " "; print "Parity Check Matrix:"; print H;
print " "; print G*Transpose(H); Q:=[0];
print "===========================================";
for j in [1..n-2] do
Q:=Q cat [0]; end for;
U:= [-a^-1] cat P cat [1]; Z:=R!U; Z; h5;
Quotrem(R!Z,R!h5); return "end";
end function;
Consider the finite field $\F_4=\{ 0,1,w,w^2\}$, where $w^2+w+1=0$, and the Frobenius automorphism $\theta :\F_4\to\F_4$ defined by $\theta(x)=x^2$. Let $\mathscr{C}$ be the skew $w$-cyclic $[8,2,4]_4-$code generate by the polynomial $x^6+wx^4+w^2 x^2+1$ (command +SD(8,w)+ in Program $1$) with generator matrix $$\begin{pmatrix}
1 & 0 & w^2 & 0 & w & 0 & 1 & 0 \\
0 & 1 & 0 & w & 0 & w^2 & 0 & 1\\
\end{pmatrix}.$$ By the command +DualCode(8,w,\[1,0,w\^2,0,w,0,1\])+ in Program $2$ we see that the dual code $\mathscr{C}^\perp$ is the skew $w^2$-cyclic $[8,6,2]_4-$code ($w^2=w^{-1}$) generated by the polynomial $x^2+w^2$ and with generator matrix $$\begin{pmatrix}
1&0&w&0&0&0&0&0\\
0&1&0&w^2&0&0&0&0\\
0&0&1&0&w&0&0&0\\
0&0&0&1&0&w^2&0&0\\
0&0&0&0&1&0&w&0\\
0&0&0&0&0&1&0&w^2\\
\end{pmatrix}.$$
The following two results show a geometric property of the dual codes of skew constacyclic codes. More precisely, we will see that the columns of a parity check matrix of a skew constacyclic code can be considered as points in a projective space which are particular orbits under the action of the semi-linear map associated to the code. Therefore, in line with [@M1], in the non-commutative case we obtain the following two results.
\[GMaruta:2\] Let $\mathscr{C}$ be a linear $[n,n-k]$-code over $\F_q$. Then $\mathscr{C}$ is a skew $\alpha$-cyclic code if and only if $\mathscr{C}$ has a parity check matrix of the form $$[P_t, (P\tau )_t, (P\tau ^2)_t,\dots,(P\tau ^{n-1})_t]$$ such that $P\tau^n=\alpha P$, where $P\in\F_q^k$, $\tau=\Theta\circ T$ and $T\in GL(k,q)$.
“$\Rightarrow$” Let $\mathscr{C}$ be a skew $\alpha$-cyclic $[n,n-k]$-code over $\F_q$ with a parity check matrix $H=[(P_1)_t,(P_2)_t,\dots,(P_n)_t]$, where $(P_i)\in\F_q^k$. Then by Theorem \[GMaruta:1e\] we see that $H'=[\alpha^{-1}
\Theta((P_n)_t),\Theta((P_1)_t),\Theta((P_2)_t),\dots,\Theta((P_{n-1})_t)]$ is also a parity check matrix for $\mathscr{C}$. Thus there is a matrix $T_t\in GL(k,q)$ such that $H= T_t\cdot H'$. This gives $$\begin{aligned}
(P_1)_t&=T_t(\alpha^{-1}\Theta((P_n)_t))=(\alpha^{-1}(P_n)\Theta\circ T)_t\\
(P_2)_t&=T_t\Theta((P_1)_t)=((P_1)\Theta\circ T)_t\\
(P_3)_t&=T_t\Theta((P_2)_t)=((P_2)\Theta\circ T)_t=((P_1)(\Theta\circ T)^2)_t\\
&~~\vdots\\
(P_n)_t&=T_t\Theta((P_n)_t)=((P_{n-1})\Theta\circ T)_t=((P_1)(\Theta\circ T)^{n-1})_t\end{aligned}$$
Furthermore, we have $$P_1=\alpha^{-1}(P_n)(\Theta\circ T)\Rightarrow P_1=\alpha^{-1}(P_1)(\Theta\circ T)^n
\Rightarrow \alpha P_1 = P_1(\Theta\circ T )^n$$ So, by putting $\tau :=\Theta\circ T$ and $P:=P_1$, we obtain that $$H=[P_t, (P\tau )_t, (P\tau ^2)_t,\dots,(P\tau
^{n-1})_t]$$ with $P\tau^n=\alpha P$.
“$\Leftarrow$” Let $\mathscr{C}$ be a code with parity check matrix $$H=[P_t, (P\tau )_t, (P\tau ^2)_t,\dots,(P\tau ^{n-1})_t]$$ with $P\in\F_q^k$, $\tau=\Theta\circ T$ and $T\in GL(k,q)$ such that $P\tau^n=\alpha P$. Then, for any $\vec{c}=(c_0,c_1,\dots,c_{n-1})
\in\mathscr{C}$ we have $\vec{c}H_t=\vec{0}$. This implies $$\begin{aligned}
\sum_{i=0}^{n-1}c_i(P\tau^i)=\vec{0}&\Longrightarrow \left( \sum_{i=0}^{n-1}c_i(P\tau^i)\right)\tau=(\vec{0})\tau\\
&\Longrightarrow \sum_{i=0}^{n-2}\theta(c_i)(P\tau^{i+1})+\theta(c_{n-1})(P\tau^n)=\vec{0}\\
&\Longrightarrow \sum_{i=0}^{n-2}\theta(c_i)(P\tau^{i+1})+\theta(c_{n-1})(\alpha P)=\vec{0}\\
&\Longrightarrow (\alpha \theta(c_{n-1}),\theta(c_0),\dots,\theta(c_{n-2}))H_t=\vec{0},\end{aligned}$$ i.e. $(\alpha
\theta(c_{n-1}),\theta(c_0),\dots,\theta(c_{n-2}))\in\mathscr{C}$ and $\mathscr{C}$ is a skew $\alpha$-cyclic code.
\[GMaruta:3\] Let $g(x)=\sum_{i=0}^k a_ix^i$ with $a_k=1$ be a skew polynomial of degree $k$ in $\F_q[x;\theta]$ that divides on the right $x^n-\alpha$, where $\alpha\in\F_q^\ast$. Then $\mathscr{C}\subset\F_q^n$ is a skew $\alpha$-cyclic $[n,n-k]$-code over $\F_q$ with generator polynomial $g(x)$ if and only if $\mathscr{C}$ is a code with partity check matrix $$[P_t,(P\tau)_t,(P\tau^2)_t,\dots,(P\tau^{n-1})_t],$$ where $P=(1,0,\dots,0)$, $\tau=\Theta \circ T_g$ and $T_g$ is the companion matrix of $g(x)$, i.e. $$T_g=\begin{pmatrix}
0&1&0&\cdots &0\\
0&0&1&\cdots &0\\
\vdots&\vdots&\vdots&\ddots &\vdots\\
0&0&0&\cdots&1\\
-a_0&-a_1&-a_2&\cdots&-a_{k-1}\\
\end{pmatrix}.$$
Consider the following linear map $$\label{cod:pi4}
\begin{aligned}
\pi \colon \phantom{aaaaat}\F^k_q\phantom{aaaaa} &\longrightarrow \phantom{aaa} R/Rg\\
(c_0,c_1,\dots,c_{k-1})&\longmapsto c_0+c_1x+\cdots c_{k-1}x^{k-1}\\
\end{aligned}$$ with $R=\F_q[x;\theta]$. Then, we can see that $\pi(P\tau^i )=x^i$ with $P=(1,0,\dots,0)$, for all $i\in\Z_{\geq 0}$. Thus, we have $$\begin{aligned}
\pi( a_0P+\cdots+a_{k-1}P\tau^{k-1}+P\tau^k)
&=\pi(a_0P)+\cdots+\pi(a_{k-1}P\tau^{k-1})+\pi(P\tau^k)\\
&=a_0\pi(P)+\cdots+a_{k-1}\pi(P\tau^{k-1})+\pi(P\tau^k)\\
&=a_0 (1)+a_1(x)+\cdots+a_{k-1}(x^{k-1})+(x^k)\\
&=g(x) = 0 \in R/Rg(x),\end{aligned}$$ i.e. $a_0P+a_1P\tau+\cdots+a_{k-1}P\tau^{k-1}+P\tau^k=(0,\dots,0)=\vec{0}$. Furthermore, as $g(x)$ is a right divisor of $x^n-\alpha$, we have $$\pi(P\tau^n-\alpha P)=\pi(P\tau^n)-\alpha\pi(P)=x^n-\alpha=0\in R/Rg(x),$$ and this implies that $P\tau^n-\alpha P=(0,0,\dots,0)$, that is, $P\tau^n=\alpha P$.\
Take $H=[(P_0)_t,(P_1)_t,\dots,(P_{n-1})_t]$, $P_i=P\tau^i$, $P_0=P=(1,0,\dots,0)$ and $\tau=\Theta \circ T_g$.
\
“$\Leftarrow$” Note that $\vec{g}=(a_0,a_1,\dots,a_{k-1},1,0,\dots,0)\in\F_q^n$ is an element of $\mathscr{C}$, since $\vec{g}H_t=\vec{0}$. Furthermore, we get $\vec{g}\tau^iH_t=\vec{0}$ for every $i=0,\dots,n-k-1$, i.e. $\{ \vec{g},\vec{g}\tau,\dots,\vec{g}\tau^{n-k-1}\}$ are $n-k$ linear independent elements of $\mathscr{C}$. Thus $\mathscr{C}$ has a parity check matrix $G$ given by $$\label{MG}
\begin{pmatrix}
a_0 & a_1 & \cdots & a_{k-1}&1 & 0 & 0 & \cdots & 0 \\
0 & \theta(a_0) & \theta(a_1) & \cdots & \theta(a_{k-1}) &1& 0 & \cdots & 0 \\
\vdots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots &\ddots & \vdots \\
0 & 0 & 0 & \theta^{n-k-1}(a_0) & \theta^{n-k-1}(a_1) & \cdots & \cdots & \theta^{n-k-1}(a_{k-1})&1 \\
\end{pmatrix}.$$
As $P\tau^n=\alpha P$, by Theorem \[GMaruta:2\] we see that $H$ is a parity check matrix of a skew $\alpha$-cyclic $[n,n-k]$-code. Finaly, since $g(x)$ is a monic polynomial which corresponds to the vector $\vec{g}\in\mathscr{C}$, we conclude that $\mathscr{C}$ is a skew $\alpha$-cyclic $[n,n-k]$-code over $\F_q$ with generator polynomial $g(x)$.\
“$\Rightarrow$” Let $\mathscr{C}$ be a skew $\alpha$-cyclic $[n,n-k]$-code over $\F_q$ with generator polynomial $g(x)=a_0+a_1x+\dots+a_{k-1}x^{k-1}+x^k$. Then the generator matrix of $\mathscr{C}$ in the canonical form is as in (\[MG\]). Let us show now that $H$ is the parity check matrix of $\mathscr{C}$. From $a_0P+a_1P\tau+\cdots+a_{k-1}P\tau^{k-1}+P\tau^k =\vec{0}$, it follows that $$\begin{aligned}
(a_0P+\cdots+P\tau^k )\tau=\vec{0}&\Leftrightarrow (\theta(a_0)P\tau+\theta(a_1)P\tau^2 +\cdots+P\tau^{k+1})=\vec{0}\\
(a_0P+\cdots+P\tau^k )\tau^2=\vec{0}&\Leftrightarrow (\theta^2(a_0)P\tau^2+\theta^2(a_1) P\tau^3+\cdots+P\tau^{k+2})=\vec{0}\\
\vdots\phantom{aaaaaaaaaa}&\phantom{aaaaaaaaaaaaaaa}{\vdots}\\
(a_0P+\cdots+P\tau^k
)\tau^{n-k-1}=\vec{0}&\Leftrightarrow
(\theta^{n-k-1}(a_0)P\tau^{n-k-1}+
\cdots+P\tau^{n-1})=\vec{0}.\end{aligned}$$ This implies that $G H_t=O$, i.e. $HG_t=O$ and $H$ is a parity check matrix of $\mathscr{C}$.
As an application of Theorem \[GMaruta:3\], the following MAGMA Program $4$ defines the function +PCM(n,g)+ which gives the parity check matrix in standard form of a skew a-cyclic codes in $\mathbb{F}_q^n$ with generator polynomial g :
\[Prog2\]
PCM:=function(n,g)
V:=VectorSpace(F,n); g1:=R!g;
if Degree(g1) ge 1 then
d:=Degree(g1); CM:=VerticalJoin(HorizontalJoin(
ZeroMatrix(F,d-1,1), ScalarMatrix(F,d-1,1)), Matrix(F,1,d,
[-Eltseq(g1)[k]: k in {1..d}])); P:=HorizontalJoin(
Matrix(F,1,1,[1]), ZeroMatrix(F,1,d-1)); TT:=R![0,1];
PP:=P; PCM1:=P; PCM2:=P; for m in {1..n-1} do
PCM1:=Matrix(F,1,d,[SpecialEvaluate(TT,PCM1[1][i]):
i in {1..d}])*CM; PCM2:=VerticalJoin(PCM2,PCM1);
end for; PCM3:=Transpose(PCM2); print
"==========================================="; print " ";
print "Polynomial generator:"; print g1; print " ";
print "Companion Matrix:"; print CM; print " ";
print "Parity Check Matrix:"; print PCM3; print " ";
print "===========================================";
end if; return "end"; end function;
Consider the finite field $\F_4=\{ 0,1,w,w^2\}$, where $w^2+w+1=0$, and the Frobenius automorphism $\theta$ defined by $\theta(x)=x^2$ for any $x\in\F_4$. The skew $w$-cyclic $[7,3]_4$-code generate by the polynomial $g(x)=x^4+x^2+w^2 x+1$ has a generator matrix $$G:=\begin{pmatrix}
1 & w^2 & 1 & 0 &1 & 0 & 0 \\
0 & 1 & w & 1 &0 & 1 & 0 \\
0& 0 &1 & w^2 &1& 0 & 1 \\
\end{pmatrix}.$$ By the command +PCM(7,\[1,w\^2,1,0,1\])+ in Program $4$, we obtain that its parity check matrix is $H:=[P_t,(P\tau)_t,(P\tau^2)_t,\dots,(P\tau^{6})_t]$, where $P=(1,0,0,0)$, $\tau=\Theta \circ T_g$ and $T_g$ is the companion matrix of $g(x)$: $$T_g=\begin{pmatrix}
0&1&0&0\\
0&0&1&0\\
0&0&0&1\\
1& w^2&1&0 \\
\end{pmatrix}, \qquad H=\begin{pmatrix}
1&0&0&0 &1&0&1\\
0&1&0&0 &w^2&1&w^2\\
0&0&1&0 &1&w&0\\
0&0&0&1 &0&1&w^2\\
\end{pmatrix}.$$
A linear code $\mathscr{C}\subseteq \F_q^n$ with parity check matrix of the form $$[P_t, (P\tau)_t,
(P\tau^2)_t,\dots,(P\tau^{n-1})_t],$$ with $P\in\F_q^k$, $\tau=\Theta\circ T$ y $T\in GL(k,q)$ is called *a code $\mathscr{C}$ defined by $(\tau,P,n)$*. Furthermore, one can define also the following set $$\Gamma_k^\alpha :=\{ (\tau,P,n) \mid \exists \text{ a skew $\alpha$-cyclic $[n,n-k]_q$-code
defined by } (\tau,P,n) \}.$$
\[GMaruta:4\] Let $\mathscr{C}_i$ be the code defined by $(\tau_i,P_i,n)\in
\Gamma_k^\alpha$ for $i=1,2$. Then, $\mathscr{C}_1=\mathscr{C}_2$ if and only if there exists a matrix $S\in GL(k,q)$ such that $\tau_1=S\cdot\tau_2\cdot S^{-1}$ and $P_1S=P_2$.
“$\Rightarrow$” Since $\mathscr{C}_1=\mathscr{C}_2$, there exists a matrix $S\in GL(k,q)$ such that $$\begin{aligned}
S_t [(P_1)_t,(P_1\tau_1)_t,\dots,(P_1\tau_1^{n-1})_t]&=
[(P_1S)_t,(P_1\tau_1S)_t,\dots,(P_1\tau_1^{n-1}S)_t]\\
&=[(P_2)_t,(P_2\tau_2)_t,(P_2\tau_2^2)_t,\dots,(P_2\tau_2^{n-1})_t]
\end{aligned}$$
From the first columns we deduce that $P_1S=P_2$, that is, $P_2S^{-1}=P_1$. Furthermore, $P_2\tau_2^i=P_1\tau_1^iS=\left(P_2S^{-1}\right)\tau_1^iS
=P_2\left(S^{-1}\tau_1S\right)^i$, with $i=1,\dots,n-1$.
Thus $\{ P_2,P_2\tau_2,\dots,P_2\tau_2^{k-1}\}$ are linearly independent vectors of $\F^n_q$ and a vector $\vec{v}\in\F_q^n$ can be written as $\vec{v}=\sum_{i=0}^{k-1}\lambda_{ij}\left(P_2\tau_2^i\right)$. So we have [$$\begin{aligned}
\vec{v}\tau_2&=\sum\limits_{i=0}^{k-1}\lambda_{ij}\left(P_2\tau_2^i\right) \tau_2 & &,& \vec{v}\left(S^{-1}\tau_1S\right)&=\sum\limits_{i=0}^{k-1}\lambda_{ij}\left(P_2\tau_2^i\right)\left(S^{-1}\tau_1S\right) \\
&=\sum\limits_{i=0}^{k-1}\lambda_{ij}\left(P_2\tau_2^{i+1}\right) & &&
& =\sum\limits_{i=0}^{k-1}\lambda_{ij}\left(P_2\left(S^{-1}\tau_1S\right)^i\right)\left(S^{-1}\tau_1S\right) \\
&=\sum\limits_{i=0}^{k-1}\lambda_{ij}\left(P_2\left(S^{-1}\tau_1S\right)^{i+1}\right) & & &
& =\sum\limits_{i=0}^{k-1}\lambda_{ij}\left(P_2\left(S^{-1}\tau_1S\right)^{i+1}\right)\end{aligned}$$]{}
Since $\vec{e_j}\tau_2=\vec{e_j}S^{-1}\tau_1S$ for any canonical vector $\vec{e_j}$ with $j=1,\dots,k$, we conclude that $S^{-1}\tau_1S=\tau_2$.\
“$\Leftarrow$” Since there is a matrix $S\in GL(k,q)$ such that $\tau_2=S^{-1}\cdot\tau_2\cdot S$ y $P_1=P_2S^{-1}$, we see that $$\begin{aligned}
[(P_2)_t,(P_2\tau_2)_t,\dots,(P_2\tau_2^{n-1})_t]&=[(P_1S)_t,(P_1\tau_1S)_t,(P_1\tau_1^2S)_t,\dots,(P_1\tau_1^{n-1}S)_t]\\
&=S_t [(P_1)_t,(P_1\tau_1)_t,(P_1\tau_1^2)_t,\dots,(P_1\tau_1^{n-1})_t].
\end{aligned}$$ By Theorem \[GMaruta:2\], the matrices defined by $(\tau_1,P_1,n)$ and $(\tau_2,P_2,n)$ are parity check matrices of skew $\alpha$-cyclic codes which differ by an invertible matrix, i.e. they correspond to the same code. Hence $\mathscr{C}_1=\mathscr{C}_2$.
Unlike the commutative case (see [@M1 Theorem 4]), when $\theta\not= id$ we have to consider the minimal polynomial $m_\tau$ of a semi-linear map $\tau$ (see [@TT2 Proposition 3.2]) instead of the characteristic polynomial. On the other hand, a minimal polynomial may be associated with two different codes. In fact, if $\mathscr{C}_1=\mathscr{C}_2$ then $m_{\tau_1}=m_{\tau_2}$, but the reverse is not true in general, as the following example shows.
Consider $R:=\F_4[x;\theta]$, $g_1=1+wx+x^2+x^3$ and $g_2=1+w^2 x+x^2+x^3$ in $R$, where $\theta$ is the Frobenius automorphism. Let $\mathscr{C}_i:=Rg_i$ be a skew $1$-cyclic $[14,11]_4$-codes for $i=1,2$. Since $g_1(x)\not=
g_2(x)$ in $R/R(x^{14}-1)$, we have $\mathscr{C}_1\not=\mathscr{C}_2$. On the other hand, we get $$\begin{aligned}
\tau_1&=\Theta\circ\begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha&1\\
\end{pmatrix}
&&,&
\tau_2&=\Theta\circ\begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha^2&1\\
\end{pmatrix}
\end{aligned}$$ with minimal polynomials $m_{\tau_1}(x)=m_{B_1}(x^2)=x^6+x^2+1$ and $m_{\tau_2}(x)=m_{B_2}(x^2)=x^6+x^2+1$ respectively (e.g., see [@TT2]), where $$\begin{aligned}
B_1&=\begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha&1\\
\end{pmatrix}_\theta \begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha&1\\
\end{pmatrix}
&,&&
B_2&=\begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha^2&1\\
\end{pmatrix}_\theta \begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha^2&1\\
\end{pmatrix} \\
&=\begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha^2&1\\
\end{pmatrix} \begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha&1\\
\end{pmatrix}
&&&
&=\begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha&1\\
\end{pmatrix} \begin{pmatrix}
0 & 1 &0 \\
0&0&1\\
1&\alpha^2&1\\
\end{pmatrix} \\
&=\begin{pmatrix}
0 & 0 &1 \\
1&\alpha&1\\
1&\alpha^2&\alpha\\
\end{pmatrix}
&&&
&=\begin{pmatrix}
0 & 0 &1 \\
1&\alpha^2&1\\
1&\alpha&\alpha^2\\
\end{pmatrix}, \\
m_{B_1}(t)&=det\begin{pmatrix}
-t & 0 & 1\\
1&\alpha-t&1\\
1&\alpha^2&\alpha-t\\
\end{pmatrix}
&,&&
m_{B_2}(t)&=det\begin{pmatrix}
-t & 0 & 1\\
1&\alpha^2-t&1\\
1&\alpha&\alpha^2-t\\
\end{pmatrix}\\
&=t^3+t+1 & &&&=t^3+t+1.\end{aligned}$$ This shows that there exist two different linear codes $\mathscr{C}_i:=Rg_i$, $i=1,2$, with the same associated minimal polynomial $m_{\tau_i}(x)$.
Finally, let us consider also $1$-generator skew quasi-twisted codes which can be easily defined from the notion of $1$-generator QT codes (see [@M2 $\S 1$]).
So, similarly to [@M2], we give first the following
\[skew QT codes\] Take $\alpha\in\F_q^\ast$ and let $\theta$ be an automorphism of $\F_q$. Denote by $R=\F_q[x;\theta]/(x^N-\alpha)$ the polynomial ring $\F_q[x;\theta]$ over $\F_q$ module $x^N-\alpha$. For $$\boldsymbol{g}=(g_1(x),g_2(x),\dots,g_m(x))\in R^N,$$ the set $$\boldsymbol{\mathscr{C}_g}=\{ (r(x)g_1(x),r(x)g_2(x),\dots,r(x)g_m(x)) \mid r(x)\in R \}$$ is called the *$1$-generator Skew Quasi-Twisted (SQT) code* with generator $\boldsymbol{g}$. Moreover, if $\alpha=1$ then $\boldsymbol{\mathscr{C}_g}$ is called the *$1$-generator Skew Quasi-Cyclic (SQC) code* with generator $\boldsymbol{g}$.
From Theorem \[GMaruta:3\] we know that an $[n,n-k]$-code over $\F_q$ is a skew $\alpha$-cyclic code with generator polynomial $g(x)=\sum_{i=0}^k a_ix^i$ with $a_k=1$ if and only if $\mathscr{C}$ is a linear code with parity check matrix $$[g^n]:=[P_t,(P\tau)_t,(P\tau^2)_t,\dots,(P\tau^{n-1})_t],$$ where $P=(1,0,\dots,0)$, $\tau=\Theta \circ T_g$ and $T_g$ is the companion matrix of $g(x)$. Now, let $\mathcal{T}$ be the projective map of $\P^{k-1}(\F_q)$ defined by $\tau$. We can say that $\mathcal{T}$ is defined by $g(x)$. Then the columns of $[g^n]$ can be considered as points in $\P^{k-1}(\F_q)$ of an orbit of $\mathcal{T}$. Conversely, we can obtain similarly a skew constacyclic code from an orbit of a projective map of $\P^{k-1}(\F_q)$.
Now, consider $m$ orbits $\mathcal{O}_1,\mathcal{O}_2,\dots,\mathcal{O}_m$ of $\mathcal{T}$ with large $N$ and starting points $P_i\in\mathcal{O}_i$ for $i=1,...,m$. For simplicity, take $P_1\equiv P=(1,0,\dots,0)$ and define the matrix $$\begin{gathered}
[P_t,(P\tau)_t,\dots,(P\tau^{n_1-1})_t;\
(P_2)_t,(P_2\tau)_t,\dots,\\ \dots, (P_2\tau^{n_2-1})_t;\
\cdots ;(P_m)_t,(P_m\tau)_t,\dots,(P_m\tau^{n_m-1})_t] \end{gathered}$$ by $[g^{n_1}]+P_2^{n_2}+\cdots+P_m^{n_m}$. Then, the matrix $[g^N]+P_2^N+\cdots+P_m^N$ generates a $SQT$ code whose generator $\boldsymbol{g}$ is given by the following result.
\[GMaruta:QT\] If $P_i\in(\F_q^{\theta})^k$ for $i=1,\dots, m$ are as above, then $[g^N]+P_2^N+\cdots+P_m^N$ generates a $1$-generator Skew Quasi-Twisted (SQT) $[mN,k]_q$-code with generator $$\boldsymbol{g}=\left(h^{\ast}(x),b_2(x^{-1})h^{\ast}(x),\dots,b_m(x^{-1})h^{\ast}(x)\right)\in R^N,$$ where $h^\ast(x)$ is as $h(x)$ in [*Proposition \[asterisco\]*]{} and $b_i(x)$ is the polynomial given by $(1,x,\dots,x^{k-1})P_i$ for $2\leq i\leq m$.
Define $H:=[g^N]+P_2^N+\cdots+P_m^N$ and note that $$\begin{gathered}
H=[P_t,(P\tau)_t,\dots,(P\tau^{N-1})_t;\ (P_2)_t,(P_2\tau)_t,\dots,\\ \dots,(P_2\tau^{N-1})_t;\ \cdots ;\ (P_m)_t,(P_m\tau)_t,\dots,(P_m\tau^{N-1})_t] \end{gathered}$$ with $P_i\in(\F_q^{\theta})^k$, $\tau=\Theta\circ T_g$ and $T_g\in GL(k,q)$ such that $P\tau^N=\alpha P$. By putting $H_i:=[(P_i)_t,(P_i\tau)_t,\dots,(P_i\tau^{N-1})_t]$ for all $i=1,\dots,n$, we get $H=[H_1|H_2|\cdots |H_m]$.
Take $P_1\equiv P:=(1,0,\dots,0)=\vec{e_1}$ and note that $P\tau^{j-1}=\vec{e_{j}}$ for all $j=1,\dots, k$, where $e_j$ is the $j$-th canonical vector of $\F_q^k$. Then, for $P_i\in(\F_q^{\theta})^k$ and $\lambda_{ij}\in \F_q^{\theta}$, we have $$P_i=\sum_{j=0}^{k-1}\lambda_{ij} P\tau^j=(P)\left(\sum_{j=0}^{k-1}\lambda_{ij} \tau^j\right)
=:P\cdot P_i(\tau),$$ where $P_i(x):=\sum_{j=0}^{k-1}\lambda_{ij}
x^j$. Thus $$P_i\tau^h=\sum_{j=0}^{k-1}\lambda_{ij}
P\tau^{j+h}=P\tau^h \cdot P_i(\tau),\quad \forall h=0,\dots,N-1$$ and, for $i=1,\dots,m$, this implies that $$\begin{aligned}
H_i&=[(P_i)_t,(P_i\tau)_t,\dots,(P_i\tau^{N-1})_t]\\
&=[(P\cdot P_i(\tau))_t, (P\tau \cdot P_i(\tau))_t,\dots,(P\tau^{N-1} \cdot P_i(\tau))_t]\\
&=P_i(\tau)_t \cdot [P_t, (P\tau)_t ,\dots,(P\tau^{N-1})_t ] \\
&= P_i(\tau)_t\cdot H_1\end{aligned}$$ Let $H_1^\ast$ be the generator matrix of $\mathscr{C}^\perp$ with $\mathscr{C}=<g(x)>$ written as $$H_1^\ast:= \begin{pmatrix} h^\ast (x)\\ xh^\ast (x)
\\ \vdots \\ x^{k-1}h^\ast(x)
\end{pmatrix} =[ J \mid
\widehat{H}_1^\ast ]$$ with $det(J)\not= 0$, where $h^\ast(x)=\theta^{N-k}(h_0)^{-1}x^{N-k}h(x^{-1})$ with $x^N-\alpha=g(x)h(x)$. By hypothesis, $H_1$ is also a parity check matrix of $\mathscr{C}$ which can be written as $H_1=[I_k\mid \widehat{H}_1]$, where $I_k$ is the identity matrix of dimension $k$. This implies that $J^{-1}H_1^\ast=H_1$ and $J^{-1}\widehat{H}_1^\ast =\widehat{H}_1$.\
Let $A$ be a matrix with $N$ columns and define the linear operator $\diamond$ as follows: $$x^{-1}\diamond A := A\cdot \left( \begin{array}{ c c c | c }
0 & \cdots &0 & \alpha \\ \hline
1 & \cdots &0 & 0 \\
\vdots & \ddots & \vdots & \vdots \\
0 &\cdots & 1 & 0 \\
\end{array}\right).$$ Observe that $x^{-1}\diamond (A'\cdot B')=A'\cdot
(x^{-1}\diamond B')$ for any two matrices $A'$ and $B'$ such that $A'\cdot B'$ is well defined and $B'$ has $N$ columns.\
Furthermore, for $h=0,\dots,N-1$, we get $$\begin{aligned}
x^{-1}\diamond H_1 &= x^{-1}\diamond [P_t,(P\tau)_t,\dots,(P\tau^{N-1})_t]\\ &
= [(P\tau)_t,\dots,(P\tau^{N-1})_t, \alpha P_t]\\&
=[(P\tau)_t,\dots,(P\tau^{N-1})_t, (P\tau^N)_t]\\&=(\tau)_t\cdot H_1\\
x^{-2}\diamond H_1&=x^{-1}\diamond (x^{-1}\diamond H_1) \\
&=x^{-1}((\tau)_t \cdot H_1)\\
&=(\tau^2)_t\cdot H_1\\
&\ \ \vdots\\
x^{-h}\diamond H_1&=(\tau^h)_t\cdot H_1\\\end{aligned}$$
Hence, we have $$\begin{aligned}
H_i&= P_i(\tau)_t \cdot H_1=
P_i(x^{-1})\diamond (J^{-1}\cdot H_1^\ast)
=J^{-1}\cdot (P_i(x^{-1})\diamond H_1^\ast)=\\
&=J^{-1}\cdot \left(P_i(x^{-1})\diamond \begin{pmatrix}
h^\ast (x)\\ xh^\ast (x) \\ \vdots \\ x^{k-1}h^\ast(x)
\end{pmatrix}\right)=J^{-1}\cdot \begin{pmatrix}
P_i(x^{-1})\cdot h^\ast (x)\\ P_i(x^{-1})\cdot xh^\ast (x) \\ \vdots \\ P_i(x^{-1})\cdot x^{k-1}h^\ast(x)
\\
\end{pmatrix} \\
& =J^{-1}\cdot \begin{pmatrix}
P_i(x^{-1})\cdot h^\ast (x)\\ x\cdot P_i(x^{-1})h^\ast (x) \\
\vdots \\ x^{k-1}\cdot P_i(x^{-1})h^\ast(x)\end{pmatrix}
=:J^{-1}\cdot
H^\ast_i \\\end{aligned}$$
Therefore, we conclude that $$H=[H_1|H_2|\cdots |H_m]=[J^{-1}H_1^\ast|J^{-1}H_2^\ast|\cdots |J^{-1}H_m^\ast]
=J^{-1}[H_1^\ast|H_2^\ast|\cdots |H_m^\ast],$$ where $\begin{pmatrix}
P_i(x^{-1})\cdot h^\ast (x)\\ x\cdot P_i(x^{-1})h^\ast (x) \\
\vdots \\ x^{k-1}\cdot P_i(x^{-1})h^\ast(x)\end{pmatrix}$, i.e. $[H_1|H_2|\cdots |H_m]$ and $[H_1^\ast|H_2^\ast|\cdots |H_m^\ast]$ are two generator matrices of the same $[mN,k]_q$-code.
The following MAGMA Program $5$ defines the function +SQT(m,N,g)+ which gives an application of the above result :
\[Prog3\]
SQT:=function(m,N,g)
g1:=R!g; d:=Degree(g1); if d ge 1 then
TT:=R![0,1]; CM:=VerticalJoin(HorizontalJoin(ZeroMatrix(F,
d-1,1), ScalarMatrix(F,d-1,1)), Matrix(F,1,d,[-Eltseq(g1)[k]:
k in {1..d}])); PS<[x]>:=ProjectiveSpace(F,d-1);
PointsPS:={PS!p : p in Points(Scheme(PS,[0]))};
PointsPS2:=PointsPS; OrbN:={}; repeat
S:=Random(PointsPS2); PP:=Matrix(F,1,d,[S[k]: k in {1..d}]);
jj:=0; repeat
jj:=jj+1; PP:=Matrix(F,1,d,[SpecialEvaluate(TT,PP[1][i]):
i in {1..d}])*CM; PPP:=PS![PP[1][i]: i in {1..d}];
PointsPS2:=PointsPS2 diff {PPP}; until S eq PPP; if jj eq N then
OrbN:=OrbN join {S}; end if;
until PointsPS2 eq {}; P1:=HorizontalJoin(Matrix(F,1,1,[1]),
ZeroMatrix(F,1,d-1)); V1:=PS![P1[1][i]: i in {1..d}];
OrbN2:=OrbN diff {V1}; if #OrbN2 ge m then
S2:={}; for l in {1..m-1} do
RR:=Random(OrbN2 diff S2); S2:=S2 join {RR}; end for;
S1:={@ q : q in S2 @}; PCM1:=P1; PCM2:=P1;
for h in {1..N-1} do
PCM1:=Matrix(F,1,d,[SpecialEvaluate(TT,PCM1[1][i]): i in
{1..d}])*CM; PCM2:=VerticalJoin(PCM2,PCM1); end for;
PCM3:=Transpose(PCM2); for ii in {1..m-1} do
P:=Matrix(F,1,d,[S1[ii][j]: j in {1..d}]); PP:=P; PCM1:=P;
PCM2:=P; for h in {1..N-1} do
PCM1:=Matrix(F,1,d,[SpecialEvaluate(TT,PCM1[1][i]): i in
{1..d}])*CM; PCM2:=VerticalJoin(PCM2,PCM1); end for;
PCM3:=HorizontalJoin(PCM3, Transpose(PCM2)); end for;
LC:=LinearCode(PCM3); print
"==========================================="; print " ";
print "Skew Quasi Twisted Code of type: ",
N, N-d, MinimumWeight(LC); print "Polynomial generator:";
print g1; print " "; print "Companion Matrix:"; print CM;
print " "; print "Parity Check Matrix:"; print PCM3; print " ";
print "==========================================="; else
print "There are not enough orbits of length", N ;
end if; end if; return "end"; end function;
Consider $\F_8=\{ 0,1,w,w^2,\dots,w^6\}$. From Program \[Prog0\], by using the command +SD(7,w\^3)+ it follows that $g(x)=x^4-(w^3+x+w^2x^2)\in\F[x;\theta]$ divides $x^7-w^3$, where $\theta(x)=x^4$ for every $x\in\F_8$. Moreover, by Program \[Prog3\] with the command +SQT(2,7,\[w\^3,1,w\^2,0,1\])+ we see that the point $P_2:=[0:w^4:0:1]\in\P^3(\F_8)$ has an orbit of length $7$. So, we get a SQT $[14,4,7]_8$-code whose generator matrix is $$\begin{aligned}
\begin{pmatrix}
1 & 0 & 0 & 0 & w^3 & 0 & 1 & 0 & w^3 & 0 & 0 &w^3& w & 1 \\
0 & 1 & 0 & 0 & 1 &w^5&w^4&w^4 & 1 & w^5 & 0 &1 &0 &0 \\
0 & 0 &1 & 0 &w^2 &1 &0 &0 &0 &1 &w^6&w^2&0 &w^6\\
0 & 0 &0 &1 &0 & w & 1 &1 &0 &0 &1 &w^3&w &0\\
\end{pmatrix}. \end{aligned}$$ For simplicity of notation, let us denote $w,w^2,\dots,w^6$ with $2,3,\dots,7$, respectively. With this notation, the above parity-check matrix can be written simply as $[g]^7+P_2^7=[4130]^7+0501^7.$
Consider $\F_9=\{ 0,1,\beta,\beta^2,\dots,\beta^7\}$, where $\beta$ is a root of $x^2+2x+2\in\F_3[x]$. We denote $0,1,\beta,\beta^2,\dots,\beta^7$ by $0,1,2,3,\dots,8$, respectively. We set that $g(x)=x^4+(1+8x+5x^2+4x^3)=x^4-(5+4x+x^2+8x^3) \in\F_9[x;\theta]$ divides $x^5-2$ with $\theta(a)=a^3$ for every $a\in\F_9$. Let $\mathcal{T}$ be the semi-linear map defined by $\theta$ and $g(x)$. Consider the following three points $P_1$, $P_2$ and $P_3$ of $\P^3(\F_9)$: $$P_1=[1:0:0:0], P_2=[1:6:8:1], P_3=[4:7:4:1].$$ Then, under $\mathcal{T}$, the orbit of $P_1$ is of length $5$ and the orbits of $P_2$ and $P_3$ are both of length $10$. Thus we deduce that $$G=[4518]^5+1681^{10}+4741^{10}$$ generate a SQT $[25,4,19]_9$-code. By comparing the distance of this code with that in the database of MAGMA, we can observe that it reaches the best known distance (BKLC) for a linear $[25,4]_9$-code.
Finally, as an application of Theorem \[GMaruta:QT\] and Program \[Prog3\], Table \[arco:generalizadoinferior1\] gives an example of how to construct some known linear codes with BKLC distance as SQT codes for small values of $q$.
$[n,k,d]_q$ Generator Matrix $N$ $\alpha$ $\theta$
--------------- --------------------------------------------------------------------------------------------------- ----- ---------- ----------
$[21,6,12]_4$ $[131313]^7+220211^{14}$ $7$ $2$ 2
$[25,4,17]_4$ $[1313]^5+3331^{10}+2310^{10}$ 5 $2$ 2
$[35,6,22]_4$ $[131313]^7+123210^{14}+113110^{14}$ 7 2 2
$[35,4,24]_4$ $[1212]^5+0321^{10}+1031^{10}+1301^{10}$ 5 3 2
$[40,4,28]_4$ $[1313]^5+0321^{10}+3301^{10}+1301^{10}+1010^5$ 5 2 2
$[45,4,32]_4$ $[1212]^5+2100^{10}+1311^{10}+0101^{10}+3321^{10}$ 5 2 2
$[49,6,32]_4$ $[131313]^7+233101^{14}+312221^{14}+231310^{14}$ 7 2 2
$[50,4,36]_4$ $[1212]^5+1231^{10}+1101^{10}+2231^{10}+3321^{10}+3100^5$ 5 3 2
$[55,4,36]_4$ $[1313]^5+2011^{10}+1131^{10}+2221^{10}+1331^{10}+0310^5$ 5 2 2
$[60,4,44]_4$ $[1313]^5+1110^{10}+0231^{10}+0031^{10}+3111^{10} +3311^{10}+2021^5$ 5 2 2
$[65,4,48]_4$ $[1212]^5+2201^{10}+1011^{10}+2131^{10} +1310^{10}+0211^{10}+3100^5+3031^{5}$ 5 3 2
$[85,4,64]_4$ $[1212]^5+3211^{10}+0011^{10}+3301^{10}+ 1301^{10}+1331^{10}+3121^{10}+1331^{10}+1010^{5}+0210^5$ 5 3 2
$[34,3,28]_8$ $[175]^4+721^{12}+610^{12}+111^6$ 4 2 4
$[50,4,41]_8$ $[6156]^5+5331^{15}+5610^{15}+6321^{15}$ 5 2 4
$[65,5,56]_8$ $[6156]^5+5541^{15}+7310^{15}+3110^{15}+1531^{15}$ 5 1 2
$[25,4,19]_9$ $[4518]^{5}+1681^{10}+4741^{10}$ 5 6 3
$[30,4,24]_9$ $[4518]^{5}+4801^{10}+8771^{10}+5810^{5}$ 5 6 3
$[42,4,34]_9$ $[2030]^{6}+8121^{12}+6821^{12}+0531^{12}$ 6 7 3
: Skew Quasi-Twisted Codes with BKLC distance.[]{data-label="arco:generalizadoinferior1"}
A lower bound for the distance {#Sec3}
==============================
In [@L] the author shows how a factorization of a skew polynomial can be made in a suitable commutative polynomial ring. By this method, we can transfer some properties of constacyclic codes to skew constacyclic codes and vice versa. In particular, a factorization of a skew polynomial in $\F_q[x;\theta]$ can be made in $\F_q[x]$ by a Leroy’s algorithm (see Theorem 2.5, [@L]). For any prime number $p$ and an integer $i\geq 1$, he defines $[i]:=\dfrac{p^i-1}{p-1}$ and $$\F_q[x^{[]}]:=\left\{ \sum_{i = 0}^{m}\alpha_i x^{[i]}\in\F_q[x] \ : \ m\in\mathbb{Z}_{\geq 0} \right\}.$$ Note that the Leroy’s algorithm and the above definitions can be generalized by considering a power of the Frobenius automorphism. More precisely, let $\F_q$ be a finite field with $q=p^t$ and $p$ prime, and consider the automorphism $\theta(a)=a^{p^s}$ for all $a\in\F_q$ with $0\leq s\leq t-1$.
By putting $[i]_s:=\dfrac{(p^s)^i-1}{p^s-1}$ instead of $[i]$, the polynomial ring $\F_q[x^{[]}]$ can be replaced by $$\F_q[x^{[]_s}]:=\left\{ \sum_{i\geq 0}\alpha_i x^{[i]_s}\in\F_q[x] \ : \ m\in\mathbb{Z}_{\geq 0} \right\}\subseteq \F_q[x] .$$ Finally, note that $[i]_s=[i]$ and $\F_q[x^{[]_s}]=\F_q[x^{[]}]$ for $s=1$.
A [*$[p^s]$-code $\mathscr{C}^{[]_s}\subseteq\F_q^{[n]_s}$*]{} is a linear code generated by a $[p^s]$-polynomial in $\F_q[x^{[]_s}]$.
Let $\mathscr{C}\subseteq\F_q^n$ be a skew $\alpha$-cyclic code generated by $f(t)\in\F_q[t;\theta]$ for some $\alpha\in\F_q^\ast$. Then the $[p^s]$-code $\mathscr{C}^{[]_s}\subseteq\F_q^{[n]_s}$, generated by the associated $[p^s]$-polynomial $f^{[]_s}(t)$, is an $\alpha$-cyclic code.
From Theorem \[GMaruta:1\], we know that $f(t)$ is a right divisor of $x^n-\alpha$. Moreover, by Theorem 2.5 [@L], we see that the associated $[p^s]$-polynomial $f^{[]_s}(x)\in\F_q[x^{[]_s}]$ to $f(t)$ is a right divisor of $x^{[n]_s}-\alpha$. Therefore the code $\mathscr{C}^{[]_s}\subseteq\F_q^{[n]_s}$ is also an $\alpha$-cyclic code.
\[remark\] More in general, if $\mathscr{C}\subseteq\F_q^n$ is a skew $\theta$-module code generated by $g(t)\in\F_q[t;\theta]$, then by [@L Theorem 2.5] we see that the code $\mathscr{C}^{[]_s}\subseteq\F_q^{[n]_s}$ is also a module code generated by $g^{[]}(x)\in\F_q[x]$.
\[corchete:2\] Let $\mathscr{C}\subseteq\F_q^n$ be a linear code generated by $f(t)\in\F_q[t;\theta]$ and let $\mathscr{C}^{[]_s}\subseteq\F_q^{[n]_s}$ be the linear code generated by $f^{[]_s}(x)\in\F_q[x^{[]_s}]$. Then $$d(\mathscr{C}^{[]_s})\leq d(\mathscr{C})$$ and equality holds if and only if there exists a polynomial of minimal weight in $\mathscr{C}^{[]_s}$ which belongs to $\F_q[x^{[]_s}]$.
Let $w(t)\in\mathscr{C}$ be a polynomial of minimal weight. By Theorem \[GMaruta:1\] we have $w(t)=a(t)\cdot f(t)$ for some $a(x)\in\F_q[t;\theta]$. Then by Theorem 2.5 [@L] we get $w^{[]_s}(x)=b(x)\cdot f^{[]_s}(x)$ for some $b(x)\in\F_q[x]$, i.e. $w^{[]_s}(x)\in\mathscr{C}^{[]_s}$. Hence $d(\mathscr{C}^{[]_s})\leq wt( w^{[]_s}(x))=wt( w(t))=d(\mathscr{C})$. This proves the first part of the statement.
Suppose now that $d(\mathscr{C}^{[]_s}) = d(\mathscr{C})$ and let $v(t)\in\mathscr{C}$ be a polynomial such that $wt(v(t))=d(\mathscr{C})$. Note that $v(t)=q(t)\cdot f(t)$ for some $q(t)\in\F_q[x;\theta]$. Thus, by Leroy’s algorithm we have $v^{[]_s}(x)=g(x)\cdot f^{[]_s}(x)$ for some $g(x)\in\F_q[x]$, i.e. $v^{[]_s}(x)\in\mathscr{C}^{[]_s}$. Since $d(\mathscr{C}^{[]_s})=d(\mathscr{C})=wt(v(t))=wt(v^{[]_s}(x))$, we conclude that $v^{[]_s}(x)\in\F_q[x^{[]_s}]$ is a polynomial in $\mathscr{C}^{[]_s}$ of minimal weight.
Finally, let $w^{[]_s}(x)\in\F_q[x^{[]_s}]$ be a minimal weight polynomial in $\mathscr{C}^{[]_s}$. Then $w^{[]_s}(x)=a(x)\cdot f^{[]_s}(x)$ for some $a(x)\in\F_q[x]$. This shows that $w(t)=b(t)\cdot f(t)$, i.e. $w(t)\in\mathscr{C}$. Therefore, we can conclude that $$d(\mathscr{C}^{[]_s})\leq d(\mathscr{C})\leq wt(w(t))=wt(w^{[]_s}(x))=d(\mathscr{C}^{[]_s}),$$ that is, $d(\mathscr{C}^{[]_s})= d(\mathscr{C})$.
A $[p^s]$-code $\mathscr{C}^{[]_s}\subseteq\F_q^{[n]_s}$ with dim$(\mathscr{C}^{[]_s})<[n]_s$ cannot be a MDS code.
Suppose there is a MDS code $\mathscr{C}^{[]_s}\subset \F_q^{[n]_s}$. Then $$d(\mathscr{C}^{[]_s})=[n]_s-\text{dim}(\mathscr{C}^{[]_s})+1
=\text{deg}( f^{[]_s}(x) )+1=[ \text{deg}(f(t)) ]+1,$$ where $f^{[]_s}(x)$ is the generator polynomial of $\mathscr{C}^{[]_s}$. Furthermore, by the Singleton Bound we have $d(\mathscr{C})\leq \text{deg}(f(t))+1$, where $\mathscr{C}$ is the code generated by $f(t)$. Note that deg$(f(t))\geq 1$ and $[i]_s>i$ for all $i\in\Z_{>0}$. Thus by Theorem \[corchete:2\] we deduce that $$[ \text{deg}(f(t)) ]_s+1=d(\mathscr{C}^{[]_s})\leq d(\mathscr{C})\leq \text{deg}(f(t))+1,$$ i.e. $[ \text{deg}(f(t)) ]_s\leq \text{deg}(f(t))$, but this gives a numerical contradiction. Thus the code $\mathscr{C}^{[]_s}$ cannot be a MDS code.
\[Lema\[k\]\] Let $\F_{q^{[k]_s}}^*=\langle w\rangle$ and $\F_{q^{[k]_s}} \subseteq \F_{q^{[n]_s}}$. Then $$[n]_s\leq q^{[k]_s}-1\ \iff \ rank\left( V_{[n]_s}^{Id}(w,w^2,\dots,w^{[n]_s-1}) \right)=[n]_s.$$
\
Let $\mathcal{N}:=\left( V_{[n]_s}^{Id}(w,w^2,\dots,w^{[n]_s-1}) \right)=$ $$=
\begin{pmatrix}
1&1&1&\cdots&1\\
1&w&w^2&\cdots&w^{[n]_s-1}\\
1&w^2&\left(w^2\right)^2&\cdots&\left(w^{[n]_s-1}\right)^2\\
\vdots&\vdots&\vdots& &\vdots\\
1&w^{[n]_s-1}&\left(w^2\right)^{[n]_s-1}&\cdots&\left(w^{[n]_s-1}\right)^{[n]_s-1}\\
\end{pmatrix}.$$ “$\Rightarrow$” Suppose that $rank\left( \mathcal{N} \right)\neq [n]_s$. Then we have $$(*)\qquad \ \ det(\mathcal{N}):=\displaystyle\prod\limits_{0\leq i< j\leq [n]_s-1}(w^j-w^i)=0,$$ that is, $w^{j-i}=1$ for some $i,j$ such that $0\leq i< j\leq [n]_s-1$. So the order of $w$ must be less than or equal to $[n]_s-1$. Thus we get $[n]_s\leq q^{[k]_s}-1= ord(w)\leq [n]_s-1$, which is impossible. Hence $det(\mathcal{N})\neq 0$, i.e. $rank\left( V_{[n]_s}^{Id}(w,w^2,\dots,w^{[n]_s-1}) \right) = [n]_s$.
\
“$\Leftarrow$” Note that $rank\left( V_{[n]_s}^{Id}(w,w^2,\dots,w^{[n]_s-1}) \right)=[n]_s$ implies by $(*)$ that $w^{j-i}\not=1$ for any $i,j$ such that $0\leq i< j\leq [n]_s-1$. Thus $q^{[k]_s}-1=ord(w)> [n]_s-1$, i.e. $[n]_s\leq q^{[k]_s}-1$.
Finally, let us give here an application of the above arguments to obtain a lower bound for the distance of a skew constacyclic code by finding roots of polynomials in $\F_q[x]$ instead of $\F_q[t;\theta]$.
\[corchete:log\] With the same notation as above, let $\mathscr{C}\subseteq \F^n_q$ be a skew $\theta$-module code with generator polynomial $g(t)\in\F_q[t;\theta]$ of degree $k$. If $$k\geq \log_{p^s}\left[1+\left(\frac{p^s-1}{r}\right)\cdot \log_p\left(\frac{(p^s)^n+p^s-2}{p^s-1}\right)\right]$$ and $g^{[]_s}(\omega^{l+ci})=0$ for $i=0,\dots,\Delta-2$, some $l\in\Z_{\geq 0}$ and $c$ such that $(c,q^{[k]_s}-1)=1$, where $\F_{q^{[k]_s}}^*=\langle \omega \rangle$, then $d(\mathscr{C})\geq \Delta$.
Note that
$$k\geq \log_{p^s}\left[1+\left(\frac{p^s-1}{r}\right)\cdot \log_p\left(\frac{(p^s)^n+p^s-2}{p^s-1}\right)\right]
\Longleftrightarrow$$ $$r\left(\frac{(p^s)^k-1}{p^s-1}\right)\geq \log_p\left(1+[n]_s \right)
\Longleftrightarrow
p^{r\cdot [k]_s} \geq \left(1+[n]_s \right)\Longleftrightarrow
q^{[k]_s}-1 \geq [n]_s\ .$$
Thus $ord(\omega)=q^{[k]_s}-1\geq [n]_s$. Furthermore, since $(c,q^{[k]_s}-1)=1$, we have $$\mathcal{N}_j^{id}(\omega^c)=(\omega^c)^j\not= 1, \text{ for }j=1,\dots,[n]_s-1.$$ From Remark \[remark\] we know that $\mathscr{C}^{[]_s}\subseteq\F_q^{[n]_s}$ is a $\theta$-module code with generator polynomial $g^{[]_s}(x)\in\F_q[x]$, $\F_{q^{[k]_s}}^*=\langle \omega \rangle$ with $\omega\in\F_{q^{[k]_s}}$ and by hypothesis $g^{[]_s}(\omega^{l+ci})=0$, for $i=0,\dots,\Delta-2$ and some $l\in\Z_{\geq 0}$. Then by Theorem \[corchete:2\] and [@TT2 Theorem 3.9] we conclude that $d(\mathscr{C})\geq d(\mathscr{C}^{[]_s})\geq \Delta$.
Conclusion
==========
In this work, we consider skew constacyclic codes and some of the main algebraic and geometric properties of their dual codes. First of all, we prove again in an easy and direct way that the dual code of a skew $\alpha$-cyclic code is a skew $\alpha^{-1}$-cyclic code and we show explicitly how to build its generator polynomial. Secondly, we observe that the parity check matrix of a skew $\alpha$-cyclic code has a very special form because its columns are related to orbits of points in projective spaces via semi-linear maps. Moreover, a generalization to the non-commutative case of a result on $1$-generator Quasi Twisted (QT) codes is given and by a commutative method due to A. Leroy we prove two lower bounds for the distance of skew $\theta$-module codes, in particular for skew constacyclic codes. Finally, from a computer point of view, some MAGMA programs are given as application of the main theoretical results and by them a table is constructed, where for small fields some examples of known linear codes with BKLC distance are constructed as $1$-generator skew QT codes.
[2]{} W. Bosma, J. Cannon and C. Playoust, *The Magma algebra system. I. The user language*, J. Symbolic Comput. [**24**]{} (1997), no. 3-4, 235–265. D. Boucher, W.Geiselmann and F. Ulmer, *Skew-cyclic codes*, Appl. Algebra Engrg. Comm. Comput. [**18**]{} (2007), no. 4, 379–389. D. Boucher and F. Ulmer, *Codes as modules over skew polynomial rings*, Cryptography and coding, Lecture Notes in Comput. Sci. [**5921**]{}, Springer, Berlin, 2009, 38–55. D. Boucher and F. Ulmer, *A note on the dual codes of module skew codes*, Cryptography and coding, Lecture Notes in Comput. Sci. [**7089**]{}, Springer, Heidelberg, 2011, 230–243. D. Boucher and F. Ulmer, *Self-dual skew codes and factorization of skew polynomials*, J. Symbolic Comput. [**60**]{} (2014), 47–61. N. Fogarty and H. Gluesing-Luerssen, *A circulant approach to skew-constacyclic codes*, Finite Fields Appl. [**35**]{} (2015), 92–114. T. Maruta, *A geometric approach to semi-cyclic codes*, Advances in finite geometries and designs (Chelwood Gate, 1990), Oxford Sci. Publ., Oxford Univ. Press, New York (1991), 311–318. T. Maruta, M. Shinohara and M. Takenaka, *Constructing linear codes from some orbits of projectivities*, Discrete Math. [**308**]{} (2008), no. 5–6 , 832–841. A. Leroy, *Noncommutative polynomial maps*, J. Algebra Appl. [**11**]{} (2012), no. 4, 1250076, 16 pp. L.F. Tapia Cuitiño and A.L. Tironi, *Dual codes of product semi-linear codes*, Linear Algebra Appl. [**457**]{} (2014), 114–153. L.F. Tapia Cuitiño and A.L. Tironi, *Some properties of skew codes over finite fields*, (2015), arXiv:1507.02726.
[^1]: During the preparation of this paper in the framework of the Project Anillo ACT 1415 PIA CONICYT, the authors were partially supported by Proyecto VRID N. 214.013.039-1.OIN. Moreover, the first author was partially supported by CONICYT-PCHA/Magíster Nacional año 2013 - Folio: 221320380
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We solve Briot and Bouquet problem on the existence of non-monodromic (multivalued) solutions for singularities of differential equations in the complex domain. The solution is an application of hedgehog dynamics for indifferent irrational fixed points. We present an important simplification by only using a local hedgehog for which we give a simpler and direct construction of quasi-invariant curves which does not rely on complex renormalization.'
address: 'CNRS, IMJ-PRG, Paris 7, Boîte courrier 7012, 75005 Paris Cedex 13, France'
author:
- 'Ricardo Pérez-Marco'
date: 'January, 2018'
title: Solution to Briot and Bouquet problem on singularities of differential equations
---
Introduction.
=============
We prove the following Theorem:
Let $f(z)=e^{2\pi i {\alpha}} z +{{\mathcal O}}(z^2)$, ${\alpha}\in {{\mathbb R}}-{{\mathbb Q}}$ be a germ of holomorphic diffeomorphism with an indifferent irrational fixed point at $0$.
There is no orbit of $f$ distinct from the fixed point at $0$ that converges to $0$ by positive or negative iteration by $f$.
This Theorem solves the question of C. Briot and J.-C. Bouquet on singularities of differential equations from 1856 ([@BB]), as well as questions of H. Dulac (1904, [@D1], [@D2]), É. Picard (1896, [@P]), P. Fatou (1919, [@F]), and two more recent conjectures of M. Lyubich (1986, [@Lyu]).
The Theorem is trivial when the fixed point is linearizable, so, for the rest of the article, we assume that $f$ is not linearizable.
The main difficulty is to understand the non-linearizable dynamics. The proof relies on hedgehogs and their dynamics discovered by the author in [@PM5]. More precisely, we have from [@PM5] the existence of hedgehogs:
\[thm\_hedgehogs\] Let $U$ be a Jordan neighborhood of $0$ such that $f$ and $f^{-1}$ are defined and univalent on $U$, and continuous on $\bar U$.
There exists a hedgehog $K$ with the following properties:
- $0\in K \subset \bar U$
- $K$ is a full, compact and connected set.
- $K\cap \partial U \not= \emptyset$.
- $f(K)=f^{-1}(K)=K$.
Moreover, $f$ acts continuously on the space of prime-ends of ${{\mathbb C}}-K$ and defines an homeomorphism of the circle of prime-ends with rotation number ${\alpha}$.
In the proof we only need to consider local hedgehogs, i.e. a hedgehog associated to a small disk $U={{\mathbb D}}_{r_0}$ with $r_0>0$ small enough. Let $K_0$ be the hedgehog associated to ${{\mathbb D}}_{r_0}$. The two following Theorems imply our main Theorem.
\[thm\_unif\] Let $(p_n/q_n)_{n\geq 0}$ be the sequence of convergents of $\alpha$. We have $$\lim_{n\to +\infty} f_{/K_0}^{\pm q_n} ={\hbox {\rm id}}_{K_0} \ ,$$ where the convergence is uniform on $K_0$.
Therefore all points of the hedgehog are uniformly recurrent, and no point on the hedgehog distinct from $0$ converges to $0$ by positive or negative iteration by $f$.
\[thm\_oscullating\] Let $z_0 \in U-K_0$ such that the positive, resp. negative, orbit $(f^n(z_0))_{n\geq 0}$, resp. $(f^{-n}(z_0))_{n\geq 0}$, accumulates a point on $K_0$. Then this orbit accumulates all $K_0$, $$K_0\subset {\overline{(f^{n}(z_0))_{n\geq 0}}} \ \ ({\hbox{\rm {resp.}}} K_0\subset {\overline{(f^{-n}(z_0))_{n\geq 0}}} ) \ .$$
In particular this implies that if such an orbit $(f^{n}(z_0))_{n\geq 0}$ (resp. $(f^{-n}(z_0))_{n\geq 0}$) accumulates $0\in K_0$ then it cannot converge to $0$. Note that if $f$ is not linearizable then it is clear that $0\in \partial K_0$. Indeed one can prove that the hedgehog $K_0$ has empty interior and $K_0=\partial K_0$, but we don’t need to use this fact. We can just prove the previous Theorem for $\partial K_0$.
The proof of these two Theorems are done by constructing quasi-invariant curves near the hedgehog. These are Jordan curves surrounding the hedgehog and almost invariant by high iterates of the dynamics. The quasi-invariance property is obtained for the Poincaré metric of the complement of the hedgehog in the Riemann sphere.
Therefore, it is enough to carry out the construction for local hedgehogs, and for these we have a direct and simpler construction of quasi-invariant curves, that does not rely on complex renormalization techniques. Classical one real dimensional estimates for smooth circle diffeomorphism combined with an hyperbolic version of Denjoy-Yoccoz Lemma in order to control the complex orbits for analytic circle diffeomorphisms, are enough. This gives an important simplification for local hedgehogs of the proof of the main Theorem that was announced in [@PM2].
Historical introduction on Briot and Bouquet problem.
=====================================================
In 1856 C. Briot and J.-C. Bouquet published a foundational article [@BB] on the local solutions of differential equations in the complex domain. They are particularly interested in how a local solution determines uniquely the holomorphic function through analytic continuation. They consider a first order differential equation of a differential equation of the form $$\frac{d y}{dx} = f(x,y) \ ,$$ where $f$ is a meromorphic function of the two complex variables $(x,y)\in {{\mathbb C}}^2$ in a neighborhood of a point $(x_0,y_0)$. A. Cauchy proved his fundamental Theorem on existence and uniqueness of local solutions[^1]: If $f$ is finite and holomorphic in a neighborhood of $(x_0,y_0)$ then there exists a unique holomorphic local solution $y(x)$ satisfying the initial conditions $$y(x_0) = y_0 \ .$$
In their terminology, Briot and Bouquet talk about “solutions monogènes et monodromes”, “monogène” or monogenic meaning ${{\mathbb C}}$-differentiable, i.e. holomorphic, and “monodrome” or monodromic meaning univalued, since they also consider multivalued solutions with non-trivial monodromy at $x_0\in {{\mathbb C}}$.
Briot and Bouquet start their article by giving a simple proof of Cauchy Theorem by the majorant series method. Then they consider the situation where $f$ is infinite or has a singularity at $(x_0, y_0)$. They observe that even in Cauchy’s situation, we may get to such a point by a global analytic continuation of any solution. We assume for now on that $(x_0,y_0)=(0,0)$. Writing down $f$ as the quotient of two holomorphic germs $$f(x,y)=\frac{A(x,y)}{B(x,y)} \ ,$$ they study the situation when $A(0,0)=B(0,0)=0$ (they call these singularities “of the form $\frac{0}{0}$”). This is done in Chapter III, starting in section 75 of [@BB]. After a simple change of variables, the equation reduces to $$x\frac{dy}{dx}= a y+bx +{{\mathcal O}}(2) \ ,$$ and a discussion starts considering the different cases for different values of the coefficients $a,b \in {{\mathbb C}}$. They prove the remarkable Theorem that if $a$ is not a positive integer, then there always exists a holomorphic solution $y(x)$ in a neighborhood of $0$ vanishing at $0$ (Theorem XXVIII in section 80 of [@BB]). They show that this holomorphic solution is the only monodromic one and in their proof of uniqueness (in section 81) the equation is put in the form $$x\frac{dy}{dx}=y (a+{{\mathcal O}}(2)) \ .$$ In this last form the holomorphic solution corresponds to $y=0$.
After that they proceed to show that when the real part of $a$ is positive there are infinitely many non-monodromic solutions (section 82 in [@BB]), i.e. holomorphic solutions $y(x)$ that are multivalued around $0\in {{\mathbb C}}$.
They make the claim in section 85 in [@BB] that when the real part of $a$ is negative there are no other solutions, not even non-monodromic, other than the holomorphic solution found.
The proof of this statement contains a gap. Starting with the new form of the differential equation $$x\frac{dy}{dx}=y(a +{{\mathcal O}}_y(1))+xy \varphi(x,y) \ ,$$ they transform it into $$\frac{dy}{y} + (A+By+\ldots)\, dy = a\, \frac{dt}{t} + \psi (x,y) \, dt \ ,$$ where $A+By+\ldots$ is a holomorphic function of $y$ near $0$ and $\psi$ is holomorphic near $(0,0)$. Assuming by contradiction the existence of another solution, integration of the equation over a path from $x_1$ to $x$, $y_1=y(x_1)$, gives $$\log \left (\frac{y}{y_1} \right )+ (A(y-y_1)+\ldots )=\log \left ( \frac{x}{x_1}\right )^a
+\int_{x_1}^x \psi(x, y(x)) \, dx \ .$$ They pretend that this is of the form $$\log \left (\frac{y(x)}{y_1} \right )=\log \left ( \frac{x}{x_1}\right )^a
+{\epsilon},$$ where ${\epsilon}$ is a small quantity, vanishing for $x=x_1$, and very small when $x\to 0$, to get a contradiction using that for $\Re a<0$, $\Re \log (x/x_1)^a\to +\infty$ when $x\to 0$ but $\Re \log (y/y_1) \to -\infty$ if $y(x)\to 0$.
Unfortunately ${\epsilon}$ is not small because since $y(x)$ is not monodromic, the integral $$\int_{x_1}^x \psi(x, y(x)) \, dx$$ is not monodromic either, and if the path of integration spirals around $0$ it can get arbitrarily large.
É. Picard observes ([@P] Vol. II p.314 and p.317, 1893, see also Vol. III p.27 and 29, 1896) that with some implicit assumptions (that are not in [@BB]) the argument is correct if we approach $x=0$ along a path of finite length where the argument of $y(x)$ stays bounded or with a tangent at $0$, trying (not very convincingly) to rebate L. Fuchs that pointed out the error in [@Fu]. H. Poincaré does not mention the error in his article [@Po1] where he states Briot and Bouquet result without any restriction, and in his Thesis [@Po2] where he studies the case where the real part of $a$ is positive (and carefully avoids discussing further the other problematic case).
Picard, in his first edition of his “Traité d’Analyse” ([@P], Vol. III, page 30, 1896), casts no doubt about the correction of Briot and Bouquet statement:
*“Il resterait à démontrer que ces deux intégrales sont, en dehors de toute hypothèse, les seules qui passent par l’origine ou qui s’en rapprochent indéfiniment. Je dois avouer que je ne possède pas une démonstration rigoureuse de cette proposition, qui ne paraît cependant pas douteuse.”*[^2]
He refers to the two Briot-Bouquet holomorphic solutions $y(x)$ and $x(y)$. His belief is probably reinforced by the saddle picture for real solutions that clearly only exhibit two real solutions in ${{\mathbb R}}^2$ passing through the singularity.
A major progress came with the Thesis of H. Dulac published in 1904 in the Journal of the École Polytechnique [@D1]. He proves the existence of an infinite number of distinct non-monodromic solutions when $a$ is a negative rational number, thus proving than Briot and Bouquet original claim is always false in the rational situation. From the introduction of [@D1] we can read
*“$\ldots $ on sait depuis bien longtemps, qu’il n’existe que deux courbes intégrales réelles passant par l’origine. En est-il de même dans le champ complexe ? C’est une question qui restait en suspens et que les géomètres penchaient à trancher par l’affirmative (Picard, Traité d’Analyse, II (sic)[^3], p. 30). Or je prouve, au contraire, tout au moins dans le cas où $\alpha$ est rationnel, qu’il existe une infinité d’intégrales $y(x)$ s’annulant avec $x$ ($x$ tendant vers zéro suivant une loi convenable) $\ldots $”* [^4]
After Dulac’s result Picard changed the quoted text in later editions of his Traité d’Analyse ([@P], Vol. III, 3rd edition, p.30, 1928) into:
*“On a longtemps présumé que ces intégrales sont, en dehors de toute hypothèse, les seules qui passent par l’origine ou qui s’en rapprochent indéfiniment. Dans un excellent travail sur les points singuliers des équations différentielles, M. Dulac a démontré que la question était très complexe. Prenons, par exemple, l’équation $$x\frac{dy}{dx}+y(\nu +\ldots )=0 \ ,$$ où $\nu$ est positif, équation à laquelle peut toujours se ramener le cas où $\lambda$ est négatif. M. Dulac examine particulièrement le cas où $\nu$ est un nombre rationnel $p/q$, et montre qu’il y a alors, en général, une infinité d’intégrales pour lesquelles $x$ et $y$ tendent vers $0$.”* [^5]
Dulac insisted in his Thesis that he had no answer for the irrational case ([@D1] p.4):
*“1. $\nu$ est irrationnel. On a un col. $H(x,y)$ existe formellement, mais est divergent, au moins dans certains cas. S’il y a des intégrales pour lesquelles $x$ et $y$ tendent simultanément vers $0$, et si l’on désigne par $\omega$ et $\theta$ les arguments de $x$ et $y$, quels que soient $m$ et $n$, $|x^my^n\omega|$ et $|x^m y^n\theta |$ croissent indéfiniment. Je ne puis me prononcer sur l’existence de ces intégrales.”*[^6]
The expression $yx^\nu H(x,y)$ is a formal first integral of the solutions and he discuss its convergence in p.20. It is well known to Dulac that convergence of $H$ solves the problem.
Then $30$ years later he recalls that the problem remains unsolved ([@D2] p.31):
*“Dans le cas $2$ ($\nu$ irrationnel, $h(x,y)$ divergent), on ne sait s’il existe des solutions nulles autres que $x=0$, $y=0$. Ce sont là deux questions qu’il y aurait grand intérêt à élucider.”*[^7]
Many results obtained by these distinguished geometers of the XIXth century where rediscovered in modern times, sometimes with a different point of view or language. The original problem of singularities of differential equations of the form $\frac{0}{0}$ (according to Briot and Bouquet terminology) is equivalent to study solutions of the holomorphic vector field $X=(B,A)$ near $(0,0) \in {{\mathbb C}}^2$,
$$\begin{cases}
\dot x &= B(x,y) \\
\dot y &= A(x,y)
\end{cases}$$
The local geometry corresponds also to the study the holomorphic foliations on ${{\mathbb C}}^2$ near the singular point $(0,0)$ defined by the differential form $$A(x,y) dx - B(x,y) dy =0 \ .$$ The situation of Briot and Bouquet problem corresponds to an irreducible singularity with a non-degenerate linear part, $$({\alpha}y+{{\mathcal O}}(2)) dx + (x +{{\mathcal O}}(2)) dy =0 \ ,$$ where $-{\alpha}=a$ is Briot and Bouquet coefficient.
When ${\alpha}\in {{\mathbb C}}-{{\mathbb R}}_+$, and ${\alpha}$ is neither a negative integer nor the inverse of a negative integer, we are in the Poincaré domain and the singularity is equivalent to the linear one. When ${\alpha}$ is a negative integer or its inverse, then we can conjugate the singularity to a finite Poincaré-Dulac normal form (see [@Ar] section 24). We assume ${\alpha}$ real and positive ${\alpha}>0$, which defines, in modern terminology, a singularity in the Siegel domain. The singularity is formally linearizable, but the convergence of the linearization presents problems of Small Divisors. Precisely in this situation Dulac already proved in his Thesis the existence of non-linearizable singularities in section 12. This is a notable achievement that anticipates in several decades the non-linearization results for indifferent fixed points. The existence of Briot and Bouquet holomorphic solution proves the existence of two leaves of the holomorphic foliation crossing transversally at $(0,0)$. This means that the singularity can be put into the form $${\alpha}y (1+{{\mathcal O}}(2)) dx + x (1+{{\mathcal O}}(2)) dy =0 \ .$$ Again $y=0$ corresponds to the Briot and Bouquet holomorphic solution. It is now easy to make the link with the original Briot and Bouquet $\frac{0}{0}$ singulatities of differential equations. Each solution $y(x)$, distinct from the only monodromic solution $y(x)=0$, with initial data $(x_0,y_0)$ close to $(0,0)$, has a graph over the $x$-axes that corresponds to the leaf of the foliation passing through the point $(x_0,y_0)$. The multivaluedness or non-monodromic character of the solution can be seen in the intersection of that leaf with a transversal $\{x=x_0\}$. The $y$-coordinates of these points of intersection give the different values taken by the non-monodromic solution that are obtained by following a path in the leaf that projects onto the $x$-axes into a path circling around $x=0$.
The topology of the foliation is understood through a holonomy construction (see [@MM], and for the rational case see [@C]): Taking a transversal $\{x=x_0\}$ and lifting the circle $C(0, |x_0|) \subset \{y=0\}$ in nearby leaves, the return map following this lift in the negative orientation, defines a germ of holomorphic diffeomorphism in one complex variable with a fixed point at $(x_0,0) \subset \{x=x_0\}$. Taking a local chart in this complex line, we have a local holomorphic diffeomorphism $f\in {\hbox{\rm Diff}} ({{\mathbb C}}, 0)$, $f(0)=0$, and linearizing the equations we can compute its linear part at $0$, $$f(z) =e^{2\pi i {\alpha}} z +{{\mathcal O}}(z^2) \ .$$ (to see this, note that $yx^{{\alpha}}$ is a first integral of the linearized differential form, thus is invariant of the solutions in the first order) Thus we get a germ of holomorphic diffeomorphism with an indifferent irrational fixed point. It is obvious from the classical point of view that the local dynamics near $0$ of this return map contains the information about the non-monodromic solutions starting at $x=x_0$. Thus we transform our original problem into a problem of holomorphic dynamics. Note that we can also reconstruct all the foliation and a neighborhood of $(0,0)\in {{\mathbb C}}^2$ minus the leave $\{y=0\}$ by continuing the continuing the complex leaves from the transversal. J.-F. Mattei and P. Moussu proved in [@MM] that two singularities in the Siegel domain with conjugated holonomies are indeed conjugated in ${{\mathbb C}}^2$ by “pushing” the conjugacy along these leaves and using Riemann removability Theorem in ${{\mathbb C}}^2$. J.-Ch. Yoccoz and the author proved in [@PMY] that the set of dynamical conjugacy classes of holonomies is in bijection with the set of conjugacy classes of singularities in the Siegel domain. The rational case was previously treated by J. Martinet and J.-P. Ramis ([@MR1], [@MR2]) by identifying the conjugacy invariants. This establish a full dictionary of the two problems. In particular, an interesting corollary is that Brjuno diophantine condition is optimal for analytic linearization of the singularity.
For our problem, the existence of non-monodromic solutions vanishing with $x$ when $x\to 0$ following an appropriate path is equivalent to finding a leave that accumulates the singularity $(0,0)$ but distinct from the Briot and Bouquet leaves $\{x=0\}$ and $\{y=0\}$ and a path $\gamma$ on this leave converging to $(0,0)$. This path $\gamma$ projects properly in the $\{y=0\}$ plane into a spiral around $(0,0)$ and converging to $(0,0)$. The path $\gamma$ is homothopic in the leave to a path above $C(0,|x_0|)$ such that the iterates of the return map converve to $(x_0,0)$. Since $\pi_1({{\mathbb C}}^*) \approx {{\mathbb Z}}$, this gives an orbit of the return map that has a positive or negative orbit converging to the indifferent fixed point. Conversely, if we have such an orbit of the return map, we can push homothopicaly the path in the leave close to $\{x=0\}$ to make it converging to $(0,0)$ (just using continuity of the foliation).
When ${\alpha}\in {{\mathbb R}}_+-{{\mathbb Q}}$, Briot and Bouquet non-existence of non-monodromic solutions vanishing at $0$ is equivalent to the existence of an orbit distinct from $0$ that converges to $0$ by iteration by the return map $f$ or $f^{-1}$.
Since linearizable dynamics don’t have this property, we see that C-L. Siegel linearization theorem ([@S], 1942) shows that Briot and Bouquet statement is true when ${\alpha}\in {{\mathbb R}}_+-{{\mathbb Q}}$ satisfies the arithmetic linearization condition that was improved later by A.D. Brjuno ([@Br]) to the so called Brjuno’s condition $$\sum_{n=0}^{+\infty } \frac{\log q_{n+1}}{q_n} <+\infty \ .$$ The sequence of $(p_n/q_n)_{n\geq 0}$ are the convergents of ${\alpha}$. The positive answer to Briot and Bouquet question in the linearizable case, that corresponds to $H(x,y)$ being convergent in Dulac’s notation, was already well known to Dulac in [@D1].
Indeed the non-existence of non-monodromic for singularities of differential equations were well understood in the linearizable case, since H. Poincaré [@Po1] because linearization is equivalent to the existence of a first integral of the system of the form (see [@D1] section \[\]) $$I(x,y)=yx^{\alpha}H(x,y) \ .$$
Note also that to have non-monodromic solutions $y(x)$ that accumulate into (but not converge to) $0$ when $x\to 0$ is a simpler problem that is equivalent for the monodromy dynamics to have an orbit that accumulates $0$ by positive or negative iteration. This was solved in general in [@PM5] by the discovery of hedgehogs and the result that almost all points in the hedgehog for the harmonic measure have a dense orbit in the hedgehog.
What remains to be elucidated for the Briot and Bouquet problem is the non-linearizable case, and more precisely the following problem:
Let ${\alpha}\in {{\mathbb R}}-{{\mathbb Q}}$ and $f(z)=e^{2\pi i {\alpha}} z +{{\mathcal O}}(z^2)$ be a germ of holomorphic diffeomorphism with an indifferent fixed point at $0$. Does there exists $z_0 \not= 0$ such that $$\lim_{n\to +\infty } f^n(z_0) =0 \ .$$
P. Fatou was confronted to this problem in his pioneer study of the dynamics of rational functions ([@F], 1919) without knowing the relation to Briot and Bouquet problem. About fixed points (“points doubles” in Fatou’s terminology) of holomorphic germs, which are indifferent, irrational and non-linearizable, Fatou writes [@F] p.220-221:
*“Il reste à étudier les points doubles dont le multiplicateur est de la forme $e^{i\alpha}$, $\alpha$ étant un nombre réel inconmensurable avec $\pi$. Nous ne savons que fort peu de choses sur ces points doubles, dont l’étude du point de vue qui nous occupe paraît très difficile. ($\ldots$) Existe-t-il alors des domaines dont les conséquents tendent vers le point double ? Nous ne pouvons actuellement ni en donner d’exemple, ni prouver que la chose soit impossible $\ldots$”* [^8]
Fatou’s question is related to the question of the non-existence of wandering components of the Fatou set for rational functions. This was only proved in 1985 by D. Sullivan [@Su]. Note that we do indeed have domains (that are not Fatou components) converging by iteration to rational indifferent fixed point as the local analysis of the rational case shows.
The non-existence of domains converging to an indifferent irrational fixed point was also conjectured by M. Lyubich in [@Lyu] p.73 (Conjecture 1.2), apparently unaware of Fatou’s question. Lyubich also conjectured (Conjecture 1.5 (a) [@Lyu] p.77) that for any indifferent irrational non-linearizable fixed point there is a critical orbit that converges to the fixed point.
The author proved in [@PM5] the Moussu-Dulac Criterium : $f$ is not linearizable if and only if $f$ has an orbit accumulating the fixed point $0$. We may think that this could give support to the existence of a converging orbit. The discovery of hedgehogs gave new tools for the understanding of the non-linearizable dynamics. Indeed, hedgehogs are the central tool in the final solution of all this problems:
\[thm\_main\] There is no orbit converging by positive or negative iteration to an indifferent irrational fixed point of an holomorphic map and distinct from the fixed point.
Therefore, the Briot and Bouquet problem has a positive solution in the irrational case. The questions of Dulac, Picard, Fatou are solved. Lyubich’s Conjecture 1.2 in [@Lyu] has a positive answer, but conjecture 1.5 (a) in [@Lyu] is false: For a generic rational function, there is no critical point converging to an indifferent irrational non-linearizable periodic orbits. There may be pre-periodic critical points to this orbit, but this is clearly non-generic. We may formulate a proper conjecture that has better chances to hold true:
Let $f$ be a rational function of degree $2$ or more, with an indifferent irrational non-linearizable fixed point $z_0$. There exists a critical point $c_0$ of $f$, such that $$\lim_{n\to +\infty} \frac1n \sum_{j=0}^{n-1} \delta_{f^j(c_0)} \to \delta_{z_0} \ .$$
Theorem \[thm\_main\] was announced in [@PM2] and a complete proof was given in the unpublished manuscript [@PM6]. The proof given here concentrates on this particular Theorem and the solution of Briot and Bouquet problem, and not the many other properties of general hedgehog’s dynamics. The proof follows the same lines as in [@PM6], but we have incorporated several new ideas that greatly improve and simplify the technical part of construction of quasi-invariant curves that are fundamental in the study of the hedgehog dynamics. It was recently noticed in [@PM7] an hyperbolic interpretation of Denjoy-Yoccoz Lemma that controls orbits of an analytic circle diffeomorphism $g$ in a complex neighborhood of the circle. Then, when we control the non-linearity $||D\log Dg||_{C^0}$ of $g$, we can construct directly the quasi-invariant curves without complex renormalization. The second observation if that in the proof of Theorem \[thm\_main\] we can work with local hedgehogs (small hedgehogs). Then the associated circle diffeomorphism has a small non-linearity and the construction of quasi-invariant curves is easier.
Analytic circle diffeomorphisms.
================================
Notations.
----------
We denote by ${{\mathbb T}}={{\mathbb R}}/{{\mathbb Z}}$ the abstract circle, and ${{\mathbb S}}^1 =
E ({{\mathbb T}})$ its embedding in the complex plane ${{\mathbb C}}$ given by the exponential mapping $E(x)=e^{2\pi i x}$.
We study analytic diffeomorphisms of the circle, but we prefer to work at the level of the universal covering, the real line, with its standard embedding ${{\mathbb R}}\subset {{\mathbb C}}$. We denote by $D^\omega ({{\mathbb T}})$ the space of non decreasing analytic diffeomorphisms $g$ of the real line such that, for any $x\in {{\mathbb R}}$, $g(x+1)=g(x)+1$, which is the commutation to the generator of the deck transformations $T(x)=x+1$. An element of the space $D^\omega ({{\mathbb T}})$ has a well defined rotation number $\rho (g)\in {{\mathbb R}}$. The order preserving diffeomorphism $g$ is conjugated to the rigid translation $T_{\rho(g)}: x\mapsto x+\rho(g)$, by an orientation preserving homeomorphism $h:{{\mathbb R}}\to {{\mathbb R}}$, such that $h(x+1)=h(x)+1$.
For $\Delta >0$, we note $B_{\Delta} =\{ z\in {{\mathbb C}}; |\Im z | < \Delta \}$, and $A_\Delta = E(B_\Delta )$. The subspace $D^\omega ({{\mathbb T}}, \Delta )\subset D^\omega ({{\mathbb T}})$ is composed by the elements of $D^\omega ({{\mathbb T}})$ which extend analytically to a holomorphic diffeomorphism, denoted again by $g$, such that $g$ and $g^{-1}$ are defined on a neighborhood of $\bar B_\Delta$.
Real estimates.
---------------
We refer to [@Yo3] for the results on this section. We assume that the orientation preserving circle diffeomorphism $g$ is $C^3$ and that the rotation number $\alpha =\rho (g)$ is irrational. We consider the convergents $(p_n /q_n)_{n\geq 0}$ of $\alpha$ obtained by the continued fraction algorithm (see [@HW] for notations and basic properties of continued fractions).
For $n\geq 0$, we define the map $g_n (x) =g^{q_n}(x)-p_n$ and the intervals $I_n(x)=[x, g_n(x)]$, $J_n(x)=I_n(x) \cup I_n(g_n^{-1}(x)) =[g_n^{-1}(x), g_n(x)]$. Let $m_n (x)=g^{q_n} (x)-x-p_n=\pm |I_n(x)|$, $M_n =\sup_{{{\mathbb R}}} |m_n (x)|$, and $m_n =\min_{{{\mathbb R}}} |m_n (x)|$. Topological linearization is equivalent to $\lim_{n\to +\infty } M_n =0$. This is always true for analytic diffeomorphisms by Denjoy’s Theorem, that holds for $C^1$ diffeomorphisms such that $\log Dg$ has bounded variation.
Since $g$ is topologically linearizable, combinatorics of the irrational translation (or the continued fration algorithm) shows:
\[lem\_comb\] Let $x\in {{\mathbb R}}$, $0\leq j < q_{n+1}$ and $k\in {{\mathbb Z}}$ the intervals $g^j\circ T^k(I_n(x))$ have disjoint interiors, and the intervals $g^j\circ T^k(J_n(x))$ cover ${{\mathbb R}}$ at most twice.
We have the following estimates on the Schwarzian derivatives of the iterates of $f$, for $0\leq j\leq q_{n+1}$, $$\left |S g^j (x)\right | \leq \frac{M_n e^{2V}S}{|I_n(x)|^2} \ ,$$ with $S=||Sg||_{C^0({{\mathbb R}})}$ and $V=\hbox{\rm {Var}} \log Dg$.
This implies a control of the non-linearity of the iterates (Corollary 3.18 in [@Yo3]):
For $0\leq j\leq 2q_{n+1}$, $c=\sqrt{2S} e^V$, $$|| D \log Dg^j ||_{C^0({{\mathbb R}})} \leq c \, \frac{M_n^{1/2}}{m_n} \ .$$
These give estimates on $g_n$. More precisely we have (Corollary 3.20 in [@Yo3]):
\[prop\_estimate\] For some constant $C >0$, we have $$||\log Dg_n ||_{C^0({{\mathbb R}})}\leq C M_n^{1/2} \ .$$
\[cor\_1\] For any ${\epsilon}>0$, there exists $n_0\geq 1$ such that for $n\geq n_0$, we have $$||Dg_n -1||_{C^0({{\mathbb R}})}\leq {\epsilon}\ .$$
Take $n_0 \geq 1$ large enough so that for $n\geq n_0$, $C M_n^{1/2} <\min (\frac{2}{3}{\epsilon}, \frac12) $, then use Proposition \[prop\_estimate\] and $\left |e^w-1 \right | \leq \frac{3}{2} |w|$ for $|w|<1/2$.
\[cor\_estimate\] For any ${\epsilon}>0$, there exists $n_0\geq 1$ such that for $n\geq n_0$, for any $x\in {{\mathbb R}}$ and $y\in I_n(x)$ we have $$1-{\epsilon}\leq \frac{m_n(y)}{m_n(x)}\leq 1+{\epsilon}\ .$$
We have $D m_n(x)=Dg_n(x)-1$, and $$\left | m_n(y)-m_n(x)\right | \leq ||Dm_n ||_{C^0({{\mathbb R}})} |y-x| \leq ||Dg_n -1||_{C^0({{\mathbb R}})} |m_n(x)| \ .$$ We conclude using Lemma \[cor\_1\].
Hyperbolic Denjoy-Yoccoz Lemma.
===============================
With these real estimates for the iterates, and, more precisely, a control on the non-linearity, we can use them to control orbits in a complex neighborhood. We give here a version of Denjoy-Yoccoz lemma (Proposition 4.4 in [@Yo3]) that is convenient for our purposes.
Given ${\Delta}>0$, we consider $g\in D^\omega ({{\mathbb T}}, {\Delta})$ such that $\inf_{B_{\Delta}} \Re Dg >0$ so that $\log Dg$ is a well defined univalued holomorphic function in $B_{\Delta}$. Given $g\in D^\omega ({{\mathbb T}})$ we get always this for a ${\Delta}>0$ small enough (as in [@Yo3]), but here we don’t need to make the assumption that for a given $g$, ${\Delta}$ is small enough.
We do assume that we have a small non-linearity in $B_{\Delta}$, more precisely, $$\tau=|| D\log Dg ||_{C^0(B_{\Delta})} <1/9 \ .$$
\[lem\_DenjoyYoccoz\] Let $n_0 \geq 1$ large enough such that for all $n\geq n_0$, $M_n <{\Delta}/2$.
For $x_0\in {{\mathbb R}}$, let $0<y_0 \leq 1$ and $$z_0= x_0+im_n(x_0) y_0 \ .$$ Then for $0\leq j\leq q_{n+1}$, $y_j\in {{\mathbb C}}$, $\Re y_j>0$, is well defined by $$z_j=g^j(z_0)=g^j(x_0) + i m_n(g^j(x_0)) y_j \ ,$$ and we have $$|y_j-y_0|\leq \frac{3}{4} \, y_0 \ .$$
For $0<t\leq 1$ we define more generally $$z_{0,t}= x_0+im_n(x_0) t y_0 \ ,$$ and we prove that $y_{j,t}\in {{\mathbb C}}$, $\Re y_{j,t}>0$, is well defined by $$z_{j,t}=g^j(z_{0,t})=g^j(x_0) + i m_n(g^j(x_0)) y_{j,t} \ ,$$ and that we have $$|y_{j,t}-y_{0,t}|\leq \frac{3}{4} \, y_{0,t} \ .$$ Note that this last inequality implies $\Re y_{j,t} \leq \frac{7}{4} y_{0,t}$. The lemma corresponds to the case $t=1$.
We prove this result by induction on $0\leq j < q_{n+1}$ starting from $j=0$ for which the result is obvious. Assuming it has been proved up to $0\leq j-1 < q_{n+1}$, then we have $$0<\Im z_{j-1,t} \leq M_n \Re y_{j-1,t} \leq M_n \frac{7}{4} \, t y_0 < \frac{7}{8} \, {\Delta}<{\Delta}\ ,$$ so $z_{j-1,t} \in B_{\Delta}$ and we can iterate once more and $z_{j,t} = g(z_{j-1,t})$ is well defined. We need to prove the estimate for $y_{j,t}$. By the chain rule we have $$\log D g^j (z_{0,t}) =\sum_{l=0}^{j-1} \log D g (z_{l,t}) \ .$$ Therefore, we have $$\begin{aligned}
\left | \log D g^j (z_{0,t}) -\log D g^j (x_0)
\right | &\leq \sum_{l=0}^{j-1}
\left | \log D g (z_{l,t}) -\log D g (x_l) \right | \\
&\leq \tau \sum_{l=0}^{j-1} |z_{l,t} -x_l| \\
&\leq \tau \sum_{l=0}^{j-1} |m_n(x_l)| |y_{l,t}| \\
&\leq \frac{7}{4} \tau ty_{0} \sum_{l=0}^{j-1} |m_n (x_l)| \\
&\leq \frac{7}{4} \tau \sum_{l=0}^{j-1} |m_n (x_l)| \ .\end{aligned}$$
Considering the $j$-iterate of $g$ on the interval $I_n(x_0)$, we obtain a point $\zeta \in ]x_0, g^{q_n} (x_0)-p_n[$ such that, $$Dg^j (\zeta ) =\frac {m_n (x_j)}{m_n (x_0)} \ ,$$ and $$\left | \log D g^j (\zeta ) -\log D g^j (x_0)
\right | \leq \tau |m_n (x_0)|\leq \tau
\sum_{l=0}^{j-1} |m_n (x_l)| \ .$$ Adding the two previous inequalities, we have $$\left | \log D g^j (z_{0,t}) -\log \frac {m_n (x_j)}
{m_n (x_0)} \right | \leq \frac{11}{4} \tau
\sum_{l=0}^{j-1} |m_n (x_l)| \ .$$ The intervals $I_n(x_l)$, $0\leq l < q_{n+1}$, being disjoint modulo $1$, we have $$\sum_{l=0}^{q_{n+1}-1} |m_n (x_l)| < 1 \ .$$ So we obtain $$\left | \log D g^j (z_{0,T}) -\log \frac {m_n (x_j)}{
m_n (x_0) } \right | \leq \frac{11}{4} \, \tau \ ,$$ and taking the exponential (using $|e^w-1|\leq 3/2 |w|$, for $|w| <1/2$, since $\tau <1/9$ and $\frac{11}{4}\, \tau <\frac12$), we have $$\left | D g^j (z_{0,t})-\frac {m_n (x_j)}{m_n (x_0) }
\right | \leq \frac{33}{8} \, \tau \frac{m_n (x_j)}{ m_n (x_0) }.$$ Now, integrating along the vertical segment $[x_0, z_{0,t}]$ we get $$\left | g^j (z_{0,t}) -g^j (x_0) -i y_0 m_n (x_j)
\right | \leq \frac{33}{8} \, \tau y_{0,t} |m_n (x_j)| \ ,$$ which, using $\tau <1/9$, finally gives $$| y_{j,t} - y_{0,t} | \leq \frac{11}{24} \, y_{0,t} <\frac34 \, y_{0,t} \ .$$
Flow interpolation in ${{\mathbb R}}$.
--------------------------------------
Since $g$ is analytic, from Denjoy’s Theorem we know that $g_{/{{\mathbb R}}}$ is topologically linearizable, i.e. there exists an non-decreasing homeomorphism $h: {{\mathbb R}}\to {{\mathbb R}}$, such that for $x\in {{\mathbb R}}$, $h(x+1)=h(x)+1$, and $$h^{-1} \circ g \circ h =T_{\alpha} \ ,$$ where $T_\alpha : {{\mathbb R}}\to {{\mathbb R}}$, $x\mapsto x+\alpha $.
We can embed $g$ into a topological flow on the real line $(\varphi_t)_{t\in {{\mathbb R}}}$ defined for $t \in {{\mathbb R}}$ by $\varphi_t =h\circ T_{t\alpha }\circ h^{-1}$. When $g$ is analytically linearizable the diffeomorphisms of this flow are analytic circle diffeomorphisms, but in general, when $g$ is not analytically linearizable the maps $\varphi_t$ are only homeomorphism of the real line, although for $t \in {{\mathbb Z}}+ \alpha^{-1} {{\mathbb Z}}$, $\varphi_{t}$ is analytic since $\varphi_t$ is an iterate of $g$ composed by an integer translation. This can happen that for other values of $t$, where $\varphi_t$ can be an analytic diffeomorphism from the analytic centralizer of $g$ since $\varphi_t\circ g =g\circ \varphi_t$. We refer to [@PM3] for more information on this fact and examples of uncountable analytic centralizers for non-analytically linearizable dynamics. Now $(\varphi_t)_{t\in [0,1]}$ is an isotopy from the identity to $g$. The flow $(\varphi_t)_{t\in {{\mathbb R}}}$ is a one parameter subgroup of homeomorphisms of the real line commuting to the translation by $1$.
Flow interpolation in ${{\mathbb C}}$.
--------------------------------------
There are different complex extensions of the flow $(\varphi_t)_{t\in {{\mathbb R}}}$ suitable for our purposes. For each $n\geq 0$, we can extend this topological flow to a topological flow ${{\mathcal F}}_n$ in ${{\mathbb C}}$ by defining, for $z_0 =x_0 +i \, | m_n (x_0)| y_0 \in {{\mathbb C}}$, with $x_0, y_0
\in {{\mathbb R}}$, $$\varphi_{t}^{(n)} (z_0)=z_0(t)= \varphi_t(x_0) +i \, |m_n (\varphi_t(x_0))| y_0 \ .$$ We denote $\Phi^{(n)}_{z_0}$ the flow line passing through $z_0$, $$\Phi^{(n)}_{z_0} = (\varphi_{t}^{(n)} (z_0))_{t\in {{\mathbb R}}}.$$
Hyperbolic Denjoy-Yoccoz Lemma.
-------------------------------
We are now ready to give a geometric version of Denjoy-Yoccoz Lemma. We denote by $d_P$ the Poincaré distance in the upper half plane.
\[lem\_DYhyp\] Let $\Delta >0$ and $g\in D^{{\omega}} ({{\mathbb T}}, {\Delta})$ such that $$||D \ \log \ Dg ||_{C^0 (B_{{\Delta}} )} < 1/9 \ .$$ Let $n_0 \geq 1$ large enough such that for all $n\geq n_0$, $M_n <{\Delta}/2$. Let $z_0= x_0+i |m_{n}(x_0)| y_0$, with $0<y_0<1$, so $z_0 \in B_{{\Delta}}$. Then for $0\leq j\leq q_{n+1}$ we have that the $(g^j(z_0))$ piece of orbit follows at bounded distance the flow ${{\mathcal F}}_n$ for the Poincaré metric of the upper half plane. More precisely we have $$d_P(g^j(z_0), \varphi^{(n)}_j(z_0)) \leq C_0 \ ,$$ for some constant $C_0 >0$ (we can take $C_0=3$).
We just use Lemma \[lem\_DenjoyYoccoz\] reminding that the Poincaré metric in the upper half plane is given by $|ds| =\frac{|d\xi|}{\Im \xi}$ and $$\begin{aligned}
d_P(z_j , \varphi_j^{(n)}(z_0)) &\leq \int_{[z_j , \varphi_j^{(n)}(z_0)]} \frac{|d\xi|}{\Im \xi} \\
&\leq |m_{n}(x_j)|\, . \,|y_j-y_0|
\, \frac{1}{\inf_{\xi \in [z_j , \varphi_j^{(n)}(z_0)]} \Im \xi} \\
&\leq |m_{n}(x_j)|\, . \, |y_j-y_0| \, \frac{4}{|m_n(x_j)| \, y_0} \\
&\leq 4\, \frac{|y_j-y_0|}{y_0} \leq 3=C_0
\end{aligned}$$ where in the second inequality we used that $\Re y_j \geq \frac{1}{4} y_0$ which follow from $|y_j-y_0|\leq \frac{3}{4} y_0$ that we also used in the last inequality.
Quasi-invariant curves for local hedgehogs. {#sec_quasi}
===========================================
Now we construct quasi-invariant curves for $g$ under the previous assumptions: $g\in D^{\omega}({{\mathbb T}}, {\Delta})$ and $$\tau=|| D\log Dg ||_{C^0(B_{\Delta})} <1/9 \ .$$
\[Quasi-invariant curves\] \[thm\_quasi\] Let $g$ be an analytic circle diffeomorphism with irrational rotation number ${\alpha}$. Let $(p_n/q_n)_{n\geq 0}$ be the sequence of convergents of $\alpha$ given by the continued fraction algorithm.
Given $C_0>0$ there is $n_0 \geq 0$ large enough such that there is a sequence of Jordan curves $(\gamma_n)_{n\geq n_0}$ for $g$ which are homotopic to ${{\mathbb S}}^1$ and exterior to $\overline{{{\mathbb D}}}$ such that all the iterates $g^j$, $0\leq j\leq q_{n}$, are defined in a neighborhood of the closure of the annulus $U_n$ bounded by ${{\mathbb S}}^1$ and $\gamma_n$, and we have $${{\mathcal D}}_{P}(g^j(\gamma_n), \gamma_n) \leq C_0 \ ,$$ where ${{\mathcal D}}_P$ denotes the Hausdorff distance between compact sets associated to $d_P$, the Poincaré distance in ${{\mathbb C}}-\overline{{{\mathbb D}}}$. We also have for any $z\in \gamma_n$, $d_P(g^{q_{n}} (z), z) \leq C_0$, that is, $$||g^{q_{n}} -{\hbox{\rm id}}||_{C^O_P(\gamma_n)} \leq C_0 \ .$$
We choose the flow lines $\gamma_{n+1}=\Phi^{(n)}_{z_0}$, with $y_0 > 1/2$ and $n\geq n_0$ for $n_0\geq 1$ large enough, for the quasi-invariant curves of the Theorem. These flow lines are graphs over ${{\mathbb R}}$. Given an interval $I\subset {{\mathbb R}}$, we label $\tilde I^{(n)}$ the piece of $\Phi^{(n)}_{z_0}$ over $I$.
\[lem\_bounded\] There is $n_0\geq 1$ such that for $n\geq n_0$ and for any $x\in {{\mathbb R}}$, the piece $\tilde I_n^{(n)}(x)$ has bounded Poincaré diameter.
Let $z=x+i \, |m_n(x)| y_0$ be the current point in $\tilde I_n^{(n)}(x)$. We have $$dz=\left ( 1 \pm i\, (Dg_n(x)-1) y_0 \right ) \, dx \ .$$ For any ${\epsilon}_0 >0$, choosing $n_0\geq 1$ large enough, for $n\geq n_0$, according to Lemma \[cor\_1\] we have $$\left |\frac{dz}{dx} -1 \right | \leq {\epsilon}_0 \ .$$ Therefore, we have $$l_P (\tilde I_n^{(n)}(x_0))=\int_{\tilde I_n^{(n)}(x_0)} \frac{1}{|m_n(x)|\, y_0} \, |dz| \leq \int_{I_n(x_0)} \frac{1}{|m_n(x)|\, y_0} \, (1+{\epsilon}_0 )\, dx \ .$$ Now using Lemma \[cor\_estimate\] with ${\epsilon}= {\epsilon}_0$ and increasing $n_0$ if necessary, we have $$l_P (\tilde I_n^{(n)}(x))\leq \int_{I_n(x_0)} \frac{1}{|m_n(x_0)|\, y_0} \, \frac{1+{\epsilon}_0}{1-{\epsilon}_0} \, dx \leq \frac{1}{y_0} \frac{1+{\epsilon}_0}{1-{\epsilon}_0}
\leq 2 \, \frac{1+{\epsilon}_0}{1-{\epsilon}_0}\leq C\ .$$
We assume $n\geq n_0$ from now on in this section and the next one.
\[lem\_cover\] For $0\leq j < q_{n+1}$ and any $x\in {{\mathbb R}}$, the pieces $(g^j\circ T^k(\tilde J_n^{(n)}(x)))_{0\leq j\leq q_{n+1}, k\in {{\mathbb Z}}}$ have bounded Poincaré diameter and cover $\Phi^{(n)}_{z_0}$.
From Lemma \[lem\_bounded\] any $\tilde I^{(n)}_n(x)$ has bounded Poincaré diameter, thus also any $\tilde J^{(n)}_n(x)= \tilde I^{(n)}_n(x) \cup \tilde I^{(n)}_n(g_n^{-1}(x))$. Moreover, we have $g^j\circ T^k(J_n(x))=J_n(g^j\circ T^k(x))$, and all $\tilde J^{(n)}_n(g^j\circ T^k(x)) $ have also bounded Poincaré diameter. From Lemma \[lem\_comb\] these pieces cover $\Phi^{(n)}_{z_0}$.
\[cor\_dense\] For some $C_0>0$, the flow orbit $(\varphi_{j,k}^{(n)} (z_0))_{0\leq j< q_{n+1}, k\in {{\mathbb Z}}}$ is $C_0$-dense in $\Phi^{(n)}_{z_0}$ for the Poincaré metric.
We prove the first property stated in Theorem \[thm\_quasi\]:
Let $\gamma_n =\Phi^{(n-1)}_{z_0}$ for some $z_0$ from the previous lemma, then we have, for $0\leq j\leq q_n$, $${{\mathcal D}}_P(g^j(\gamma_n), \gamma_n)\leq 2C_0$$
We prove this Proposition for $n+1$ instead of $n$ (the proposition is stated to match $n$ in Theorem \[thm\_quasi\]). It follows from the hyperbolic Denjoy-Yoccoz Lemma that the orbit $(g^j\circ T^k (z_0))_{0\leq j<q_{n+1}, k\in {{\mathbb Z}}}$ is $C_0$-close to flow orbit $(\varphi_{j,k}^{(n)} (z_0))_{0\leq j<q_{n+1}, k\in {{\mathbb Z}}}$, and from Corollary \[cor\_dense\] we have that a $2C_0$-neighborhood of $g^j(\gamma_{n+1})$ contains $\gamma_{n+1}$. Conversely, since we can chooose any $z_0 \in \gamma_{n+1}$, we also have that $g^j(\gamma_{n+1})$ is in a $C_0$-neighborhood of $\gamma_{n+1}$.
We prove the second property of Theorem \[thm\_quasi\]. We observe that $g^{q_{n+1}}(z_0) \in \tilde J_n^{(n)}(x_0)$, that $z_0 \in \tilde J_n^{(n)}(x_0)$, and that $\tilde J_n^{(n)}(x_0)$ has a bounded Poincaré diameter by Lemma \[lem\_cover\]. Thus we get (taking a larger $C_0 >0$ if necessary):
For any $z_0\in \Phi^{(n)}$ , we have $$d_P(z_0, g^{q_{n+1}}(z_0)) \leq C_0 \ .$$
Osculating orbit. {#sec_osc}
=================
We prove the existence of an osculating orbit.
\[Oscullating orbit\]\[thm\_osc\] With the above hypothesis, for $n\geq n_o$ there exists a quasi-invariant curves $\gamma_n=\Phi^{(n-1)}_{z_0}$ such that the orbit $(g^j(z_0))_{0\leq j\leq q_n}$ is such that the union of Poincaré balls $$U_n=\bigcup_{0\leq j< q_{n}, k\in {{\mathbb Z}}} B_P(g^j(z_0)+k, C_0) \ ,$$ separates ${{\mathbb R}}$ from $\{\Im z > H\}$ with $H>0$ large enough, and any orbit $(g^j(w_0))_{j\in {{\mathbb Z}}}$ with $\Im w_0 >H$ with an iterate between $\gamma_n$ and ${{\mathbb R}}$ has, for any $0\leq j\leq q_n$, an iterate in $$\bigcup_{k\in {{\mathbb Z}}} B_P(g^j(z_0)+k, C_0) \ .$$
From Lemma \[lem\_cover\] we get the property that the hyperbolic balls $B_P(\varphi^{(n)}_{t+k}(z_0), C_0)$ cover $\Phi^{(n)}_{z_0}$.
\[lem\_cov\] We have that $$U_n=\bigcup_{0\leq j< q_{n+1}, k\in {{\mathbb Z}}} B_P(\varphi^{(n)}_{t+k}(z_0), C_0)$$ is a neighborhood of the flow line $\Phi^{(n)}_{z_0}$
We prove Theorem \[thm\_osc\]. In the following argument $C_0$ will denote several universal constants. Enlarging the constant $C_0$, and using Lemma \[lem\_DYhyp\] we can replace the points $\varphi^{(n)}_{t+k}(z_0)$ by the points $g^j(z_0)+k$ in the orbit of $z_0$ in Lemma \[lem\_cov\]. Also, any orbit that jumps over $\gamma_n$ (by positive or negative iteration) as in Theorem \[thm\_osc\] has to visit a $C_0$-neighborhood of $\gamma_n$ , and will be $C_0$-close to a point $z_1\in \gamma_n$ and then will be $C_0$-close to the $q_n$-orbit of $z_1$ modulo $1$. Finally we can replace $z_1$ by $z_0$ using that each point of the $q_n$-orbit of $z_1$ is $C_0$-close to a point in the $q_n$-orbit of $z_0$ modulo $1$ (enlarge $C_0$ if need be).
Proof of the main Theorem.
==========================
We prove Theorems \[thm\_unif\] and \[thm\_oscullating\] that imply the main Theorem. We prove first the following preliminary Lemma that will allow us to work only with local hedgehogs.
\[lem\_bigmodulus\] Let $g_n\in D^\omega ({{\mathbb T}}, {\Delta}_n )$ with $\rho (g_n)=\alpha$ and ${\Delta}_n\to +\infty$. Then $g_n\to R_{\alpha}$ uniformly on compact sets of ${{\mathbb C}}^*$ and $$\lim_{n\to +\infty} ||D \log D g_{n}||_{C^0({{\mathbb R}})} =0 \ .$$
Let $\tilde g_n$ be the associated circle diffeomorphism. The sequence $(\tilde g_n)$ is a normal family in ${{\mathbb C}}^*$ (bounded inside ${{\mathbb D}}$, and outside is the reflection across the unit circle), and any accumulation point is not constant since the unit circle is in the image of all $g_n$. Then by Hurwitz theorem any limit is an automorphism of ${{\mathbb C}}^*$, that extends to $0$ by Riemann’s theorem, and so gives an automorphism of the plane leaving the unit circle invariant. The rotation number on the circle depends continuously on $\tilde g_n$ and is constant equal to $\alpha$, therefore the only possible limit of the sequence $(g_n)$ is $R_{\alpha}$. Since $D \log D R_{\alpha}=0$ we get the last statement.
We consider now the hedgehog $K_0$ given by Theorem \[thm\_hedgehogs\] for the domain $U={{\mathbb D}}_{r_0}$, and we use the relation between hedgehogs and analytic circle diffeomorphisms presented in [@PM5] to construct a circle diffeomorphism $g_0$.
We consider a conformal representation $h_0: {{\mathbb C}}-\overline{{{\mathbb D}}} \to {{\mathbb C}}-K_0$ (${{\mathbb D}}$ is the unit disk), and we conjugate the dynamics to a univalent map $g_0$ in an annulus $V$ having the circle ${{\mathbb S}}^1=\partial {{\mathbb D}}$ as the inner boundary, $$g_0=h_0^{-1}\circ f\circ h_0 : V \to {{\mathbb C}}\ .$$ The topology of $K_0$ is complex ([@B1], [@B2], [@PM3]) and in particular $K_0$ is never locally connected, and $h_0$ does not extend to a continuous correspondence between ${{\mathbb S}}^1$ and $\partial K_0$. Nevertheless, $f$ extends continuously to Caratheodory’s prime-end compactification of ${{\mathbb C}}-K_0$. This shows that $g_0$ extends continuously to ${{\mathbb S}}^1$ and its Schwarz reflection defines an analytic map of the circle defined on $V\cup {{\mathbb S}}^1 \cup \bar V$, where $\bar V$ is the reflected annulus of $V$. Then it is not difficult to see that $g_0$ is an analytic circle diffeomorphism. We can also prove that $g_0$ has rotation number ${\alpha}$. This is harder to prove in general (for an aribtrary hedgehog), but it is not difficult to show that we can pick $K_0$ so that the rotation number of $g_0$ is ${\alpha}$ (see [@PM5] Lemma III.3.3) that is enough for our purposes. We choose such a $K_0$. Therefore, the dynamics in a complex neighborhood of $K_0$ corresponds to the dynamics of an analytic circle diffeomorphism with rotation number ${\alpha}$.
There is no risk of confusion and we denote also $g_0$ the lift to ${{\mathbb R}}$.
Let ${\epsilon}_0 >0$ and ${\Delta}>0$ be given
For $r_0>0$ small enough we have $g_{0} \in D^\omega({{\mathbb T}}, {\Delta})$ and $$||D \log D g_{0}||_{C^0({{\mathbb R}})} <{\epsilon}_0 \ .$$
When $r_0\to 0$, we have $K_{0} \to \{ 0\}$ and the annulus where $g_{0}$ and $g_0^{-1}$ are defined has a modulus $M_0\to +\infty$. Therefore, by Grötsch extremal problem, for $r_0>0$ small enough we have $g_0 \in D^\omega({{\mathbb T}}, {\Delta})$. From Lemma \[lem\_bigmodulus\] $$\lim_{r_0\to 0} ||D \log D g_{0}||_{C^0({{\mathbb R}})} =0 \ ,$$ and the result follows.
Let ${\epsilon}_0=1/9$ and ${\Delta}>0$ be as in Section \[sec\_quasi\] and Section \[sec\_osc\]. We fix now $r_0>0$ small enough such that $g_{0} \in D^\omega({{\mathbb T}}, {\Delta})$, $\rho(g_0) ={\alpha}$, and $$||D \log D g_{0}||_{C^0({{\mathbb R}})} <{\epsilon}_0 \ ,$$ so that the hypothesis of Theorem \[thm\_quasi\] are fulfilled for $g_{0}$. Now we can apply Theorem \[thm\_osc\] and find a sequence $(\gamma_n)_{n\geq n_0}$ of quasi-invariant curves for $g_{0}$. We transport them by $h_{0}$ to get a sequence of Jordan curves $(\eta_n)_{n\geq n_0}$ $$\eta_n =h_{0} (\gamma_n) \ .$$ We have $$||g_{0}^{q_n}-{\hbox{\rm id}}||_{C_P^0(\gamma_n)} \leq C_0\ ,$$ therefore, for the Poincaré metric of the exterior of the hedgehog, $$||f^{q_n}-{\hbox{\rm id}}||_{C_P^0(\eta_n)} \leq C_0\ ,$$ and, since $\eta_n \to K_0$, for the euclidean metric, we have $$||f^{q_n}-{\hbox{\rm id}}||_{C^0(\eta_n)} ={\epsilon}_n \to 0\ .$$ Thus, if $\Omega_n$ is the Jordan domain bounded by $\eta_n$, by the maximum principle we have $$||f^{q_n}-{\hbox{\rm id}}||_{C^0(\Omega_n)} ={\epsilon}_n \to 0\ .$$ Since $\Omega_n$ is a neighborhood of $K_0$, $K_0\subset \bar \Omega_n$, we have $$||f^{q_n}-{\hbox{\rm id}}||_{C^0(K_0)} ={\epsilon}_n \to 0\ .$$ This proves Theorem \[thm\_unif\] for the positive iterates (same proof for the negative ones, or just apply the result to $f^{-1}$).
We prove Theorem \[thm\_oscullating\] for $K_0$, or more precisely for $\partial K_0$ that was noted before that is enough for proving the Main Theorem (the hedgehog $K_0$ has empty interior and $K_0=\partial K_0$, but we don’t need to use this fact). For the proof of Theorem \[thm\_oscullating\] we transport by $h_{0}$ the Poincaré $C_0$-dense orbit $(g^j_{0}(z_0))_{0\leq j\leq q_n}$ given by Theorem \[thm\_osc\]. Let $\zeta_0=h_{0}(z_0)$ and ${{\mathcal O}}_n=(f^j(\zeta_0))_{0\leq j\leq q_n}$ be this orbit. Since $\eta_n \to \partial K_0$, we have, for ${\epsilon}_n\to 0$ $${{\mathcal D}}({{\mathcal O}}_n,\partial K_0) \leq {{\mathcal D}}({{\mathcal O}}_n , \eta_n)+ {{\mathcal D}}(\eta_n , \partial K_0) \leq {\epsilon}_n \ ,$$ where ${{\mathcal D}}$ denotes the Hausdorff distance for the euclidean metric.
Then any orbit starting outside of $\eta_n$ with an iterate inside $\eta_n$ must visit ${\epsilon}_n$-close for the euclidean metric any point of $\partial K_0$ that is strictly larger than $\{0\}$.
[9]{}
L. AUTONNE, *Sur l’équation différentielle du premier ordre et sur les singularités de ses intégrales algébriques*, Journal de l’Ecole Polytechnique, II série, cahier II, p.51-169, 1897.
V.I. ARNOLD, *Geometrical methods in the theory of ordinary differential equations*, 2nd edition, Springer, 1988.
G.D. BIRKHOFF, *Surface transformations and their dynamical applications*, Acta Mathematica, [**43**]{}, 1920.
K. BISWAS, *Nonlinearizable holomorphic dynamics and hedgehogs*, Eur. Math. Soc. Newsl., **73**, p.11-15, 2009.
K. BISWAS, *Positive area and inaccessible fixed points for hedgehogs*, Ergodic Theory Dynam. Systems, **36**, 6, p.1839-1850, 2016.
K. BISWAS, R. PÉREZ-MARCO, *Log-Riemann surfaces*, ArXiv:1512.03776, 2015.
C. BRIOT, T. BOUQUET, *Recherches sur les proprietés des équations différentielles*, J. École Impériale Polytechnique, **21 : 36**, p.133-198, 1856.
A.D. BRJUNO, *Analytic form of differential equations*, Trans. Moscow Math. Soc. , **25**, p.131-288, 1971; **26**, p.199-239, 1972.
C. CAMACHO, *On the local structure of conformal mappings and holomorphic vector fields in ${{\mathbb C}}^2$*, Astérisque, **59-60**, Soc. Math. France, p.83–94, 1978.
H. DULAC, *Recherches sur les points singuliers des équations différentielles*, Journal de l’Ecole Polytechnique, II série, cahier IX, p.1-125, 1904.
H. DULAC, *Points singuliers des équations différentielles*, Mémorial Sciences Mathématiques, **61**, Gathier-Villars, Paris, 1934.
P. FATOU, *Sur les équations fonctionnelles*, Bull. Soc. Math. Fr. [**47**]{}, p.161-271, 1919; p.33-94, 1920; [**48**]{}, p.208-304, 1920.
L. FUCHS, *Über die Werthe, welche die Integrale einer Differentialgleichung erster Ordnung in singulären Punkten annehmen können*, Akademie der Wissenschaften zu Berlin, p.279-300, 1886.
G.H.HARDY, E.M. Wright, *An introduction to the theory of numbers*, 4th Edition, Oxford, 1960.
M. R. HERMAN, *Sur la conjugaison différentiable des difféomorphismes du cercle à des rotations*, Publ. I.H.E.S. [**49**]{}, 1979.
M. LYUBICH, *The dynamics of rational transforms : The topological picture* Russian Math. Surveys, **41 : 4**, p. 43-117, 1986.
J.-F. MATTEI, R. MOUSSU, *Holonomie et intégrales premières*, Ann. Sc. E.N.S. $4^{eme}$ série, **13**, p.469-523, 1980.
J. MARTINET, J.-P. RAMIS, *Problèmes de modules pour les équations différentielles non linéaires du premier ordre*, Publ. Math. I.H.E.S., **55**, p.63-164, 1982.
J. MARTINET, J.-P. RAMIS, *Classification analytique des équations non linéaires résonnantes du premier ordre*, Ann. Sc. E.N.S. $4^{eme}$ série, **16**, p. 671-625, 1983.
R. PÉREZ-MARCO, *Sur les dynamiques holomorphes non linéarisables et une conjecture de V. I. Arnold*, Ann. Scient. Ec. Norm. Sup. 4 serie, **26**, p.565-644, 1993.
R. PÉREZ-MARCO, *Sur une question de Dulac et Fatou*, Comptes Rendus Académie des Sciences de Paris, **321** , Série I, p.1045-1048, 1995.
R. PÉREZ-MARCO, *Topology of Julia sets and hedgehogs*, preprint Université Paris-Sud, 94-48, 1994.
R. PÉREZ-MARCO, *Uncountable number of symmetries for non-linearizable holomorphic dynamics*, Inventiones Mathematicae, **119**, **1**, p.67-127, 1995.
R. PÉREZ-MARCO, *Fixed points and circle maps* Acta Mathematica, Acta Mathematica, **179**, p.243-294 ,1997.
R. PÉREZ-MARCO, *Hedgehog dynamics*, Manuscript, 1998.
R. PÉREZ-MARCO,*On quasi-invariant curves*, 2018.
R. PÉREZ-MARCO, J.-C. YOCCOZ, *Germes de feuilletages holomorphes à holonomie prescrite*m “Complex methods in dynamical systems” Astérisque, **222**, p.345-371, 1994.
É. PICARD, *Traité d’analyse*, vol. I, II, II, various editions, 1891, 1893, 1896, 1908, 1928.
H. POINCARÉ, *Note sur les propriétés des fonctions définies par les équations différentielles*, Journal de l’Ecole Polytechnique, XXVIII, p.13-26, 1878.
H. POINCARÉ, *Sur les propriétés des fonctions définies par les équations aux différences partielles*, Thèse, 1879.
C.-L. SIEGEL, *Iterations of analytic functions*, Ann. Math., **43**, p.807-812, 1942.
C.-L. SIEGEL, J.K. MOSER, *Lectures on celestial mechanics*, Springer, 1971.
D. SULLIVAN, *Quasiconformal homeomorphisms and dynamics. I. Solution of the Fatou-Julia problem on wandering domains*, Annals of Mathematics, **122**, 3, p.401-418, 1985.
J.-C. YOCCOZ,*Conjugation différentiable des difféomorphismes du cercle dont le nombre de rotation vérifie une condition diophantienne*, Ann. Scient. Ec. Norm. Sup., 4eme serie, **17**, p.333-359, 1984.
J.-C. YOCCOZ, *Linéarisation des difféomorphismes analytiques du cercle*, manuscript, 1989.
J.-C. YOCCOZ, *Analytic linearization of circle diffeomorphisms*, Dynamical systems and small divisors (Cetraro, 1998), Lecture Notes in Math., **1784**, Springer, Berlin, p.125–173, 2002.
[^1]: What is called today in Calculus books Cauchy-Lipschitz Theorem.
[^2]: *“It remains to prove that these two solutions are, without any assumption, the only ones passing through the origin or accumulating it. I have to admit that I don’t have a proof of this fact but it doesn’t seem doubtful.”*
[^3]: Volume III is the correct reference.
[^4]: “$\ldots $ from long time ago we know that there are only two real solutions passing through the origin. Is it the same in the complex? This is a question that remained open and that the geometers were inclined to decide in the affirmative (Picard, Traité d’Analyse, II (sic), p. 30). But, on the contrary, I prove, at least in the case when ${\alpha}$ is rational, that there are infinitely many solutions $y(x)$ vanishing with $x$ ($x$ converging to $0$ under a suitable law) $\ldots$”
[^5]: “For a long time it was believed that, without any further condition, these are the only solutions passing through or accumulating the origin.
In an excellent work on the singular points of differential equations, M. Dulac has proved that the question is very complex. Take for instance the equation $$x\frac{dy}{dx}+y(\nu +\ldots )=0 \ ,$$ where $\nu$ is positive, equation that we can always reduce the case where $\lambda$ is negative.
M. Dulac examines specially the case where $\nu$ is a rational number $p/q$, and proves that in general there are an infinite number of solutions for which $x$ and $y$ converge to $0$.”
[^6]: “1.$\nu$ is irrational. We have a saddle. $H(x,y)$ exists formally, but is divergent, at least in certain cases. If there are solutions $x$ and $y$ which tend simultaneously to $0$, and if we note $\omega$ and $\theta$ the arguments of $x$ and $y$, then for all $m$ and $n$, $|x^my^n\omega|$ and $|x^m y^n\theta |$ must grow indefinitely. I cannot decide on the existence of such solutions.”
[^7]: “In case $2$ ($\nu$ irrational, $h(x,y)$ divergent), we don’t know if there are null solutions other than $x=0$, $y=0$.”
[^8]: *“It remains to study fixed points with a multiplier of the form $e^{i\alpha}$, $\alpha$ being a real number incommensurable with $\pi$. We know little about these fixed points, and their study from our point of view appears very hard. ($\ldots$) Are there any domain such that the positive iterates converge to the fixed point? We cannot give examples nor rule out this possibility.”*
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Classical age-structured mass-action models such as the McKendrick-von Foerster equation have been extensively studied but they are structurally unable to describe stochastic fluctuations or population-size-dependent birth and death rates. Stochastic theories that treat semi-Markov age-dependent processes using *e.g.*, the Bellman-Harris equation, do not resolve a population’s age-structure and are unable to quantify population-size dependencies. Conversely, current theories that include size-dependent population dynamics (*e.g.*, mathematical models that include carrying capacity such as the Logistic equation) cannot be easily extended to take into account age-dependent birth and death rates. In this paper, we present a systematic derivation of a new fully stochastic kinetic theory for interacting age-structured populations. By defining multiparticle probability density functions, we derive a hierarchy of kinetic equations for the stochastic evolution of an ageing population undergoing birth and death. We show that the fully stochastic age-dependent birth-death process precludes factorization of the corresponding probability densities, which then must be solved by using a BBGKY-like hierarchy. However, explicit solutions are derived in two simple limits and compared with their corresponding mean-field results. Our results generalize both deterministic models and existing master equation approaches by providing an intuitive and efficient way to simultaneously model age- and population-dependent stochastic dynamics applicable to the study of demography, stem cell dynamics, and disease evolution.'
author:
- 'Chris D. Greenman$^{1,2}$ and Tom Chou$^{3}$'
bibliography:
- 'refs1.bib'
title: 'A kinetic theory for age-structured stochastic birth-death processes'
---
Introduction
============
Age is an important controlling feature in populations of living organisms. Processes such as birth, death, and mutation are typically highly dependent upon an organism’s chronological age. Age-dependent population dynamics, where birth and death probabilities depend on an organism’s age, arise across diverse research areas such as demography [@KEYFITZBOOK], biofilm formation [@AYATI2007], and stem cell proliferation and differentiation [@SUN2013; @ROSHAN2014]. In this latter application, not only does a the cell cycle give rise to age-dependent processes [@QU2003; @CELLCYCLE2014], but the often small number of cells requires a stochastic interpretation of the population. Despite the importance of age structure (such as that arising in the study of cell cycles [@QU2003; @CELLCYCLE2014; @AGING0]), there exists no theoretical method to fully quantify the stochastic dynamics of aging and population-dependent processes.
Past work on age-structured populations has focussed on [ *deterministic*]{} models through the analysis of the so-called McKendrick-von Foerster equation, first studied by McKendrick [@MCKENDRICK; @KEYFITZ] and subsequently von Foerster [@VONFOERSTER], Gurtin and MacCamy [@Gurtin1; @Gurtin2], and others [@IANNELLI1995; @WEBB2008]. In these classic treatments, $\rho(a,t)\dd a$ is used to define, at time $t$, the density of noninteracting agents with age between $a$ and $a+\dd a$. The total number of particles in the system at time $t$ is thus $n(t) =
\int_{0}^{\infty} \rho(a,t)\dd a$. If $\mu(a; n(t))$ is the death rate for individuals of age $a$, the McKendrick-von Foerster equations are [@Gurtin1; @Gurtin2]
$${\partial \rho(a,t)\over \partial t} + {\partial \rho(a,t)\over \partial a} =
-\mu(a;n(t))\rho(a,t),
\label{MCKENDRICK0}$$
with $\rho(a,t=0) = g(a)$ and
$$\rho(a=0,t) = \int_{0}^{\infty}
\beta(a;n(t))\rho(a,t)\dd a
\label{MCKENDRICKBC}$$
for initial and boundary conditions, respectively. The boundary condition (Eq. \[MCKENDRICKBC\]) reflects the fact that birth gives rise to age-zero individuals. Note that the birth and death rates $\beta$ and $\mu$ are usually simply assumed to be functions of the total population $n(t)$.
The population dependence of $\beta(a;n(t))$ and $\mu(a;n(t))$ in Eqs. \[MCKENDRICK0\] and \[MCKENDRICKBC\] are assumed without explicit derivation and it is not clear whether such simple expressions are self-consistent. Moreover, the McKendrick-von Foerster equation is expected to be accurate exact only when the dynamics of each individual are not correlated with those of any other. Therefore, a formal derivation will allow a deeper understanding of how population dependence and correlations arise in a fully stochastic age-structured framework.
Two approaches that have been used for describing stochastic populations include Master equations [@VANKAMPEN2011; @FPTREVIEW] and evolution equations for age-dependent branching process such as the Bellman-Harris process [@BELLMANHARRIS; @REID1953; @JAGERS1968; @SHONKWILER1980; @CHOUJTB]. Master-equation approaches can be used to describe population-dependent birth or death rates [@KENDALL1948; @Gurtin1; @Gurtin2; @LJSALLEN] but implicitly assume exponentially distributed waiting times between events [@FPTREVIEW]. On the other hand, age-dependent models such as the Bellman-Harris branching process [@BELLMANHARRIS] allow for arbitrary distributions of times between birth/death events but they cannot resolve age-structure of the entirte population nor describe population-dependent dynamics that arise from *e.g.,* regulation or environmental carrying capacities.
A number of approaches attempt to incorporate ideas of stochasticity and noise into age-dependent population models, [@SUN2013; @REID1953; @THESIS1998; @DIFFUSION2009; @GETZ1984; @COHEN; @LESLIE1945; @LESLIE1948]. For example, stochasticity can be implemented by assuming a random rate of advancing to the next age window (by [*e.g.*]{}, stochastic harvesting [@GETZ1984; @COHEN] or a fluctuating environment [@LANDE1988; @LANDE2005]). However, such models do not account for the intrinsic stochasticity of the underlying birth-death process that acts differently on individuals at each different age. One alternative approach might be to extend the mean-field, age-structured McKendrick-von Foerster theory into the stochastic domain by considering the evolution of $P(n(a);t)$, the probability density that there are $n$ individuals within age window $[a,a+\dd a]$ at time $t$ [@SUN2013; @POLLARD1966]. This approach is meaningful only if a large number of individuals exist in each age window, in which case a large system size van Kampen expansion within each age window can be applied [@VANKAMPEN2011]. However, such an assumption is inconsistent with the desired small-number stochastic description of the system.
A mathematical theory that addresses the age-dependent problem of constrained stochastic populations would provide an important tool for quantitatively investigating problems in demography, bacterial growth, population biology, and stem cell differentiation and proliferation. In this paper, we develop a new kinetic equation that intuitively integrates population stochasticity, age-dependent effects (such as cell cycle), and population regulation into a unified theory. Our equations form a hierarchy analogous to that derived for the BBGKY (Bogoliubov-Born-Green-Kirkwood-Yvon) hierarchy in kinetic theory [@MCQUARRIE; @ZANETTE], allowing for a fully stochastic treatment of age-dependent process undergoing population-dependent birth and death.
Kinetic equations for aging populations
=======================================
To develop a fully stochastic theory for age-structured populations that can naturally describe both age- and population size-dependent birth and death rates, we invoke multiple-particle distribution functions such as those used in kinetic theories of gases [@ZANETTE]. Our analysis builds on the Boltzmann kinetic theory of D. Zanette and yields a BBGKY-like hierarchy of equations. Here, the positions of ballistic particles will represent the ages of individuals.
Changes in the total population require that we consider a family of multiparticle distribution functions, each with different dimensionality corresponding to the number of individuals. In this picture, birth and death are represented by transitions between the different distribution functions residing on different fixed particle-number “manifolds." Processes that generate newborns (particles of age zero) manifest themselves mathematically through boundary conditions on higher dimensional distribution functions.
To begin, we define
$$f_{n}(x_{1}, x_{2},x_{3}, \ldots, x_{n}; t)\dd x_{1}\dd x_{2}\ldots\dd x_{n}$$
as the probability that at time $t$, one observes $n$ distinguishable (by virtue of their order of birth) individuals, such that the youngest one has age within $(x_{1}, x_{1}+\dd x_{1})$, the second youngest has age within $(x_{2}, x_{2}+\dd x_{2})$, and so on. If the individuals are identical (except for their ages) and one does not distinguish which are in each age window, one can define $\rho_{n}(x_{1}, x_{2}, x_{3}, \ldots,x_{n};t)\dd x_{1}\dd
x_{2}\ldots\dd x_{n}$ as the probability that after randomly selecting individuals, the first one chosen has age in $(x_1,x_{1}+\dd x_{1})$, the second has age in $(x_{2}, x_{2}+\dd x_{2})$, and so on. For example, if there are three individuals with ordered ages $x_1 < x_2 <
x_3$, the probability of making any specific random selection, such as choosing the individual with age $x_{2}$ first, the one with age $x_{1}$ second, and the one with age $x_{3}$ third, is $\frac{1}{3!}$. More generally, when the ages $\x_{1,n}\equiv \x_{n} =
(x_1,x_2,\hdots, x_{n})$ are unordered, the associated probability density is
$$\rho_n(\x_{n};t)=\frac{1}{n!}f_n({\cal T}(\{x_{i}\});t),
\label{ORDERFN}$$
in which ${\cal T}$ is the time-ordering permutation operator such that, for example, ${\cal T}(x_2,x_1,x_3)=(x_1,x_2,x_3)$. Note that in this formulation, $\rho_n(\x_{n};t)$ is invariant under interchange of the elements of $\x_{n}$.
To derive kinetic equations for $\rho_n(\x_{n};t)$, we first define an ordered cumulative probability distribution
$$Q_{n}(\a_{n};t) =
\int_{0}^{a_{1}}\dd x_{1}\int_{x_{1}}^{a_{2}}\dd x_{2}\cdots
\int_{x_{n-1}}^{a_{n}}\!\!\!\dd x_{n} f_{n}(\x_{n};t),$$
where $\a_{n} = a_{1,n} = (a_{1},\ldots,a_{n})$. $Q_{n}(\a_{n};t)$ describes the probability that there are $n$ existing individuals at time $t$ and that the youngest individual has age $x_{1}$ less than or equal to $a_{1}$, the second youngest individual has age $x_{1}\leq
x_{2} < a_{2}$, and so on. The oldest individual has age $x_{n-1} \leq
x_{n} \leq a_{n}$.
We now compute the change in $Q_{n}(\a_{n}; t)$ over a small time increment $\ve$: $Q_{n}(\a_{n} + \ve; t+\ve) = Q_{n}(\a_{n}; t) +
\int_{t}^{t+\ve}J(\a_{n};t')\dd t'$, where $J(\a_{n};t') =
J^{+}(\a_{n};t') - J^{-}(\a_{n};t')$ is the net probability flux at time $t'$. The probability flux which increases the cumulative probability is denoted $J^{+}$ while that which decrease the cumulative probability is labelled $J^{-}$. Each of the $J^{\pm}$ include contributions from different processes that remove or add individuals. A schematic of our birth-death process, starting from a single parent, is depicted in Fig. \[MODELS\]A.
In the $\ve \to 0$ limit, we find the conservation equation
$${\partial Q_{n}(\a_{n};t)\over \partial t} +
\sum_{i=1}^{n} {\partial Q_{n}(\a_{n};t)\over \partial a_{i}} =
J^{+}(\a_{n};t)- J^{-}(\a_{n};t).
\label{Q0}$$
Eq. \[Q0\] is a “weak form” integral equation for the probability density which allows us to systematically derive an evolution equation and the associated boundary conditions for $f_{n}(\x_{n};t)$. The probability fluxes can be decomposed into components representing age-dependent birth and death
$$J^{\pm}(\a_{n};t) = J^{\pm}_{\beta}(\a_{n};t) + J^{\pm}_{\mu}(\a_{n};t),
\label{JPM}
$$
where the birth and death that reduce probability can be expressed as
$$\begin{aligned}
\displaystyle J^{-}_{\beta}(\a_{n};t) = \displaystyle
\int_{0}^{a_{1}}\!\!\dd x_{1}
\int_{x_{1}}^{a_{2}}\!\!\dd x_{2}
\cdots\!\int_{x_{n-1}}^{a_{n}}\!\!\!\!\!\dd x_{n}f_{n}(\x_{n};t)\sum_{i=1}^{n}\beta_{n}(x_{i}), \label{JMINUS_B}\\
\displaystyle J^{-}_{\mu}(\a_{n};t) = \displaystyle
\int_{0}^{a_{1}}\!\!\dd x_{1}
\int_{x_{1}}^{a_{2}}\!\!\dd x_{2}
\cdots\!\int_{x_{n-1}}^{a_{n}}\!\!\!\!\!\dd x_{n}f_{n}(\x_{n};t)\sum_{i=1}^{n}
\mu_{n}(x_{i}).\label{JMINUS_MU}
$$
Similarly, the probability fluxes that increase probability are
$$\begin{aligned}
\displaystyle J^{+}_{\beta}(\a_{n};t) = &
\int_{0}^{a_{2}}\!\!\dd x_{1}\cdots\!
\int_{x_{j-1}}^{a_{j+1}}\!\!\!\dd x_{j}\cdots\!\int_{x_{n-2}}^{a_{n}}\!\!\!\!\dd x_{n-1}
f_{n-1}(\x_{n-1};t)\sum_{i=1}^{n-1}\beta_{n-1}(x_{i}), \label{JPLUS_B}\end{aligned}$$
$$\begin{aligned}
\displaystyle J^{+}_{\mu}(\a_{n};t) \displaystyle = & \sum_{i=0}^{n}
\int_{0}^{a_{1}}\!\!\dd x_{1}\cdots\!\int_{x_{i-1}}^{a_{i}}\!\!\dd x_{i}
\int_{x_{i}}^{a_{i+1}}\!\!\!\dd y \int_{y}^{a_{i+1}}\!\!\!\!\dd x_{i+1} \cdots\!
\int_{x_{n-1}}^{a_{n}}\!\!\!\dd x_{n}\,\mu_{n+1}(y)
f_{n+1}(\x_{i},y,\x_{i+1,n};t),\label{JPLUS_MU}
$$
in which $\x_{i,j}\equiv (x_{i},x_{i+1},\ldots,x_{j}), x_{0} \equiv
0$, $a_{n+1}\equiv \infty$, and the age- and population-dependent birth and death rates for individual $i$ are denoted $\beta_{n}(x_{i})$ and $\mu_{n}(x_{i})$, respectively. The probability flux into $Q_{n}(\a_{n};t)$ arising from birth of the $n-1$ individuals of age $\a_{2,n}\equiv (a_{2}, a_{3},\ldots,a_{n})$ generates an individual of age zero. Hence, a key feature of $J_{\beta}^{+}(\a_{n};t)$ is that it does not depend on $a_{1}$.
 A simple age-dependent birth-death process. Each parent gives birth with an age-dependent rate $\beta_{n}(a)$, which may also depend on the total population size $n$. Individuals can also die (open circles) at an age- and population-dependent rate $\mu_{n}(a)$. (B) Age trajectories in the upper ($a>t$) octant are connected to those in the lower one ($a<t$) through the birth processes. Individuals that exist at time $t=0$ can be traced back and defined by their time of birth $b_{i}$. Here, the labeling ordered according to increasing age. The pictured trajectories define characteristics $a_{i}(t)$ that can be used to solve Eq. \[F0\].](birth4ab.eps){width="4.2in"}
We can now describe the fully stochastic aging process in terms of the ordered distribution function $f_{n}(\x_{n};t)$ by using Eqs. \[JPM\]-\[JPLUS\_MU\] in Eq. \[Q0\] and applying the operator ${\partial
\over \partial a_{n}}\cdots{\partial \over \partial a_{2}}{\partial
\over \partial a_{1}}$ to find
$$\begin{array}{l}
\displaystyle {\partial f_{n}(\a_{n};t)\over \partial t} + \sum_{j=1}^{n}
{\partial f_{n}(\a_{n};t)\over \partial a_{j}} =
-f_{n}(\a_{n};t)\sum_{i=1}^{n} \gamma_{n}(a_{i})+ \sum_{i=0}^{n}\int_{a_{i}}^{a_{i+1}}\!\!\!\!
\mu_{n+1}(y)f_{n+1}(\a_{i},y,\a_{i+1,n};t)\dd y,
\label{F0}
\end{array}$$
where $a_{0}\equiv 0$, $a_{n+1}\equiv \infty$, and the total age-dependent transition rate is $$\gamma_{n}(a_{i}) = \beta_{n}(a_{i})+ \mu_{n}(a_{i}).
$$ Note that the $a_{1}-$independent source term $J^{+}_{\beta}$ that had contributed to the ordered cumulative (Eq. \[Q0\]) does not contribute to the bulk equation for $f_{n}(\a_{n};t)$. Rather, it arises in the boundary condition for $f_{n}$, which can be found by setting $a_{1} = 0$ in Eq. \[Q0\]. Since $Q(0,a_{2}, \ldots,
a_{n};t) = 0$ and $J_{\beta}^{+}(\a_{n};t)$ are independent of $a_{1}$, the remaining terms are
$$\int_{0}^{a_{2}}\!\!\dd x_{2}\cdots\int_{x_{n-1}}^{a_{n}}\!\!\!\!
\dd x_{n} f_{n}(x_{1}=0,\x_{2,n};t) =
J^{+}_{\beta}(\a_{n};t).
\label{BCQ}$$
Further taking the derivatives ${\partial \over \partial
a_{n}}\cdots{\partial \over \partial a_{2}}$ of Eq. \[BCQ\], we find the boundary condition
$$f_{n}(a_{1}=0,\a_{2,n};t) =
f_{n-1}(\a_{2,n};t)\sum_{i=2}^{n} \beta_{n-1}(a_{i}).
\label{BCF}$$
We now consider [*indistinguishable*]{} individuals as described by the density defined in Eq. \[ORDERFN\]. Equation \[F0\] can then be expressed in terms of $\rho_{n}(\a_{n};t)$: the probability density that if we [*randomly*]{} label individuals, the first one has age between $a_{1}$ and $a_{1}+\dd a_{1}$, the second has age between $a_{2}$ and $a_{2}+\dd a_{2}$, and so on. The kinetic equation for $\rho_{n}$ can then be expressed in the form
$$\begin{aligned}
\displaystyle {\partial \rho_{n}(\a_{n};t)\over \partial t} +
& \sum_{j=1}^{n}{\partial \rho_{n}(\a_{n};t)\over \partial a_{j}} =
-\rho_{n}(\a_{n};t)\sum_{i=1}^{n} \gamma_{n}(a_{i})
+ (n+1)\!\int_{0}^{\infty}\!\!\mu_{n+1}(y)
\rho_{n+1}(\a_{n},y;t)\dd y, \label{RHO0} \end{aligned}$$
and the boundary condition becomes
$$\begin{array}{l}
n \rho_{n}(a_{1},\ldots,a_{\ell}=0,\ldots, a_{n};t) = \rho_{n-1}(a_{1},\ldots,
\hat{a}_{\ell},\ldots,a_{n};t)\sum_{i(\ne \ell)=1}^{n}\beta_{n-1}(a_{i}),
\label{BCRHO}
\end{array}$$
where the sum precludes the $i=\ell$ term and $\hat{a}_{\ell}$ indicates that the variable $a_{\ell}$ is omitted from the sequence of arguments [@ZANETTE]. Equation \[RHO0\] and the boundary conditions of Eq. \[BCRHO\], along with an initial condition $\rho_{n}(\a_{n};t=0)$, fully define the stochastic age-structured birth-death process and is one of our main results. Eq. \[RHO0\] is analogous to a generalized Boltzmann equation for $n$ particles [@ZANETTE; @PETERS]. The evolution operator corresponds to that of free ballistic motion in one dimension corresponding to age. However, instead of particle collisions typically studied in traditional applications of the Boltzmann equation, our problem couples density functions for $n$ particles to those of $n+1$ and $n-1$ (through the boundary condition).
Solutions and equation hierarchies
==================================
Equation \[RHO0\] defines a set of coupled linear integro-differential equations. We would like to find solutions for $\rho_n(\a_n;t)$ expressed in terms of an initial condition $g_{n}(\a_{n}-t;t=0)$. However, we will see below that the presence of births during the time interval $[0,t]$ prevents a simple solution to Eq. \[RHO0\] due to interference from the boundary condition in Eq. \[BCRHO\]. Instead, we will obtain a solution for $\rho_{n}(\a_{n};t)$ at time $t$ in terms of the distribution $\rho_{n}(\a_{n}-(t-t_0);t_0)$ at an earlier time $t_0$ selected such that no births occur during the time interval $(t_0,t]$. That is, if $b_i=t-a_i$ represents the time of birth of the $i^{\rm th}$ individual (see Fig. \[MODELS\]B), we have the condition $t_0 \ge
b_{i} \,\forall\, i$. The dynamics described by Eq. \[RHO0\] are then unaffected by the boundary condition (Eq. \[BCRHO\]) and can be solved using the characteristics $a_i = t - b_i$ indexed by individual times of birth $b_{i}$. Note that any individual initially present (at time $t=0$) has a projected negative time of birth. We can then solve $\rho_{n}(t-\b_{n};t)$ explicitly along each characteristic and then re-express them in terms of $\a_{n}$, to obtain $$\begin{aligned}
\rho_{n}(\a_{n};t)& = U_{n}(\a_{n};t_0;t)\rho_{n}(\a_{n}-(t-t_0);t_0)
+(n+1)\int_{t_0}^{t}U_{n}(\a_{n};t';t)
\left[\int_{0}^{\infty}\!\!\!\mu_{n+1}(y)
\rho_{n+1}\dd y\right]\dd t',\label{SOLN0}\end{aligned}$$ where $\rho_{n+1} \equiv \rho_{n+1}(\a_{n}-(t-t'),y;t')$ above, and $$\begin{aligned}
U_{n}(\a_{m};t';t) & =
\exp\left[-\sum_{i=1}^{m}\int_{t'}^{t}\!\gamma_{n}(a_{i}-(t-s))\dd s\right]
\equiv U_{n}^{-1}(\a_{m};t_0;t')U_{n}(\a_m;t_0;t)
\label{PROP}\end{aligned}$$ is the propagator for any set of $m \leq n$ individuals from time $t'$ to $t$.
In the case of a pure death process where no births occur ($\beta_{n}=0$), allowing us to set $t_0=0$. A complete solution can be found through successive iteration of Eq. \[SOLN0\]. We further simplify matters by assuming an initial condition that factorizes into an initial total number distribution $\rho(n)$ and common initial age probability densities $g(a)$: $\rho_n(\a_n-t;0)=\rho(n)\prod\limits_{i=1}^{n}g(a_i-t)$. If we further assume a death rate $\mu_n(a)=\mu(a)$ that is independent of population size, Eq. \[SOLN0\] can be solved, after some algebra, to yield
$$\begin{aligned}
\rho_n(\a_n;t) = & U(\a_n;0;t)\prod\limits_{i=1}^{n} g(a_i-t)
\displaystyle\sum\limits_{k=0}^\infty{n+k \choose k}\rho(n+k) \left[\int\limits_0^{t} g(y-s)\int\limits_s^{\infty} U(y;0;s)\mu(y) \dd y \dd s \right]^k.
\label{DEATHEQNSOL}\end{aligned}$$
For a pure birth process where $\mu_{n}=0$, the second integral term in Eq. \[SOLN0\] disappears. In this case, we must use the boundary condition (Eq. \[BCRHO\]) to successively bootstrap the solution by applying the propagator $U$ between birth times. Assume a starting time $t=0$ with an initial condition consisting of $m$ individuals with corresponding ages $a >t$. The symmetry of $\rho_n(\a_n;t)$ and $U_n(\a_n;t';t)$ implies that, without loss of generality, ages can be arranged in decreasing order: $a_1 > a_2 >
\hdots > a_{m} > t > a_{m+1} > \hdots > a_n$, where the youngest was born most recently at time $t-a_n>0$. If we select $t_0$ to be the moment of birth at time $b_n=t-a_{n}$ of the most recently born ($n^{\rm th}$) individual, the density over all individuals is propagated forward according to
$$\rho_{n}(\a_{n};t) = U_{n}(\a_{n};b_{n};t)\rho_{n}(\{\a_{n-1}-a_{n},0\};t-a_{n}),$$
where $\rho_{n}(\{\a_{n-1}-a_{n},0\};t-a_{n})$ is the initial condition immediately after the birth of the $n^{\rm th}$ individual and can be related to $\rho_{n-1}$ through the boundary condition in Eq. \[BCRHO\]. The density function thus obeys $$\begin{aligned}
\rho_{n}(\a_{n};t) = & \frac{1}{n}U_{n}(\a_{n};b_{n};t)
\rho_{n-1}(\a_{n-1}-a_{n};t-a_{n})\sum_{i=1}^{n-1}\beta_{n-1}(a_{i}-a_{n}).
\label{ITERATION}\end{aligned}$$ Eq. \[ITERATION\] can then be iterated back to $t=0$ to find the solution for randomly selected individuals. For the case in which $\gamma_{n} = \gamma$ is independent of the population size, the propagator can be separated into a product across individuals. If $\beta_{n} = \beta$ is also independent of $n$, the solution takes the simple form
$$\begin{aligned}
\rho_{n}(\a_{n};t) = & g_{m}(\a_{m}-t) U(\a_{m};0;t)
{m! \over n!} \!\!\!\prod_{k=m+1}^{n}U
(a_k;b_k;t)
\sum_{\ell=1}^{k-1}\beta(a_{\ell}-a_{k}),
\label{BIRTHSOL}\end{aligned}$$
where $b_{k} = t-a_{k}$ and $g_{m}$ is the initial distribution of ages for the $m$ individuals born before $t=0$. The above solutions for $\rho_{n}(\a_{n};t)$ allow us to explicitly compare differences between the fully stochastic theory and the deterministic McKendrick-von Foerster model. As an example, consider the expected number of individuals at time $t$ that have age between $0$ and $a$,
$$P(a,t) = \int_{0}^{a}\!\!\rho(y,t)\dd y,
\label{PAT}$$
where $\rho(y,t)$ is found from Eqs. \[MCKENDRICK0\] and \[MCKENDRICKBC\]. We wish to compare this quantity with the [ *probability*]{} $P_m(a,t)$ that there are $m$ individuals at time $t$ with age between $0$ and $a$. The probability $P_m(n,a,t)$ that there are $n$ total individuals of which exactly $m$ have age between $0$ and $a$ can be constructed from our fully stochastic theory via
$$P_m(n,a,t) = {n \choose m}\prod_{j=1}^{m}\int_{0}^{a}\dd a_{j}\!\!\!
\prod_{\ell=m+1}^{n}\!\!\!\int_{a}^{\infty}\!\!\!\!\dd a_{\ell}\, \rho_{n}(\a_{n};t).
\label{PMNAT}$$
The marginal probability $P_m(a,t)$ is then found by summing over $n\geq m$:
$$P_m(a,t) = \sum_{n=m}^{\infty}P_m(n,a,t).
\label{STOCHCUMUL}$$
The comparison can be made more explicit by considering simple cases such as an age-independent birth-only process with fixed birth rate $\beta$. If the process starts with precisely $N$ individuals, standard methods [@IANNELLI1995; @WEBB2008] yields a simple solution of the McKendrick-von Foerster equation which when used in Eq. \[PAT\] gives $P(a<t;t)=Ne^{\beta t}\left(1-e^{-\beta
a}\right)$. Substituting the pure birth solution of Eq. \[BIRTHSOL\] into Eqs. \[PMNAT\] and \[STOCHCUMUL\] yields $$P_m(a,t)={m+N-1 \choose m}\frac{e^{-N\beta t}\left(1-e^{-\beta a}\right)^m}
{\left(1-e^{-\beta a}+e^{-\beta t}\right)^{m+N}}.$$
![\[Birth\_Example\]Comparison of $P(a,t)$ (Eq. \[PAT\]) derived from the McKendrick-von Foerster equation with $P_{m}(a,t)$ of a fully stochastic pure birth process with constant $\beta=0.1$. We start with $N=10$ individuals and analyze our quantities at time $t=10$ for ages $a<t$. (A) Each of the 100 grey lines count the number of individuals younger than age $a$ in one simulation. The solid black curve indicates the deterministic (McKendrick-von Foerster) solution $P(a,t)=\int_0^a \rho(y,t)\dd y$, which can also be obtained through $P(a,t)=\sum_{m=1}^{\infty}mP_{m}(a,t)$. The shaded region represents the inter-quartile range of $P_{m}(a,t)$. (B) Distribution constructed from 1000 simulations (bars) and theoretical distribution $P_m(a=5,t=10)$ (black curve).](Fig2A.eps "fig:"){width="2.1in"}![\[Birth\_Example\]Comparison of $P(a,t)$ (Eq. \[PAT\]) derived from the McKendrick-von Foerster equation with $P_{m}(a,t)$ of a fully stochastic pure birth process with constant $\beta=0.1$. We start with $N=10$ individuals and analyze our quantities at time $t=10$ for ages $a<t$. (A) Each of the 100 grey lines count the number of individuals younger than age $a$ in one simulation. The solid black curve indicates the deterministic (McKendrick-von Foerster) solution $P(a,t)=\int_0^a \rho(y,t)\dd y$, which can also be obtained through $P(a,t)=\sum_{m=1}^{\infty}mP_{m}(a,t)$. The shaded region represents the inter-quartile range of $P_{m}(a,t)$. (B) Distribution constructed from 1000 simulations (bars) and theoretical distribution $P_m(a=5,t=10)$ (black curve).](Fig2BL.eps "fig:"){width="2.1in"}
In Fig. \[Birth\_Example\]A we compare the expected value $P(a,t)$ derived from solutions to the McKendrick-von Foerster equation with stochastic simulations that sample the stochastic result $P_m(a,t)$. The fully stochastic nature of the process is clearly shown by the spread of the population about the expected value. Fig. \[Birth\_Example\]B plots the corresponding number distribution $P_{m}(5,10)$.
Finally, to connect our general kinetic theory with statistically-reduced (and deterministic) descriptions, we consider reduced $k-$dimensional distribution functions defined by integrating $\rho_n(\a_n;t)$ over $n-k$ age variables:
$$\rho_{n}^{(k)}(\a_{k};t) \equiv \int_{0}^{\infty}\!\dd a_{k+1}
\ldots\int_{0}^{\infty}\!\dd a_{n}\, \rho_{n}(\a_{n};t).$$
The symmetry properties of $\rho_{n}(\a_{n};t)$ indicate that it is immaterial which of the $n-k$ age variables are integrated out. If we integrate Eq. \[RHO0\] over all ages ($k=0$), and assume $\rho_{n}^{(1)}(a = \infty;t) = 0$, we find
$$\begin{aligned}
{\partial \rho_{n}^{(0)}(t) \over \partial t} = & n\rho_{n}^{(1)}(a=0;t)
-n \int_{0}^{\infty}\!\!\!\gamma_{n}(y)\rho_{n}^{(1)}(y;t)\dd y +(n+1)\int_{0}^{\infty}\!\!\!\mu_{n+1}(y)\rho_{n+1}^{(1)}(y;t)\dd y.
\label{ZERO0}\end{aligned}$$
Furthermore, integrating Eq. \[BCRHO\] over $a_{i\neq \ell}$ yields $n\rho_{n}^{(1)}(a=0;t) = (n-1)
\int_{0}^{\infty}\beta_{n-1}(y)\rho_{n-1}^{(1)}(y;t)\dd y$. Thus, Eq. \[ZERO0\] can be written in the form
$$\begin{aligned}
\displaystyle {\partial \rho_{n}^{(0)}(t) \over \partial t} = & (n-1)\int_{0}^{\infty}
\beta_{n-1}(y)\rho_{n-1}^{(1)}(y;t)\dd y - n \int_{0}^{\infty}(\beta_{n}(y)+\mu_{n}(y))\rho_{n}^{(1)}(y;t)\dd y +(n+1)\int_{0}^{\infty}\!\mu_{n+1}(y)\rho_{n+1}^{(1)}(y;t)\dd y.
\label{ZERO1}\end{aligned}$$
Eq. \[ZERO1\] describes the evolution of the probability $\rho_{n}^{(0)}(t)$ that the system contains $n$ individuals at time $t$ and it contains the single-particle marginal density $\rho_{n}^{(1)}(y;t)$. Upon deriving equations for $\rho_{n}^{(1)}(y;t)$, one would find that they depend on $\rho_{n}^{(2)}(y_{1}, y_{2};t)$, and so on. Therefore, the marginal probability densities form a hierarchy of equations, as is typically seen in classic settings such as the kinetic theory of gases [@MCQUARRIE] and the statistical theory of turbulence [@FRISCH]. Note that if the birth and death rates $\beta_{n}$ and $\mu_{n}$ are age-independent, they are constants with respect to the integral and Eq. \[ZERO1\] reduces to the familiar constant birth and death rate master equation for the simple birth-death process:
$$\begin{aligned}
{\partial \rho_{n}^{(0)}(t) \over \partial t} = & (n-1)\beta_{n-1}\rho_{n-1}^{(0)}(t)
-n(\beta_{n}+\mu_{n})\rho_{n}^{(0)}(t) + (n+1)\mu_{n+1}\rho_{n+1}^{(0)}(t),
\label{ZERO2}\end{aligned}$$
where $\rho_{n}^{(0)}(t)$ is the probability the system contains $n$ individuals at time $t$, regardless of their ages.
In general, integration of Eq. \[RHO0\] over $n-k\geq 0$ age variables leaves $k$ remaining independent age variables. The resulting kinetic equation for $\rho_n^{(k)}(\a_{k}; t)$ involves both $\rho_{n+1}^{(k+1)}(\a_{k},y;t)$ and boundary terms $\rho_{n}^{(k+1)}(\a_{k},a_{k+1}=0;t)$. These boundary terms can be eliminated by using the result obtained from integration of the boundary condition (Eq. \[BCRHO\]) over $n-k-1$ age variables. By exploiting the symmetry properties of the marginals $\rho_n^{(k)}$, we find
$$\begin{aligned}
\displaystyle {\partial \rho_{n}^{(k)}(t) \over \partial t} +
\sum_{i=1}^{k}{\partial \rho_{n}^{(k)}(\a_{k};t)\over \partial a_{i}} = &
+ \left({n-k \over n}\right)\rho_{n-1}^{(k)}(\a_{k};t)
\sum_{i=1}^{k}\beta_{n-1}(a_{i}) +{(n-k)(n-k-1)\over n}\int_{0}^{\infty}\beta_{n-1}(y)
\rho_{n-1}^{(k+1)}(\a_{k},y;t)\dd y \nonumber \\
\: & \displaystyle -\rho_{n}^{(k)}(\a_{k};t)\sum_{i=1}^{k}\gamma_{n}(a_{i})
-(n-k)\int_{0}^{\infty}\gamma_{n}(y)
\rho_{n}^{(k+1)}(\a_{k},y;t)\dd y \label{RHOK1} \\
\: & \displaystyle + (n+1)\int_{0}^{\infty}\mu_{n+1}(y)
\rho_{n+1}^{(k+1)}(\a_{k},y;t)\dd y.\nonumber
\label{RHOK1}\end{aligned}$$
Each function $\rho_{n}^{(k)}$ in the hierarchy not only depends on the functions in the $n \pm 1$ subspace, but is connected to functions with $k+1$ and $k-1$ variables. The latter coupling arises through the boundary condition for $\rho_{n}^{(k)}$ which involves densities $\rho_{n}^{(k-1)}$. As with similar equations in physics, the hierarchy of equations cannot be generally solved, and either factorization approximations or truncation (such as moment closure) must be used.
We now show that the $k=1$ equation explicitly leads to the classic McKendrick-von Foerster equation and its associated boundary condition. For $k=1$, $\rho_{n}^{(1)}(a;t)\dd a$ is the probability that there are $n$ individuals and that if one is randomly chosen, it will have age between $a$ and $a+\dd a$. Therefore, the probability that we have $n$ individuals of which any one has age between $a$ and $a+\dd a$ is $n\rho_{n}^{(1)}(a; t)\dd a$. Summing over all possible population sizes $n\geq 1$ gives us the probability $\rho(a,t)\dd a$ that the system contains an individual with age between $a$ and $a+\dd a$:
$$\rho(a,t) \equiv \sum_{n=0}^{\infty}n\rho_{n}^{(1)}(a;t).$$
Multiplying Eq. \[RHOK1\] (with $k=1$) by $n$ and summing over all positive integers $n$, we find after carefully cancelling like terms
$$\displaystyle {\partial \rho(a,t) \over \partial t} +
{\partial \rho(a,t)\over \partial a} = -\sum_{n=1}^{\infty}n\mu_{n}(a)\rho_{n}^{(1)}(a;t).
\label{MVF0}$$
Equation \[MVF0\] generalizes the McKendrick-von Foerster model to allow for population-dependent death rates, but does not reduce to the simple form shown in Eq. \[MCKENDRICK0\]. Population-dependent effects in equation for $\rho(a,t)$ requires requires knowing the “single-particle” density function $\rho_{n}^{(1)}(a;t)$ and subsequently all higher order distribution functions.
A boundary condition is naturally recovered by integrating over all ages but $a_{\ell}$ in Eq. \[BCRHO\] and summing over all $n$: $$\begin{aligned}
\sum_{n=1}^{\infty}n\rho_{n}^{(1)}(a=0;t) & \equiv \rho(a=0,t) = \sum_{n=2}^{\infty}(n-1)
\!\int_{0}^{\infty}\!\!\!\beta_{n-1}(y)\rho_{n-1}^{(1)}(y;t)\dd y.
\label{BCMVF0}\end{aligned}$$ These equations show that the McKendrick-von Foerster equation is recovered only if both $\mu_{n}(a) = \mu(a)$ and $\beta_{n}(a) =
\beta(a)$ are independent of population size. In this case, $\mu(a)$ can be pulled out of the sum in Eq. \[MVF0\] and $\sum_{n=1}^{\infty}n\mu_{n}(a)\rho_{n}^{(1)}(a;t)=
\mu(a)\rho(a,t)$. Similarly, $\int_{0}^{\infty}\beta(y)\left[\sum_{n=2}^{\infty}(n-1)
\rho_{n-1}^{(1)}(y;t)\right]\dd y =
\int_{0}^{\infty}\beta(y)\rho(y,t)\dd y$, which is the simple boundary condition associated with the classic McKendrick-von Foerster model. This derivation clearly shows that population-dependent birth and death rates cannot be readily incorporated into an age-dependent model, even one that is deterministic, without considering the hierarchy of population densities.
Discussion and conclusions
==========================
We have developed a complete kinetic theory for age-structured birth-death processes. To stochastically describe the age structure of a population requires a higher dimensional probability density. The evolution of this high-dimensional probability density mirrors that found in the Boltzmann equation for one-dimensional, ballistic, noninteracting gas dynamics. However, one crucial difference is that the number of individuals can increase or decrease according to the age-dependent birth and death rates. Thus, the dynamics are determined by a phase-space-conserving Liouville operator so long as the number of individuals does not change [@MCQUARRIE]. Once an individual is born or dies, the system jumps to another manifold in a higher or lower dimensional phase-space, immediately after which conserved dynamics resume until the next birth or death event. Such variable number dynamics share similarities with the kinetic theory of chemically reacting gases [@REACTING0].
Our main mathematical results are Eqs. \[RHO0\] and \[BCRHO\]. These equations show that birth-death dynamics couple densities associated with different numbers $n$ and describes the process in terms of ballistically moving particles all moving with unit velocity in the age “direction.” The individual particles can die at rates that depend on their distance from their origin (birth). Particles can also give birth at rates dependent on their age. The injection of newborns at the origin (zero age) is described by the boundary condition (Eq. \[BCRHO\]).
One important advantage of our approach is that it provides a natural framework for incorporating both age- and population-dependent birth and death rates into a stochastic description, which has thus far not been possible with other approaches. In general, our kinetic equations need to be solved numerically; however, we found analytic expressions for $\rho_{n}(\a_{n};t)$ when either birth or death vanishes and the other is independent of population. Furthermore, we define marginal density functions and develop a hierarchy of equations analogous to the BBGKY hierarchy (Eq. \[RHOK1\]). These equations for the marginal densities allow one to construct any desired statistical measure of the process and are also part of our main results. We explicitly showed how a zeroth order equation leads to the equation for the marginal probability of observing $n$ individuals in the standard [*age-independent*]{} birth-death processes (Eq. \[ZERO2\]) [@LJSALLEN]. The first-order equation is also used to derive a hybrid equation for the mean density $\rho(a,t)$ that involves the single-particle density function $\rho_{n}^{(1)}(a;t)$ (which ultimately depends on higher-dimensional densities $\rho_{n}^{(k>1)}(\a_{k};t)$ through the hierarchy). Only when death is independent of population does the theory reduce to the deterministic McKendrick-von Foerster equation (Eq. \[MVF0\]) and the associated boundary condition (Eq. \[BCMVF0\]).
Extensions of our high-dimensional age-structured kinetic theory to more complex birth-death mechanisms such as sexual reproduction and renewal/branching processes can be straightforwardly investigated. The simple birth-death process we analyzed allows for the birth of only a single age-zero daughter from a parent at a time. We note that the Bellman-Harris process described via generating functions [@JAGERS1968; @SHONKWILER1980] (which can describe age-dependent death and branching, but cannot be used to model population-dependent dynamics) assumes self-renewal at each branching event. That is, two (or more) daughters of zero age are simultaneously produced from a parent. Such differences in the underlying birth process can lead to qualitative differences in important statistical measures beyond mean-field, such as first passage times [@CHOUJTB]. The branching/renewal process, as well as sexual reproduction, requires nontrivial extensions of our kinetic theory and will be explored in a future investigation.
Acknowledgements
================
This research was supported in part at KITP by the National Science Foundation under Grant No. PHY11-25915. TC is also supported by the NIH through grant R56 HL126544 and the Army Research Office through grant W911NF-14-1-0472.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The antiferromagnetic pyrochlore material [[NaCaCo$_2$F$_7$]{}]{} is a thermal spin liquid over a broad temperature range ($\approx$ 140 K down to $T_F = 2.4 $K), in which magnetic correlations between Co$^{2+}$ dipole moments explore a continuous manifold of antiferromagnetic XY states [@ross2016static]. The thermal spin liquid is interrupted by spin freezing at a temperature that is $\sim 2$ % of the mean field interaction strength, leading to short range static XY clusters with distinctive relaxation dynamics. Here we report the low energy inelastic neutron scattering response from the related compound [[NaSrCo$_2$F$_7$]{}]{}, confirming that it hosts the same static and dynamic spin correlations as [[NaCaCo$_2$F$_7$]{}]{}. We then present the single-ion levels of Co$^{2+}$ in these materials as measured by inelastic neutron scattering. An intermediate spin orbit coupling model applied to an ensemble of trigonally distorted octahedral crystal fields accounts for the observed transitions. The single-ion ground state of Co$^{2+}$ is a Kramers doublet with a strongly XY-like $g$-tensor ($g_{xy}/g_{z} \sim 3$). The local disorder inherent from the mixed pyrochlore $A$ sites (Na$^{+}$/Ca$^{2+}$ and Na$^{+}$/Sr$^{2+}$) is evident in these measurements as exaggerated broadening of some of the levels. A simple model that reproduces the salient features of the single-ion spectrum produces approximately 8.4% and 4.1% variation in the $z$ and $xy$ components of the $g$-tensor, respectively. This study confirms that an $S_{\text{eff}} =1/2$ model with XY antiferromagnetic exchange and weak exchange disorder serves as a basic starting point in understanding the low temperature magnetic behavior of these strongly frustrated magnets.'
author:
- 'K.A. Ross'
- 'J.M. Brown'
- 'R.J. Cava'
- 'J.W. Krizan'
- 'S. E. Nagler'
- 'J.A. Rodriguez-Rivera'
- 'M. B. Stone'
title: |
Single-ion properties of the $S_{\text{eff}}$ = 1/2 XY antiferromagnetic pyrochlores,\
Na$A^{\prime}$Co$_2$F$_7$ ($A^{\prime} = $ Ca$^{2+}$, Sr$^{2+}$)
---
Introduction {#sec:intro}
============
Magnetism on the pyrochlore lattice is a rich field of study that encompasses many unusual magnetic phenomena such as spin ice, spin liquids, and Order by Disorder (ObD) [@gardner2010magnetic]. Experimental access to these phenomena has been granted mainly by rare earth oxide pyrochlore materials. These have the general chemical formula $A_2B_2$O$_7$, with $A$ a magnetic trivalent rare earth cation and $B$ a non-magnetic tetravalent cation such as Ti, Sn, Zr, Ge or Pt [@gardner2010magnetic; @shannon1968synthesis; @dun2015antiferromagnetic; @cai2016high]. Both the $A$ and $B$ sites independently form a pyrochlore-type sublattice, a highly-frustrated three dimensional network composed of corner-sharing tetrahedra (Figure \[fig:struct\] a). The anisotropic nature of the rare earth magnetic moments, arising due to strong spin-orbit coupling (SOC) combined with crystal electric field (CEF) effects, has offered fascinating variations of magnetic behavior. For instance, local Ising anisotropy is required for generating spin ice and its emergent magnetic monopoles in Ho$_2$Ti$_2$O$_7$ and Dy$_2$Ti$_2$O$_7$ [@castelnovo2012spin], while XY anisotropy can lead to ObD [@bramwell1994order; @wong2013ground; @mcclarty2014order], as suggested for the pyrochlore material Er$_2$Ti$_2$O$_7$ [@champion2003er; @champion2003er; @zhitomirsky2012quantum; @savary2012order; @oitmaa2013phase; @maryasin2014order; @andreanov2015order]. However, the $4f$ electrons which are responsible for this magnetism have weak interaction strengths on the order of 1 K, requiring experiments to be done at millikelvin temperatures in order to access their magnetic correlations. This can limit the exploration of the lowest temperature states in these frustrated systems. Furthermore, the ground states selected by these interactions can be quite sensitive to small amounts of chemical disorder, as has been observed in Yb$_2$Ti$_2$O$_7$ [@ross2012lightly; @yaouanc2011single] and Tb$_2$Ti$_2$O$_7$ [@taniguchi2013long].
The recently discovered $3d$ transition metal pyrochlore fluoride [[NaCaCo$_2$F$_7$]{}]{} (NCCF) has previously been suggested as a strongly interacting version of the antiferromagnetic (AFM) XY pyrochlore model [@ross2016static], potentially allowing for a detailed investigation of this model and its variations at lower effective temperatures. The XY anisotropy of effective spin 1/2 ($S_{\text{eff}} = 1/2$) magnetic moments in NCCF and the related compound [[NaSrCo$_2$F$_7$]{}]{} (NSCF) are confirmed here through the modeling of Co$^{2+}$ single-ion levels by fitting to inelastic neutron scattering measurements. Furthermore, both materials are shown to have nearly identical spin correlations in their spin frozen states, consisting of short range ordered XY clusters and associated dynamics, despite differences in the strength of chemical disorder.
NCCF and NSCF belong to a family of $3d$ transition metal fluoride pyrochlores [@krizan2014nacaco; @krizan2015nasrco2f7; @krizan2015nacan; @sanders2016nasrmn2f7]. The chemical formula and structure are similar to the rare earth oxide pyrochlores, taking the form $A_2B_2$F$_7$, where $A$ is a split site of Na$^+$/$A^{\prime}$ ($A^\prime$ = Ca$^{2+}$, Sr$^{2+}$) and $B$ is six-fold coordinated Co$^{2+}$, Ni$^{2+}$, Fe$^{2+}$ or Mn$^{2+}$ (Figure \[fig:struct\]). Unlike the previously studied fluoride-pyrochlore CsNiCrF$_6$ [@zinkin1997short], the magnetic $B$-site is chemically uniform, hosting just one magnetic species. The average structures of these compounds as measured by x-ray diffraction show well-ordered pyrochlores (space group $Fd\bar{3}m$), consistent with random distribution of cations on the $A$ site. Despite the materials being well-ordered on average, the local structure near the $B$ cations will be distorted away from the average trigonal $D_{3d}$ point group due to the $A$ site disorder, leading to variations in exchange parameters and single-ion anisotropy (Figure \[fig:struct\] b). In light of this, perhaps it is not surprising that all of the materials in this series eventually display spin freezing transitions. However, the spin freezing occurs at very low *effective* temperatures; their Curie-Weiss temperatures range from $\theta_{CW} \sim$ -70 [@sanders2016nasrmn2f7] to -140 K [@krizan2014nacaco], but the spins freeze only at 2 to 4 K. This gives large frustration indices of $f = \frac{|\theta_{CW}|}{T_F} \approx 19$ to 58, suggesting that the strength of the disorder, the presumed cause of the spin freezing, is weak compared to the overall interaction strength. An interesting comparison in this regard can be made to the $B$-site disordered Yb-based pyrochlore Yb$_2$GaSbO$_7$; with its $\theta_{CW}= -1.15$ K and lack of a freezing transition down to the lowest measured temperature of 20 mK, this material demonstrates similarly “ineffective” exchange disorder relative to its (much weaker) average interaction strength ($f\ge 58$)[@blote1969heat; @hodges2011magnetic].
Unlike the rare earth oxides, nearly ideal Heisenberg moments with large spins should be expected for most of the members in this new series of transition metal fluoride pyrochlores. This is due to the quenching of orbital angular momentum expected for most octahedrally coordinated $3d$ transition metals. This means that phenomenology like spin ice or ObD will likely not pertain to most members of this series. The exceptions to this are the Co$^{2+}$ compounds which we study here: NCCF and NSCF. The free ion Co$^{2+}$ ($3d^7$) forms a $^4F$ ground term with $S=3/2$ and $L=3$, and when placed in a octahedral coordination the CEF and SOC conspire to form an $S_{\text{eff}}=1/2$ single-ion ground state [@buyers1971excitations; @abragam2012electron]. A distorted octahedral environment, such as the average environment of NCCF and NSCF, will lead to single-ion anisotropy in the $S_{\text{eff}} = 1/2$ states (Fig. \[fig:struct\] c)[@lines1963magnetic; @goff1995exchange; @abragam2012electron]. At temperatures low enough that the ground doublet states are the only relevant degrees of freedom, it is then possible model the interactions between effective spin 1/2 operators, in which the single-ion anisotropy encoded by the $g$-tensor is projected into the effective exchange interactions [@lines1963magnetic; @white2006quantum; @ross2011quantum; @guitteny2013palmer]. Thus, it seems that NCCF and NSCF could serve as new “high temperature” examples in which this successful method, well-developed for the rare earth oxide series, could be used [@ross2011quantum; @savary2012order; @yan2016general].
The static and dynamic spin correlations in NCCF were previously measured by inelastic neutron scattering [@ross2016static]. It was found that below the freezing temperature ($T_F = 2.4$ K), short range order (SRO) of the spins develops with a correlation length of 16 Å, corresponding to XY AFM configurations from the $\Gamma_5$ irreducible representation of the tetrahedral point group. This manifold is spanned by two basis states called $\psi_2$ and $\psi_3$, shown in Figure \[fig:struct\] d). In the XY AFM pyrochlore model, the $q=0$ long range ordered (LRO) magnetic structures based on these two states are accidentally degenerate [@bramwell1994order; @champion2004soft; @wong2013ground; @mcclarty2014order], and either one can be selected by various ObD mechanisms, as has been discussed at length in the context of Er$_2$Ti$_2$O$_7$. The observed selection of the non-coplanar $\psi_2$ state in that material [@poole2007magnetic] has been argued to occur either via quantum and thermal fluctuations (ObD) [@champion2003er; @zhitomirsky2012quantum; @savary2012order; @oitmaa2013phase], or from a non-ObD mechanism involving virtual excitations to higher crystal field levels [@petit2014order]. *Quenched* disorder in the form of dilution [@andreanov2015order] or exchange disorder [@maryasin2014order] has recently been predicted to select the coplanar $\psi_3$ state instead, and this has been studied in yttrium-diluted Er$_2$Ti$_2$O$_7$ [@gaudet2016magnetic]. In NCCF, despite a clear mechanism for exchange disorder and the presence of XY AFM correlations, an LRO state is not selected.
The low energy inelastic neutron scattering (INS) response of NCCF shows non-dispersive, diffusive excitations with a distinctive intensity vs. $\vec{Q}$ pattern, extending to approximately $E =$ 10 meV [@ross2016static], as well as a broad distribution of relaxation times in the $\mu$eV range, as probed by NMR [@sarkar2016unconventional]. These dynamic signatures persist above the freezing temperature up to at least $T=$14 K (INS [@ross2016static]) and 20 K (NMR [@sarkar2016unconventional]). These data, combined with the high Curie-Weiss temperature ($\theta_{CW}$ = -140 K) relative to the freezing temperature ($T_F = 2.4$ K) giving a frustration index of $f = 58$, show that NCCF hosts an XY thermal spin liquid, i.e., a strongly correlated but disordered state dominated by entropy, over a large temperature range. Surprisingly, NCCF resists ordering or freezing to much lower effective temperatures than the canonical XY pyrochlore Er$_2$Ti$_2$O$_7$ ($f = 20$)[@ruff2008spin], suggesting that this material could be closer to a classical phase boundary in the general anisotropic exchange model developed for the rare earth oxides [@yan2016general]. To make further progress in modeling NCCF, it is crucial to establish the relevance of the general anisotropic pseudo-spin 1/2 model. This relies upon the understanding of the single-ion Hamiltonian for Co$^{2+}$ in these materials.
In this article we first present low energy inelastic and elastic neutron scattering measurements on the compound NSCF and compare it to the previously reported measurements on NCCF, demonstrating that despite their different $A^{\prime}$ sites the two compounds share the same experimental signatures and could be treated by the same theoretical approach. We then present the observed single-ion levels for both compounds, measured by INS over the energy range of $E$ = 30 - 1500 meV. The observed INS response as well as the dc magnetic susceptibility is well described by a disorder-averaged intermediate SOC model. This model confirms that the single-ion ground state in both materials is a well-isolated Kramers doublet, i.e., $S_{\text{eff}} = 1/2$, and provides the average $g$-tensors, which are observed to be strongly XY-like and have approximately 8% variation due to local disorder.
The contents of the paper are as follows: in Section \[sec:experimental\] we give the experimental details of the INS measurements. In Section \[sec:single-ion\] we describe single-ion model used to fit the high energy INS data. Section \[sec:lowEresults\] demonstrates the equivalence of the spin correlations in the two materials, and presents detailed low energy INS measurements at temperatures above $T_F$ for the first time (i.e., the thermal spin liquid regime). Section \[sec:singleionresults\] details the single-ion results and fits. The results are further discussed in Section \[sec:discussion\], and conclusions from this study are presented in Section \[sec:conclusions\].
Experimental Method {#sec:experimental}
===================
Single crystals of NCCF and NSCF (space group $Fd\bar{3}m$, room temperature lattice constants $a = 10.4189$ Å and 10.545 Å, respectively) were grown via the Bridgman-Stockbarger method in an Optical Floating Zone furnace, as previously reported [@krizan2014nacaco].
We studied a 3.527 g single crystal of NSCF using the MACS spectrometer at the NIST Center for Neutron Research [@rodriguez2008macs]. The dynamic structure factor, $S(\vec{Q},E)$, was measured in the \[$HHL$\] reciprocal lattice plane. Neutrons with a final energy $E_f = 3.7$ meV were selected, and post-sample BeO filters were used to remove higher harmonic contamination and reject neutrons for which $E_f>3.7$ meV. For elastic scattering, a Be filter preceded the sample, while for inelastic scattering no incident filters were used. The resulting energy resolution was $\delta E$ = 0.17 meV at the elastic line. These data are compared to those previously published on a 0.87 g crystal of NCCF, which was also taken using the MACS spectrometer in a similar configuration [@ross2016static]. The relative masses cannot be used to compare the intensities directly due to differences in crystal mounting. The NSCF and NCCF data were scaled relative to each other to match the inelastic intensities near (002) and (111) (Fig. \[fig:MACS\]). The same scaling also produced matching (220) magnetic intensities below $T =$ 1.7 K.
In order to investigate the Co$^{2+}$ single ion levels in NCCF and NSCF, higher energy INS experiments were performed using the SEQUOIA time-of-flight chopper spectrometer at the Spallation Neutron Source, Oak Ridge National Laboratory [@granroth2006sequoia]. The crystals were oriented with \[HHL\] in the horizontal scattering plane. Data were taken at $T$ = 5 K and 200 K, using incident energies of $E_i =$ 60, 250, 700, and 2500 meV. This wide range of incident energies was employed to probe the large dynamic range expected for single ion energy levels of Co$^{2+}$. For these energies, the T$_0$ Chopper and Fermi Chopper (FC1 or FC2) speeds were set to the following values: for $E_i$ = 60 meV, 60 Hz (T$_0$) and 420 Hz (FC2); for $E_i$ = 250 meV, 120 Hz (T$_0$) and 480 Hz (FC1); for $E_i$ = 700 meV, 150 Hz (T$_0$) and 600 Hz (FC1); and for $E_i$ = 2500 meV, 180 Hz (T$_0$) and 600 Hz (FC1). The elastic energy resolutions in these configurations were $\delta E$ = 1.92, 12.71, 44.96 and 276.44 meV, respectively. Note that $\delta E$ decreases as the neutron energy transfer increases for each $E_i$, and the full energy-dependent resolution for SEQUOIA [@granroth2006sequoia] was employed in our fits.
Single-ion calculations for Co$^{2+}$ {#sec:single-ion}
=====================================
In this section we present a method for calculating the single-ion levels in Co$^{2+}$ within an intermediate spin orbit coupling scheme, as well as the resulting INS response and dc magnetic susceptibility. This method was used to fit the data presented in Section \[sec:singleionresults\]. The method is similar to that used for the rare earth based magnets [@gaudet2015neutron; @babkevich2015neutron], and results in similar phenomenology. The single ion levels for Co$^{2+}$ in a trigonally distorted octahedron are Kramers doublets with anisotropic moments described by a $g$-tensor.
Intermediate spin orbit coupling Hamiltonian {#sec:hamiltonian}
--------------------------------------------
The single-ion Hamiltonian is approximated by two contributions,
$$H_\text{ion} = H_\text{CEF}(\vec{L}) + H_{\text{SOC}}(\vec{L},\vec{S}),
\label{eqn:Hion}$$
where $H_\text{CEF}(\vec{L})$ is the crystal electric field (CEF) Hamiltonian that acts only on the orbital angular momentum subspace, and H$_\text{SOC}$ is the spin orbit coupling (SOC) term. The CEF Hamiltonian can be written in general as:
$$H_{\text{CEF}} = \sum_{l,m} B_{l,m} \hat{O}_{l,m}
\label{eqn:Hcef}$$
where $\hat{O}_{l,m}$ are the Stevens operator equivalents [@stevens1952matrix; @hutchings1964point; @mcphase_manual]. For transition metal ions, these operators are written in terms of the orbital angular momentum matrix operators $\hat{L_+}$, $\hat{L_-}$ and $\hat{L_z}$ (as opposed to the total angular momentum operators $\hat{J_+}$, $\hat{J_-}$ and $\hat{J_z}$ relevant for $f$-electron systems)[@hutchings1964point]. For the trigonal point group symmetry relevant to the average local environments of Co$^{2+}$ in NCCF and NSCF, $(l,m)= (2,0),(4,0)$ and $(4,3)$ are the only nonzero terms [@hutchings1964point]. The $B_{l,m}$ values can be extracted by fitting to INS data. The spin orbit coupling term is given by,
$$H_{\text{SOC}} = p \lambda (\vec{S}\cdot \vec{L}) = p \lambda(\hat{S}_x \hat{L}_x + \hat{S}_y \hat{L}_y + \hat{S}_z \hat{L}_z),
\label{eqn:Hsoc}$$
where for the free ion Co$^{2+}$ $\lambda = -22.32$ meV [@kanamori1957theory], and $p$ is the “orbital reduction parameter” that can be used to account for changes in effective SOC strength due to covalency [@buyers1971excitations]. The $x$ and $y$ spin operators are linear combinations of $\hat{S}_+$ and $\hat{S}_-$ as usual: $\hat{S}_x = \frac{1}{2}(\hat{S}_+ + \hat{S}_-)$ and $\hat{S}_y = \frac{1}{2i}(\hat{S}_+ - \hat{S}_-)$ and similarly for $\hat{L}_x$ and $\hat{L}_y$.\
\
The full single-ion Hamiltonian (Eqn. \[eqn:Hion\]) can be diagonalized within the $28 \times 28$ manifold of states formed by the $|L_z, S_z\rangle$ basis of the $S=3/2$ and $L=3$ free ion term ($^4F$) for Co$^{2+}$. A trigonally distorted octahedral coordination, as found in NCCF and NSCF, produces 14 Kramers doublets. The presence of a doublet ground state permits the description of the magnetic moments as pseudo-spin 1/2 ($S_{\text{eff}} = 1/2$) at sufficiently low temperatures. This doublet can be described as an anisotropic magnetic moment with strength given by the $g$-tensor [@abragam2012electron]. In the trigonal symmetry appropriate to NCCF and NSCF, the $g$-tensor has two independent components, $g_z$ and $g_{xy}$, which refer, respectively, to the local \[111\] direction pointing into the center of the tetrahedron, and the plane perpendicular to that (Fig. \[fig:struct\]c). The $g$-tensor components are given by the matrix elements of the magnetic moment operator in the subspace of the two ground state wavevectors ($|\nu_1 \rangle$ and $|\nu_2\rangle$),
$$g_{xy} = -2 \langle\nu_1| (\hat{L}_x + 2 \hat{S}_x)|\nu_2\rangle,$$ and $$g_{z} = 2 \langle\nu_2| (\hat{L}_z + 2 \hat{S}_z)|\nu_2\rangle.$$
The resulting saturated magnetic moment sizes in the principal directions are $\mu_z = \frac{g_z}{2} \mu_B$ and $\mu_{xy} = \frac{g_{xy}}{2} \mu_B$.
Comparison to inelastic neutron scattering {#sec:INS_comparison}
------------------------------------------
The dynamic structure factor, $S(\vec{Q},E)$, is related to the observed intensity of inelastic neutron scattering via $I(\vec{Q},E) = \frac{k_f}{k_i} f(|Q|)^2 S(\vec{Q},E)$, where $k_i$ and $k_f$ are the initial and final wavenumbers of neutrons scattered with energy transfer $E = \frac{\hbar}{2m} (k_i^2 - k_f^2)$, and $f(|Q|)$ is the magnetic form factor [@squires2012introduction]. The single-ion dynamic structure factor at constant $|Q|$ associated with Eqns. \[eqn:Hion\] to \[eqn:Hsoc\] can be calculated as follows [@jensen1991rare]:
$$S(E) = C \sum_{n,n',\alpha}\frac{e^{-\beta E_n}}{Z}\frac{\Gamma_{\langle n,n' \rangle} |\langle \nu_n | \hat{L}_\alpha + 2 \hat{S}_\alpha | \nu_{n'} \rangle|^2}{([E_{n'} - E_{n}] - E)^2 + \Gamma_{\langle n,n' \rangle}^2},
\label{eqn:SofE}$$
where $\alpha = x,y,z$, while $n$ and $n'$ label eigenstates of Eqn. \[eqn:Hion\]. $C$ is a scale factor applied to match the measured intensity of constant $|Q|$ data (note that at constant $|Q|$ the form factor $f(|Q|)^2$ can be absorbed into $C$). The partition function is $Z = \sum_n \exp{(-\beta E_n)}$ with $\beta = 1/k_B T$. $|\nu_n\rangle$ is the wavevector from the single ion calculation corresponding to energy eigenvalue $E_n$. A Lorentzian half width at half maximum (HWHM), $\Gamma_{\langle n,n' \rangle}$, accounts for line broadening due to instrumental resolution, finite excitation lifetimes, or averaged dispersion of the transitions between state $n$ and $n'$. To fit the data presented in Section \[sec:singleionresults\], we set $\Gamma_{\langle n,n' \rangle}$ to the energy transfer-dependent instrument resolution for all transitions except those involving the ground doublet, $n = 1,2$, to the first excited doublet $n' = 3,4$. In that case, $\Gamma_{\langle n,n' \rangle} = 13.5$ meV accounts for the increased width due to the finite dispersion of the first excited level (Fig. \[fig:firstlevel\]). An elastic peak with a full width at half maximum (FWHM) fixed at the instrumental resolution was also added to account for all sources of nuclear elastic scattering, both coherent and incoherent.
To account for disorder in the local environment of Co$^{2+}$ brought about by the randomly mixed $A$ site in the crystal structures of NCCF and NSCF, the average $S(E)$ can be calculated over an ensemble of ions with slightly varying $B_{l,m}$ parameters in Eqn. \[eqn:Hcef\]. If the local site symmetry remains trigonal, as expected based on results of vibrational spectroscopy from the related compound NaCaMg$_2$F$_7$ [@oliveira2004crystal], there remain only three parameters that enter into $H_{\text{CEF}}$. For simplicity we chose a linear distribution of the $B_{4,0}$ parameter, allowing the width of the distribution ($\Delta B_{4,0}$) to be fit. The reasoning for this choice is described in Section \[sec:discussion\].
Magnetic Susceptibility {#sec:suscept}
-----------------------
The eigenvectors and eigenvalues of Eqn. \[eqn:Hion\] also permit the calculation of the (powder averaged) single ion magnetic susceptibility as a function of temperature [@jensen1991rare; @kimura2014experimental],
$$\begin{gathered}
\chi_{\text{ion}}(T) = \frac{N_A \mu_B^2}{3 k_B Z} \sum_\alpha \bigg(\frac{ \sum_{n} |\langle \nu_n | \hat{L}_\alpha + 2\hat{S}_\alpha | \nu_n \rangle|^2 e^{-E_n/T}}{T} \\ + \sum_n\sum_{m\neq n} |\langle \nu_m | \hat{L}_\alpha + 2\hat{S}_\alpha | \nu_n \rangle|^2 \frac{e^{-E_n/T} - e^{-E_m/T}}{E_m - E_n} \bigg).
\label{eqn:suscept}\end{gathered}$$
This expression includes the van Vleck susceptibility, but excludes the diamagnetic susceptibility $\chi_{\text{dia}}$, and does not account for the mean field interaction between magnetic moments. To compare to the measured dc susceptibility, the following mean field approximation can be used [@white2006quantum],
$$\chi_{\text{MF}} = \chi_{\text{dia}} + \frac{\chi_{\text{ion}}}{1 + \lambda_W \chi_{\text{ion}}}.
\label{eqn:chiMF}$$
Here, $\lambda_W$ is the Weiss molecular field constant, which accounts for the mean exchange interactions. With this sign convention, a positive value of $\lambda_W$ indicates AFM interactions.
Results {#sec:results}
=======
Low energy dynamic structure factor in [[NaSrCo$_2$F$_7$]{}]{} {#sec:lowEresults}
--------------------------------------------------------------
The low temperature ($T = 1.7$ K) and low energy ($E < 5$ meV) magnetic neutron scattering response from NSCF is compared to that from NCCF in Figure \[fig:MACS\]. The NCCF data was previously published in Ref. . For the elastic scattering presented for both materials, the equivalent high temperature ($T = 14$ K $> T_F$) elastic maps were subtracted in order to isolate the magnetic scattering. For the inelastic scattering, an empty can subtraction was performed. The neutron scattering intensity of NCCF has been scaled such that the (220) magnetic elastic peak intensity matches NSCF, which also produces equivalent intensities for the inelastic scattering. The neutron scattering patterns are nearly identical over the whole energy range, indicating that the static and dynamic spin correlations are equivalent in both compounds. An interpretation of this scattering in terms of static and dynamic XY correlations has been presented in Ref. ; this interpretation is expected to qualitatively apply to both materials. The sharp features near the (002) and (00$\bar{2}$) positions in NSCF seen in Fig. \[fig:MACS\] a) appear in the elastic channel using several choices of incident energies, but can be made to vanish for other choices, signifying that their origin is multiple scattering. Such (002) scattering has been observed before in rare earth titanates and was previously thought to be caused by symmetry lowering, but has recently been identified as multiple scattering in those materials as well [@baroudi2015symmetry].
The data shown in Fig. \[fig:MACS\] was collected below the freezing temperatures of NCCF and NSCF (2.4 K and 3.0 K, respectively). Above the freezing transition, it was shown previously in NCCF that the elastic magnetic scattering vanishes, while inelastic magnetic scattering at E = 0.5 meV persists to at least $T=14$ K and retains the same distinctive pattern in reciprocal space [@ross2016static]. In NSCF we have confirmed the persistence of this inelastic scattering above $T_F$ at $T=14$ K. The inelastic scattering at $E=0.25$ meV is broader at 14 K compared to 1.7 K, implying shorter ranged dynamic correlations (Fig. \[fig:Tdep\] a). The full energy dependence of the scattering up to 5 meV is compared for the two temperatures in Fig. \[fig:Tdep\] b), which shows that the spectrum has similar features above and below $T_F$, but with enhanced low energy spectral weight at $T=$14 K. The energy integrated scattering at $T = 14$ K in NSCF ($E = 0.2$ to $5.0$ meV) over the \[HHL\] plane is shown in Fig. \[fig:Tdep\] c). This scattering could be compared to equal time correlation functions, such as those calculated in Ref. .
Single ion levels and average $g$-tensors of [[NaSrCo$_2$F$_7$]{}]{} and [[NaCaCo$_2$F$_7$]{}]{} {#sec:singleionresults}
------------------------------------------------------------------------------------------------
Transitions between single ion levels of Co$^{2+}$ were observed in NCCF and NSCF using the SEQUOIA time of flight neutron spectrometer. The energies of the observed transitions were found range from 28 to 908 meV, requiring the use of multiple incident energies. The first excited level measured using $E_i$ = 60 meV in both NCCF and NSCF is shown in Figure \[fig:firstlevel\], at both $T = 5$ and 200 K. In both materials this level is located near 30 meV, persists to temperatures above $|\theta_{CW}|$ (demonstrating its single-ion origin), and displays the characteristic decreasing intensity vs. $|Q|$ dependence of magnetic excitations (Fig. \[fig:CEF\]a)). At $T = 5$ K the first excited level displays distinct dispersion, suggesting exchange-induced mixing of the ground state and first excited level [@buyers1971excitations]. Note that the low energy spin excitations presented in Section \[sec:lowEresults\] are also identifiable in Fig. \[fig:firstlevel\] below $\sim$ 10 meV; these are due to the magnetic correlations discussed in Ref. , rather than single-ion levels.
The full set of observed single-ion levels in NCCF at $T=5$ K are shown in Figure \[fig:CEF\] along with results of fitting the model described in Section \[sec:single-ion\]. Some of these features are significantly broader than the energy resolution as expected, since the local variation in Co environment should produce finite energy ranges over which transitions are observed. The measured energies of the single-ion levels and corresponding energy widths are tabulated in Table \[tab:measured\_peaks\] for both NCCF and NSCF.[^1] All neutron scattering data in Fig. \[fig:CEF\] have a corresponding background subtracted and are corrected for the $k_f/k_i$ factor described in Section \[sec:INS\_comparison\], thus representing the quantity $f(|Q|)^2 S(Q,E)$ in arbitrary units. The orientational average of these single crystal data are presented in order to provide a concise overview of the excitations; no significant dispersion was detectable in any but the first excited level, and this dispersion is well-accounted for in the model by an increased FWHM, as discussed below. For a uniform Co$^{2+}$ environment, 14 levels (including the ground state) are expected in total, each of them doubly degenerate according to Kramers theorem. Note that at $T = $ 5 K, a low temperature compared to the energy of the first excited level ($\sim 300$ K), only transitions from the ground state to the excited states will be observable. Additionally, some of these transitions may have weak intensities, depending on the strength of the transition matrix elements between eigenstates of Eqn. \[eqn:Hion\]. The intensities are weighted by the square of the transition matrix elements of the magnetic moment operator, $M^2_{\langle n,n' \rangle} = \sum_\alpha |\langle \nu_n | \hat{L}_\alpha + 2 \hat{S}_\alpha | \nu_{n'} \rangle|^2$, which enter into Eqn. \[eqn:SofE\].
----------------------------- ------------------------ ----------------------------- ------------------------
Energy (meV) FWHM (meV) Energy (meV) FWHM (meV)
28.05(2)$^a$ 27.50(7) 29.31(2)$^a$ 29.53(8)
$\sim$ 46 - 101$^a$ broad $\sim$ 41 - 100$^a$ broad
121.9(1)$^a$ 9.1(4) 120.7(2)$^a$ 7.2(9)$^*$
$\sim$ 139$^a$ broad $\sim$ 134$^a$ broad
364(1)$^b$ 24(3) – –
9.1(3)$\times$10$^{2}$ $^c$ 3.2(6)$\times$10$^{2}$ 9.0(3)$\times$10$^{2}$ $^c$ 3.7(6)$\times$10$^{2}$
----------------------------- ------------------------ ----------------------------- ------------------------
: Measured single-ion levels in [[NaCaCo$_2$F$_7$]{}]{} (NCCF) and [[NaSrCo$_2$F$_7$]{}]{} (NSCF), with central peak positions and full width at half maxima (FWHM). Due to time constraints, the peak expected near 360 meV in NSCF was not measured. The quoted error on the peak centers is the standard deviation of the fitted peak center. The FWHM are marked with an asterisk ($^*$) when they are resolution-limited. Instrument configurations are indicated by $a : E_i = 250$ meV, $b : E_i = 700$ meV, $c : E_i = 2500$ meV. []{data-label="tab:measured_peaks"}
Fig. \[fig:CEF\] a) shows measurements at $E_i = 250$ meV, with the intensity presented on a logarithmic color scale. This shows several of the magnetic excitations, including the first excited level near 30 meV (as in Fig. \[fig:firstlevel\]), a broad band of magnetic features spanning approximately 46 to 100 meV (indicated by a rectangle), and two higher energy features near 120 and 140 meV. Note that phonon scattering is also visible below 30 meV, but its intensity increases as function of $|Q|$, while the magnetic features show the opposite trend in intensity vs. $|Q|$ due to the magnetic form factor (see also Appendix \[sec:Qdep\]). Intensity vs. energy cuts from $Q=[4,5]$ Å$^{-1}$ are compared for NCCF and NSCF in Fig. \[fig:CEF\] b), showing the overall similarity between the single ion levels in the two compounds. The first excited level is at a slightly higher energy for NSCF (29.31(2) meV) compared to NCCF (28.05(2) meV). Two higher energy levels are shown in Fig. \[fig:CEF\] d) and e), located at $364 \pm 1$ meV and $908 \pm 30$ meV. In these high energy ranges, multi-phonon scattering creates a strong, sloping background and obscures these magnetic features. Nevertheless, their magnetic nature is confirmed by their $|Q|$-dependence. For the data shown in panel d), the single-ion level energy was extracted using a fit that included a sloping background plus a Gaussian. For the 908 meV feature, intensity vs. $E$ for two $|Q|$ ranges were compared, as shown in panel e). The precision of the energy determination for these higher modes is relatively poor due to decreased energy resolution of the higher incident energy instrument configurations.
Fits to these measured single-ion levels were carried out using a least squares minimization routine. The dynamic structure factor at constant $|Q|$, i.e, $S(E)$, was calculated from the single-ion model with local trigonal disorder described in Section \[sec:single-ion\], and was compared to the $E_i = 250$ meV data averaged from $|Q| = 4-5$ Å$^{-1}$ (Fig. \[fig:CEF\] c)). The measured energies and uncertainties of the higher excited levels ($364 \pm 1$ meV and $908 \pm 30$ meV) were included as constraints on the minimization, but no attempt was made to compare their relative intensities, which are difficult to accurately determine due to instrumental effects. The HWHM parameters $\Gamma_{\langle n,n' \rangle}$ in Eqn. \[eqn:SofE\] were held fixed to the calculated instrumental resolution for most values of $n$ and $n'$, but were increased to 13.5 meV for $n = 1,2$ and $n'=3,4$ to account for the dispersion of the first excited level. The coherent and incoherent nuclear elastic scattering was modeled by a Gaussian with FWHM fixed to the instrumental resolution ($\delta E = 12.7$ meV), and the area of that elastic signal was used as an additional fitting parameter ($C_E$). In total there were seven fitting parameters; the three average crystal field parameters ($B_{2,0}$, $B_{4,0}$, and $B_{4,3}$), a width of the linear distribution of $B_{4,0}$ ($\Delta B_{4,0}$), the orbital reduction parameter $p$, the elastic line area $C_E$, and an overall scale factor $C$. These fitted parameters and the associated average $g$-tensors with their percent variation are listed in Table \[tab:fitparams\]. The energy eigenvalues for the average $B_{l,m}$, their variation over the full $\Delta B_{4,0}$ range, as well as transition matrix elements, are listed in Table \[tab:eigenvalues\]. The energy levels as a function of $B_{4,0}$ are visualized in Figure \[fig:Elevels\]. Note that due to the overall similarity between the single-ion levels of NCCF and NSCF, as well as the limited accuracy of the simple model of local disorder (as discussed in Section \[sec:discussion\]), only one set of parameters are presented. Although this fit is not likely to be unique in the sense that the broad observed transitions and minimal modeling of disorder prevent a precise determination of the parameters, the obtained parameters adequately capture the main features of the single-ion levels in both materials and provide the key physical insights that can be gained from these data.
Finally, the high temperature dc magnetic susceptibility can be compared to the measured susceptibility of NCCF from Ref. , using the average $B_{l,m}$ parameters shown in Table \[tab:fitparams\]. Fitting the inverse susceptibility from $T=$100-300 K using Eqns. \[eqn:suscept\] and \[eqn:chiMF\], the two parameters $\chi_{\text{dia}} = -3.1(1)\times10^{-4}$ emu/mol and $\lambda_W = 23.99(5)$ mol/emu were determined, and excellent agreement is found with the high temperature data (Fig. \[fig:CEF\] f). This agreement further confirms the validity of the single-ion scheme determined from the neutron scattering results.
------------------ ---------------- --
Parameter Fitted value
$B_{2,0}$ 25.8(5) (meV)
$B_{4,0}$ 0.455(5) (meV)
$B_{4,3}$ 14.01(7) (meV)
$\Delta B_{4,0}$ 0.127(4) (meV)
$p$ 1.021(4)
$g_z$ 1.87
$g_{xy}$ 6.08
------------------ ---------------- --
: Parameters for the single-ion Hamiltonian, fit to inelastic neutron scattering data in NCCF. The values of $g_z$ and $g_{xy}$ are those obtained from the mean value of the CEF parameters (i.e., when $\Delta B_{40} = 0$).[]{data-label="tab:fitparams"}
State ($\nu_{n'}$) $E_{n'}$ for $\Delta B_{4,0} = 0$ Range of $E_{n'}$ $M^2_{\langle 1,n' \rangle}$+$M^2_{\langle 2,n' \rangle}$
----------------------- ----------------------------------- ------------------- -----------------------------------------------------------
$\nu_1$,$\nu_2$ 0.0 0.0 19.48
$\nu_3$,$\nu_4$ 27.7 3.5 17.06
$\nu_5$,$\nu_6$ 69.4 36.1 3.55
$\nu_7$,$\nu_8$ 90.0 25.4 0.37
$\nu_9$,$\nu_{10}$ 135.4 3.03 0.12
$\nu_{11}$,$\nu_{12}$ 154.3 13.0 0.04
$\nu_{13}$,$\nu_{14}$ 355.6 22.6 0.51
$\nu_{15}$,$\nu_{16}$ 367.6 28.6 0.67
$\nu_{17}$,$\nu_{18}$ 389.3 35.6 0.83
$\nu_{19}$,$\nu_{20}$ 411.1 41.1 0.17
$\nu_{21}$,$\nu_{22}$ 863.9 13.5 0.11
$\nu_{23}$,$\nu_{24}$ 896.7 12.2 0.71
$\nu_{25}$,$\nu_{26}$ 1103.9 16.5 0.02
$\nu_{27}$,$\nu_{28}$ 1133.3 15.8 0.01
: Calculated energy levels (in meV) from the fit shown in Fig. \[fig:CEF\], and the corresponding energy ranges arising from the variation in a crystal field parameter ($\Delta B_{4,0}$) which was used to model the local disorder at the Co$^{2+}$ site. The sum of the square of the transition matrix elements from the ground doublet to each excited doublet are listed, for the average CEF parameters ($M^2_{\langle n,n' \rangle} = \sum_\alpha |\langle \nu_n | \hat{L}_\alpha + 2 \hat{S}_\alpha | \nu_{n'} \rangle|^2$). Note that not all of the transitions listed in this table have been observed; refer to Fig. \[fig:CEF\] for the correspondence between the measured and calculated peak positions.[]{data-label="tab:eigenvalues"}
Discussion {#sec:discussion}
==========
### Single-ion model
The single-ion fits suggest that the reduction of spin orbit coupling (quantified by $p$) is not significant, since it refines to 1.021(4), which is nearly 1.0. This is in contrast to KCoF$_3$, for which $p = $0.93 [@buyers1971excitations]), indicating orbital moment reduction due to covalency. Thus, naively, NCCF and NSCF seem to have strongly ionic bonds.
The model used to fit the single-ion levels, presented in Section \[sec:single-ion\], accounts for the local disorder in a minimal way. The local symmetry at the Co$^{2+}$ site is assumed to remain trigonal, based on vibrational spectroscopy of the non-magnetic structural analog NaCaMg$_2$F$_7$ from Ref. . In that study it was found that the number and selection rules of vibrational modes was inconsistent with the full $Fd\bar{3}m$ symmetry of the average structure, but could be accounted for using the $F4\bar{3}m$ subgroup. Using the online software package [isodistort]{}[@campbell2006isodisplace], and assuming atomic displacements consistent with this symmetry lowering, we find that the point group of the Co site remains trigonal, but is reduced from $D_{3d}$ to $C_{3v}$ (note that on average, these local distortions would cancel out and produce the $Fd\bar{3}m$ space group that is observed for the average structure). This implies that only three CEF parameters are required, but that there may be some variation in their values from site to site, as the Co-F bond angles and distances (and therefore the CEF) are modified by the local disorder. In the absence of a more realistic model for the distribution of local disorder, we initially allowed a linear variation in each of the crystal field parameters in Eqn. \[eqn:Hcef\] and averaged the calculated $S(E)$ for each set of parameters. We found that only a variation in $B_{4,0}$ (which we call $\Delta B_{4,0}$) was necessary to reproduce the observed widths and positions of the peaks observed via INS (Fig. \[fig:CEF\]). Specifically, the energies of the $\nu_5$, $\nu_6$ and $\nu_7$, $\nu_8$ doublets are more widely distributed compared to the other levels, consistent with the broad band of magnetic intensity observed between $\sim$ 40 and 100 meV in INS, without creating excessive widening (beyond the measured widths) of the higher energy levels. It may be noted that this model does not perfectly account for the intensities and positions of all features, particularly near the modes near 121 meV. This is not unexpected, as the model we are using for the disorder is simplified. The numerical values we report for the $g$-tensor and its variation ($g_z = 1.87 \pm 8.4$ % and $g_{xy} = 6.08 \pm 4.1$%) can nevertheless be used as good estimates.
A more accurate model of the disorder could perhaps be developed based on additional experimental information about the local structure of these materials. To this end, a $^{23}$Na NMR study of NCCF has revealed an unusual high temperature response suggesting that there are two main crystallographic environments for Na [@sarkar2016unconventional]. So far no model to explain this observation has been proposed. A measurement of the neutron or x-ray Pair Distribution Function (PDF) could provide additional information.
Although there remains some uncertainty in the distribution of $g$-tensor values, the results of our single-ion study unequivocally indicate that NSCF and NCCF are XY pyrochlores with well-isolated Kramers doublets. This doublet nature of the low energy degrees of freedom is consistent with estimates of the spin entropy from specific heat which reach nearly $R$ln2 per mole of Co [@krizan2014nacaco; @krizan2015nasrco2f7], as well as with the sum rule analysis of low energy neutron scattering intensities [@ross2016static]. Our results show that interactions in these materials may be treated within a pseudo-spin 1/2 model with XY effective exchange. If the dominant interactions occur between nearest-neighbors, this would produce a model similar to the one used to describe Er$_2$Ti$_2$O$_7$ and Er$_2$Sn$_2$O$_7$ [@zhitomirsky2012quantum; @savary2012order; @guitteny2013palmer]. The variation in the $g$-tensor implies that exchange disorder is needed in the effective model.
### Susceptibility
The dc susceptibility in both NSCF and NCCF can be well-accounted for by the single-ion model using with the average $B_{l,m}$ parameters reported in Table \[tab:fitparams\]. The model naturally explains the high effective moments of these two materials ($\sim$ 6 $\mu_B$) as determined from Curie-Weiss fits of the inverse susceptibility over the range 100 K to 300 K [@krizan2014nacaco]. Another feature of the model is a downturn in $1/\chi$ near 50 K; this experimentally observed feature was previously assumed to be due to ferromagnetic correlations, but it can now be understood as changes in the thermal population of excited single-ion levels. The fitted Weiss molecular field constant, $\lambda_W$ = 23.99(4) mol/emu, confirms that AFM interactions are dominant, in agreement with the low energy INS results.
### Comparison of [[NaCaCo$_2$F$_7$]{}]{} and [[NaSrCo$_2$F$_7$]{}]{}
The neutron scattering results presented in Section \[sec:results\] indicate that the magnetic correlations and single ion properties of NSCF and NCCF are nearly identical. NCCF and NSCF have freezing temperatures of 2.4 K and 3.0 K, respectively. The increased freezing temperature of NSCF seems likely to be related to the larger differential ionic radius on the $A$ site: for 8-fold coordinated Na$^{+}$, $r = 1.18$ Å, for Ca$^{2+}$ $r = 1.12$ Å and for Sr$^{2+}$ $r = 1.26$ [@shannon1976revised]. The bigger this size difference, the more severe the local disorder is expected to be. The measured single-ion levels are very similar, with the only notable difference being the position of the first excited level, which is slightly higher in energy for NSCF compared to NCCF. However, within the accuracy of our model, there is no discernible difference in the single-ion ground state wavefunctions. Thus, it seems likely that both NSCF and NCCF could be treated within the same low energy effective model, but the strength of exchange disorder should be enhanced for NSCF.
### Co$^{2+}$ pyrochlores
Given the promise of extending the phenomenology of anisotropic rare earth pyrochlores to materials with higher interaction strengths, searching for other Co$^{2+}$-based pyrochlores seems desirable. To the best of our knowledge, only one other material has been reported with a Co$^{2+}$ pyrochlore sublattice: the spinel GeCo$_2$O$_4$. This material has been discussed in terms of its $S_{\text{eff}}=1/2$ single-ion ground state, and some unusual details of the higher energy excitation spectrum have been modeled in terms of molecular magnetism involving neighboring tetrahedra of these $S_{\text{eff}}=1/2$ moments [@tomiyasu2011molecular]. In GeCo$_2$O$_4$ the CoO$_6$ octahedra are not distorted, and the moment is therefore isotropic.
An interesting question for future exploration is whether there exists a Co$^{2+}$-based *spin ice* material (either quantum or classical). Such a material would have the advantage that the emergent monopole dynamics would occur at temperatures much easier to access, possibly even high enough to be useful for applications. If the U(1) quantum spin liquid phase of *quantum* spin ice could be produced in a Co$^{2+}$ pyrochlore, its emergent photon modes would have a more accessible bandwidth compared to the rare earth-based candidates, making them more easily identifiable with low energy inelastic neutron scattering and thermodynamic measurements. To achieve the necessary Ising anisotropy within a Co$^{2+}$ pyrochlore, the octahedral environment would need to be *elongated* along the $\langle 111 \rangle$ directions rather than compressed as it is for NCCF and NSCF.
Conclusions {#sec:conclusions}
===========
We have presented inelastic neutron scattering results on the pyrochlore materials [[NaCaCo$_2$F$_7$]{}]{} and [[NaSrCo$_2$F$_7$]{}]{}. The low energy response of [[NaSrCo$_2$F$_7$]{}]{} (below 10 meV) confirms that it hosts the same type of spin correlations as [[NaCaCo$_2$F$_7$]{}]{}, which are fully dynamic above the freezing temperatures (2.4 - 3.0 K) but static and short range correlated below. These correlations are well-described in terms of static and dynamic antiferromagnetic XY spin clusters, as discussed in Ref. . Below $T_F$, the XY configurations are frozen, short range correlated versions of the $\psi_2$ and $\psi_3$ states which are known to be selected through various types of order-by-disorder in the XY AFM models of the pyrochlore lattice.
Our main result is the measurement and analysis of the single-ion levels of Co$^{2+}$ in these materials. These are well-modeled by an intermediate spin orbit coupling Hamiltonian. We incorporated a distribution of crystal field parameters to account for local disorder brought about by the split non-magnetic $A$ site. The ground state of Co$^{2+}$ in these environments is always a Kramers doublet with an XY $g$-tensor. The average crystal field parameters produce $g_z = 1.87$ and $g_{xy} = 6.08$, with a variation of 8.4% and 4.1%, respectively. The single-ion ground state doublets are separated from the first excited states by $\sim$300 K. Thus, at low temperatures ($T<\theta_{CW} \sim150$ K), where spin correlations develop, a low energy effective theory built from $S=1/2$ operators can be used to describe these materials. Due to the XY g-tensor and its variation, the effective exchange interactions are likely to be XY-like with bond (exchange) disorder. A basic model of the AFM XY pyrochlore with exchange disorder has been predicted to lead to a long range ordered state. However, [[NaCaCo$_2$F$_7$]{}]{} and [[NaSrCo$_2$F$_7$]{}]{} do not seem to conform to this prediction.
In summary, our results show that these materials can be thought of as “high temperature” versions of the $S_{\text{eff}} = 1/2$ AFM XY pyrochlore, which has been studied in relation to the rare earth oxide series of pyrochlores. Their strong interactions allow the exploration of a wider temperature range of this model. The inherent local disorder arising from the split non-magnetic $A$-site is likely to produce exchange disorder, in part through the local variations of the $g$-tensor that we have estimated here. The role of this exchange disorder in a model appropriate to these materials requires more investigation, as the ground states appear to be contrary to the prediction that *quenched* order by disorder selects long range order. Finally, [[NaCaCo$_2$F$_7$]{}]{} and [[NaSrCo$_2$F$_7$]{}]{} are more frustrated ($f \sim 50$) versions of the AFM XY pyrochlore than the canonical example, Er$_2$Ti$_2$O$_7$ ($f = 20$). This may indicate that they are closer to a phase boundary in the general anisotropic exchange model developed for the pyrochlore lattice; this proximity would make the materials more sensitive to quantum fluctuations, and may even put them in proximity to a quantum disordered phase.
The authors gratefully acknowledges discussions with K. Kimura, the use of software adapted from work by G. Hester and J. O’Brien, and the encouragement of C. L. Broholm. The neutron scattering data were reduced using the Mantid [@arnold2014mantid] and DAVE[@azuah2009dave] software packages. The crystal growth activities at Princeton were supported by the US Department of Energy, office of Basic Energy Sciences, Division of Material Sciences and Engineering under grant DE-FG02-08ER46544. This work utilized neutron scattering facilities supported in part by the National Science Foundation under Agreement No. DMR-1508249. Research using ORNL’s Spallation Neutron Source was sponsored by the Scientific User Facilities Division, Office of Basic Energy Sciences, U.S. Department of Energy.
$Q$-dependence of modes observed with $E_i$ = 700 meV and 250 meV {#sec:Qdep}
=================================================================
The inelastic modes shown in Figure \[fig:CEF\] a)-d) can be identified as being magnetic in origin due to the $|Q|$-dependence of their intensities. The intensity of magnetic features (such as single-ion levels) should decrease with increasing $|Q|$ due to the magnetic form factor [@lovesey1984theory]. Meanwhile the intensity of features arising from phonons (whether single- or multi-phonon processes) should show the opposite trend. For the mode identified near 360 meV in NCCF (using $E_i$ = 700 meV), the $|Q$-dependence of intensity reveals a combination of both magnetic and multi-phonon contributions. This is demonstrated in Figure \[fig:Qdep\] a), which compares intensity near 360 meV to that in energy ranges just above and just below the observed mode. This demonstrates that the peak observed near 360 meV is magnetic in origin. The intensity of the features observed using $E_i$ = 250 meV are also shown for three different $Q$ ranges in panel b), demonstrating their magnetic nature.
[59]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1146/annurev-conmatphys-020911-125058) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [**]{}, ed., edited by (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [“,” ](http://www2.cpfs.mpg.de/~rotter/homepage_mcphase/manual/manual.html) @noop [****, ()]{} @noop [**]{} (, ) @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{}, Vol. (, )
[^1]: The $E \sim 360$ meV mode was measured only in NCCF due to time limitations. Given the overall similarity if the other single-ion levels in the two materials, this mode is assumed to lie in same energy range in NSCF.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Aspartate carbamoyltransferase (ATCase) is a large dodecameric enzyme with six active sites that exhibits allostery: its catalytic rate is modulated by the binding of various substrates at distal points from the active sites. A recently developed method, bond-to-bond propensity analysis, has proven capable of predicting allosteric sites in a wide range of proteins using an energy-weighted atomistic graph obtained from the protein structure and given knowledge only of the location of the active site. Bond-to-bond propensity establishes if energy fluctuations at given bonds have significant effects on any other bond in the protein, by considering their propagation through the protein graph. In this work, we use bond-to-bond propensity analysis to study different aspects of ATCase activity using three different protein structures and sources of fluctuations. First, we predict key residues and bonds involved in the transition between inactive (T) and active (R) states of ATCase by analysing allosteric substrate binding as a source of energy perturbations in the protein graph. Our computational results also indicate that the effect of multiple allosteric binding is non linear: a switching effect is observed after a particular number and arrangement of substrates is bound suggesting a form of long range communication between the distantly arranged allosteric sites. Second, cooperativity is explored by considering a bisubstrate analogue as the source of energy fluctuations at the active site, also leading to the identification of highly significant residues to the T $\leftrightarrow$ R transition that enhance cooperativity across active sites. Finally, the inactive (T) structure is shown to exhibit a strong, non linear communication between the allosteric sites and the interface between catalytic subunits, rather than the active site. Bond-to-bond propensity thus offers an alternative route to explain allosteric and cooperative effects in terms of detailed atomistic changes to individual bonds within the protein, rather than through phenomenological, global thermodynamic arguments.'
author:
- Maxwell Hodges
- Mauricio Barahona
- 'Sophia N. Yaliraki'
bibliography:
- 'ATCase\_paper.bib'
title: 'Allostery and cooperativity in multimeric proteins: bond-to-bond propensities in ATCase'
---
Introduction
============
Much has been written about allostery, the process through which binding of a molecule distal to the active site of a protein causes an attenuation or an enhancement in the catalytic rate of that protein [@Nussinov2016; @Guo2016; @Ribeiro2016]. Yet the physical mechanisims underpinning this effect are still not well understood at the microscopic level, thus limiting the potential for chemical design and intervention. Most of the previous work on allostery has focussed on thermodynamic models linking changes in catalytic rates to modifications in the conformation of the protein. Such an outlook led to the traditional models of allostery: the Monod-Wyman-Changeaux (MWC) model [@Monod1963], whereby binding of allosteric substrates causes a concerted conformational shift of the protein subunits towards the active state, and the Koshland-Nemethy-Filmer (KNF) model [@koshland1966comparison], which proposed that binding of an allosteric substrate to a subunit drives the latter towards the active state and the overall transition to the full active state is sequential. More recently, Hilser and coworkers proposed the ensemble allosteric model (EAM) [@Motlagh2014], which rationalises allosteric outcomes according to the effect of the substrates on the entire conformational ensemble of the protein. Furthermore, there is a growing appreciation of the role of dynamics in allostery [@Guo2016; @Volkman2001], including the role that entropy plays in the modelling of the energy landscape, which has led to the design of protein switches [@Choi2015].
![**ATCase comprises of six catalytic and six regulatory subunits, shown in green and gold respectively, with more than 43000 atoms. PALA (red) is a bisubstrate analogue of the reaction substrates (carbamoyl phosphate and aspartate) and sits in the active site, while ATP and CTP bind to the regulatory subunits and are shown in silver.**[]{data-label="ATCase_cartoon"}](ATCase_cartoon){width="80.00000%"}
Whilst thermodynamic models of allostery provide understanding of the equilibrium effects of substrate binding, they are unable to provide a detailed description of how a signal is transmitted between the allosteric binding site and the active site at the microscopic scale. The so called *structural view* of allostery posits that some form of propagation pathway between the allosteric and active sites must exist as a condition for allostery, though the existence of such a pathway does not imply allosteric behaviour by itself. Tsai and Nussinov [@Tsai2014] argue that both a structural and a thermodynamic component is required for a complete description of allostery. Leitner notes [@Leitner2008] that there are two alternate descriptions of this energy transfer: the traversing of energy from one residue to another along structural pathways (often utilised in discussions of energy dispersion after photoexcitation [@Koyama2006404; @doi:10.1021/jp034558f]) or energy transfer between the normal modes of the protein [@doi:10.1021/jp9813286; @Kholodenko200071]. A significant advance regarding the possibility of structural pathways in allosteric proteins was the work of Lockless and Ranganathan, who used a statistical approach on evolutionary data to demonstrate coupling between residues in the PDK family of proteins situated between a binding and a distal site [@Lockless1999]. However, the study of energy propagation in proteins is far from trivial experimentally, particularly on large proteins. Dyer *et al* [@Li2014] used ultrafast infrared spectroscopy to examine the flow of energy within albumin and found that the flow was ballistic and anisotropic rather than diffusive, supporting the idea that structural pathways exist within proteins that allow for efficient energy transfer between coupled sites. One of the challenges involved in experimental studies of allostery is that structural changes upon ligand binding can be subtle. Falk *et al* [@Falk2016] exploited the extreme sensitivity of NMR chemical shifts to small structural changes to address this problem. By using mutational studies to create singly bound thymidylate synthase dimers, they demonstrated that binding of the first allosteric effector *primes* the enzyme for the binding of the second effector, such that both effectors are required for the allosteric response.
Due to the challenges inherent to the experimental studies of allostery, a wide range of computational methods have been developed to model allosteric behaviour [@Ribeiro2016]. Perhaps the most common methods, particularly for larger proteins, are those that employ elastic network models (ENMs). In general, ENMs model a protein as a system of balls and springs, where usually each ball represents a $C_{\alpha}$ atom, such that the protein is coarse-grained at the residue level. The potential energy function for the springs is assumed to be a quadratic function and the overall dynamics is then described via normal mode analysis (NMA) [@Lopez-Blanco2016]. Improvements to the spring force constants have been found by direct comparison with molecular dynamics (MD) simulations [@Ahmed2010], whilst MD simulations themselves have been shown to provide insight into communication pathways in proteins [@Ghosh2007]. However, MD simulations suffer from high computational cost when applied to fully atomistic descriptions, and, moreover, it is often difficult to understand the coupling of dynamics across the relevant scales. This issue is particularly problematic for the large range transitions that are often involved in allosteric proteins.
Network based approaches have become increasingly common, typically at the residue level of description. Examples include studying changes in residue contacts upon allosteric substrate binding [@Daily2008], identifying residues involved in shortest paths[@DelSol2009] or using MD trajectories to build up so called protein energy networks (PENs) [@Ribeiro2015]. An important insight provided by Ribeiro and Ortiz is that when residue motion correlations are used to create the network, statistical errors render the results less accurate than when interaction energies are used, as a result of the high sensitivity of the signalling pathways to the network topology [@Ribeiro2014]. Recently, strain analysis of residues was applied to crystal structures [@Mitchell2016], which showed that sites of shear strain correspond to binding sites on the protein, thus providing physical insight into how allosteric substrates may transfer energy through a protein structure. Network-theoretic machine learning tools have also been applied to fully atomistic protein graphs [@Delmotte2011; @Amor2014] demonstrating that a wealth of information can be obtained from static structures, avoiding the time consuming calculations often involved in molecular dynamics or Monte Carlo approaches.
Bond-to-bond propensity analysis is a recently developed method [@Amor2016], which has previously been used to predict allosteric sites in a wide set of proteins through knowledge only of the active site of those proteins. This graph-theoretical method was initially introduced to study flow redistribution in electrical networks and can be thought of as a graph-theoretical analogue of a Green’s function in edge space [@Schaub2014]. A number of features of the method stand out. Firstly, it uses a fully atomistic, energy-weighted graph description of the protein [@Delmotte2011; @Amor2014], and hence it does not rely on any coarse-graining techniques to reduce the complexity of the protein structure. Secondly, despite maintaining atomistic detail, the method remains computationally efficient; the calculations are carried out in almost linear time with respect to the number of edges as a result of recent work in algorithmic matrix theory [@Spielman2004; @Kelner]. Furthermore, in contrast to most network approaches, bond-to-bond propensity analysis is focused on the interactions between *edges* in a network (i.e., the *bonds* in a biomolecular system), and can thus reveal the significance of individual bonds in response to energy propagation. It is through bonds that energy fluctuations and flow propagate in a protein and this appears to be the key link between the mathematical basis of the method and the physical processes in the protein, as shown by the robustness of the method in identifying the allosteric sites for a set of 20 benchmark proteins [@Amor2016].
The success of bond-to-bond propensity analysis in predicting allosteric sites motivates our work here, in which we study the ‘reverse’ process; that is, we use bound allosteric substrates as the source of perturbation so as to replicate how fluctuations spreads physically throughout the protein structure. This allows us to identify the residues and bonds that are particularly crucial to energy transport. By comparing the ’reverse’ and ’forward’ processes on both active and inactive states of the protein, we aim to explain how the different energy propagation processes may affect the transition between the two states and the allosteric effect altering the catalytic rate of the protein.
Aspartate carbamoyltransferase (ATCase) is a dodecameric enzyme with six catalytic and six regulatory subunits (Fig. \[ATCase\_cartoon\]). ATCase catalyses the reaction between L-aspartate and carbamoyl phosphate to form N-carbamyl-L-aspartate, the first step in the pyrimidine biosynthesis pathway. ATCase has two main states: an inactive *tense* state (T state) and an active *relaxed* state (R state) [@Lipscomb2012]. The transition between the T and R states produces a major change in shape and symmetry in this protein complex. ATCase has been extensively studied for over 50 years [@Gerhart2014; @Changeux2013] as a paradigmatic example of two important phenomena in protein activity (for an extensive review see Ref. [@Kantrowitz2012]). Firstly, ATCase exhibits cooperativity, whereby binding of the reaction substrates at the active site results in a shift in the equilibrium towards the active R state. Experimentally, this shift can be achieved by addition of the bisubstrate analogue N-(phosphonoacetyl)-L-aspartate (PALA) [@Jin1999]. Secondly, ATC exhibits allostery, with allosteric effectors ATP and CTP as part of a negative feedback mechanism that controls the levels of pyrimidine in the cell. ATP increases the catalytic rate whilst CTP decreases it, and both substrates bind at the same position on the protein, though in slightly different configurations. Whilst both ATP and CTP bind to both forms and cause slight changes in the quaternary structure, the binding of ATP to the inactive T state and CTP to the active R state is not sufficient to cause a population shift to the opposite state [@Kantrowitz2012].
Although ATCase has been extensively studied, the microscopic mechanism that underpins allostery and cooperativity in this large protein complex is still not fully understood. In particular, the role that each of the molecules play in the different subunits in bringing about the allosteric transition remains elusive. Bond-to-bond propensity analsysis is well suited to probing these questions as it is computationally efficient to model the [*whole*]{} complex at the atomistic level , thus allowing us to gauge the individual effect each perturbation has on any other bond in the multimer. ATCase is a large protein, consisting of 43,134 atoms and to the best of the authors’ knowledge, the dodecameric ATCase structure has not previously been studied using a fully atomistic method.
Material and Methods
====================
Structural data
---------------
The three X-ray crystal structures of ATCase used in this work were downloaded from the Protein Data Bank (PDB) [@Berman2000]. We studied two active state structures: 4KGV, the R state bound to ATP (obtained at 1.2Å resolution [@Cockrell2013]); and 1D09, the unligated active state (resolved at 2.1Å [@Jin1999]). We also used one inactive structure: 5AT1, the T state bound to CTP (obtained at 2.6Å resolution [@Stevens1990]).
Construction of the atomistic protein graph
-------------------------------------------
The initial step in the method is the conversion of the 3-dimensional coordinates of the atoms of the protein from the PDB file to a *weighted graph*; that is, a collection of nodes (here representing the atoms) and edges (bonds, interactions) that link them. The *weight* of an edge between two nodes corresponds to the interaction energy of a bond or weak interaction obtained through atomic potentials. The procedure for the atomistic graph construction has been described in detail in Refs. [@Delmotte2011; @Amor2014; @Amor2016] and here we summarise the main features. The crystal structures typically do not contain hydrogen atoms and so the program Reduce (v.3.23) [@Word19991735] is used to add these. Following this, the software FIRST [@PhysRevLett.75.4051; @jacobs2000computer] is used to identify covalent bonds and non-covalent interactions (hydrogen bonds using a threshold of 0.01 $\text{kcal mol}^{-1}$, hydrophobic tethers with a distance cutoff of 8Å and salt bridges). Covalent bonds are weighted using standard bond energies [@HuheeyJamesEandKeiterEllenAandKeiterRichardLandMedhi2006]; hydrogen bonds according to the potential in Ref. [@Dahiyat1997]; and hydrophobic interactions using the potential developed by Lin *et al* [@Lin2007]. Finally, electrostatic interactions of ions and ligands recorded by the LINK entries are accounted for using a standard Coulomb potential with atomic charges for the residues assigned using the OPLS-AA force field [@jorgensen1988opls].
Bond-to-bond propensity
-----------------------
The formulation of *bond-to-bond propensity* was presented in detail in Ref. [@Amor2016] and thus is briefly summarised here. The key matrix that defines bond-to-bond propensities is $M$, the $m \times m$ bond-to-bond transfer matrix, where $m$ is the number of edges or bonds. The element $M_{ji}$ describes how a perturbation at bond $i$ is transmitted to bond $j$ via a propagation that includes the entire graph structure [@Schaub2014]. $M$ is shown to be given by $$M = \frac{1}{2} W B^T L^{\dagger} B
\label{propensity}$$ where $B$ is the $n \times m$ incidence matrix for the graph with $n$ nodes and $m$ edges and $L^{\dagger}$ is the pseudo-inverse of the weighted Laplacian matrix $L$, which governs the diffusion dynamics on the energy-weighted graph [@lambiotte2014random]. The weighted Laplacian is given by: $$L=\begin{cases}
-\omega_{ij}, & \text{$i\neq j$}.\\
\sum_{j} \omega_{ij}, & \text{$i = j$},
\end{cases}$$ where $\omega_{ij}$ corresponds to the interaction energy between atoms $i$ and $j$. More compactly, the Laplacian can be rewritten as $L=B W B^T$ where $W=\text{diag}(\omega_{ij})$ is a $m \times m$ diagonal matrix that contains the energy of interactions of all edges on the diagonal.
To evaluate the effect of perturbations from a group of bonds $b'$ (e.g., belonging to a ligand) on another bond $b$ we select the corresponding columns of the matrix $M$ and compute the sum of the absolute values in the $b^{th}$ row of the selected columns: $$\Pi_b^\text{ raw} = \sum_{b' \in \, \text{ligand}} |M_{bb'}|$$ where $b'$ includes all the weak bonds between the protein and the source (i.e., the ligand).
The *bond propensity* is then defined as: $$\label{bond_score}
\Pi_b = \dfrac{\Pi_b^\text{ raw}}{\sum_{b}\Pi_b^\text{ raw}},$$ which is normalised by the total propensity score of all the bonds in the system.
The results presented in this paper are often in the form of the *residue propensity*, which is calculated by summing over the normalised bond propensities of the bonds belonging to the residue $R$: $$\Pi_R = \sum_{b \in R} \Pi_b.
\label{residue_score}$$
Quantile regression
-------------------
As is physically expected, the propensity of a bond within the protein decays away from the perturbation source. To detect significant effects in the protein structure, we need to compare bond propensities at a similar distance from the source, thus taking into account the expected effect of distance. This is achieved using *conditional quantile regression* (QR) [@Koenker1978], which allows us to identify high propensity bonds at the tail of the highly non-normal distribution [@Amor2016].
The distance of a bond from the perturbation source is taken to be the minimum distance between that bond $b$ and any of the bonds of the chosen source residues: $$d_b = \underset{b' \in \text{source bonds}}{\text{min}} |\mathbf{x}_b - \mathbf{x}_{b'}|,$$ where $\mathbf{x}_b$ holds the cartesian coordinates of the midpoint of bond $b$. Because propensity scores are seen to generally fall away exponentially with distance, the logarithm of the propensity is used to generate the parameters in the QR minimisation problem: $$\hat{\boldsymbol{\beta}}_b^\text{prot}(p) = \underset{(\beta_{b,0}, \ \beta_{b,1})}{\text{argmin}} \sum_{b}^{\text{protein}} \rho_{p} \left( \log(\Pi_b) - (\beta_{b,0} + \beta_{b,1}d_{b}) \right)$$ where $$\rho_{p}(y) = \left \vert y\ (p - \mathbbm{1}(y < 0)) \right \vert$$ is the QR loss function to be minimised for each quantile $p$ and $\mathbbm{1}$ denotes the indicator function. The result of this optimisation is the model $\hat{\boldsymbol{\beta}}^\text{prot} =(\hat{\beta}^\text{prot}_{b,0} (p), \hat{\beta}^\text{prot}_{b,1} (p))$ that describes the quantiles of the propensities for all bonds in the protein.
The *bond quantile score* can then be calculated for each bond in the protein by finding the quantile $\rho_p$ such that: $$p_{b} = \underset{p \in [0,1]}{\text{argmin}} \left \vert \log(\Pi_{b}) - (\hat{\beta}_{b,0}^\text{prot}(p) + \hat{\beta}_{b,1}^\text{prot}(p)d_{b}) \right \vert$$ for bond $b$ with propensity $\Pi_b$ at a distance $d_b$ from the source bonds. The corresponding *residue quantile score* ($p_R$) is similarly defined, instead using residue propensities and the minimum distance between the atoms of each residue and those of the source bonds: $$\hat{\boldsymbol{\beta}}_R^\text{prot}(p) = \underset{(\beta_{R,0}, \ \beta_{R,1})}{\text{argmin}} \sum_{R}^{\text{protein}} \rho_{p} \left( \log(\Pi_R) - (\beta_{R,0} + \beta_{R,1}d_{R}) \right)$$ and $$p_{R} = \underset{p \in [0,1]}{\text{argmin}} \left \vert \log(\Pi_{R}) - (\hat{\beta}_{R,0}^\text{prot}(p) + \hat{\beta}_{R,1}^\text{prot}(p)d_{R}) \right \vert
\label{residue_QR}$$
We can then use this bond quantile score (and its corresponding residue analogue $p_R$) to establish which bonds (and residues) have significantly propensities once the distance effect has been regressed out.
Our quantile regression calculations make use of the R library *quantreg* written by R. Koenker [@koenker2015quantreg].\
The datasets generated during and analysed during the current study are available from the corresponding author on request.
Results
=======
Allostery: Active R State with ATP sources at allosteric sites
--------------------------------------------------------------
### Identification of key residues and bonds under full allosteric occupation
ATP is an allosteric activator of ATCase, able to increase the activity of the enzyme by 180% at a 2mM concentration [@Kantrowitz2012]. ATP does not affect the maximal rate of the enzyme; instead, it induces a shift from the inactive T state to the active R state. From a global thermodynamic perspective, the MWC and EAM models would suggest this shift is caused by a preferential stabilisation of the active R state over the inactive T state, whilst the KNF model would attribute the shift to the binding of ATP to the inactive state driving it towards the active state.
Here, we instead focus on the changes in energy flow within the protein as a result of different substrate binding.
We analyse the full atomistic graph obtained from the crystal structure of ATCase in the active R state (4KGV) with six allosteric binding sites for ATP.
Bond-to-bond propensity analysis is then used to identify significant bonds and residues under allosteric perturbation sources, as well as the effect of changing the number of ATP sources, as a proxy for ATP concentration.
![**Residue ranking of the active R state of ATCase with 6 ATPs as the source by bond-to-bond propensities and conditional quantile regression. All residues are ranked (shown from a red to blue scale) and can be seen either directly on the structure (a) or plotted against distance from the source (b). Here, we further focus on the top 1% as the most significant and plot them on the protein structure (c). Thus (c) displays entirely equivalent results to (a) but the method allows us to highlight those residues that are *particularly* important to energy distribution without making any changes to the underlying data.**[]{data-label="R_state_6_ATP_method"}](R_state_6_ATP_method){width="80.00000%"}
We first consider the propensities with all six ATP allosteric substrates as source residues in order to identify residues and bonds that score highly and could thus be in some way significant to energy transfer in the active R state. Figure \[R\_state\_6\_ATP\_method\] demonstrates the output of the method. The residue propensity for each residue in the protein is computed from their corresponding bond propensities and all residues are then ranked by conditional quantile regression taking account of the distance of the residue from the source sites. Figure \[R\_state\_6\_ATP\_method\]a shows residues coloured according to a scale where red corresponds to those rank highly down to blue if ranked low. Our results show a strong link between the allosteric and active sites: all six instances of PALA, the bisubstrate analogue that sits in the active site, score highly (average $p_R = 0.996$). Indeed, it can be seen starkly from Figure \[R\_state\_6\_ATP\_method\] that the highest scoring residues are concentrated at both the allosteric and active sites.
In order to investigate the effect of energy flow in relation to allostery, we are also interested in the highest scoring residues according to $p_R$, the residue quantile score.
The highest scoring residue is Tyr240, with each of the six residues scoring $p_R = 1$ (Table \[4KGV\_state\_table\]). Tyr240 is known to play an important role in the $T \leftrightarrow R$ transition: each pair of tyrosine residues forms bonds between their phenyl rings in the R state across the gap between the two catalytic trimers (Fig. \[R\_state\_6\_ATP\_method\]c), as opposed to an hydrogen bond to Asp271 in the T state [@Kantrowitz1989; @Burley23]. In fact, Cherfils *et al* [@Kantrowitz1989] used site directed mutagenesis to substitute Tyr240 for phenylalanine, which has the effect of removing the hydroxyl group that forms the hydrogen bond in the T state, and the resulting mutated enzyme shifted strongly towards the R state upon addition of ATP in contrast to the wild-type protein.
![**Bonds ranking in the active R state of ATCase with 6 ATPs as the source by bond-to-bond propensities and conditional quantile regression. Each bond receives a propensity score, which is then ranked by conditional quantile regression, (a) and (b). We can clearly highlight the highest scoring bonds by only selecting those that have scored above the $99^{th}$ percentile and display those bonds that are disproportionately affected by the perturbation at the six allosteric sites (c).**[]{data-label="R_state_6_ATP_bonds"}](R_state_6_ATP_bonds){width="80.00000%"}
Our analysis also provides detailed bond information and we now turn to considering the bond scores directly (Fig. \[R\_state\_6\_ATP\_bonds\]), as key bonds within residues may be missed if other low scoring bonds in the residue ’average out’ the overall residue score. One of the key bonds that emerges as significant is the hydrogen bond between Lys164 and Glu239, a bond that forms in the R state but is not present in the T state [@Kantrowitz2012] (a different Lys164–Glu239 interaction exists in the T state). All six instances of this bond (from each of the six catalytic subunits) score very highly (average score of $p_b = 0.997$, where $p_b$ is the bond quantile score). In fact, it has been shown experimentatally that when either of Lys164 or Glu239 is substituted with glutamine and lysine respectively, the mutant ATCase protein exists in the R state even in the absence of PALA and does not exhibit cooperative or allosteric effects [@Newell1990], highlighting the importance of this interaction. Similarly, Asn111 in the regulatory chain forms a bond with Glu109 in the catalytic chain when in the R state, and again the six instances of this hydrogen bond all score very highly. Interestingly, however, there is a slight asymmetry across the two trimers. In chains C, G and K (See Figure \[ATCase\_cartoon\]), the average bond score is $p_b = 0.997$, whilst it is slightly lower for chains A, E and I on the other catalytic trimer ($p_b = 0.985$). Experimental mutation of Asn111 to alanine also leads to the absence of homotropic and heterotropic effects and a shift to the R state [@Eisenstein1989]. Another interdomain interaction identified as being highly important for stabilisation of the active R state is the Glu50–Arg234 [@Stieglitz2004] interaction, and we find that two different hydrogen bonds score very highly (0.995 for one set of six hydrogen bonds and 0.994 for the other) across all six catalytic chains, suggesting that the link between these two residues is particularly important for energy transfer. This approach highlights the importance of modelling proteins at the bond level, as even course-graining to the residue level may remove crucial information.
------------------------ ---------------- ------------------------ ----------------
Residue Name and Chain Quantile Score Residue Name and Chain Quantile Score
\[0.5ex\] Tyr240 A 1 Arg105 C 0.992
Tyr240 I 1 Arg105 K 0.992
Pala401 C 1 Asp19 B 0.992
Tyr240 E 1 Arg65 I 0.992
Tyr240 C 1 Arg65 E 0.992
Tyr240 K 1 Asp19 J 0.992
Tyr240 G 1 Arg65 A 0.992
Asn84 B 1 Asn84 H 0.992
Asn84 F 1 Arg105 G 0.992
Pala401 K 0.996 Arg56 K 0.988
Pala401 G 0.996 Glu109 C 0.988
Pala401 A 0.996 Asp19 F 0.988
Arg65 C 0.996 Arg56 G 0.988
Arg65 K 0.996 Glu239 I 0.988
Pala401 E 0.996 Arg56 A 0.988
Pala401 I 0.996 Arg56 C 0.988
Asn84 J 0.996 Glu109 G 0.988
Asn84 L 0.996 Glu109 K 0.988
Asn84 D 0.996 Val83 J 0.988
Arg65 G 0.992 Glu50 C 0.988
------------------------ ---------------- ------------------------ ----------------
: **The top 40 residues by *quantile score* (as defined in Eq. ) in the active R state (4KGV) with six ATP sources (Fig. \[R\_state\_6\_ATP\_method\]). All six active site substrate PALA residues and all six Tyr240 residues score above the 99.5% quantile.**[]{data-label="4KGV_state_table"}
### Switching of the allosteric effect is magnified by three ATP sources in cyclic formation
In the previous section, all six ATP molecules were used as the source of the perturbation. The computational efficiency of our method allows us to carry out *in silico* computations with different numbers of allosteric ATP molecules as perturbation sources. We can then model the effect of progressively adding more ATP molecules to ATCase to investigate how energy flow is modified when different numbers of ligand sites are occupied. Fig.\[R\_state\_ATP\_switch\] shows the results of this analysis, starting with a single ATP source on chain B and adding further ligands on chain F followed by chain J.
![**a) Binding of the first two ATP molecules does not appear to show communication between the allosteric source sites and the active site (identified by the gold PALA residues) in the active R state by bond-to-bond propensities. However, binding of a third ATP ligand leads to a switching effect, at which point all six active site PALA residues score within the top 20 residues out of 2790. b) Scatter plot showing the average rank of the two highest scoring residues (out of 2790) from the 6 ATP case.**[]{data-label="R_state_ATP_switch"}](R_state_ATP_switch){width="80.00000%"}
When a single ATP source in regulatory chain B is used, the ranking of the residues in the protein is starkly different to the result when all six ATP molecules are used as the source (compare Figs. \[R\_state\_6\_ATP\_method\] and \[R\_state\_ATP\_switch\]). For example, the active site residue PALA decreases its average quantile score from 0.996 (with 6 ATP molecules) to 0.941 (with 1 ATP molecule). Instead, most of the highest scoring residues when 1 ATP is present are located near the allosteric site on chain L, which is situated across the chain B source in the multimer (Fig \[ATCase\_cartoon\]). For instance, Asp19 (which binds to ATP) on chain L scores $p_R = 1$ and Lys56 scores $p_R = 0.996$. Indeed, experimental mutation of Lys56 to alanine led to the disappearance of homotropic cooperativity in the presence of ATP, but not CTP [@Corder1989], suggesting it is involved in the communication pathway between ATP and the active site. It is also interesting to consider Tyr240, the other highly significant residue in the case of full allosteric occupation with six ATP sources. Under single ATP occupation, the pair of Tyr240 residues in chains E and K still score highly ($p_R = 0.993$ and 1 respectively) whilst the Tyr240 residues in the other four catalytic chains score lower (average of 0.932 across the four catalytic chains). As Fig \[ATCase\_cartoon\] shows, catalytic chains E and K are in fact situated on the other side of the protein to the chain B source, again suggesting that communication within ATCase is long range. When a second ATP molecule (on chain F) is included in the perturbation source, the results are similar with significant residues appearing again in the region of the allosteric sites on chains J and L distal to both source sites on chains B and F. The PALA score on the active site is again similar to the single ATP source case ($p_R = 0.946$) showing little change in the communication to the active site upon ‘binding’ of a second ligand. Tyr240 scores slightly higher on average here ($p_R = 0.955$), though no single Tyr240 residues scores as high as in the six ATP case. Overall, there is no significant change in the propensities in the active and allosteric sites to source perturbations when comparing between one or two ATP molecules binding.
In contrast, a significant change occurs upon addition of a third ATP molecule in chain J to the perturbation source, as seen in Figure \[R\_state\_ATP\_switch\]. The average score for PALA now jumps to $p_R = 0.996$, the same as in the six ATP case, whilst Tyr240 achieves a score of $p_R = 0.998$. Importantly, if the third ATP source is added instead to chain D (see Fig. \[ATCase\_cartoon\]), the increases are not as pronounced (PALA = 0.948 and Tyr240 = 0.962), suggesting that the symmetric distribution of the ATP sources introducing cycles in the protein may be important for facilitating communication with the active site.
Cooperativity: Active unligated R state with PALA sources at the active sites
-----------------------------------------------------------------------------
![**To investigate homotropic cooperativity, the six PALA substrates were selected as the source on the active unligated R state. The structure on the right shows just one half on the protein for clarity and here it is clear that the highest scoring regions (in red) of ATCase are located around the active and allosteric sites.**[]{data-label="R_state_No_ligand_1d09"}](R_state_No_ligand_1d09){width="80.00000%"}
To investigate energy flow in relation to homotropic cooperativity we analyse the full atomistic graph obtained from the crystal structure of ATCase in the active R state (1D09) with PALA molecules bound to the active sites. PALA acts as a bisubstrate analogue and, as previously, all six PALA residues are included as source residues in order to identify residues that are significant with respect to energy distribution and thus may be implicated in the cooperative mechanism.
Figure \[R\_state\_No\_ligand\_1d09\] shows the overall effect of a perturbation applied at the six active sites. Similarly to the ‘reverse’ process when ATP is used as the source (compare to Fig. \[R\_state\_6\_ATP\_method\]), the highest scoring regions are clustered around both the allosteric and active sites. The result reinforces the idea that there is a form of communication between these distal sites and there do not appear to be obvious, individual pathways between the two types of site.
As seen in Table \[1D09\_state\_table\], the highest scoring residue is Glu50, with all six instances scoring the maximum of $p_R = 1$. As mentioned earlier, Glu50 is a crucial residue for the stability of the R state: substitution of glutamic acid for alanine leads to dramatic changes in the activity of the enzyme, which is reduced 15-fold, whereas cooperativity is completely lost [@Newton1990]. Significant communication from the active sites to the allosteric sites is also seen, with Asp19 (one of the residues that interacts with ATP and CTP) scoring $p_R = 0.992$ over the six sites, whilst Lys60, another allosteric residue, scores highly ($p_R = 0.989$) over the regulatory chains on one side of the protein (chains D, H and L in Fig. \[ATCase\_cartoon\]), again demonstrating asymmetry over the structure. It is known experimentally that Glu233 forms a salt link with Arg229 only in the R state, which orients Arg229 into the active site [@Ke1988]. The removal of the salt link via mutation of glutamic acid to serine leads a significant decrease in both catalytic activity and cooperativity [@Biol1989]. We find that Glu233 scores highly overall ($p_R = 0.985$), though once again a difference is seen between the two catalytic trimers (trimer AEI scores $p_R = 0.993$ vs $p_b = 0.977$ for the opposite trimer CGK).
Analysis of the *bond level* data reveals further information. As expected, the previously mentioned Glu233–Arg229 salt bridge ranks very highly ($p_b = 0.996$), whilst the Glu50 interaction with Arg167 (which itself interacts with PALA in the active site, being positioned correctly by its association with Glu50) involving two types of bonds scores above $p_b = 0.995$ across all units. The Asp19–Lys56 link scores an average of $p_b = 0.999$ over its six instances and it was found that substitution of lysine by alanine affected not only cooperativity but also removed the ability of ATP to activate the enzyme [@Corder1989]. Therefore, as Asp19 is one of the allosteric residues, it appears that this bond to Lys56 may be crucial in communicating with the active site.
------------------------ ---------------- ------------------------ ----------------
Residue Name and Chain Quantile Score Residue Name and Chain Quantile Score
\[0.5ex\] Asp19 D 1 Lys60 H 0.993
Asp19 H 1 Leu58 F 0.993
Asp19 L 1 Ile18 D 0.993
Glu50 E 1 Ile18 H 0.993
Glu50 A 1 Asp90 K 0.993
Glu50 K 1 Asp90 I 0.993
Glu50 G 1 Glu233 I 0.993
Glu50 I 1 Glu233 E 0.993
Glu50 C 1 Glu233 A 0.993
Met1 H 0.996 Arg167 E 0.993
Ile44 L 0.996 Asp19 F 0.989
Ile44 H 0.996 Lys60 D 0.989
Ile44 D 0.996 Leu58 B 0.989
Asp90 E 0.996 Leu58 J 0.989
Asp90 A 0.996 Asp90 G 0.989
Asp90 C 0.996 Arg269 A 0.989
Arg105 G 0.996 Glu86 E 0.989
Arg105 C 0.996 Arg167 I 0.989
Arg105 K 0.996 Arg167 G 0.989
Met1 D 0.993 Arg167 K 0.989
------------------------ ---------------- ------------------------ ----------------
: **The top 40 residues by *quantile score* in the active R state (1D09) with six PALA sources (see Fig. \[R\_state\_No\_ligand\_1d09\]). Each of the six Glu50 residues score the maximum value of 1.**[]{data-label="1D09_state_table"}
Stabilisation of the catalytic trimer: inactive T state with allosteric CTP
---------------------------------------------------------------------------
Finally, we study the inactive state through the analysis of the graph obtained from the crystal structure of ATCase in the T state (5AT1) with CTP molecules bound to the allosteric sites. CTP acts as an inhibitor for ATCase reducing its catalytic rate by causing a shift towards the inactive T state. When using the full allosteric occupation scenario (6 CTP molecules as perturbation sources) in the T state, the highest scoring regions of the protein appear most strongly located at the C1–C2 interface (Fig. \[T\_state\_CTP\_6\_source\]) instead of the active site. This is in stark contrast with the results for the R state, both under ATP (allosteric site) and PALA (active site) perturbation sources, as shown in Figs. \[R\_state\_6\_ATP\_method\] and \[R\_state\_No\_ligand\_1d09\]. The two catalytic trimers each move as essentially rigid units during the T $\leftrightarrow$ R transition so there is little structural change within the trimer. There is thus only minor differences between the inactive T state and the active R state in this region [@Krause1987].
Two residues in particular stand out, as seen in Table \[T\_state\_table\]: Arg65 (average $p_R = 0.999 $) and Arg56 (average $p_R = 0.998 $). It can be seen from Figure \[T\_state\_CTP\_6\_source\] that both these residues bridge the C1–C2 interface, though they do not form links to each other. Looking in more detail at the bond level data, one of the key interactions made by Arg65 is with Asp100 (average $p_b = 0.999$). This specific interaction was identified experimentally as being important for the stability of the catalytic trimer [@Baker1993] and replacement of Asp for either Asn or Ala reduces the half life of inactivation of the catalytic subunit. Arg65 additionally forms a hydrogen bond to His41, another residue implicated in catalytic subunit stability and this interaction also scores highly ($p_b = 0.983$), though once again there is a significant difference between the two catalytic subunits, with the interactions in the AEI trimer scoring $p_b = 0.999$, compared to $p_b = 0.968$ in trimer CGK (Fig. \[ATCase\_cartoon\]). There is possibly a link here with experimental data showing that in the R state, only half (i.e. three) of the His41–Glu37 interactions are broken [@Gouaux1990; @Stevens1991a] during the transition from the T state, demonstrating an intriguing asymmetry that appears to be captured by our computed *bond-to-bond propensities*. In fact, the residue results for Glu37 are even starker, with the average quantile score across chains A, E and I 0.990 versus 0.262 for chains C, G and K, a remarkable difference between essentially symmetrically equivalents sets of residues. Glu37 itself has been associated with stabilising the catalytic trimer [@Baker1993].
Conversely, there appears to be little experimental data on Arg56, nor on the two highest scoring links it makes: to Gly72 ($p_b = 0.999$) and Gln60 ($p_b = 0.986$), though the Gly72 interaction occurs across the C1–C2 interface [@Jin1999; @Stevens1991a] so it would seem possible that this interaction is also involved in stability of the trimer. Perhaps less surprisingly, a number of residues located close to the CTP site also rank highly: Ile86, which forms a non-polar interaction with the nucleotide [@Kantrowitz2012]) and Asn84, which interacts with the phosphate part of CTP [@Stevens1991a] score $p_R = 0.993$ and $0.985$ respectively. Val17 also forms a non-polar interaction with CTP, though scores slightly lower with an average bond quantile score of $0.978$.
![**When six CTP molecules are used as the perturbation source on the inactive T state, the highest scoring residues appear at the C1-C2 interface, which is the boundary between catalytic subunits within the catalytic trimer. Arg56 and Arg65 are two of the highest scoring residues, shown on the right, forming a link across the C1-C2 interface.**[]{data-label="T_state_CTP_6_source"}](T_state_CTP_6_source){width="80.00000%"}
------------------------ ---------------- ------------------------ ----------------
Residue Name and Chain Quantile Score Residue Name and Chain Quantile Score
\[0.5ex\] Arg65 G 1 Val71 G 0.993
Arg56 G 1 Val71 A 0.993
Arg65 C 1 Val71 K 0.993
Arg65 A 1 Glu37 A 0.993
Arg56 A 1 Val71 C 0.993
Arg65 K 1 Gln60 E 0.993
Asn84 H 1 Gln60 A 0.993
Asn84 D 1 Val83 H 0.993
Ile86 J 1 Val17 H 0.993
Arg56 K 0.996 Val17 D 0.993
Arg65 I 0.996 Ile86 D 0.993
Arg65 E 0.996 Arg105 A 0.989
Arg56 C 0.996 Gln60 I 0.989
Arg56 I 0.996 Phe73 K 0.989
Arg56 E 0.996 Glu37 E 0.989
Arg85 H 0.996 Arg105 E 0.989
Ile86 F 0.996 Glu37 I 0.989
Ile86 B 0.996 Arg85 D 0.989
Asn84 L 0.996 Val83 D 0.989
Ile86 H 0.996 Ile12 J 0.989
------------------------ ---------------- ------------------------ ----------------
: **T state with CTP (5AT1). The top 40 residues by quantile score are listed and both Arg56 and Arg65 appear six times each. These residues sit at the C1-C2 interface within the catalytic subunits.**[]{data-label="T_state_table"}
### Sequential binding of CTP and ATP show similar switching patterns but on different regions of the protein
Whilst the identity of the highest scoring residues when CTP is used as the perturbation source is different to the ones that appeared when ATP is used, there is a similar ’switching effect’ observed when a third CTP molecule is included as a source in a cyclic arrangement around the ATCase structure. As seen in Figure \[T\_state\_CTP\_switch\], inclusion of a third ligand leads to the clustering of high scoring residues in the region of C1–C2 interface between the catalytic subunits within a trimer, leading to a similar pattern observed under the full occupation six CTP source case (Fig. \[T\_state\_CTP\_6\_source\]).
Once again, it appears to be the interaction between the CTP ligands located in such a symmetric, cyclic arrangement around the ATCase protein that leads to energy flow amplifying the effect on residues identified as particularly significant. This non linear effect is similar to the one observed for ATP allosteric occupation in Figure \[R\_state\_ATP\_switch\], albeit in a different location in the protein. This effect is illustrated numerically by focusing on two of the highest scoring residues: Arg56 and Arg65. Starting from a single CTP source, the scores for Arg65 progress from $p_R = 0.904$ to 0.961 and then to 0.989 when a third ligand is included; equivalently, for Arg56, the scores are 0.779, 0.916 and 0.982 as each of the CTP ligands is added. The increases in scores of these two highest scoring residues in this case are actually more linear than in the case of ATP but it is still only when a third CTP ligand is included cyclically that the results from the six CTP case are replicated. When the third ligand is instead added to chain D, such that the three CTP ligands are now bound to chains B, D and F (Fig. \[ATCase\_cartoon\]), the increase in score upon addition of the third ligand is smaller for both Arg65 ($p_R = 0.973$) and Arg56 ($p_R = 0.922$) which again suggests that it is a particular feature of the symmetric arrangement of the allosteric ligands that facilitates communication to the key residues within the protein by creating cyclic reinforcement of energy flows in the protein.
![**a) The top 2% of residues displayed when varying numbers of CTP molecules are included as source residues. In contrast to the ATP case, there does not appear to be as much communication with the distal allosteric sites for one or two source ligands but again inclusion of a third ligand on chain J leads to the results resembling the six CTP case described previously. b) Scatter plot showing the average rank of the two highest scoring residues (out of 2790) from the 6 CTP case.**[]{data-label="T_state_CTP_switch"}](T_state_CTP_switch){width="80.00000%"}
Discussion
==========
In this work, we have demonstrated how *bond-to-bond propensities* can be used to investigate the energy flow process of a heterotropic ligand binding to an allosteric site, as well as the homotropic case of substrates binding to the active site. We have focused on a large (43,000 atom) well studied multimeric enzyme, ATCase which shows both allostery and cooperativity. In the active R state, using ATP allosteric binding sites as perturbation sources, reveals a number of residues and specific bonds as being particularly significant, including Tyr240, which links the two sides of the ATCase protein, and PALA, which sits in the active site. There is thus a clear communication pathway between the allosteric and active sites in ATCase but in accordance with other computational studies of ATCase [@Mitchell2016], this communication does not appear to occur through individual, discrete pathways of residues but instead via a collective of lower scoring residues.
Furthermore, we find that the geometrical distribution of the ligands bound is important leading to a switching of the allosteric effect in our computations. Only when three ATP residues arranged cyclically around the ATCase structure are used as the perturbation source do we recover the results of the case with full allosteric site occupation. In contrast, when a single or two ATP molecules are used as sources, no strong link between the allosteric and active sites is observed, though there does seem to be communication between distal allosteric sites. A feature of our results is a consistent asymmetry in the scores between the two sets of catalytic trimers, despite the symmetrical structure of ATCase, demonstrating the ability of *bond-to-bond propensities* to capture subtle structural features at the atomistic level.
Homotropic cooperativity was investigated by using the six PALA substrates bound at the active sites as the perturbation source. The regions that scored most highly in this case were the allosteric sites and around the bound active site, reinforcing the idea that the two types of sites are highly coupled in the active state and also hinting that homotropic and heterotropic cooperative are not orthogonal phenomena and are, instead, closely intertwined.
Finally, allosteric inhibition of ATCase by CTP was studied by using the CTP molecules as the perturbation source. Interestingly, rather than the active site region being identified as significant, the C1–C2 interface of the catalytic trimers was instead found to be signficantly coupled to the allosteric sites. Interestingly, the boundary between the catalytic subunits has been found to be important for stability of the enzyme but not particularly vital for catalytic activity; hence it is possible that different allosteric ligands may play subtly different roles when binding to the active and inactive states of the enzyme.
Our results highlight that both the atomistic nature of the methodology and the long-range effects made possible by the global properties of the graph-theoretical approach are essential to understanding the effects of allostery and cooperativity in multimeric protein. Given its computational efficiency and generality we hope that this method can be useful to the study of other such protein systems of broad relevance.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Ben Amor for useful discussions. This work was funded by an EPSRC Centre for Doctoral Training Studentship from the Institute of Chemical Biology (Imperial College London) awarded to MH. MB acknowledges funding from the EPSRC project EP/N014529/1 supporting the EPSRC Centre for Mathematics of Precision Healthcare.
Contributions {#contributions .unnumbered}
=============
M.H., S.N.Y. and M.B. designed the experiments, M.H. carried out the experiments and analyzed data, M.H., S.N.Y. and M.B. wrote and approved the manuscript.
Competing interests {#competing-interests .unnumbered}
===================
The authors declare no competing interests.
Corresponding author {#corresponding-author .unnumbered}
====================
Correspondence to Sophia N. Yaliraki ([email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The $k$-Hessian operator $\sigma_k$ is the $k$-th elementary symmetric function of the eigenvalues of the Hessian. It is known that the $k$-Hessian equation $\sigma_k(D^2u)=f$ with Dirichlet boundary condition $u=0$ is variational; indeed, this problem can be studied by means of the $k$-Hessian energy $-\int u\sigma_k(D^2u)$. We construct a natural boundary functional which, when added to the $k$-Hessian energy, yields as its critical points solutions of $k$-Hessian equations with general non-vanishing boundary data. As a consequence, we prove a sharp Sobolev trace inequality for $k$-admissible functions $u$ which estimates the $k$-Hessian energy in terms of the boundary values of $u$.'
address:
- |
Jeffrey S. Case\
109 McAllister Building\
Penn State University\
University Park, PA 16802
- 'Yi Wang, Department of Mathematics, Johns Hopkins University, 404 Krieger Hall, 3400 N. Charles Street, Baltimore, MD 21218,'
author:
- 'Jeffrey S. Case'
- Yi Wang
title: A fully nonlinear Sobolev trace inequality
---
Introduction {#sec:intro}
============
Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with boundary $M=\partial X$. The usual sharp Sobolev trace inequality states that $$\label{eqn:model_trace}
-\int_X u\Delta u\, dx + \oint_M fu_n d\mu\geq \oint_M f(u_f)_n d\mu$$ for all $f\in C^\infty(M)$ and all $u\in C^\infty(\overline{X})$ such that $u{\rvert}_M=f$, where $u_n$ denotes the derivative of $u$ with respect to the outward-pointing normal along $M$, $u_f$ is the harmonic function in $X$ such that $u_f{\rvert}_M=f$, and $dx$, $d\mu$ are the volume forms on $X$ and $M$, respectively. A standard density argument implies that the trace $u\mapsto u{\rvert}_M=:\operatorname{tr}u$ extends to a bounded linear operator $\operatorname{tr}\colon W^{1,2}(\overline{X})\to W^{1/2,2}(M)$, while the extension $f\mapsto u_f=:E(f)$ extends to a bounded linear operator $E\colon W^{1/2,2}(M)\to W^{1,2}(\overline{X})$ such that $\operatorname{tr}\circ E$ is the identity.
The sharp Sobolev trace inequality is a useful tool in many analytic and geometric problems. For example, the Dirichlet-to-Neumann map $f\mapsto(u_f)_n$ is a pseudodifferential operator with principle symbol $(-\Delta)^{1/2}$; indeed, it is the operator $(-\Delta)^{1/2}$ when $\Omega={\mathbb{R}}_+^n$ is the upper half-plane. Thus relates the energy of the local operator $-\Delta$ to the energy of the nonlocal Dirichlet-to-Neumann operator, providing a useful tool for establishing estimates for PDEs stated in terms of the latter operator. This strategy provides a key motivation for the approach of Caffarelli and Silvestre [@CaffarelliSilvestre2007] for studying fractional powers of the Laplacian. As another example, Escobar [@Escobar1988; @Escobar1990] proved an analogue of on compact manifolds with boundary for which both sides of the inequality are conformally invariant. In particular, this recovers when $X={\mathbb{R}}_+^n$. Using conformal invariance, he also proved a sharp Sobolev trace inequality which yields the continuous embedding $W^{1,2}(\overline{{\mathbb{R}}_+^n})\subset L^{\frac{2(n-1)}{n-2}}({\mathbb{R}}^{n-1})$ when $n\geq3$. This work has important implications for the Yamabe Problem on manifolds with boundary [@Escobar1992]. By considering weights or higher-order operators, analogues of have been established with implications for the energies of fractional powers of the Laplacian of all non-integral orders [@CaffarelliSilvestre2007; @Yang2013] as well as for the energies of conformally covariant fractional powers of the Laplacian [@Case2015fi; @CaseChang2013; @ChangGonzalez2011; @ChangYang2015] and the fractional Yamabe problem [@GonzalezQing2013].
The purpose of this article is to establish an analogue of in terms of the $k$-Hessian energy $\sigma_k(D^2u)$. Here $D^2u$ denotes the Hessian of $u$ and the $k$-th elementary symmetric function $\sigma_k(A)$ of a symmetric matrix $A$ is defined by $$\sigma_k(A) := \sum_{i_1<\dotsb<i_k} \lambda_{i_1}\dotsm\lambda_{i_k}$$ for $\lambda_1,\dotsc,\lambda_n$ the eigenvalues of $A$. The Dirichlet problem $$\label{eqn:dirichlet_problem}
\begin{cases}
\sigma_k(D^2u) = F(x,u), & \text{in $X$}, \\
u = f(x), & \text{on $M$}
\end{cases}$$ has been well-studied for functions $u$ in the elliptic $k$-cone $$\label{eqn:kcone}
\Gamma_k^+ := \left\{ u\in C^\infty({\overline{X}}) {\mathrel{}\middle|\mathrel{}}\sigma_j(D^2u)>0, 1\leq j\leq k \right\} ;$$ e.g. [@Caff4; @IvochkinaTrudingerWang2004; @Urbas1990; @Wang2; @Wang]. Note that the existence of a solution to requires that $M$ be $(k-1)$-convex [@Caff4]; i.e. the second fundamental form $L$ of $M$ must satisfy $\sigma_j(L)>0$ for $1\leq j\leq k-1$. Indeed, provided $M$ is $(k-1)$-convex, X.-J. Wang proved [@Wang2] the fully nonlinear Sobolev inequality $$\label{eqn:wang_sobolev}
\int_X -u\sigma_k(D^2u) dx\geq C(X)\left(\int_X {\lvert}u{\rvert}^{\frac{n(k+1)}{n-2k}}dx\right)^{\frac{n-2k}{n}}$$ for all $u\in\Gamma_k^+$ such that $u{\rvert}_M=0$. In a sense, the Sobolev inequality is dual to the desired fully nonlinear analogue of : in the extremal functions are “flat” on the boundary, in the sense $u{\rvert}_M=0$, while in the extremal functions are “flat” in the interior, in that $\Delta u=0$.
To establish a fully nonlinear analogue of requires us to both know that the purported minimizers of the inequality exist and to identify what boundary terms to add to the interior term $-\int u\sigma_k(D^2u) dx$. The first problem is settled: existence and uniqueness of a solution $u\in\overline{\Gamma_k^+}$ of the degenerate Dirichlet problem with $F=0$ is known [@IvochkinaTrudingerWang2004; @WangXu]; here $\overline{\Gamma_k^+}$ is the closure of the elliptic $k$-cone with respect to the $C^{1,1}$-norm in ${\overline{X}}$. The second problem is addressed in this article. This is accomplished via the following proposition.
\[prop:main\_prop\] Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with boundary $M=\partial X$ and let $k\in{\mathbb{N}}$. Then there is a multilinear differential operator $$\label{eqn:boundary_requirement}
B_k\colon\left(C^1(\overline{X}) \cap C^2(M)\right)^k \to C^0(M)$$ such that the multilinear form ${\mathcal{Q}}_k\colon\left(C^2(X)\cap C^2(M)\cap C^1(\overline{X})\right)^{k+1}\to{\mathbb{R}}$ defined by $$\label{eqn:k-energy}
{\mathcal{Q}}_k(u,w^1,\dotsc,w^k) := -\int_X u\,\sigma_k(D^2w^1,\dotsc,D^2w^k)dx + \oint_M u\,B_k(w^1,\dotsc,w^k)d\mu$$ is symmetric, where $\sigma_k(D^2w^1,\dotsc,D^2w^k)$ is the polarization of the $k$-linear map $w\mapsto\sigma_k(D^2w)$.
The notation specifies that the operators $B_k$ depend on at most second-order tangential derivatives and at most first-order transverse derivatives of their inputs along the boundary $M$.
An explicit formula for such operators $B_k$ can be deduced from Section \[sec:polarized\] and Section \[sec:adjusted\]. From we see that $B_1(u)=u_n$ satisfies the conclusions of Proposition \[prop:main\_prop\]. The following result gives a boundary operator which satisfies the conclusions of Proposition \[prop:main\_prop\] when $k=1$.
\[prop:main\_prop2\] Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with boundary $M=\partial X$. Define $B_2\colon\left(C^1({\overline{X}})\cap C^2(M)\right)^2\to C^0(M)$ by $$\label{eqn:k2_boundary_operator}
B_2(v,w) = \frac{1}{2}\left(v_n{\overline{\Delta}}w + w_n{\overline{\Delta}}v + L({\overline{\nabla}}v,{\overline{\nabla}}w) + Hv_nw_n \right) .$$ Then the multilinear form ${\mathcal{Q}}_2\colon\left(C^2({\overline{X}})\right)^3\to{\mathbb{R}}$ given by $${\mathcal{Q}}_2(u,v,w) = -\int_X u\sigma_2(D^2v,D^2w)dx + \oint_M uB_2(v,w)d\mu$$ is symmetric.
Here ${\overline{\Delta}}$ and ${\overline{\nabla}}$ denote the tangential Laplacian and tangential gradient, respectively; i.e. the Laplacian and the gradient defined with respect to the induced metric on the boundary $M$.
Denote by ${\mathcal{E}}_k(u):={\mathcal{Q}}_k(u,\dotsc,u)$ the energy associated to ${\mathcal{Q}}_k$ as in Proposition \[prop:main\_prop\]. The fact that defines a symmetric $(k+1)$-linear form implies that if $v\in C^\infty(\overline{X})$ is such that $v{\rvert}_M=0$, then $$\left.\frac{d^j}{dt^j}\right|_{t=0}{\mathcal{E}}_k(u+tv) = -\frac{(k+1)!}{(k+1-j)!}\int_X v\,\sigma_k\Bigl(\overbrace{D^2v,\dotsc,D^2v}^{j-1},\overbrace{D^2u,\dotsc,D^2u}^{k+1-j}\Bigr) dx$$ for all $1\leq j\leq k+1$. That is, within a class $${\mathcal{C}}_f := \left\{ u\in C^\infty(\overline{X}) {\mathrel{}\middle|\mathrel{}}u{\rvert}_M = f \right\}$$ of functions with fixed trace $f\in C^\infty(M)$, the derivatives of the energies ${\mathcal{E}}_k$ depend only on the interior integrals. In particular, it is straightforward to identify the critical points of ${\mathcal{E}}_k$ and deduce the convexity of ${\mathcal{E}}_k$ within the positive cone $\Gamma_k^+$. This leads to the following family of fully nonlinear Sobolev trace inequalities.
\[thm:main\_thm\] Fix $k\in{\mathbb{N}}$ and let $X\subset{\mathbb{R}}^n$ be a bounded $(k-1)$-convex domain with boundary $M=\partial X$. Let $B_k$ be as in Proposition \[prop:main\_prop\]. Given $f\in C^\infty(M)$, let $${\mathcal{C}}_{f,k} := \left\{ u\in {\mathcal{C}}_f {\mathrel{}\middle|\mathrel{}}D^2u\in\Gamma_k^+ \right\} .$$ Then it holds that $$\label{eqn:k-sobolev-trace}
{\mathcal{E}}_k(u) \geq {\mathcal{E}}_k(u_f)$$ for all $u\in\overline{{\mathcal{C}}_{f,k}}$, where $u_f$ is the unique solution to the Dirichlet problem $$\label{eqn:degenerate_dirichlet_problem}
\begin{cases}
\sigma_k(D^2u) = 0, & \text{in $X$}, \\
u = f, & \text{on $M$} ,
\end{cases}$$ and $\overline{{\mathcal{C}}_{f,k}}$ is the closure of ${\mathcal{C}}_{f,k}$ with respect to the $C^{1,1}$-norm in ${\overline{X}}$.
Note that ${\mathcal{E}}_k(u_f)=\oint f\,B_k(u_f,\dotsc,u_f)d\mu$, so that Proposition \[prop:main\_prop\] implies that the right-hand side of depends only on $f$, the tangential gradient ${\overline{\nabla}}f$, the tangential Hessian $\bar D^2 f$, and the normal derivative $(u_f)_n$ of the extension $u_f$. This is consistent with the expected regularity $u_f\in C^{1,1}(\overline{X})$. One may regard as a norm inequality for part of the trace embedding $W^{\frac{2k}{k+1},k+1}(X)\subset W^{\frac{2k-1}{k+1},k+1}(M)$.
We conclude this introduction with a few additional comments on the boundary operators $B_k$ of Proposition \[prop:main\_prop\]. Given $f\in C^\infty(M)$ and $k\in{\mathbb{N}}$, define $${\mathcal{B}}_k(f) := B_k(u_f,\dotsc,u_f)$$ for $u_f$ the solution to . The specification of the domain of the boundary operators $B_k$ implies that ${\mathcal{B}}_k$ is a well-defined function; it should be regarded as a fully nonlinear analogue of the Dirichlet-to-Neumann map. Theorem \[thm:main\_thm\] yields a relationship between the energy of ${\mathcal{B}}_k$ and the energy associated to the $\sigma_k$-curvature. Motivated by the similar relationship between the energies associated to fractional order operators and the Laplacian induced by , we propose the study of the operators ${\mathcal{B}}_k$ as an interesting family of fully nonlinear pseudodifferential operators. In particular, it seems interesting to ask if there exists a constant $C(M)>0$ such that $$A(M)\oint_M f\,{\mathcal{B}}_k(f)d\mu + B(M)\oint_M {\lvert}f{\rvert}^{k+1} d\mu\geq \left( \oint_M {\lvert}f{\rvert}^{\frac{(k+1)(n-1)}{n-2k}}d\mu\right)^{\frac{n-2k}{n-1}} .$$ If true, this would provide a fully nonlinear analogue of the sharp Sobolev inequality of X.-J. Wang [@Wang]. Note that this is already known in the case $k=1$; cf. [@LiZhu1997].
The conditions of Proposition \[prop:main\_prop\] do not uniquely determine the boundary operators $B_k$ of Proposition \[prop:main\_prop\]; indeed, the operators are not unique even if we require additionally that the operators $B_k$ commute with diffeomorphisms, as do the operators constructed in the proof of Proposition \[prop:main\_prop\]. A trivial source of nonuniqueness comes from the freedom to add symmetric zeroth-order terms to $B_k$. For example, if $B_k$ satisfies the conclusions of Proposition \[prop:main\_prop\], so too does the operator $$(w^1,\dotsc,w^k) \mapsto B_k(w^1,\dotsc,w^k) + cHw^1\dotsm w^k$$ for any $c\in{\mathbb{R}}$, where $H$ is the mean curvature of the boundary $M$. More generally, one may add to the boundary operators $B_k$ any symmetric multilinear operator which is also symmetric upon pairing with integration. For example, consider the operator $D\colon\left(C^1(\overline{X})\right)^2\to C^\infty(M)$ defined by $$D(v,w) = {\overline{\delta}}\left( L({\overline{\nabla}}(vw))\right) - L({\overline{\nabla}}v,{\overline{\nabla}}w) .$$ It is readily verified that $(u,v,w)\mapsto\oint u\,D(v,w)d\mu$ is a symmetric trilinear form, and thus $D$ can be added to the operator to yield another operator $\tilde B_2$ which satisfies the conclusions of Proposition \[prop:main\_prop\].
This article is organized as follows. In Section \[sec:bg\] we collect some useful facts involving the $k$-Hessian and the elliptic cones. In Section \[sec:polarized\] and Section \[sec:adjusted\] we prove Proposition \[prop:main\_prop\] by explicitly constructing a suitable boundary operator. In Section \[sec:variation\] we prove Theorem \[thm:main\_thm\]. In Section \[sec:example\] we discuss in more detail the case $k=2$.
Preliminaries {#sec:bg}
=============
The $\Gamma_k^+$-cone
---------------------
In this subsection, we describe some properties of the elementary symmetric functions and their associated convex cones.
The $k$-th elementary symmetric function for $\lambda=(\lambda_1,\dotsc,\lambda_n)\in \mathbb{R}^n$ is $$\sigma_{k}(\lambda):=\sum_{i_1<\dotsb<i_k}\lambda_{i_1}\dotsm \lambda_{i_k}.$$
The elementary symmetric functions are special cases of hyperbolic polynomials [@Garding]. As such, they enjoy many nice properties in their associated positive cones.
\[cone\] The positive $k$-cone is the connected component of $\left\{\lambda {\mathrel{}\middle|\mathrel{}}\sigma_k(\lambda)>0\right\}$ which contains $(1,\dotsc,1)$. Equivalently, $$\Gamma_k^+=\left\{\lambda\in \mathbb{R}^n{\mathrel{}\middle|\mathrel{}}\sigma_{1}(\lambda)>0,\dotsc,\sigma_{k}(\lambda)>0\right\}.$$
For example, the positive $n$-cone is $$\Gamma_n^+ = \left\{ \lambda\in{\mathbb{R}}^n {\mathrel{}\middle|\mathrel{}}\lambda_1,\dotsc,\lambda_n>0 \right\}$$ and the positive $1$-cone is the half-space $$\Gamma_1^+ = \left\{ \lambda\in{\mathbb{R}}^n {\mathrel{}\middle|\mathrel{}}\lambda_1+\dotsb+\lambda_n>0 \right\} .$$ Note that $\Gamma_k^+$ is an open convex cone and that $$\Gamma_{n}^+ \subset \Gamma_{n-1}^+\dotsb \subset\Gamma_{1}^+.$$
Applying G[å]{}rding’s theory of hyperbolic polynomials [@Garding], one concludes that $\sigma_{k}^{\frac{1}{k}}$ is a concave function in $\Gamma_{k}^{+}$.
A symmetric matrix $A$ is in the $\tilde{\Gamma}_{k}^+$ cone if its eigenvalues $$\lambda(A)=(\lambda_1(A),\dotsc,\lambda_n(A))\in \Gamma_{k}^+.$$
Suppose $f$ is a function on $\Gamma_{k}^+$. Denote by $F=f(\lambda(A))$ the function on $\tilde{\Gamma}_{k}^+$ induced by $f$. It is known [@Caff4] that if $f$ is concave in $\Gamma_{k}^{+}$, then the induced function $F$ is concave in $\tilde{\Gamma}_{k}^{+}$. For this reason, we shall denote $\tilde{\Gamma}_k^+$ by $\Gamma_k^+$ and $\sigma_k(\lambda(A))$ by $\sigma_k(A)$ when there is no possibility of confusion.
Notice that $\sigma_n(A)=\det (A)$. An equivalent definition of $\det (A)$ is $$\det A:=\frac{1}{n!}\delta^{i_1\dotsc i_n}_{j_1\dotsc j_n}A_{i_1j_1} \dotsm A_{i_nj_n},$$ where $\delta^{i_1\dotsc i_n}_{j_1\dotsc j_n} $ is the generalized Kronecker delta; it is zero if $\{i_1,\dotsc,i_n\}\neq\{j_1,\dotsc,j_n\}$ and equals $1$ (resp. equals $-1$) if $(i_1,\dotsc,i_n)$ and $(j_1,\dotsc,j_n)$ differ by an even (resp. odd) permutation. Similarly, an equivalent definition of $\sigma_k(A)$ is $$\sigma_{k}(A):=\frac{1}{k!}\delta^{i_1\dotsc i_k}_{j_1\dotsc j_k}A_{i_1j_1} \dotsm A_{i_kj_k}.$$ The Newton transformation tensor is defined as $$T_{k}(A)_{ij}:=\frac{1}{k!}
\delta^{ii_1\dotsc i_k}_{jj_1\dotsc j_k}{(A)}_{i_1j_1} \dotsm {(A)}_{i_kj_k}.$$
\[2.3\] The polarization of $\sigma_k$ is $$\sigma_{k}(A_1,\dotsc,A_k):=\frac{1}{k!}\delta^{i_1\dotsc i_k}_{j_1\dotsc j_k}{(A_1)}_{i_1j_1} \dotsm {(A_k)}_{i_kj_k}.$$
It is called the polarization of $\sigma_k$ because $\sigma_k(A_1,\dotsc,A_k)$ is the symmetric multilinear form such that $\sigma_k(A)=\sigma_k(A,\dotsc,A)$.
\[2.4\] The polarized Newton transformation tensor is $$T_{k}(A_1,\dotsc,A_k)_{ij}:=\frac{1}{k!}\delta^{ii_1\dotsc i_k}_{jj_1\dotsc j_k}{(A_1)}_{i_1j_1} \dotsm {(A_k)}_{i_kj_k}.$$
When some components in the polarizations are the same, we adopt the notational conventions $$\begin{aligned}
\sigma_{k}(\overbrace{B,\dotsc,B}^{l},C,\dotsc,C) & :=\sigma_{k}(\overbrace{B,\dotsc,B}^{l},\overbrace{C,\dotsc,C}^{k-l}), \\
T_{k}(\overbrace{B,\dotsc,B}^{l},C,\dotsc,C)_{ij} & :=T_{k}(\overbrace{B,\dotsc,B}^{l},\overbrace{C,\dotsc,C}^{k-l})_{ij}.\end{aligned}$$
Some useful relations between the Newton transformation tensor $T_{k}$ and $\sigma_k$ are as follows. For any symmetric matrix $A$, if we denote the trace by $\operatorname{Tr}$, then $$\begin{aligned}
\sigma_{k}(A)&=\frac{1}{n-k}\operatorname{Tr}(T_{k}(A)_{ij}), \\
\sigma_{k+1}(A)&=\frac{1}{k+1}\operatorname{Tr}(T_{k}(A)_{im}A_{mj} ).\end{aligned}$$
Many useful algebraic inequalities for elements of $\Gamma_{k}^+$ can be deduced from G[å]{}rding’s theory of hyperbolic polynomials [@Garding]. For us, the important such inequality is the fact that if $A_1,\dotsc,A_k\in\bar{\Gamma}_{k+1}^{+}$, then $T_k(A_1,\dots,A_k)_{ij}$ is a nonnegative matrix.
Construction of the polarized functional {#sec:polarized}
========================================
We begin our construction of the boundary integrals of Proposition \[prop:main\_prop\]. Define $$\label{eqn:S0}
\begin{split}
S_0(u,w^1,\dotsc,w^k)&:= -2\sum_p \int_X u_i w_j^p T_{k-1}(D^2 w^{{\wedge}p})_{ij}dx\\
&\quad-\sum_{p\neq q} \int_X w_i^p w_j^q T_{k-1}(D^2 u, D^2 w^{{\wedge}p,q})_{ij}dx.\\
\end{split}$$ where $D^2w^{{\wedge}p}$ denotes the list $(D^2w^1,\dotsc,D^2w^{p-1},D^2w^{p+1},\dotsc,D^2w^k)$ obtained from $(D^2w^1,\dotsc,D^2w^k)$ by removing the entry $D^2w^p$, and likewise $D^2 w^{{\wedge}p,q}$ denotes the list obtained from $(D^2w^1,\dotsc,D^2w^k)$ by removing the entries $D^2w^p$ and $D^2w^q$. Similar notation will be used to remove more elements from the list. Using integration by parts to rewrite as a sum of an interior and a boundary integral, both of which have integrands which factor through $u$, yields the following first step towards proving Proposition \[prop:main\_prop\].
\[prop:pre\_main\_prop\] There exists a symmetric ${\mathbb{R}}$-multilinear function $A_k\colon C^\infty(\overline{X})\to C^\infty(M)$ such that $$\label{eqn:L}
L(u,w^1,\dotsc, w^k):= \int_X u \sigma_k(D^2 w^1,\dotsc, D^2 w^k) dx + \oint_M u A_k(w^1,\dotsc,w^k)d\mu$$ is symmetric in $u,w^1,\dotsc, w^k$.
The operators $A_k$ constructed by our proof depend on at most $4$ derivatives of their inputs.
Note that $S_0$ is symmetric. Our objective is to rewrite in the desired form . To that end, writing as a sum over pairs $p\not=q$ and then integrating by parts in $X$ yields $$\begin{aligned}
S_0&=\sum_{p\neq q}\Bigl[-\frac{2}{k-1}
\int_X u_i w_j^p T_{k-1}(D^2 w^{{\wedge}p})_{ij} dx -\int_X w_i^p w_j^q T_{k-1}(D^2 u, D^2 w^{{\wedge}p,q})_{ij} dx \Bigr]\\
&=\sum_{p\neq q}\Bigl[\frac{2}{k-1}
\int_X u w_{ij}^p T_{k-1}(D^2 w^{{\wedge}p})_{ij}dx+\int_X w^p u_{ij} T_{k-1}( D^2 w^{{\wedge}p})_{ij}dx\\
&\quad-\frac{2}{k-1}\oint_M u w_{j}^p T_{k-1}(D^2 w^{{\wedge}p})_{jn} d\mu -\oint_M w^p w_{j}^q T_{k-1}(D^2 u, D^2 w^{{\wedge}p,q })_{jn} d\mu\Bigr].\end{aligned}$$ Integrating by parts in $X$ once more yields $$\begin{aligned}
S_0&=\sum_{p\neq q}\Bigl[\frac{k+1}{k-1}
\int_X u w_{ij}^p T_{k-1}(D^2 w^{{\wedge}p})_{ij}dx-\frac{k+1}{k-1}\oint_M u w_{j}^p T_{k-1}(D^2 w^{{\wedge}p})_{jn}d\mu \\
&\quad-\oint_M w^p w_{j}^q T_{k-1}(D^2 u, D^2 w^{{\wedge}p,q })_{jn}d\mu
+\oint_M w^p u_j T_{k-1}(D^2 w^{{\wedge}p})_{jn}d\mu\Bigr].\end{aligned}$$ Denote the boundary integral by $T$: $$\label{6.1}
\begin{split}
T &= \sum_{p\neq q}\Bigl[\oint_M w^p u_j T_{k-1}( D^2 w^{{\wedge}p})_{jn} d\mu- \oint_M w^p w_j^q T_{k-1}(D^2 u, D^2 w^{{\wedge}p,q})_{jn}d\mu \Bigr] \\
&\quad-(k+1)\sum_p\oint_M uw_j^pT_{k-1}(D^2w^{{\wedge}p})_{jn} d\mu.
\end{split}$$ Thus $$S_0 = k^2(k+1) \int_X u \sigma_{k}(D^2 w^{1},\dotsc, D^2 w^k)dx + T .$$
We aim to write $T$ as the sum of a symmetric term and a boundary integral of the form $\oint uB(w^1,\dotsc, w^k)d\mu$. To that end, consider the symmetrization of the second term of : $$\label{6.2}
\begin{split}
S_1&:=\sum_{p\neq q}\Bigl[ - \oint_M w^p w_j^q T_{k-1}(D^2 u, D^2 w^{{\wedge}p,q})_{jn}d\mu\\
&\quad-\frac{1}{k-1}\oint_M w^p u_j T_{k-1}( D^2 w^{{\wedge}p})_{jn}d\mu-\frac{1}{k-1}\oint_M u w^p_j T_{k-1}( D^2 w^{{\wedge}p})_{jn}d\mu\Bigr].\\
\end{split}$$ Note that $S_1$ is symmetric with respect to $u, w^1, \dotsc, w^k$. Combining and yields $$T=S_1-k\sum_{p}\oint_M u w_j^p T_{k-1}(D^2 w^{{\wedge}p})_{jn} d\mu+ k\sum_{p}\oint_M w^p u_j T_{k-1}( D^2 w^{{\wedge}p})_{jn}d\mu .$$ We define $$\begin{aligned}
U_1 & := -k\sum_{p}\oint_M u w_j^p T_{k-1}(D^2 w^{{\wedge}p})_{jn} d\mu, \\
Q & := k\sum_{p} \oint w^p u_j T_{k-1}( D^2 w^{{\wedge}p})_{jn} d\mu,\end{aligned}$$ so that $$T= U_1+ S_1+ Q.$$ $U_1$ is of the correct form $\oint u B(w^1,\cdots w^p) d\mu$. We continue with the term $Q$. Observe that $$Q=k\sum_{p}\Bigl[\oint_M w^p u_{\alpha} T_{k-1}( D^2 w^{{\wedge}p})_{\alpha n}d\mu + \oint_M w^p u_{n} T_{k-1}( D^2 w^{{\wedge}p})_{n n} d\mu\Bigr] .$$ where Greek indices $\alpha,\beta\in\{1,\dotsc,n-1\}$ denote tangential directions and $n$ denotes the outward-pointing normal along $M$. By the definition of Newton tensor, $T_{k-1}( D^2 w^{{\wedge}p})_{n n}=\sigma_{k-1}(D^2 w{\rvert}_{TM}^{{\wedge}p} )$, where $D^2w{\rvert}_{TM}^{{\wedge}p}$ denotes the list of the restrictions $D^2w^1{\rvert}_{TM},\dotsc,D^2w^n{\rvert}_{TM}$ with the $p$-th element removed. Thus $$Q=k\sum_{p}\Bigl[ \oint_M w^p u_{\alpha} T_{k-1}( D^2 w^{{\wedge}p})_{\alpha n}d\mu + \oint_M w^p u_{n} \sigma_{k-1}(D^2 w{\rvert}_{TM}^{{\wedge}p} ) d\mu\Bigr] .$$ Define $$\begin{aligned}
U_2 & := k\sum_{p} \oint_M w^p u_\alpha T_{k-1}(D^2w^{{\wedge}p})_{\alpha n}d\mu, \\
Q_1 & := k\sum_{p} \oint_M w^pu_{n} \sigma_{k-1}(D^2w{\rvert}_{TM}^{{\wedge}p}) d\mu,\end{aligned}$$ so that $$Q = U_2 + Q_1 .$$ Integrating by parts along $M$ shows that $$U_2= -k\sum_{p} \oint u ( w^p T_{k-1}( D^2 w^{{\wedge}p})_{\alpha n})_\alpha d\mu.\\$$ Thus $U_2$ is of the correct form $\oint uB(w^1,\dotsc,w^p)d\mu$. Therefore we need only consider $Q_1$.
Consider the symmetrization of $Q_1$: $$\begin{aligned}
S_2&:=\sum_{p\neq q}\Bigl[ \frac{k}{k-1}\oint_M w^p u_{n} \sigma_{k-1}(D^2 w{\rvert}_{T_xM}^{{\wedge}p} )d\mu + \frac{k}{k-1}\oint_M u w^p_{n} \sigma_{k-1}(D^2 w{\rvert}_{TM}^{{\wedge}p} )d\mu\\
&\quad+k \oint_M w^p w^q_{n} \sigma_{k-1}(D^2 u{\rvert}_{TM} , D^2 w{\rvert}_{TM}^{{\wedge}p,q} ) d\mu\Bigr].\\\end{aligned}$$ Note that $S_2$ is symmetric with respect to $u, w^1, \dotsc, w^k$. Moreover, $$\label{eqn:Q1_sym}
\begin{split}
Q_1 & =S_2 - \frac{k}{k-1}\sum_{p\neq q} \oint_M u w^p_{n} \sigma_{k-1}(D^2 w{\rvert}_{TM}^{{\wedge}p} ) d\mu\\
& \quad - k\sum_{p\neq q} \oint_M w^p w^q_{n} \sigma_{k-1}(D^2 u{\rvert}_{TM} , D^2 w{\rvert}_{TM}^{{\wedge}p,q} )d\mu .
\end{split}$$ Denote by $\bar{D}^2$ the Hessian with respect to the induced metric of $M$ and by $L_{\alpha\beta}$ the second fundamental form of $M$. Given $v\in C^\infty({\overline{X}})$, it holds that $$\label{eqn:tangential_hessian}
D^2 v{\rvert}_{TM} = \bar{D}^2 v + v_nL$$ along $M$. Define $$\begin{aligned}
U_3 & :=-\frac{k}{k-1}\sum_{p\neq q} \oint_M u w^p_{n} \sigma_{k-1}(D^2 w{\rvert}_{TM}^{{\wedge}p} ) d\mu, \\
U_4 & :=-k\sum_{p\neq q} \oint_M w^p w^q_{n} \sigma_{k-1}(\bar{D}^2 u , D^2 w{\rvert}_{TM}^{{\wedge}p,q} )d\mu .\end{aligned}$$ Integrating by parts along $M$ yields $$U_4=-\frac{k}{k-1}\sum_{p\neq q} \oint_M u \big(w^p w^q_{n} T_{k-2}( D^2 w{\rvert}_{TM}^{{\wedge}p,q} )_{\alpha\beta}\big)_{\bar\alpha\bar\beta} d\mu ,$$ where the bars on $\alpha$ and $\beta$ denote covariant derivatives with respect to the induced metric on $M$. In particular, both $U_3$ and $U_4$ are of the form $\oint uB(w^1,\dotsc,w^k)d\mu$. Define $$Q_2:=-k\sum_{p\neq q} \oint_M w^p w^q_{n}u_n \sigma_{k-1}(L, D^2 w{\rvert}_{TM}^{{\wedge}p,q} ) d\mu.$$ It follows from , and the definitions of $U_3, U_4, Q_2$ that $$Q_1=S_2+U_3+U_4+Q_2 .$$
Now we want to write $Q_2$ in the desired form. To that end, consider the symmetrization of $Q_2$: $$\label{6.4}
\begin{split}
S_3&:=-k \sum_{p\neq q\neq r} \Bigl[\frac{1}{k-2}\oint_M w^p w^q_{n}u_n \sigma_{k-1}(L, D^2 w{\rvert}_{TM}^{{\wedge}p,q} )d\mu \\
&\quad+\frac{1}{2!(k-2)}\oint_M u w^p_n w^q_{n} \sigma_{k-1}(L, D^2 w{\rvert}_{TM}^{{\wedge}p,q} ) d\mu\\
&\quad+\frac{1}{2!}\oint_M w^p w^q_{n}w^r_n \sigma_{k-1}( L,D^2 u{\rvert}_{TM}, D^2 w{\rvert}_{TM}^{{\wedge}p,q, r} )d\mu \Bigr],\\
\end{split}$$ Note that $S_3$ is symmetric with respect to $u,w^1,\dotsc,w^k$. Define $$\begin{aligned}
U_5 & :=\frac{k}{2!(k-2)}\sum_{p\neq q\neq r}\oint_M u w^p_n w^q_{n} \sigma_{k-1}(L, D^2 w{\rvert}_{TM}^{{\wedge}p,q} ) d\mu,\\
U_6 & :=\frac{k}{2!}\sum_{p\neq q\neq r}\oint_M w^p w^q_{n}w^r_n \sigma_{k-1}( L,\bar{D}^2 u{\rvert}_{TM}, D^2 w{\rvert}_{TM}^{{\wedge}p,q,r} ) d\mu.\end{aligned}$$ As above, integration by parts along $M$ implies that both $U_5$ and $U_6$ are of the form $\oint uB(w^1,\dotsc,w^k)d\mu$. Define $$Q_3 := \frac{k}{2!}\sum_{p\neq q\neq r}\oint_M w^p w^q_{n}w^r_nu_n \sigma_{k-1}( L,L, D^2 w{\rvert}_{TM}^{{\wedge}p,q,r} )d\mu .$$ From and the definitions of $Q_2$, $U_5$, $U_6$ and $Q_3$ we deduce that $$Q_2= S_3+U_5+U_6+Q_3 .$$
Proceeding in this way, for all $2\leq i\leq k$ we make the following definitions. First, define $$\begin{aligned}
S_i & := (-1)^ik \sum_{p_1\neq \dotsb \neq p_i} \Bigl[\frac{1}{(i-2)!(k+1-i)}\oint_M w^{p_1} w^{p_2}_{n}\dotsm w^{p_{i-1}}_n u_n \\
&\qquad\qquad\qquad\qquad \times \sigma_{k-1}(\overbrace {L,\dotsc, L}^{i-2}, D^2 w{\rvert}_{TM}^{{\wedge}p_1,\dotsc, p_{i-1}} ) d\mu\\
&\quad+\frac{1}{(i-1)!(k+1-i)}\oint_M u w^{p_1}_n \dotsm w^{p_{i-1}}_n \sigma_{k-1}(\overbrace {L,\dotsc, L}^{i-2}, D^2 w{\rvert}_{TM}^{{\wedge}p_1,\dotsc, p_{i-1}} )d\mu ,\\
&\quad+\frac{1}{(i-1)!}\oint_M w^{p_1} w^{p_2}_{n}\dotsm w^{p_{i}}_n \sigma_{k-1}( \overbrace {L,\dotsc, L}^{i-2},D^2 u{\rvert}_{TM}, D^2 w{\rvert}_{TM}^{{\wedge}p_1,\dotsc, p_i} ) d\mu\Bigr] .\end{aligned}$$ Note that $S_i$ is symmetric with respect to $u,w^1,\dotsc,w^k$. Next, define $$\begin{aligned}
U_{2i-1}& := \frac{(-1)^{i+1}k}{(i-1)!(k+1-i)}\sum_{p_1\neq \dotsb \neq p_i}\oint_M u w^{p_1}_n \dotsm w^{p_{i-1}}_n \\
&\quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times \sigma_{k-1}(\overbrace {L,\dotsc, L}^{i-2}, D^2 w{\rvert}_{TM}^{{\wedge}p_1,\dotsc, p_{i-1}} ) d\mu,\\
U_{2i}& := \frac{(-1)^{i+1}k}{(i-1)!}\sum_{p_1\neq \dotsb \neq p_i} \oint_M w^{p_1} w^{p_2}_{n}\dotsm w^{p_i}_n \sigma_{k-1}(\overbrace {L,\dotsc, L}^{i-2},\bar{D}^2u, D^2 w{\rvert}_{TM}^{{\wedge}p_1,\dotsc, p_i})d\mu .\end{aligned}$$ Integration by parts along $M$ implies that both $U_{2i-1}$ and $U_{2i}$ are of the form $\oint uB(w^1,\dotsc,w^k)d\mu$. Then $$Q_i:= \frac{(-1)^{i+1} k}{(i-1)!}\sum_{p_1\neq \dotsb \neq p_i}\oint_M w^{p_1} w^{p_2}_{n}\dotsm w^{p_i}_n u_n \sigma_{k-1}(\overbrace {L,\dotsc, L}^{i-1}, D^2 w{\rvert}_{TM}^{{\wedge}p_1,\dotsc,p_i} ) d\mu$$ is such that $$Q_{i-1} = S_i + U_{2i-1} + U_{2i} + Q_i .$$
It remains to write $Q_k$ as the sum of a symmetric integral and a boundary integral whose integrand factors through $u$. To that end, define $$\begin{aligned}
S_{k+1} & := \frac{(-1)^{k+1}k}{(k-1)!}\sum_{p_1\neq \dotsb \neq p_k} \Bigl[\oint_M w^{p_1} w^{p_2}_{n}\dotsm w^{p_{k}}_n u_n\sigma_{k-1}(L)d\mu \\
& \quad +\frac{1}{k}\oint_M u w^{p_1}_n \dotsm w^{p_{k}}_n \sigma_{k-1}(L )d\mu \Bigr] .\end{aligned}$$ Note that $S_{k+1}$ is symmetric with respect to $u,w^1,\dotsc,w^k$. Also define $$U_{2k+1} := \frac{(-1)^{k}}{(k-1)!}\sum_{p_1\neq \dotsb \neq p_k}\oint_M u w^{p_1}_n \dotsm w^{p_{k}}_n \sigma_{k-1}(L)d\mu .$$ Note that $U_{2k+1}$ is of the form $\oint uB(w^1,\dotsc,w^k)d\mu$ and that $$Q_k = S_{k+1} + U_{2k+1} .$$
In summary, we have shown that $$\label{eqn:final_induction}
S_0 - \sum_{i=1}^{k+1} S_i = k^2(k+1)\int_X u\,\sigma_{k}(D^2w^1,\dotsc,D^2w^{k}) dx+ \sum_{i=1}^{2k+1} U_i$$ and observed that the left-hand side is symmetric in $u,w^1,\dotsc,w^k$ while the right-hand side is of the form $\oint uB(w^1,\dotsc,w^k)d\mu$. Dividing through by $k^2(k+1)$ yields .
Adjusted polarized functional {#sec:adjusted}
=============================
The difference between Proposition \[prop:main\_prop\] and Proposition \[prop:pre\_main\_prop\] is that in the latter result, we only ask that the boundary integrals making up the polarized functional are such that their integrands factor through $u$. In particular, it is not clear that from the proof of Proposition \[prop:pre\_main\_prop\] that the functions $A_k$ depend only on at most second-order tangential derivatives and at most first-order transverse derivatives along $M$. This arises in two ways. First, the integral $U_1$ depends on the second-order derivative $w_{\alpha n}$. Second, when written in the form $\oint uB(w^1,\dotsc,w^k) d\mu$, the integrals $U_{2i}$, $1\leq i\leq k$, depend also on third- and fourth-order derivatives of $w^p$. By more carefully considering the integration by parts along $M$ invoked in the proof of Proposition \[prop:pre\_main\_prop\], we show that the combination $\sum U_i$ only depends on at most second-order tangential derivatives and at most first-order transverse derivatives of $w^p$. This proves Proposition \[prop:main\_prop\]. To that end, we first require a few facts.
\[lem:boundary\_derivatives\] Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with boundary $M=\partial X$. Let $w^1,\dotsc,w^k\in C^\infty({\overline{X}})$. Then $$\begin{aligned}
\label{eqn:tangential_mixed} w_{\beta n} & = w_{n\bar\beta} - L_{\alpha\beta}w_\alpha, \\
\label{eqn:Tktangential_mixed} T_k(D^2w^1,\dotsc,D^2w^k)_{\alpha n} & = -\frac{1}{k}\sum_{p=1}^k T_{k-1}(D^2w{\rvert}_{TM}^{{\wedge}p})_{\alpha\beta} w_{\beta n}^p ,
\end{aligned}$$ where $\alpha,\beta\in\{1,\dotsc,n-1\}$ denote tangential directions, $n$ denotes the outward-pointing normal along the boundary, and $w_{n\bar\beta}$ denotes the tangential gradient of $w_n$. Moreover, $$\label{eqn:Tktangentialdivergence}
T_k(\overbrace{L,\dotsc,L}^{i},D^2w{\rvert}_{TM}^{{\wedge}p_1,\dotsc,p_i})_{\alpha\beta,\bar\beta} = \sum_{p\not=p_1,\dotsc,p_i} T_{k}(\overbrace{L,\dotsc,L}^{i+1},D^2w{\rvert}_{TM}^{{\wedge}p,p_1,\dotsc,p_i})_{\alpha\beta}w_{\beta n}^p,$$ where the left-hand side denotes the divergence with respect to the induced metric on $M$.
follows immediately from the definition of the second fundamental form $L$ and follows immediately from the definitions of the Newton tensors. To prove , first recall that the Newton tensors are divergence-free with respect to the flat metric in $X$. From the definition of the second fundamental form, we have that $$w_{\alpha\beta,\gamma} = w_{\alpha\beta,\bar\gamma} + L_{\alpha\gamma}w_{\beta n} + L_{\beta\gamma}w_{\alpha n} .$$ Inserting this into the definition of the Netwon tensors yields the result (cf. [@Chen2009]\*[Lemma 11]{}).
Lemma \[lem:boundary\_derivatives\] allows us to carefully perform the integration by parts argument as described above.
Denote ${\mathcal{C}}:=C^1({\overline{X}})\cap C^2(M)$. Define $$\begin{aligned}
\tilde U_1 & := -k\sum_p \oint_M uw_n^p\sigma_{k-1}(D^2w{\rvert}_{TM}^{{\wedge}p}) d\mu, \\
\hat U_1 & := -k\sum_p \oint_M uw_\alpha^p T_{k-1}(D^2w)_{\alpha n} d\mu .
\end{aligned}$$ It follows from that $\tilde U_1$ is well-defined on ${\mathcal{C}}$; i.e. $\tilde U_1$ depends on at most second-order tangential derivatives and first-order transverse derivatives of $w^1,\dotsc,w^k$ on $M$. Furthermore, we have that $$U_1 = \tilde U_1 + \hat U_1 .$$
Consider now $\hat U_1+U_2+U_4$. Define $$\begin{aligned}
W_1 & := -\frac{k}{k-1}\sum_{p\neq q}\oint_M uw_\alpha^p T_{k-2}(D^2w{\rvert}_{TM}^{{\wedge}p,q})_{\alpha\beta}L_{\beta\gamma}w_\gamma^q d\mu, \\
W_2 & := -\frac{k}{k-1}\sum_{p\neq q}\oint_M uw_n^p T_{k-2}(D^2w{\rvert}_{TM}^{{\wedge}p,q})_{\alpha\beta}w_{\bar\alpha\bar\beta}^q d\mu, \\
W_3 & := \frac{k}{k-1}\sum_{p\neq q\neq r}\oint_M uw_n^pw_\alpha^qT_{k-2}(L,D^2w{\rvert}_{TM}^{{\wedge}p,q,r})_{\alpha\beta}L_{\beta\gamma}w_\gamma^r d\mu.
\end{aligned}$$ It follows from that $W_1,W_2,W_3$ are well-defined on ${\mathcal{C}}$. Define also $$\begin{aligned}
V_1 & := \frac{k}{k-1}\sum_{p\neq q}\oint_M w^pu_\alpha T_{k-2}(D^2w{\rvert}_{TM}^{{\wedge}p,q})_{\alpha\beta}L_{\beta\gamma}w_\gamma^qd\mu, \\
V_2 & := -\frac{k}{k-1}\sum_{p\neq q\neq r} \oint_M w^pw_n^qu_\alpha T_{k-2}(L,D^2w{\rvert}_{TM}^{{\wedge}p,q,r})_{\alpha\beta} L_{\beta\gamma} w_\gamma^rd\mu,
\end{aligned}$$ Note that $V_1$ and $V_2$ still involve derivatives of $u$; this issue will be dealt with later. Integrating by parts along $M$ and using Lemma \[lem:boundary\_derivatives\] yields $$\begin{aligned}
\hat U_1 + U_2 + U_4 & = W_1 + V_1 + \frac{k}{k-1}\sum_{p\neq q}\oint_M uw_\alpha^p T_{k-2}(D^2w{\rvert}_{TM}^{{\wedge}p,q})_{\alpha\beta}w_{n\bar\beta}^q d\mu\\
& \quad - \frac{k}{k-1}\sum_{p\neq q}\oint_M w^p T_{k-2}(D^2w{\rvert}_{TM}^{{\wedge}p,q})_{\alpha\beta}(u_\alpha w_n^q)_{\bar\beta} d\mu\\
& = W_1 + V_1 + V_2 + \frac{k}{k-1}\sum_{p\neq q}\oint_M w_\alpha^p T_{k-2}(D^2w{\rvert}_{TM}^{{\wedge}p,q})_{\alpha\beta}(uw_n)_{\bar\beta}d\mu \\
& \quad + \frac{k}{k-1}\sum_{p\neq q\neq r} \oint_M w^pw_n^qu_\alpha T_{k-2}(L,D^2w{\rvert}_{TM}^{{\wedge}p,q,r})_{\alpha\beta}w_{n\bar\beta}^rd\mu \\
& = W_1 + W_2 + W_3 + V_1 + V_2 + \hat U_2 + \hat U_3 ,
\end{aligned}$$ where $$\begin{aligned}
\hat U_2 & := -\frac{k}{k-1}\sum_{p\neq q\neq r} \oint_M uw_n^pw_\alpha^q T_{k-2}(L,D^2w{\rvert}_{TM}^{{\wedge}p,q,r})_{\alpha\beta}w_{n\bar\beta}^rd\mu, \\
\hat U_3 & := \frac{k}{k-1}\sum_{p\neq q\neq r} \oint_M w^pw_n^qu_\alpha T_{k-2}(L,D^2w{\rvert}_{TM}^{{\wedge}p,q,r})_{\alpha\beta}w_{n\bar\beta}^rd\mu .
\end{aligned}$$
We continue this process by considering $\hat U_2+\hat U_3+U_6$. More generally, given $1\leq i\leq k-1$, we make the following definitions. First, define $$\begin{aligned}
W_{2i-1} & := (-1)^i\frac{k}{k-1}\sum_{p_0\neq\dotsb\neq p_i} \frac{1}{(i-1)!}\oint_M uw_\alpha^{p_0}w_n^{p_1}\dotsm w_n^{p_{i-1}}
\\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times
T_{k-2}(\overbrace{L,\dotsc,L}^{i-1},D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_i})_{\alpha\beta}L_{\beta\gamma}w_\gamma^{p_{i}}d\mu, \\
W_{2i} & := (-1)^i\frac{k}{k-1}\sum_{p_0\neq\dotsb\neq p_{i}}\frac{1}{i!}\oint_M uw_n^{p_0}\dotsm w_n^{p_{i-1}} \\
& \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\times T_{k-2}(\overbrace{L,\dotsc,L}^{i-1},D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_{i}})_{\alpha\beta}w_{\bar\alpha\bar\beta}^{p_{i}} d\mu.
\end{aligned}$$ It follows from that $W_{2i-1}$ and $W_{2i}$ are well-defined on ${\mathcal{C}}$. Next, define $$\begin{aligned}
V_i &:= (-1)^{i+1}\frac{k}{k-1}\sum_{p_0\neq\dotsb\neq p_{i}} \frac{1}{(i-1)!}\oint_M u_\alpha w^{p_0}w_n^{p_1}\dotsm w_n^{p_{i-1}} \\ &\quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times
T_{k-2}(\overbrace{L,\dotsc,L}^{i-1},D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_{i}})_{\alpha\beta}L_{\beta\gamma}w_\gamma^{p_{i}}d\mu .\\ \end{aligned}$$ Note that $V_i$ still involves derivatives of $u$; this issue will be dealt with later. Finally, define $$\begin{aligned}
\hat U_{2i} & := (-1)^i\frac{k}{k-1}\sum_{p_0\neq\dotsb\neq p_{i+1}} \frac{1}{i!}\oint_M uw_\alpha^{p_0}w_n^{p_1}\dotsm w_n^{p_i}
\\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times T_{k-2}(\overbrace{L,\dotsc,L}^i,D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_{i+1}})_{\alpha\beta}w_{n\bar\beta}^{p_{i+1}}d\mu, \\
\hat U_{2i+1} & := (-1)^{i+1}\frac{k}{k-1}\sum_{p_0\neq\dotsb\neq p_{i+1}} \frac{1}{i!}\oint_M u_\alpha w^{p_0}w_n^{p_1}\dotsm w_n^{p_i}
\\
& \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times T_{k-2}(\overbrace{L,\dotsc,L}^i,D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_{i+1}})_{\alpha\beta}w_{n\bar\beta}^{p_{i+1}}d\mu ;
\end{aligned}$$ note that $\hat U_{2k-2}=\hat U_{2k-1}=0$. Integrating by parts along $M$ and using Lemma \[lem:boundary\_derivatives\] yields $$\hat U_{2i} + \hat U_{2i+1} + U_{2i+4} = V_{i+2} + W_{2i+2} + W_{2i+3} + \hat U_{2i+2} + \hat U_{2i+3} .$$ In particular, it follows that $$\label{eqn:Usimplification}
\sum_{i=1}^{2k+1} U_i = \tilde U_1 + \sum_{i=1}^k U_{2i+1} + \sum_{i=1}^{2k-2}W_i + \sum_{i=1}^{k-1} V_i .$$ Note that $\tilde U_1$, $\sum U_{2i+1}$, and $\sum W_i$ are all well-defined on ${\mathcal{C}}$. It remains to check that, after integration by parts, $\sum V_i$ can be written as a boundary integral with integrand the product of $u$ with a function which is well-defined on ${\mathcal{C}}$.
Given $1\leq i\leq k-1$, define $$\begin{aligned}
A_i & := (-1)^i\frac{k}{(i-1)!(k-1)}\sum_{p_0\neq\dotsb\neq p_{i+1}} \oint_M uw^{p_0}w_n^{p_1}\dotsm w_n^{p_{i-1}} w_\alpha^{p_i} w_{n\bar\beta}^{p_{i+1}} \\
& \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times T_{k-2}(\overbrace{L,\dotsc,L}^i,D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_{i+1}})_{\alpha\gamma}L_{\gamma\beta} d\mu, \\
B_i & := (-1)^{i+1}\frac{k}{(i-1)!(k-1)}\sum_{p_0\neq\dotsb\neq p_{i+1}} \oint_M uw^{p_0}w_n^{p_1}\dotsm w_n^{p_{i-1}} w_\gamma^{p_i} w_\delta^{p_{i+1}}
\\
& \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times T_{k-2}(\overbrace{L,\dotsc,L}^i,D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_{i+1}})_{\alpha\beta} L_{\alpha\gamma} L_{\beta\delta} d\mu, \\
C_i & := (-1)^i\frac{k}{(i-1)!(k-1)}\sum_{p_0\neq\dotsb\neq p_i} \oint_M uw_n^{p_0}\dotsm w_n^{p_{i-2}}
\\
&\quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \times T_{k-2}(\overbrace{L,\dotsc,L}^{i-1},D^2w{\rvert}_{TM}^{{\wedge}p_0,\dotsc,p_i})_{\alpha\beta}\left(w^{p_{i-1}}w_\gamma^{p_i}L_{\alpha\gamma}\right)_{\bar\beta} d\mu.
\end{aligned}$$ Note that $B_i$ and $C_i$ are well-defined on ${\mathcal{C}}$. Moreover, integration by parts along $M$ readily yields $$V_i = A_i - A_{i-1} + B_i + C_i,$$ where we interpret $A_0=0$. Since $A_{k-1}=0$, it follows that $$\label{eqn:Vsimplification}
\sum_{i=1}^{k-1} V_i = \sum_{i=1}^{k-1} \left( B_i + C_i\right) .$$ Combining and yields the desired result.
The first and second variation {#sec:variation}
==============================
It is straightforward to compute the first and second variations of the energy functional $${\mathcal{E}}_k(u) := {\mathcal{Q}}_k(u,\dotsc,u)$$ associated to the symmetric multilinear form constructed by Proposition \[prop:main\_prop\].
\[prop:first\_variation\] Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with boundary $M=\partial X$. Let $u,v\in C^\infty({\overline{X}})$ and suppose that $v{\rvert}_M=0$. Then $$\label{eqn:first_variation}
\left.\frac{d}{dt}\right|_{t=0}{\mathcal{E}}_k(u+tv) = -(k+1)\int_X v\,\sigma_k(D^2u,\dotsc,D^2u) dx.$$
Since ${\mathcal{Q}}_k$ is symmetric, we compute that $$\left.\frac{d}{dt}\right|_{t=0}{\mathcal{E}}_k(u+tv) = (k+1){\mathcal{Q}}_k(v,u,\dotsc,u) .$$ Since $v{\rvert}_M=0$, we see that the boundary integral in vanishes. This yields .
\[prop:second\_variation\] Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with boundary $M=\partial X$. Let $u,v\in C^\infty({\overline{X}})$ and suppose that $v{\rvert}_M=0$. Then $$\left.\frac{d^2}{dt^2}\right|_{t=0}{\mathcal{E}}_k(u+tv) = (k+1)\int_X v_iv_j T_{k-1}(D^2u)_{ij}dx .$$ In particular, if $u\in\overline{\Gamma_k^+}$, then $$\left.\frac{d^2}{dt^2}\right|_{t=0}{\mathcal{E}}_k(u+tv)\geq0$$ for all $v\in C^\infty({\overline{X}})$ such that $v{\rvert}_M=0$.
Since ${\mathcal{Q}}_k$ is symmetric, we compute that $$\left.\frac{d^2}{dt^2}\right|_{t=0}{\mathcal{E}}_k(u+tv) = k(k+1){\mathcal{Q}}_k(v,v,u,\dotsc,u) .$$ Since $v{\rvert}_M=0$, it follows that $$\begin{aligned}
\left.\frac{d^2}{dt^2}\right|_{t=0}{\mathcal{E}}_k(u+tv) & = -k(k+1)\int_X v\,\sigma_k(D^2v,D^2u,\dotsc,D^2u) dx, \\
& = -(k+1)\int_X v T_{k-1}(D^2u)_{ij} v_{ij}dx \\
& = (k+1)\int_X v_iv_j T_{k-1}(D^2u)_{ij} dx.
\end{aligned}$$ The last conclusion follows from the fact that if $u\in\overline{\Gamma_k^+}$, then $T_{k-1}(D^2u)_{ij}$ is nonnegative.
We are now ready to prove Theorem \[thm:main\_thm\], which we restate here for convenience.
Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with $(k-1)$-convex boundary $M=\partial X$. Fix $f\in C^\infty(M)$ and denote $${\mathcal{C}}_{f,k} = \left\{ u\in \Gamma_k^+{\mathrel{}\middle|\mathrel{}}u{\rvert}_M = f \right\} .$$ Then $${\mathcal{E}}_k(u) \geq {\mathcal{E}}_k(u_f)$$ for all $u\in\overline{{\mathcal{C}}_{f,k}}$, where $u_f\in\overline{{\mathcal{C}}_{f,k}}$ is the solution to the Dirichlet problem $$\label{eqn:dirichlet_problem_proof}
\begin{cases}
\sigma_k(u_f) = 0, & \text{in $X$}, \\
u_f = f, & \text{on $M$}.
\end{cases}$$
By Proposition \[prop:first\_variation\], the solution $u_f$ to is a critical point of the functional ${\mathcal{E}}_k\colon C^{1,1}({\overline{X}})\to{\mathbb{R}}$. By Proposition \[prop:second\_variation\], the restriction ${\mathcal{E}}_k\colon\overline{{\mathcal{C}}_{f,k}}\to{\mathbb{R}}$ is a convex functional. Since $\overline{{\mathcal{C}}_{f,k}}$ is convex, $u_f$ realizes the infimum of ${\mathcal{E}}_k\colon\overline{{\mathcal{C}}_{f,k}}\to{\mathbb{R}}$. Indeed, if not, then there is a $u\in\overline{{\mathcal{C}}_{f,k}}$ such that ${\mathcal{E}}_k(u)<{\mathcal{E}}_k(u_f)$. Since $\overline{{\mathcal{C}}_{f,k}}$ is convex, it follows that $tu+(1-t)u_f\in\overline{{\mathcal{C}}_{f,k}}$ for all $t\in[0,1]$. Denote ${\mathcal{E}}_k(t):={\mathcal{E}}_k(tu+(1-t)u_f)$. Since ${\mathcal{E}}_k(u)<{\mathcal{E}}_k(u_f)$, there exists a $t^\ast\in[0,1]$ such that ${\mathcal{E}}_k^\prime(t^\ast)<0$. This contradicts the facts that ${\mathcal{E}}_k^\prime(0)=0$ and ${\mathcal{E}}_k^{\prime\prime}\geq0$ for all $t\in[0,1]$.
The case $k=2$ {#sec:example}
==============
We conclude this article by considering the specific case $k=2$; the case $k=1$ is covered by . First, a suitable boundary operator as in Proposition \[prop:main\_prop\] is given by Proposition \[prop:main\_prop2\], which we restate here for convenience.
\[prop:boundary2\] Let $X\subset{\mathbb{R}}^n$ be a bounded smooth domain with boundary $M=\partial X$. Define $B\colon\left(C^1({\overline{X}})\cap C^2(M)\right)^2\to C^0(M)$ by $$\label{eqn:boundary2}
B_2(v,w) = \frac{1}{2}\left(v_n{\overline{\Delta}}w + w_n{\overline{\Delta}}v + L({\overline{\nabla}}v,{\overline{\nabla}}w) + Hv_nw_n\right) .$$ Then the multilinear form ${\mathcal{Q}}_2\colon\left(C^2({\overline{X}})\right)^3\to{\mathbb{R}}$ given by $${\mathcal{Q}}_2(u,v,w) = -\int_X u\sigma_2(D^2v,D^2w) dx+ \oint_M u B_2(v,w)d\mu$$ is symmetric.
Following the proof of Proposition \[prop:main\_prop\], we see that a suitable choice of boundary operator is $$\begin{gathered}
\tilde B_2(v,w) := \frac{1}{2}\left(v_n{\overline{\Delta}}w + w_n{\overline{\Delta}}v + L({\overline{\nabla}}v,{\overline{\nabla}}w) + Hv_nw_n\right) \\ + \frac{1}{6}\left( A({\overline{\nabla}}v,{\overline{\nabla}}w) + v{\langle}A,\bar{D}^2w{\rangle}+ w{\langle}A,\bar{D}^2v{\rangle}+ v{\langle}{\overline{\nabla}}H,{\overline{\nabla}}w{\rangle}+ w{\langle}{\overline{\nabla}}H,{\overline{\nabla}}v{\rangle}\right) .
\end{gathered}$$ A straightforward computation yields $$\begin{gathered}
{\overline{\delta}}\left(vA({\overline{\nabla}}w)\right) + {\overline{\delta}}\left(wA({\overline{\nabla}}v)\right) - A({\overline{\nabla}}v,{\overline{\nabla}}w) \\ = A({\overline{\nabla}}v,{\overline{\nabla}}w) + v{\langle}A,\bar{D}^2w{\rangle}+ w{\langle}A,\bar{D}^2v{\rangle}+ v{\langle}{\overline{\nabla}}H,{\overline{\nabla}}w{\rangle}+ w{\langle}{\overline{\nabla}}H,{\overline{\nabla}}v{\rangle}\end{gathered}$$ On the other hand, $$\begin{gathered}
\oint_M u\left[\,{\overline{\delta}}\left(vA({\overline{\nabla}}w)\right) + {\overline{\delta}}\left(wA({\overline{\nabla}}v)\right) - A({\overline{\nabla}}v,{\overline{\nabla}}w)\right]d\mu \\ = -\oint_M \left[ uA({\overline{\nabla}}v,{\overline{\nabla}}w) + vA({\overline{\nabla}}w,{\overline{\nabla}}u) + wA({\overline{\nabla}}u,{\overline{\nabla}}v) \right]d\mu
\end{gathered}$$ is symmetric in $u,v,w$. Thus $B_2-\tilde B_2$, and hence ${\mathcal{Q}}_2$, is symmetric in $u,v,w$.
Applying this boundary operator in Theorem \[thm:main\_thm\] yields the following sharp Sobolev trace inequality.
\[thm:trace2\] Let $X\subset{\mathbb{R}}^n$ be a bounded smooth mean-convex domain with boundary $M=\partial X$. Given $f\in C^\infty(M)$, set $${\mathcal{C}}_f = \left\{ u\in \Gamma_2^+ {\mathrel{}\middle|\mathrel{}}u{\rvert}_M=f \right\} .$$ Then it holds that $$-\int_X u\sigma_2(D^2u) dx+ \oint_M uB_2(u,u) d\mu\geq \oint_M fB_2(u_f,u_f)d\mu$$ for all $u\in\overline{{\mathcal{C}}_f}$, where $B_2$ is the operator and $u_f\in C^{1,1}({\overline{X}})\cap\overline{\Gamma_2^+}$ is the unique solution to the Dirichlet problem $$\begin{cases}
\sigma_2(D^2u_f)=0, & \text{in $X$}, \\
u = f, & \text{on $M$} .
\end{cases}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nucleation of vortices in a superconductor below the first critical field can be assisted by transverse sound in the GHz frequency range. Vortices will enter and exist the superconductor at the frequency of the sound. We compute the threshold parameters of the sound and show that this effect is within experimental reach.'
author:
- 'Jaroslav Albert and E. M. Chudnovsky'
title: Ignition of superconducting vortices by acoustic standing waves
---
A superconducting cylinder rotated at an angular velocity ${\bf
\Omega}$ about its symmetry axis develops a magnetic moment ${\bf M}
= -(mc/2\pi e){\bf \Omega}$, where $m$ and $e$ are bare electron mass and charge, and $c$ is the speed of light. This effect predicted by London [@London] has been subsequently tested in experiment and proved with an accuracy to many significant figures. It is a consequence of a more general gyromagnetic effect predicted by Barnett [@Barnett]: “A substance which is magnetic must become magnetized by a sort of molecular gyroscopic motion on receiving an angular velocity”. Barnet effect is, in its turn, a consequence of the Larmor theorem [@Larmor]: In the rotating frame of reference the action of the rotation on charged particles is equivalent to the action of the magnetic field ${\bf H}_{\Omega}
= {\bf \Omega}/\gamma$ [@Larmor], where $\gamma$ is the gyromagnetic ratio. For electron’s orbital motion $\gamma = e/(2mc)
\approx 0.9 \times 10^7\,$(Gauss)$^{-1}$ s$^{-1}$. Thus, in practice, the fictitious field in the reference frame of a rotating macroscopic cylinder can hardly exceed a fraction of a milligauss. This would be well below the lower critical field ${\bf H}_{c1}$ when the temperature of the superconductor is not too close to $T_c$. Due to the Meissner effect [@Meissner] (considered in the frame of the rotating cylinder) such a field would be expelled from the bulk of the cylinder by a superconducting current induced at the surface. Writing ${\bf B} = {\bf H}_{\Omega} + 4 \pi {\bf M} = 0$ for the total field in the bulk, one obtains the London’s magnetic moment, ${\bf M} = -{\bf H}_{\Omega}/(4\pi) = -(mc/2\pi e){\bf
\Omega}$. Due to the symmetry of the problem it is the same in the rotating and laboratory frames.
In this Letter we would like to take this problem a little further and look at the consequence of an angular velocity well beyond the experimental limit. In particular, we are interested in the rotational velocity of a magnitude that would generate a fictitious magnetic field that exceeds ${\bf H}_{c1}$. If Larmor’s theorem still holds, than it must be the case that a superconducting vortex enters the bulk of the cylinder. This would require the angular velocity to be of order $10^9$s$^{-1}$, clearly surpassing the feasible experimental value for a mechanical rotation. While this scenario is merely a thought experiment we will use it as a motivation to study the effect of local rotations generated in a superconductor by high frequency ultrasound. Interaction of sound with vortices has been studied in the past [@Haneda; @Dom; @Sonin]. Radiation of phonons by supersonic vortices [@Ivlev; @BC], phonon contribution to the vortex mass [@Coffey; @Duan; @EC-Kuklov-PRL], and decoherence of flux qubits by phonons [@EC-Kuklov-PRB; @Albert-EC-PRB] have been investigated. In this Letter we are addressing a completely different problem – possibility of the nucleation of a vortex by sound.
Within continuous elastic theory, local deformations are described by the displacement vector field, ${\bf u}({\bf r},t)$. We will be interested in the effect of transverse sound waves. Such waves create shear deformations of the crystal lattice, such that $$\label{udot}
\nabla\cdot{\bf u}=0.$$ In the long-wave limit they do not affect the density of the ionic lattice but result in a local rotation at an angular velocity [@LL] $$\label{omega}
{\bf \Omega}({\bf r},t)=\frac{1}{2}\nabla\times\dot{\bf u}\,.$$ The frequency of ultrasound achievable in experiment with, e.g., surface acoustic waves can easily be in the ballpark of $f\sim10^{10}$s$^{-1}$ [@Santos]. According to Eq.(\[omega\]) a sound of such frequency and amplitude of a few nanometers can provide $\Omega \sim 10^9$s$^{-1}$ that can generate fictitious magnetic fields above ${\bf H}_{c1}$. For practical purposes, it may be convenient to loosen the restriction on the frequency and amplitude of ultrasound by applying an external magnetic field ${\bf H}$ near, but less than, ${\bf H}_{c1}$. We shall see that within one percent of ${\bf H}_{c1}$, vortices can be ignited by the ultrasound in the GHz frequency range.
For a vortex to enter a superconductor, the Gibbs free energy of the system must be lowered. We compute the extra free energy due to the vortex and determine the condition at which it becomes negative. It should be noted that the system under consideration is dynamical, and therefore is not at a thermodynamical equilibrium. However, we are interested in the free energy of the Cooper pairs which can adjust to the changes of state in a time scale orders of magnitude shorter than the period of the sound. This time scale is proportional to the relaxation time $\tau$ of the cooper pairs, i.e. $\tau\sim10^{-12}$s. As mentioned before, the period of the sound $T=1/f$ will be always greater than $10^{-10}$s. Under these conditions, our system is adiabatic and the thermodynamic equilibrium can be safely established. The calculation that follows is similar to the conventional calculation of ${\bf H}_{c1}$. The presence of ${\bf \Omega}$, however, introduces a new feature into this calculation so we will follow it all the way through to show how the sound enters the problem.
It is convenient to calculate the extra free energy in terms of the magnetic field and its spacial derivatives. The electric field produced by the time derivatives will be neglected. The kinetic energy of the superfluid is $\frac{1}{2}n_sm{\bf v}^2$ where $n_s$ is the number density of the superconducting electrons and $$\label{SFvel}
{\bf
v}=\frac{e^{\ast}}{m^{\ast}c}\left(\frac{\hbar}{e^{\ast}}\nabla\varphi-{\bf
A}\right)$$ is the velocity of the cooper pairs with $\varphi$ and ${\bf A}$ being the phase of the superfluid wavefunction and the magnetic vector potential, respectively. The stared quantities represent the effective mass and charge of cooper pairs. We will take them to be $m^{\ast}=2m$ and $e^{\ast}=2e$. The normal electrons experience viscous forces as they move relative to the nuclei contributing zero average normal current. The ionic charge per unit volume consisting of the nuclei and the normal electrons is therefore exactly opposite to that of the cooper pairs. The total current is then $$\label{current}
{\bf j}= en_s({\bf v}-\dot{\bf u}),$$ where $\dot{\bf u}$ is the velocity of ions. Eq. (\[current\]) reflects the fact that the electric current corresponds to the motion of electrons relative to ions. It is invariant with respect to the motion of the reference frame. With Eq. (\[SFvel\]) in mind we can write the gauge invariant current in terms of $\varphi$ and ${\bf u}$ as $$\label{GagInvJ}
{\bf j}=\frac{n_se\hbar}{2m}\left(\nabla\varphi-\frac{2e}{\hbar}{\bf
A}_{eff}\right),$$ where $$\label{Aeff}
{\bf A}_{eff}={\bf A}+\frac{mc}{e}\dot{\bf u}$$ is the effective vector potential felt by the electrons in the rotating frame of the ions [@EC-Kuklov-PRL]. In terms of the total current ${\bf j}$, the kinetic energy of the superconducting electrons may be expressed in the form $$\label{SCKE}
KE_e=\int d^3r\frac{n_sm}{2}\left(\frac{1}{n_se}{\bf j}+\dot{\bf
u}\right)^2.$$ The energy of the sound is $$\label{EneSound}
E_s=\int d^3r\frac{1}{2}(\rho_0\dot{\bf
u}^2-\lambda_{iklm}u_{ik}u_{lm})$$ in which $\rho_0$ is the combined mass density of ions and normal electrons, $\lambda_{iklm}$ is the tensor of elastic coefficients and $u_{ik}=\frac{1}{2} (\partial_iu_k+\partial_ku_i)$ is the strain tensor. Using Maxwell’s equation $\nabla\times{\bf B}=(4\pi/c){\bf
j}$ and combining Eqs. (\[SCKE\]) and (\[EneSound\]) the expression for the total Gibbs free energy yields $$\begin{aligned}
\label{freeEne}
G&=&{\cal F}_0+\frac{1}{8\pi}\int d^3r\left[{\bf
B}^2+\frac{\lambda^2}{f(r)}(\nabla\times{\bf B})^2\right]\nonumber\\
&+&\frac{1}{4\pi}\int d^3r\frac{mc}{e}\dot{\bf
u}\cdot(\nabla\times{\bf B})-
\frac{1}{4\pi}\int d^3r{\bf H}\cdot{\bf B}\nonumber\\
&+&\int d^3r\frac{1}{2}\left[\rho\dot{\bf
u}^2-\lambda_{iklm}u_{ik}u_{lm}\right],\end{aligned}$$ Here, ${\cal F}_0$ is the free energy in the absence of currents, fields, and sound, $\lambda=\sqrt{mc^2/4\pi n_se^2}$ is the London penetration depth, $f(r)=(|\psi|/|\psi_\infty|)^2$ in which $\psi$ is the complex order parameter and $|\psi_\infty|=\sqrt{n_s/2}$ is the order parameter in the absence of gradients and fields, and $\rho = \rho_0+n_sm$ is the total mass density of the superconductor. The fourth term can be recognized as the interaction of the external magnetic field with the magnetization. It is this term that is responsible for the nucleation of vortices in the absence of sound when $H\geq H_{c1}$.
Before we can calculate the free energy of Eq. (\[freeEne\]) we must first work out the magnetic field. This can be done by replacing the current in the Maxwell’s equation $\nabla\times{\bf
B}=(4\pi/c){\bf j}$ with Eq. (\[GagInvJ\]) and defining a gauge invariant vector potential ${\bf Q}={\bf A}-(\hbar
c/2e)\nabla\varphi$, so that we obtain the following equation: $$\label{Q}
\lambda^2\nabla\times(\nabla\times{\bf Q})+f(r){\bf
Q}=-\frac{mc}{e}f(r)\dot{\bf u}.$$ For $\nabla\varphi=0$ (${\bf Q}={\bf A}$) Eq. (\[Q\]) becomes equivalent to the London’s equation with a source. When a vortex enters a superconductor the phase must be quantized according to the condition $\oint\nabla\varphi\cdot d{\bf l}=2\pi$. For certainty we consider a transverse standing sound wave having one node at the center of a superconducting slab of thickness $d$ large compared to the coherence length $\xi$. The external field is applied parallel to the slab, see Fig. 1. In this case $\lambda_s=2d$. Generalization to standing waves with many nodes is straightforward. If the field is close to ${\bf H}_{c1}$, a vortex will periodically enter and exit the slab. The boundary condition on the current is ${\bf
J}_\perp \cdot {\bf n}=0$, where ${\bf n}$ is the direction of the surface. If the thickness of the slab $d$ is of order or less then $\lambda$, this boundary will distort the cylindrical symmetry of the vortex. We can satisfy the boundary condition by placing image vortices of alternating sign a distance $d$ apart on the outside of the slab. The equation for the magnetic field, in the region $r>\xi$ where $|\psi|=1$, can then be written in two parts, namely ${\bf
B}={\bf B}_0+{\bf B}_v$, such that the first term satisfies $$\label{B0}
\lambda^2\nabla\times(\nabla\times{\bf B}_0)+{\bf
B}_0=-\frac{2mc}{e}{\bf \Omega},$$ while the second is a solution of $$\label{Bv}
\lambda^2\nabla\times(\nabla\times{\bf B}_v)+{\bf B}_v=\Phi_0{\bf
e}_z\sum_{n=-\infty}^{\infty}(-1)^n\delta({\bf r}+nd{\bf e}_x),$$ where $\Phi_0=hc/2e$ is the flux quantum. Notice that Eqs. (\[B0\]) and (\[Bv\]) can be obtained by taking a curl of Eq. (\[Q\]) with the account of the vortex cores represented by the delta functions.
Since we are interested in standing sound waves we can choose the displacement vector ${\bf u}$ to be $$\label{u}
{\bf u}({\bf r},t)=u_0\sin(kx)\sin(\omega t){\bf e}_y.$$ The quantity $k=\omega/v=2\pi/\lambda_s=\pi /d$ is the wave number with $\lambda_s$ and $v$ being the wavelength and the speed of sound respectively. It is easy to see from Eq. (\[omega\]) that $\Omega$ is maximum at the nodes.
(18,5.5)
The corresponding solutions to Eqs. (\[B0\]) and (\[Bv\]) with the boundary condition ${\bf B} = {\bf H}$ at $x=\pm d/2$ are $$\begin{aligned}
{\bf B}_0(x) & = & {\bf B}_M + {\bf B}_s \label{B0Sol} \\ {\bf
B}_v({\bf r}) & = & \sum_{n=-\infty}^{\infty}(-1)^n{\bf b}({\bf
r}_n) \label{BvSol}\end{aligned}$$ where $$\begin{aligned}
& &{\bf B}_M = 2{\bf H}\,\frac{\sinh(d/2\lambda)}
{\sinh(d/\lambda)}\cosh\left(\frac{x}{\lambda}\right) \label{BM}\\
& &{\bf B}_s = -\frac{2mc}{e}\frac{\bf \Omega}{1+k^2\lambda^2}
\label{Bs} \\
& &{\bf b}({\bf r}_n)=\frac{\Phi_0}{2\pi\lambda^2}K_0(|{\bf
r}+nd{\bf e}_x|/\lambda){\bf e}_z\label{b}\,,\end{aligned}$$ and $K_0$ is a zeroth-order Hankel function of imaginary argument. The first term in Eq. (\[B0Sol\]) is the Meissner field while the second is due to the sound.
Let us now integrate by parts the third term in Eq. (\[freeEne\]) and insert ${\bf B}={\bf B}_0+{\bf B}_v$. By doing so we obtain $$\begin{aligned}
\label{freeEne3}
& & G = G_0+\Delta E+\frac{1}{4\pi}\int dr^3{\bf B}_v\cdot\left[
\frac{2mc}{e}{\bf \Omega}+{\bf B}_s\right]\nonumber\\
& &+ \frac{1}{4\pi}\int d^3r\frac{\lambda^2}{f(r)}(\nabla\times{\bf
B}_v)\cdot(\nabla\times{\bf B}_s)\end{aligned}$$ where $G_0$ is the Gibbs free energy without a vortex and $$\begin{aligned}
\label{DeltaE}
\Delta E&=&\frac{1}{4\pi}\int
d^3r\frac{\lambda^2}{f(r)}(\nabla\times{\bf
B}_M)\cdot(\nabla\times{\bf
B}_v)\nonumber\\
&+&\frac{1}{8\pi}\int d^3r\left[{\bf
B}_v^2+\frac{\lambda^2}{f(r)}(\nabla\times{\bf
B}_v)^2\right]\nonumber\\
&-&\frac{1}{4\pi}\int dr^3{\bf H}\cdot[{\bf B}_v+{\bf B}_M]\end{aligned}$$ is the vortex energy. One can simplify the volume integrals in Eq. (\[freeEne3\]) by separating the integration over the core from the integration over the volume outside of the core. When the latter is integrated by parts, the integrals outside the core cancel, and the free energy in Eq. (\[freeEne3\]) with the help of Eq. (\[B0\]) becomes $$\Delta{\cal F}=\Delta{\cal F}_1+\Delta{\cal F}_2+\Delta{\cal
F}_3+\Delta E\,,$$ where $$\begin{aligned}
& &\Delta{\cal F}_1=\frac{1}{4\pi}\int_c dr^3 {\bf
B}_v\cdot\left[\frac{2mc}{e}{\bf \Omega}+{\bf
B}_s\right]\label{DeltaF1}\\
& &\Delta{\cal F}_2=\frac{\lambda^2}{4\pi}\oint_c{\bf
B}_v\times(\nabla\times{\bf B}_s)\cdot d{\bf s} \label{DeltaF2} \\
& &\Delta{\cal F}_3=\frac{1}{4\pi}\int_c
d^3r\frac{\lambda^2}{f(r)}(\nabla\times{\bf
B}_v)\cdot(\nabla\times{\bf
B}_s).\label{DeltaF3} \nonumber\\\end{aligned}$$ The subscript $c$ indicates an integration over the core. The surface integral in Eq. (\[DeltaF2\]) is over the boundary of the normal core. Near the vortex core $f(r)=(r/a)^2$, where $a\approx\xi$. It is straightforward to check that in the limit $r\rightarrow 0$ the exact solution to Eq. (\[Q\]) for the vector potential ${\bf A}_s({\bf r})$ is $$\label{As}
{\bf A}_s({\bf r})=-\frac{2mc}{e}\Omega_0x{\bf e}_y\,.$$ Then the magnetic field ${\bf B}_s=\nabla\times{\bf A}_s({\bf r})$ generated by the sound at the center of the core is $$\label{Bs2}
{\bf B}_s({\bf r})=-\frac{2mc}{e}\Omega_0{\bf e}_z,$$ where $$\label{DeltaNout}
\Omega_0=\frac{1}{2}u_0k\omega=\frac{\pi}{2}\frac{u_0}{d}\omega.$$ It can be shown that near the vortex core, $\nabla\times B_v\propto
r^4$ and $\nabla\times B_s\propto r^5$. The expression under the integral in Eq. (\[DeltaF3\]) is therefore proportional to $r^8$ near the center of the core and to $rK_1(r/\lambda)$ at $r\gtrsim
\xi$. Thus, the integral in Eq. (\[DeltaF3\]) falls off very rapidly inside the core and can be neglected.
The case of $k\lambda \geq 1$ is rather involved as it requires explicit knowledge of the structure of the vortex core. For $k\lambda\ll1$ Eq. (\[Bs\]) provides that ${\bf
B}_s\cong-(2mc/e){\bf \Omega}$ in all regions of space, so that $\Delta{\cal F}_1\rightarrow 0$. In this limit the Meissner field ${\bf B}_{M}$ and the fields due to images can be neglected. The total interaction energy per unit length of the vortex acquires the simplest form at $\kappa=\lambda/\xi\gg1$: $$\label{DelF2Solution}
\frac{\Delta{\cal F}_2}{L} = -\frac{mc}{2\pi}\Omega_0\Phi_0
\left(\frac{k\lambda}{\kappa}\right)^2\ln\kappa\,,$$ where $L$ is the dimension of the slab in the z-direction.
If one excludes small contribution from the vortex core in Eq. (\[DeltaE\]), then the integration by parts yields $$\label{DeltaE2}
\frac{\Delta E}{L}=\frac{\lambda^2}{8\pi}\oint {\bf
B}_v\times(\nabla\times{\bf B}_v)\cdot d{\bf s} -\frac{1}{4\pi}\int
d^3r{\bf H}\cdot{\bf B}_v\,.$$\
This approximation is good if $\lambda$ and $d$ are large compared to the coherence length $\xi$. Then the vortex energy per unit length is $$\label{DeltaESol}
\frac{\Delta E}{L} = \frac{\Phi_0^2}{({4\pi\lambda})^2}\ln\kappa -
\frac{\Phi_0H}{4\pi}\,.$$ The first term in this expression is the self-energy of the vortex, while the second term is the energy of the interaction of the flux quantum with the external field.
The condition for the nucleation of the vortex, $\Delta{\cal F}_2 +
\Delta E = 0$, yields $$\label{ConFin}
\frac{2mc}{e}\Omega_0\left(\frac{k\lambda}{\kappa}\right)^2\ln\kappa=
{\epsilon}H_{c1}\,,$$ where $$\epsilon = 1 - \frac{H}{H_{c1}}$$ and $H_{c1}=\Phi_0\ln \kappa/(4\pi\lambda^2)$ is the first critical field that follows from Eq. (\[DeltaESol\]) at $\Delta E = 0$. Substituting Eq. (\[DeltaNout\]) into Eq. (\[ConFin\]), one finds the conditions on the frequency $f$ and amplitude $u_0$ of the sound needed to nucleate a vortex in the geometry shown in Fig. \[fig1\]: $$\label{ConFin2}
f = \frac{v}{2d}\,, \qquad u_0=\frac{\epsilon}{4} \left(\frac{d}{\pi
\lambda}\right)^4 \frac{\hbar\kappa^2}{mv}\,.$$ While the last formula was derived under the conditions $\pi\xi <
\pi\lambda \ll d$, our numerical analysis shows that it holds even for $d \sim \pi\lambda$ at $\kappa \gg 1$ and is true by order of magnitude for $\kappa \sim 1$. In this case the expression for $H_{c1}$ carries the signature of the surface barrier [@Bean]: $H_{c1}=\beta \Phi_0\ln \kappa/(4\pi\lambda^2)$, where $$\beta = \frac{1-{2({\ln \kappa})^{-1}\sum_1^\infty (-1)^n
K_0(dn/\lambda)}}{1 -
2{\sinh(d/2\lambda)}[{\sinh(d/\lambda)}]^{-1}}\,.$$
For the speed of the transverse sound $v \sim 3\times 10^5$cm/s, in a slab of thickness $d \sim \pi\lambda \sim 6\times 10^{-5}$cm and $\kappa \sim 2$, with $H$ within one percent of $H_{c1}$, one gets from Eq. (\[ConFin2\]) $f \sim 3$GHz and $u_0 \sim 0.2$nm. These are accessible values of frequency and amplitude of ultrasound.
As ${\bf \Omega}$ changes its sign every half a period of the sound, vortices are periodically attracted and repelled by the standing acoustic wave in Fig. \[fig1\]. Periodic entering and expulsion of vortices should result in the elevated attenuation of the ultrasound and in the ac voltage across the slab at the sound frequency. In a different experiment one can assist vortices to enter or exit the superconductor with the help of the surface acoustic waves (SAW). Like in the problem with a slab, local rotation of the crystal produced by the SAW may assist nucleation of the vortex at the field just below $H_{c1}$.
In conclusion, we have demonstrated that nucleation of a vortex in a superconductor can be assisted by ultrasound. In the presence of a standing sound wave, vortices will periodically enter and exit the superconductor. The required amplitude and frequency of ultrasound are within experimental reach.
We thank Lev Bulaevskii and Carlos Calero for useful discussions. This work has been supported by the Department of Energy through Grant No. DE-FG02-93ER45487.
[10]{}
F. London, [*Superfluids*]{} (Wiley, New York, 1950), Vol. 1.
, Phys. Rev. [**6**]{}, 239 (1915).
J. Larmor, Lond. Math. Soc. Proc. [**1**]{}, 1 (1903).
W. Meissner and R. Ochsenfeld, Naturwiss. [**21**]{}, 787 (1933).
H. Haneda and T. Ishiguro, Physica [**C 235-240**]{}, 2076 (1994).
D. Dominguez, L. Bulaevskii, B. Ivlev, M. Maley, and A. R. Bishop, Phys. Rev. Lett. [**74**]{}, 2579 (1995); Phys. Rev. B [**51**]{}, 15649 (1995); Phys. Rev. B [**53**]{}, 6682 (1996).
E. B. Sonin, Phys. Rev. Lett. [**76**]{}, 2794 (1996).
B. I. Ivlev, S. Mejia-Rosales, and M. N. Kunchur, Phys. Rev. B [**60**]{}, 12419 (1999).
L. N. Bulaevskii and E. M. Chudnovsky, Phys. Rev. B [**72**]{}, 094518 (2005).
M. W. Coffey, Phys. Rev. B [**49**]{}, 9774 (1994).
J.-M. Duan and E. Simanek, Phys. Lett. [**A 190**]{}, 118 (1994).
E. M. Chudnovsky and A. B. Kuklov, Phys. Rev. Lett. [**91**]{}, 067004 (2003).
E. M. Chudnovsky and A. B. Kuklov, Phys. Rev. B [**67**]{} 064515 (2003).
J. Albert and E. M. Chudnovsky, Phys. Rev. B [**75**]{}, 144502 (2007).
, [*Theory of Elasticity*]{} (Pergamon, New York, 1959).
, Rep. Prog. Phys. [**68**]{}, 1639 (2005).
, Phys. Rev. Lett. [**12**]{}, 1 (1963).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, the generalized Bloch Conjecture on zero cycles for the quotient of certain complete intersections with trivial canonical bundle is proved to hold.
As an application of Bloch-Srinivas method on the decomposition of the diagonal, we compute the rational coefficient Lawson homology for 1-cycles and codimension two cycles for these quotient varieties. The (Generalized) Hodge Conjecture is proved to hold for codimension two cycles (and hence also for 2-cycles) on these quotient varieties.
author:
- Wenchuan Hu
title: 'The Generalized Bloch Conjecture for the quotient of certain Calabi-Yau varieties'
---
\[section\] \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Corollary]{}
\[theorem\][Definition]{} \[theorem\][Example]{} \[theorem\][Exercise]{} \[theorem\][Conjecture]{}
\[theorem\][Remark]{}
Ł Ø ¶
Introduction
============
In this paper, all varieties are defined over $\C$. For a projective variety $X$, denote by $\cZ_p(X)$ the spaces of algebraic $p$-cycles and $\Ch_p(X)$ the Chow group of $p$-cycles on $X$, i.e, $\Ch_p(X)=\cZ_p(X)/{\hbox{\{rational equivalence\}}}$. Let $cl_p:\Ch_p(X)\to H_{2p}(X,\Z)$ be the cycle class map. Tensoring with ${{\Q}}$, we have $cl_p\otimes{\Q}:{\Ch}_p(X)\otimes{\Q}\to H_{2p}(X,\Q).$ Let $\Ch_p(X)_{hom}\subset \Ch_p(X)$ be the subgroup of $p$-cycles homologous to zero. Set $\Ch^q(X):=\Ch_{n-q}(X)$.
In 1968, D. Mumford showed that $\Ch_0(X)_{hom}$ is not finite dimensional for a smooth projective surface $X$ with non-vanishing geometric genus $p_g(X)$ (cf. [@Mumford]). This result was generalized by Roĭtman to arbitrary dimension (cf. [@Roitman]). In this situation, a nontrivial conjecture of Bloch asserts that if a smooth projective surface $X$ with $p_g(X)=0$, then $\Ch_0(X)_{hom}$ is finite dimensional ([@Bloch]). Equivalently, if $p_g(X)=0$, then there is a curve $C\subset X$ such that the natural map $\Ch_0(C)\to \Ch_0(X)$ is surjective.
This conjecture can be generalized as follows (cf. [@Paranjape-Srinivas]):
Let $X$ be a smooth projective variety satisfying $H^{i,0}(X) = 0$ for all $i>r$. Then there is a subvariety $i : Z \hookrightarrow X$, where $\dim Z = r$, such that $i_* : \Ch_0(Z) \to \Ch_0(X)$ is surjective.
Some examples are known in support of these conjectures. For example, Bloch’s conjecture is true for surfaces which are not of general type [@Bloch-K-L]. This conjecture also holds for some surfaces of general type which are quotients of some special surfaces by a free finite group action (cf. [@Inose-Mizukami], [@Voisin]). In higher dimensional case, it was proved by Roĭtman [@Roitman] that $\Ch_0(X)\cong \Z$ for smooth projective complete intersection with $H^{i,0}(X) = 0$ for all $i>0$. In [@Bloch-S], it was proved by Bloch and Srinivas that the Generalized Bloch Conjecture holds for Kummer varieties of odd dimensions.
In this paper, the Generalized Bloch Conjecture is proved to hold for quotients of certain even dimensional complete intersection with trivial canonical bundle by a free involution and for the resolution of singularities to the quotients of certain odd dimensional complete intersection with trivial canonical bundle by an involution with isolated fixed points. Based on these results, we compute the rational Lawson homology and verify the Generalized Hodge Conjecture on 1-cycles and codimension-2 cycles for these varieties.
I would like to express my gratitude to Michael Artin for helpful discussion and suggestion during the preparation of this paper.
Main results
============
Now let $X^n\subset\P^{2n+1}$ be the complete intersection of quadrics $$Q_i(z_0,z_1\cdots,z_n)+Q_i'(z_{n+1},z_{n+2}\cdots,z_{2n+1})=0, i=0,1,\cdots,n,$$ where $Q_i$ and $Q_i'$ are quadratic forms in $n+1$ variables. For our propose, we assume $$\label{eq1}
Q_i(z_0,z_1\cdots,z_n)=\sum_{j=0}^na_{ij}z_j^2$$ for $i=0,1,\cdots,n$. We also assume that $X^n$ is smooth, which holds for the generic choice of $Q_i'$ and the choice of $a_{ij}, i,j=0,1,\cdots, n$ such that $\det(a_{ij})\neq 0$. From the direct calculation we know $X^n$ is a Calabi-Yau $n$-fold, i.e., $K_{X^n}$ is trivial and hence $h^{n,0}(X^n):=\dim H^{n,0}(X^n)=1$.
For $n=2m$ a positive even integer, we define an involution of $\P^{2n+1}$ by $$\label{eq2}
\sigma:(z_0:z_1:\cdots:z_{2n+1})\mapsto (-z_0: -z_1:\cdots: -z_n, z_{n+1},\cdots, z_{2n+1})$$ which takes $X^n$ to itself. The quotient $Y^n=X^n/\langle \sigma\rangle$ is a smooth projective variety with $H^{i,0}(Y^n)=0$ for all $i\geq 1$ for $n=2m$ even (cf. Lemma \[lemma2.2\]). Denote by $\pi:X^n\to Y^n$ the projection.
Our first main result is following theorem.
\[Th1.1\] Let $n$ be a positive even integer. The Generalized Bloch Conjecture holds for $Y^n$, i.e., for the projective variety $Y^n=X^n/\langle \sigma\rangle$ above, we have $\Ch_0(Y^n)=\Z$.
If $n=2m-1$ is a positive odd integer, then we define another involution of $\P^{2n+1}$ by $$\label{eq3}
\rho:(z_0:z_1:\cdots:z_{2n+1})\mapsto (z_0: -z_1:\cdots: -z_n, z_{n+1},\cdots, z_{2n+1})$$ which takes $X^n$ to itself. The involution $\rho$ has $2^{2m}$ isolated fixed points. Hence the quotient $Y^{2m-1}=X^{2m-1}/\langle \rho\rangle$ is a projective variety with $2^{2m}$ isolated singular points, denote by $q_i, i=1,2,\cdots,2^{2m}$. Denote also by $\pi:X^{2m-1}\to Y^{2m-1}$ the projection. Each singular point is a cyclic quotient singular point. Let $\widetilde{Y^{2m-1}}\to Y^{2m-1}$ be a resolution of singularity, then the exceptional divisor $E_i$ at each singular point $q_i$ has only normal crossings in $\widetilde{Y^{2m-1}}$ and every irreducible component of $E_i$ is nonsingular and rational [@Fujiki].
Our second main result is the following theorem.
\[Th1.7\] Let $n=2m-1$ be a positive odd integer. The Generalized Bloch Conjecture holds for $\widetilde{Y^{n}}$, i.e., for $\widetilde{Y^{n}}$ above, we have $\Ch_0(\widetilde{Y^{2m-1}})\cong\Z$. Moreover, $\Ch_0(Y^{2m-1})\cong\Z$.
The application of the main results on algebraic cycles and Lawson homology is given in section \[sec4\].
The proof of main theorems {#sec3}
==========================
For the generic choice of $Q_i'$ and the choice of $a_{ij}, i,j=0,1,\cdots, n$ such that the determinant $\det(a_{ij})$ of the matrix $(a_{ij})$ is nonzero, then $X^{n}$ is a smooth projective variety of dimension $n$.
It follows from the definition of smoothness of projective variety. Note that $\det(a_{ij})\neq 0$ implies that there is no common solution for the system of equations by $Q_i=0$ and those of partial derivatives. Similarly for the generic choice of $Q_i'$, there is no common solution for the system of equations $Q_i', i=1,2,\cdots, n+1$.
\[lemma2.2\] For $n=2m$ a positive even integer, the quotient $Y^n=X^n/\langle \sigma\rangle$ by the involution $\sigma:X^n\to X^n$ is a smooth projective variety. Moreover, $H^{i,0}(Y^n)=0$ for all $i>0$.
The involution $\sigma:X^n\to X^n$ is induced by the involution of $\P^{2n+1}$ defined by $\sigma:(z_0:z_1:\cdots:z_{2n+1})\mapsto (-z_0: -z_1:\cdots: -z_n, z_{n+1},\cdots, z_{2n+1})$. By the assumption, the fixed point set of $\sigma:\P^{2n+1}\to \P^{2n+1}$ and $X^n$ have no intersection since the system of equations $Q_i=0$, $i=0,1,\cdots, n$ has no common solution in $\P^n$ by the assumption that $\det(a_{ij})\neq 0$. Similarly for a generic choice of $Q_i'$, $i=0,1,\cdots, n$. Therefore, $\sigma:\P^{2n+1}\to \P^{2n+1}$ induces a fixed point free involution on $X$ and hence the quotient $Y^n=X^n/\langle \sigma\rangle$ is a smooth projective variety. Note that $H^{i,0}(Y^n)=0$ for $i> n$ for the reason of dimension. For $i<n$, $\dim H^{i,0}(Y^n)\leq \dim H^{i,0}(X^n)$ and the latter is zero by Lefschetz hyperplane Theorem. For $i=n$, we have $2=\chi(\O_{X^n})=2 \chi (\O_{Y^n})=2(1-\dim H^{1,0}(X^n)+\cdots+(-1)^n\dim H^{n,0}(Y^n))=2(1+\dim H^{n,0}(Y^n))$, where the second equality holds since $X^n\to Y^n$ is a étale morphism (cf. Example 18.3.9 in [@Fulton]) and the last equality holds since $n$ is an even integer and so $\dim H^{n,0}(Y^n)=0$.
Since $H^{i,0}(Y^n)=0$ for all $i>0$, the Generalized Bloch Conjecture implies $\Ch_0(Y^n)\otimes\Q=\Q$. The statement in Theorem \[Th1.1\] is slightly stronger than this.
A well-known result is needed in our computation.
\[lemma2.3\] Suppose a finite group $G$ acts on a variety $X$ with nonsingular quotient variety $Y=X/G$. Let $\pi:X\to Y$ be the quotient map. Then there exist two homomorphisms $\pi_*:\Ch_0(X)_{hom}\to \Ch_0(Y)_{hom}$ and $\pi^*:\Ch_0(Y)_{hom}\to \Ch_0(X)_{hom}$ such that $$\label{eq4}
\left\{\begin{array}{lll}
\pi^*\pi_*&=& \sum_{g\in G} g_*\\
\pi_*\pi^*&=& N
\end{array}\right.$$ where $N=|G|$ means the multiplication by $N$ in $\Ch_0(X)_{hom}$. In particular, $ \Ch_0(Y)_{hom}=0$ if and only if $\sum_{g\in G} g_*=0$ in $End(\Ch_0(Y)_{hom})$.
See, e.g. [@Inose-Mizukami], Lemma 1 and Lemma 2, where $X$ is a surface. The point is that both $\pi_*$ and $\pi^*$ are well-defined. The proof works in higher dimensional case and the case that $Y$ is singular (cf. [@Fulton]).
Now we can apply Lemma \[lemma2.3\] to the quotient map $\pi:X^n\to Y^n$. Let $\pi_*:\Ch_0(X^n)\to\Ch_0(Y^n)$ be the push forward map and let $\pi^*:\Ch_0(Y^n)\to\Ch_0(X^n)$ be the pull back. Then we have $$\pi_*\pi^*=2:\Ch_0(Y)_{hom}\to \Ch_0(Y)_{hom}$$ and $$\pi^*\pi_*=\sigma_*+1:\Ch_0(X)_{hom}\to \Ch_0(X)_{hom}$$
Since $\Ch_0(Y^n)_{hom}$ is divisible (cf. [@Bloch2], [@Roitman]), it suffices to show $$2\Ch_0(Y)_{hom} = \pi_*\pi^*\Ch_0(Y)_{hom} = 0.$$ Since $\pi_*$ is surjective, it suffices to show $\pi^*\pi_*:\Ch_0(X^n)_{hom}\to \Ch_0(X^n)_{hom}$ is the zero map. That is, we need to show that $\sigma_*=-1:\Ch_0(X^n)_{hom}\to \Ch_0(X^n)_{hom}$.
Therefore, Theorem \[Th1.1\] follows from the following proposition.
\[prop2.1\] Let $\sigma:X^n\to X^n$ be induced by the involution in Equation (\[eq2\]). Then $\sigma_*=-1:\Ch_0(X^n)_{hom}\to \Ch_0(X^n)_{hom}$.
To prove this proposition, we need some auxiliary results. Let $\tau_i:X^n\to X^n$ be the automorphism of $X^n$ induced by $\sigma_i:\P^{2n+1}\to \P^{2n+1}$, where $$\sigma_i:(z_0:\cdots: z_{i-1}:z_i:z_{i+1}\cdots:z_{2n+1})
\mapsto (z_0:\cdots: z_{i-1}:-z_i:z_{i+1}:\cdots: z_{2n+1})$$ for $i=0,1,\cdots, n$. Note that $\tau_i$ maps $X^n$ to itself by the assumption of $Q_i$. Then $\tau_i:X^n\to X^n$ induces a homomorphism on Chow group of 0-cycles $\tau_{i*}:\Ch_0(X^n)\to \Ch_0(X^n)$ and so $\tau_{i*}:\Ch_0(X^n)_{hom}\to \Ch_0(X^n)_{hom}$. Then we have the following result.
\[lemma2.6\] The homomorphism $\tau_{i*}=-1:\Ch_0(X^n)_{hom}\to \Ch_0(X^n)_{hom}$ for all $i=0,1,\cdots, n$.
By the symmetry of $t_0,t_1,\cdots,t_n$, we only need to show the case for $i=0$. From the definition of $X^n$ and the assumption that the matrix $(a_{ij})$ is non-degenerated, we can make a linear transformation such that there is only one quadratic, say $Q_1+Q_1'$, depending on the variable $t_0$. Then one can see that the function field of $X^n/\langle\tau_1\rangle$ is $$\C(t_1,t_2,\cdots,t_{2n},t_{2n+1})/\langle Q_i+Q_i'=0, i=2,3,\cdots,n+1\rangle.$$
Note that the variety $Y_0\subset \P^{2n}$ defined by equations $Q_i+Q_i'=0,i=2,3,\cdots,n+1$ is a smooth complete intersection. Since the sum of the degrees of the defining equations of $Y_0$ is $\sum_{i=1}^n 2=2n$, $Y_0$ is a smooth Fano variety. This implies $Y_0$ is rationally connected (cf. [@Campana], [@Kollar-Miyaoka-Mori]). Therefore, for any two generic points $p_1,p_2$ on $Y_0$, there is a rational curve $C$ passing through $p,q$.
From the definition, the rational function field of $Y_0$ is also $$\C(t_1,t_2,\cdots,t_{2n},t_{2n+1})/\langle Q_i+Q_i'=0, i=2,3,\cdots,n+1\rangle.$$ So $Y_0$ is birational equivalent to $X^n/\langle\tau_1\rangle$. Hence, for any generic two points on $X^n/\langle\tau_1\rangle$, there also exists a curve passing through the two points, i.e., $X^n/\langle\tau_1\rangle$ is a rationally connected variety. So $\Ch_0 (X^n/\langle\tau_1\rangle) \cong
\Z$ and $\Ch_0(X^n/\langle\tau_1\rangle)_{hom} = 0$.
This together with Lemma \[lemma2.3\] implies that $\tau_{1*}+1=0 \in End(\Ch_0(X^n))_{hom}$. This completes the proof of the Lemma.
Note that from the definition we have $\sigma=\tau_0\circ \tau_1\circ\cdots\circ\tau_n$ and so $\sigma_*=\tau_{0*}\circ \tau_{1*}\circ\cdots\circ\tau_{n*} = (-1)^{n+1} = -1$ since $n=2m$ is an even integer. This completes the proof of Proposition \[prop2.1\] and hence Theorem \[Th1.1\].
In the following, we focus on the proof of Theorem \[Th1.7\]. Note that $n=2m-1$ in the below of this section.
For the generic choice of $Q_i'$ and $(a_{ij})$, the variety $Y^{2m-1}=X^{2m-1}/\langle\rho\rangle$ has exact $2^{2m}$ isolated singular points.
Note that the set of singular points on $Y^{2m-1}$ is exact the set of fixed points of the involution $\rho: X^{2m-1}\to X^{2m-1}$. This fixed points set is defined by equations $$\label{eq5}
t_1=t_2=\cdots=t_n=0 \quad and \quad Q_i+Q_i'=0, ~ i=0,1,\cdots, n,$$ i.e., the intersection of $\P^{n+1}$ and $X^n$ in $\P^{2n+1}$. Note that the degree of $X^n$ is $2^{n+1}=2^{2m}$ and for a generic choice of $Q_i'$, equation (\[eq5\]) has no solution of multiplicity bigger than 1.
By Lemma \[lemma2.6\], the map $\rho: X^{2m-1}\to X^{2m-1}$ induces the push forward map $\rho_*=(-1)^{n}=(-1)^{2m-1}=-1: \Ch_0(X^{2m-1})_{hom}\to \Ch_0(X^{2m-1})_{hom}$. By Lemma \[lemma2.3\], $\pi^*\pi_*=\rho_*+1=0:\Ch_0(Y^{2m-1})_{hom}\to \Ch_0(Y^{2m-1})_{hom}$ and $\Ch_0(Y^{2m-1})_{hom}=0$.
Let $\phi:\widetilde{Y^{2m-1}}\to Y^{2m-1}$ be a resolution of singularity. Then $\Ch_0(\widetilde{Y^{2m-1})}\cong \Z$.
For each singular point $q_i\in Sing(Y^{2m-1})$, $i=1,2,\cdots, 2^{2m}$, the exceptional divisor $E_i=\phi^{-1}(q_i)$ has normal crossings in $\widetilde{Y^{2m-1}}$ and every irreducible component of $E_i$ is nonsingular and rational (cf. Corollary after Theorem 1 in [@Fujiki]). Since each singular point in our case is a quotient singularity of type $\C^{2m-1}/\Z_2$, the exceptional divisor $E_i$ contains exactly one irreducible component, which is isomorphic to $\P^{2m-2}$ (cf. Remark \[remark2.9\]).
Set $E=\bigcup_{i=0}^{2^{2m}} E_i$. Since $\Ch_0(E_i)_{hom}=0$ and that $E_i$ are mutually disjoint to each other, we get $\Ch_0(E)_{hom}=0$. Set $U=\widetilde{Y^{2m-1}}-E\cong Y^{2m-1}-\bigcup_{i=0}^{2^{2m}} q_i$. Then the isomorphism $\Ch_0(\widetilde{Y^{2m-1})}\cong \Z$ follows from the fact that $\Ch_0(\widetilde{Y^{2m-1})}_{hom}=0$. This fact can be seen from the commutative diagram of Chow groups $$\xymatrix{\Ch_0(E)_{hom}\ar[r]\ar[d]^{\cong}& \Ch_0(\widetilde{Y^{2m-1}})_{hom}\ar[r]\ar[d]^{\phi_*}&\Ch_0(U)_{hom}\ar[r] \ar[d]^{\cong}&0\\
\Ch_0(\bigcup_{i=0}^{2^{2m}} q_i)_{hom}\ar[r]&\Ch_0({Y^{2m-1}})_{hom}\ar[r] & \Ch_0(U)_{hom}\ar[r]&0.
}$$
\[remark2.9\] Each singular point of $Y^{2m-1}$ is the quotient singularity of the same type as that of $\C^{2m-1}/\Z_2$, where $\Z_2$ acts on $\C^{2m-1}$ as $$(x_1,x_2,\cdots,x_{2m-1}) \to (-x_1,-x_2,\cdots,-x_{2m-1}).$$ The singular point of $\C^n/\Z_2$ can be resolved by one blow up with the exceptional divisor $E\cong \P^{n-1}$.
To see this, we first note that all the $\Z_2$-invariant monomials of $x_1$, $x_2$, $\cdots$, $x_{n}$ are $x_ix_j$, $1\leq i\leq j\leq n$. This gives an embedding $X:=\C^n/\Z_2\hookrightarrow \C^N$, where $N=(^{n+1}_{~2})=\frac{1}{2}n(n+1)$. Let $u_{ij}$, $1\leq i\leq j\leq n$ be the coordinates of $\C^N$. Then $\C^n/\Z_2$ is the locus of the ideal generated by all $2\times 2$ minors of the symmetric matrix $(u_{ij})_{1\leq i,j\leq n}$, where $u_{ij}:=u_{ji}$ if $i>j$.
Let $\widetilde{\C^N}$ be the blow up of $\C^N$ at the origin and let $\widetilde{X}$ be the proper transform of $X=\C^n/\Z_2$. A direct calculation shows that $\widetilde{X}$ is smooth. The explicit equations for $n=3$ will be given below while the general case is similar. The exceptional divisor $E$ of $\widetilde{X}\to X$ is just the quadric equation given by those $2\times 2$ minors in $\P^{N-1}$, i.e, the intersection of $\P^{N-1}$ and $\widetilde{X}$. Note that $E\subset \P^{N-1}$ with the above defining equations is exactly the image of the Plücker embedding $\P^{n-1}\hookrightarrow \P^{N-1}$. Therefore, $E\cong \P^{n-1}$.
Now we write down the details for the case that $n=3$. In this case $N=(^4_2)=6$. Let $\widetilde{\C^6}\subset \C^6(u_1,\cdots,u_6)\times \P^5[v_1:\cdots:v_6]$ be defined by $u_iv_j=u_jv_i$, $1\leq i\neq j\leq 6$. Note that $X=\C^3/\Z_2\subset \C^6$ is defined by $$\left\{\begin{array}{lll}
u_1u_2=u_6^2,\\
u_1u_3=u_5^2,\\
u_2u_3=u_4^2.
\end{array}\right.$$
Note that $\widetilde{\C^6}$ is covered by affine open sets $(v_i\neq 0)$, $i=1,2,\cdots,
6$. On the affine open piece $v_1\neq 0$, $\widetilde{X}$ is defined by $$\left\{\begin{array}{lll}
u_i=u_1v_i, i=2,3,\cdots, 6,\\
v_2=v_6^2,\\
v_3=v_5^2,\\
v_2v_3=v_4^2.
\end{array}\right.$$
It is easy to check by the definition of smoothness that this piece of $\widetilde{X}$ is smooth. Similarly for all other pieces of $\widetilde{X}$. Therefore $\widetilde{X}$ is smooth.
The exceptional divisor $E$ is defined by the following equations: $$\left\{\begin{array}{lll}
u_i=0, i=1,2,\cdots, 6,\\
v_1v_2=v_6^2,\\
v_2v_3=v_5^2,\\
v_2v_3=v_4^2.
\end{array}\right.$$
Hence $E$ is isomorphic to the image of Plücker embedding $\P^2\hookrightarrow \P^5$ and therefore $E\cong \P^2$.
Application to 1-cycles and codimension two cycles {#sec4}
==================================================
In this section, we deduce a sequence of results on algebraic cycles and cohomology theories for $Y^n$ as the application of the decomposition of the diagonal given by Bloch [@Bloch], Bloch and Srinivas [@Bloch-S] and the generalization by many others.
First we consider the case that $n=2m$ is an even positive integer.
\[cor1.1\] $\Ch^p(Y^{2m})$ is weakly representable for $p\leq 2$.
It follows from Theorem \[Th1.1\] and Theorem 1 in [@Bloch-S].
The [**Hodge Conjecture**]{} for codimension $p$ cycles(denote by $\HC^{q,q}(X)$): [*The rational cycle class map $$cl^q\otimes{\mathbb{Q}}:{\Ch}^q(X)\otimes{\mathbb{Q}}\rightarrow H^{q,q}(X)\cap
H^{2q}(X,\mathbb{Q})$$ is surjective.* ]{}
More generally, let $N^pH^k(X,\Q) \subset H^k(X,\Q)$ be the arithmetic filtration defined by Grothendieck [@Grothendieck] and let $F^pH^k(X,\C) \subset H^k(X,\C)$ be the Hodge filtration. Set $F^pH^k(X,\Q) : = F^pH^k(X,\C)\cap H^k(X,\Q)$ and denote by $\widetilde{F}^pH^k(X,\Q)$ the maximal sub-Hodge structure in $F^pH^k(X,\Q)$. It was shown in [@Grothendieck] that $N^pH^k(X,\Q)\subset\widetilde{F}^pH^k(X,\Q)$.
The [**generalized Hodge Conjecture**]{} can be stated as follows (denote by $\GHC(p,k,X)$): $$N^pH^k(X,\Q)=\widetilde{F}^pH^k(X,\Q).$$
\[cor1.2\] The Hodge Conjecture for $Y^4$ holds. The generalized Hodge Conjecture for $\GHC(1,4,Y^{4})$ holds. More generally, the generalized Hodge Conjecture $\GHC(1,2m,Y^{2m})$ for $Y^{2m}$ holds.
The first statement follows from Theorem \[Th1.1\] and Theorem 1 in [@Bloch-S]. Similar method can be used to prove $\GHC(1,4,Y^{4})$ and more general statement $\GHC(1,2m,Y^{2m})$. By Theorem \[Th1.1\], $
\Ch_0(Y^{2m})\cong \Z$, we have $\GHC(1,2m,Y^{2m})$ by Corollary 15.23 in [@Lewis] or Proposition 5.5 in [@Voineagu].
$\GHC(1,4,Y^{4})$ is the only non-trivial part of the generalized Hodge Conjecture for $Y^4$. The Hodge Conjecture for 2-cycles and codimension 2 cycles on $Y^{2m}$ holds, i.e., both $\HC^{2m-2,2m-2}(Y^{2m})$ and $\HC^{2,2}(Y^{2m})$ hold for all positive integer $m$. However, both $\HC^{2m-2,2m-2}(Y^{2m})$ and $\HC^{2,2}(Y^{2m})$ are trivial if $m>2$ since both $H^{4}(Y^{2m})$ and $ H^{4m-4}(Y^{2m})$ are isomorphic to $\Z$.
Recall that the **Lawson homology** $L_pH_k(X)$ of $p$-cycles is defined by $$L_pH_k(X) := \pi_{k-2p}({\cZ}_p(X)) \quad {\rm for}\quad k\geq 2p\geq 0,$$ where ${\mathcal Z}_p(X)$ is provided with a natural topology (cf. [@Friedlander1], [@Lawson1] and [@Lawson2]). For general background on Lawson homology, the reader is referred to [@Lawson2]. There are natural maps, called **cycle class maps** $ \Phi_{p,k}:L_pH_{k}(X)\rightarrow H_{k}(X). $
Define $$\begin{array}{lcl}
L_pH_{k}(X)_{hom}&:=&{\rm ker}\{\Phi_{p,k}:L_pH_{k}(X)\rightarrow H_{k}(X)\};\\
L_pH_{k}(X,\Q)_{hom}&:=& L_pH_{k}(X)_{hom}\otimes{\Q};\\
T_pH_{k}(X)&:=&{\rm Image}\{\Phi_{p,k}:L_pH_{k}(X)\rightarrow H_{k}(X)\}; \\
T_pH_{k}(X,\Q)&:=&T_pH_{k}(X)\otimes\Q.
\end{array}$$
The **Griffiths group** of $p$-cycles is defined to $${\Griff}_p(X):={\Ch}_p(X)_{hom}/{\Ch}_p(X)_{alg},$$ where ${\Ch}_p(X)_{alg}$ denotes the space of cycles in ${\Ch}_p(X)$ which are algebraically equivalent to zero. Set ${\Griff}^p(X):=\Griff_{\dim X-p}(X)$. It was shown in [@Friedlander1] that $L_pH_{2p}(X)_{hom}\cong \Griff_p(X)$ for any projective variety.
\[cor1.3\] For every positive integer $m$, we have $L_pH_{k}(Y^{2m},\Q)_{hom}=0$ for $p\leq 1$, $p\geq 2m-2$ and $k\geq 2p$. In particular, $\Griff_1(Y^{2m})\otimes \Q=0$ and $\Griff^2(Y^{2m})=0$.
Recall that a theorem of Peters [@Peters] says that if $\Ch_0(Y)\otimes\Q \cong \Q$ for a smooth projective variety $Y$, then $L_pH_*(Y)_{hom}=0$ for $p\leq 1$ and all $*$. Hence the $p\leq 1$ part follows from Theorem \[Th1.1\] and Peters’ result. It was observed, independently by M. Voineagu [@Voineagu] and the author [@Hu], that Peters’ method could be used to show $L_pH_*(Y)_{hom}=0$ for $p\geq \dim(Y)-2$ and all $*$ under the same assumption. So the $p\geq 2m-2$ part follows from Theorem \[Th1.1\] and the observation. In particular, $\Griff_1(Y^{2m})\otimes \Q=0$ and $\Griff^2(Y^{2m})\otimes \Q=0$. Since $\Griff^2(Y^n)$ has no torsion [@Bloch-S], therefore $\Griff^2(Y^{2m})=0$, i.e., homological equivalence and algebraic equivalence coincide for codimension-2 cycles on $Y^{2m}$. The completes the proof of Corollary \[cor1.3\].
As applications of Theorem \[Th1.7\], we have similar results for $\widetilde{Y^{2m-1}}$ as those in Corollary \[cor1.1\],\[cor1.2\] and \[cor1.3\].
\[cor1.8\] $\Ch^p(\widetilde{Y^{2m-1}})$ is weakly representable for $p\leq 2$.
\[cor1.9\] The Generalized Hodge Conjecture for $\GHC(1,3,\widetilde{Y^{3}})$ holds. More generally, the Generalized Hodge Conjecture $\GHC(1,2m-1,\widetilde{Y^{2m-1}})$ for $Y^{2m}$ holds.
\[cor1.10\] For every integer $m\geq 2$, we have $L_pH_{k}(\widetilde{Y^{2m-1}},\Q)_{hom}=0$ for $p\leq 1$, $p\geq 2m-3$ and $k\geq 2p$. In particular, $\Griff_1(\widetilde{Y^{2m-1}})\otimes \Q=0$ and $\Griff^2(\widetilde{Y^{2m-1}})=0$.
Recall that for $V\subset U$ a Zariski open subset of a quasi-projective variety $U$, we have the long exact sequence for Lawson homology, i.e., $$\label{eq6}
\cdots\rightarrow L_pH_k(Z)\rightarrow L_pH_k(U)\rightarrow
L_pH_k(V)\rightarrow L_pH_{k-1}(Z)\rightarrow\cdots$$ where $Z=U-V$ (cf. [@Lima-Filho]).
By Corollary \[cor1.10\] and Equation (\[eq6\]), we get
\[cor1.11\] For every integer $m\geq 2$, we have $L_pH_{k}({Y^{2m-1}},\Q)_{hom}=0$ for $p\leq 1$, $p\geq 2m-3$ and $k\geq 2p$.
Since $Y^{2m-1}$ is a singular variety, the Bloch-Srinivas method on decompositions of the diagonal does not work for $Y^{2m-1}$. So we try to compute $L_pH_{k}({Y^{2m-1}},\Q)_{hom}$ by the localization sequences for Lawson homology.
Set $V:=\widetilde{Y^{2m-1}}-E$ and $E=\cup_{i=1}^{2^{2m}} E_i$. Then $V\cong Y^{2m-1}-\cup_{i=1}^{2^{2m}} p_i$, where $p_i$, $i=1,\cdots, 2^{2m}$ are singular points of $Y^{2m-1}$ and $E_i$, $i=1,\cdots, 2^{2m}$ are the corresponding exceptional divisors. By using Equation (\[eq6\]) to the $U\subset Y^{2m-1}$, we get $L_1H_k(U)\cong L_1H_k(Y^{2m-1})$ and so $L_1H_k(U)_{hom}\cong L_1H_k(Y^{2m-1})_{hom}$.
From the following commutative diagram (cf. [@Lima-Filho2], Prop. 4.9) [$$\xymatrix{L_1H_k(E)\ar[r]\ar[d]^{\cong}& L_1H_k(\widetilde{Y^{2m-1}})\ar[r]\ar[d]^{\Phi_{1,k}}&L_1H_k(U)\ar[r]
\ar[d]^{\Psi_{1,k}}& L_1H_{k-1}(E)\ar[r]\ar[d]^{\cong}&L_1H_k(\widetilde{Y^{2m-1}})\ar[d]^{\Phi_{1,k-1}}\\
H_k(E)\ar[r]& H_k(\widetilde{{Y^{2m-1}}})\ar[r] & H_k^{BM}(U)\ar[r]&H_{k-1}(E)\ar[r]&H_k(\widetilde{Y^{2m-1}})}$$ ]{} where $H_k^{BM}(U)$ is the Borel-Moore homology of $U$, and the injectivity of $\Phi_{1,k}\otimes \Q$ (i.e., $L_1H_{k}(\widetilde{Y^{2m-1}},\Q)_{hom}=0$ by Corollary \[cor1.10\]), we get the injectivity of $\Psi_{1,k}\otimes \Q$ by the Five Lemma.
Low dimensional examples
========================
For a smooth complex projective variety, we set $h^{i,j}(X):=\dim_{\C} H^{i,j}(X)$.
The case $n=1$ is trivial. In this case $Y^1\cong \P^1$. In the case $n=2$, all $Y^2$ are Enrique surfaces. It was proved in [@Bloch-K-L] that all Enrique surfaces $S$ satisfy $\Ch_0(S)\cong \Z$.
The next case is $n=3$. In this case, $X^3$ (for simplicity denote by $X$ in this paragraph) is the complete intersection of 4 quadric hypersurfaces in $\P^7$. By the adjunction formula, the canonical bundle $K_X$ of $X$ is trivial and so $H^{3,0}(X)\cong \C$ and $h^{3,0}(X)=1$. The Euler class $\chi(X)$ of $X$ is the top Chern class of $X$ (Gauss-Bonnet Theorem). Let $h$ be the hyperplane class of $X$. The total Chern class $c(X):=1+c_1(X)+c_2(X)+c_3(X)=(1+h)^8(1+2h)^{-4}|_X$ and so $c_3(X)=-8h^3=-128$ since $h^3|_X=\deg(X)=16$. Hence $\chi(X)=-128$. This together with Lefschetz hyperplane theorem implies $b_3(X)=132$. By the Hodge decomposition of $H^3(X,\C)$, we get $h^{3,0}(X) + h^{2,1}(X) + h^{1,2}(X) + h^{0,3}(X) = 132$. So $h^{2,1}(X)=h^{1,2}(X)=65$.
Since $\pi:X^3\to Y^3$ is an involution with 16 isolated fixed point, we have $\chi(X^3)-16=2(\chi(Y^3)-16)$ and so $\chi(Y^3)=-56$. Note that $\widetilde{Y^3}\to Y^3$ is the resolution of singularity with exceptional divisor $E=\cup_{i=1}^{16} E_i$, where $E_i\cong \P^2$ since it is the exceptional divisor of the resolution of singularity on $\C^3/\Z_2$ (cf. Remark \[remark2.9\]). So $\chi(\widetilde{Y^3})-16\chi(\P^2)= \chi(Y^3)-16$, i.e., $\chi(\widetilde{Y^3}) = -24$. Since $b_2(Y^3)=1$, we get $b_2(\widetilde{Y^3}) = 17$. Hence $b_1(\widetilde{Y^3})=0$, we get $b_3(\widetilde{Y^3})=60$. We get $h^{1,2}(\widetilde{Y^3}) = h^{2,1}(\widetilde{Y^3})=30$ since $h^{3,0}(\widetilde{Y^3})=0$.
Recall that Suslin’s Conjecture on Lawson homology states that: *For any abelian group $A$ and smooth quasi-projective variety $X$ of dimension $n$, the map $L_pH_k(X,A) \to H^{BM}_k(X,A)$ is an isomorphism for $k\geq n+p$ and a monomorphism for $k=n+p-1$.* Here $H^{BM}_k(X,A)$ means Borel-Moore homology with coefficient in $A$.
By the above computation, Theorem \[Th1.7\] and the Proposition 5.1 in [@Voineagu], we have
For $\widetilde{Y^3}$ above, we have the following statements:
1. $L_1H_2(\widetilde{Y^3})\cong H_2(\widetilde{Y^3}) \cong \Z^{17}$.
2. $L_1H_3(\widetilde{Y^3})\cong H_3(\widetilde{Y^3})\cong \Z^{60}$.
3. $L_2H_4(\widetilde{Y^3})\cong L_1H_4(\widetilde{Y^3})\cong H_4(\widetilde{Y^3}) \cong \Z^{17}$.
4. $L_3H_6(\widetilde{Y^3})\cong L_2H_6(\widetilde{Y^3})\cong L_1H_6(\widetilde{Y^3})\cong H_6(\widetilde{Y^3}) \cong \Z$.
5. All other $L_pH_*(\widetilde{Y^3})$ ($p\geq 1$) are trivial.
In particular, Suslin’s Conjecture for $\widetilde{Y^3}$ holds.
Next case is $n=4$. In this case, it can be calculated that $h^{4,0}(X^4)=1$, $h^{3,1}(X^4)=h^{1,3}(X^4)=151$ and $h^{2,2}(X^4)=652$ since $X^4$ is a complete intersection. Since $h^{4,0}(Y^4)=0$, we get $h^{3,1}(Y^4)=h^{1,3}(Y^4)=75$ and $h^{2,2}(Y^4)=326$ by Riemann-Roch-Hirzebruch theorem for orbit spaces (cf. [@Atiyah-Singer], 4.7), Since $h^{3,1}(Y^4)\neq 0$, $\Ch_1(Y^4)$ is not weakly representable. In particular, $\Ch_1(Y^4)_{hom}\otimes\Q$ is nontrivial. From this and the proof of Theorem \[Th1.1\], we obtain the Chow group of 1-cycles for $X^4/\langle\tau_i\rangle$ is not weakly representable, although $X^4/\langle\tau_i\rangle$ is a rationally connected variety for each $i=0,1,\cdots, n$.
In this case we can say a little more about Lawson homology of $Y^4$. By By the above computation, Theorem \[Th1.1\] and the Proposition 5.3 in [@Voineagu], we have the following result.
For $Y^4$ above, we have the following statements
1. $L_1H_2(Y^4)_{\mathbb{Q}}{\cong}H_2(Y^4)_{\mathbb{Q}}\cong\Q.$
2. $L_2H_4(Y^4)\hookrightarrow L_1H_4(Y^4)\cong H_4(Y^4)\cong\Z^{476}.$
3. $L_3H_6(Y^4)\cong L_2H_6(Y^4)\cong L_1H_6(Y^4)\cong H_6(Y^4)\cong\Z.$
4. $L_4H_8(Y^4)\cong L_3H_8(Y^4)\cong L_2H_8(Y^4)\cong L_1H_8(Y^4)\cong H_8(Y^4)\cong\Z.$
5. All other $L_pH_*(Y^4)$ ($p\geq 1$) are trivial.
In particular, Suslin’s Conjecture for $Y^4$ holds.
One can show that $H^{1,n-1}(Y^{2m},\Q)$ is nontrivial for all $n=2m\geq 4$. Therefore $\Ch_1(Y^{2m})_{hom}\otimes\Q\neq 0$ and Theorem \[Th1.1\] is the best result we could obtain. Similarly, One can show that $H^{1,n-1}(Y^{2m-1},\Q)$ is nontrivial for all $n=2m-1\geq 3$ and Theorem \[Th1.7\] is the best result.
[AAAA]{} M. F. Atiyah and I. M. Singer, [*The index of elliptic operators. III.*]{} Ann. of Math. (2) 87 1968 546–604.
S. Bloch, [*Lectures on algebraic cycles.*]{} Duke University Mathematics Series, IV. Duke University, Mathematics Department, Durham, N.C., 1980. 182 pp. (not consecutively paged).
S. Bloch, [*Some elementary theorems about algebraic cycles on Abelian varieties.*]{} Invent. Math. 37 (1976), no. 3, 215–228.
S. Bloch, A. Kas, and D. Lieberman, [*Zero cycles on surfaces with $p\sb{g}=0$.*]{} Compositio Math. 33 (1976), no. 2, 135–145.
S. Bloch and V. Srinivas, [*Remarks on correspondences and algebraic cycles.*]{} Amer. J. Math. 105 (1983), no. 5, 1235–1253.
F. Campana, [*Connexité rationnelle des variétés de [F]{}ano.*]{} Ann. Sci. École Norm. Sup. (4) 25 (1992), no. 5, 539–545.
E. Friedlander, [*Algebraic cycles, Chow varieties, and Lawson homology.*]{} Compositio Math. 77 (1991), no. 1, 55–93.
A. Fujiki, [*On resolutions of cyclic quotient singularities.*]{} Publ. Res. Inst. Math. Sci. 10 (1974/75), no. 1, 293–328.
W. Fulton, [*Intersection theory.*]{} Second edition, Springer-Verlag, Berlin, 1998.
A. Grothendieck, [*Hodge’s general conjecture is false for trivial reasons.*]{} Topology 8 1969 299–303.
W. Hu, [*A note on Lawson homology for smooth varieties with small Chow groups.*]{} arxiv:math/0602516
H.Inose and M. Mizukami, [*Rational equivalence of $0$-cycles on some surfaces of general type with $p\sb{g}=0$*]{}. Math. Ann. 244 (1979), no. 3, 205–217.
J. Kollár, Y. Miyaoka and S. Mori, [*Rational connectedness and boundedness of Fano manifolds.*]{} J. Differential Geom. 36 (1992), no. 3, 765–779.
B. Lawson, [*Algebraic cycles and homotopy theory.*]{}, Ann. of Math. [**129**]{}(1989), 253-291.
B. Lawson, [*Spaces of algebraic cycles.*]{} pp. 137-213 in Surveys in Differential Geometry, 1995 vol.2, International Press, 1995.
J. D. Lewis, [*A survey of the Hodge conjecture.*]{} (English summary) Second edition. Appendix B by B. Brent Gordon. CRM Monograph Series, 10. American Mathematical Society, Providence, RI, 1999. xvi+368 pp. ISBN: 0-8218-0568-1
P. Lima-Filho, [*Lawson homology for quasiprojective varieties.*]{} Compositio Math. 84 (1992), no. 1, 1–23.
P. Lima-Filho, [*On the generalized cycle map.*]{} (English summary) J. Differential Geom. 38 (1993), no. 1, 105–129.
D. Mumford, [*Rational equivalence of $0$-cycles on surfaces.*]{} J. Math. Kyoto Univ. 9 1968 195–204. C. Peters, [*Lawson homology for varieties with small Chow groups and the induced filtration on the Griffiths groups.*]{} Math. Z. 234 (2000), no. 2, 209–223.
K. H. Paranjape and V. Srinivas, Algebraic cycles. Current trends in mathematics and physics, 71–86, Narosa, New Delhi, 1995.
A.A. Roĭtman, [*Rational equivalence of zero-dimensional cycles.*]{} (Russian) Mat. Sb. (N.S.) 89(131) (1972), 569–585, 671.
Mircea Voineagu, [*Semi-topological K-theory for certain projective varieties.*]{} Preprint. arxiv.org/abs/math/0601008
C. Voisin, [*Sur les zéro-cycles de certaines hypersurfaces munies d’un automorphisme* ]{} Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 19 (1992), no. 4, 473–492.
Department of Mathematics, Massachusetts Institute of Technology, Room 2-363B, 77 Massachusetts Avenue, Cambridge, MA 02139, USA
Email: [[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The detailed observations of the diffuse ionized gas through the emission lines H$\alpha$, \[NII\], and \[SII\] in the Perseus Arm of our Galaxy by the Wisconsin H$\alpha$ Mapper (WHAM)–survey challenge photoionization models. They have to explain the observed rise in the line ratios \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$. The models described here consider for the first time the detailed observational geometry toward the Perseus Arm. The models address the vertical variation of the line ratios up to height of 2 kpc above the midplane. The rising trends of the line ratios are matched. The increase in the line ratios is reflected in a rise of the temperature of the gas layer. This is due to the progressive hardening of the radiation going through the gas. However an extra heating above photoionization is needed to explain the absolute values. Two different extra heating rates are investigated which are proportional to $n^0$ and $n^1$. The models show that a combination of both are best to explain the data, where the extra heating independent of density is dominant for z $>$ 0.8 kpc.'
author:
- 'T. Elwert and R.–J. Dettmar'
title: Constraining the extra heating of the Diffuse Ionized Gas in the Milky Way
---
Introduction
============
Containing typically half the mass of ionized hydrogen in galaxies, the Diffuse Ionized Gas (DIG) is visible as an extended H$\alpha$ emitting layer in our Galaxy (e.g. the WHAM–survey) and in many other galaxies (see e.g. Tüllmann & Dettmar [@tuellmannB], Collins & Rand [@collins], Otte et al. [@otte], Hoopes & Walterbos [@hoopes]). Studies of emission line ratios such as \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$ provide information about the physical conditions of the gas. Simple energy estimations (Reynolds [@reynolds90], Reynolds [@reynolds93]) favor O and early B stars to be responsible for most of the DIG. 3D models using various methods (e.g. Miller & Cox [@miller2], Dove & Shull [@dove], Wood & Loeb [@wood2], Ciardi et al. [@ciardi], Wood et al. [@wood]) showed that it is possible for ionizing photons from O stars to penetrate from the midplane into the halo. Wood & Mathis [@wood3] noted that the line ratios increase with distance from the midplane due to the progressive hardening of the radiation. So far photoionization models made no specific attempt to model the trends of the line ratios. Models by Mathis [@mathis], Domgörgen & Mathis [@domgoergen], Sembach et al. [@sembach], Bland–Hawthorn et al. [@bland-hawthorn] used volume average models to explain the observed data. The analytical approach by Haffner et al. [@haffner], referred to as Haffner99 in the rest of this paper, treated the dependence of the line ratios with height. Haffner99 and its further application to other galaxies by e.g. Collins & Rand [@collins], Otte et al. [@otte], and Miller & Veilleux [@miller] gave evidence that an additional heating source is needed in order to explain the rise of the line ratios with increasing distance $z$ from the midplane.
We are constructing photoionization models in order to examine the trends in the observed line ratios and if photoionization can heat up the gas sufficiently in order to explain the data. We introduce specific extra heating terms, extra means in addition to photoionization, and discuss their properties. In the following section \[data\_perseus\] we introduce the observations of the Perseus Arm to which our models are compared. Section \[parameters\] deals with the model parameters and discusses geometry and sight line effects taken into account. The models are compared with the data in the next section. The last section summarizes the results.
Data of the Perseus Arm {#data_perseus}
=======================
We are using the data taken from the Wisconsin H$\alpha$ Mapper (WHAM) (e.g. Haffner99, Haffner et al. [@haffner01]) survey which were kindly provided by Ron Reynolds and Matt Haffner. The WHAM–survey mapped the northern sky in H$\alpha$ with declinations of $\delta > -
30^\circ$. The Perseus Arm ($-35^\circ < \delta < -11^\circ$ and $120^\circ < l < 150^\circ$) was additionally mapped in \[NII\]$\lambda$6583 and \[SII\]$\lambda$6716. At each pointing an averaged spectrum with a beam of $1^\circ$ is measured with a velocity resolution of 12 km $\rm s^{-1}$. The emission of the Perseus Arm can be separated in velocity space from the local emission (Haffner99) for galactic longitudes $120^\circ \, < \, \rm l \, < \, 150^\circ$. This was performed by integrating the line emission in the velocity range $-100$ km ${\rm s^{-1}}\,<\,v\,< -20$ km $\rm s^{-1}$, no line fitting was performed. We are using the intensity of the H$\alpha$ line as well as the line ratios \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$. The sensitivity limit of 0.1 Rayleighs[^1] results in an observed vertical height of up to $|z|$ = 2 kpc assuming a distance to the arm of 2.5 kpc.
Model parameters {#parameters}
================
We use the spectral simulation code CLOUDY, version 96.00 (described by Ferland [@ferland3], [@ferland1], [@ferland2]) to model the DIG. CLOUDY determines the physical conditions by balancing the heating and cooling rates, so that the energy is conserved. The results of the models are compared to the observed emission line ratios \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$, and the gas temperature as derived from \[NII\]/H$\alpha$.
In order to realize a model describing the DIG certain parameters have to be specified: The ionizing spectrum of the source is a composition of three different stellar temperature: 56% from T = 35000 K, 12% from T = 40000 K, and 32% from T = 45000 K, as used in Mathis [@mathis] and Wood et al. [@wood]. The WMbasic models (Pauldrach et al. [@pauldrach]), which include N–LTE effects, X–ray emission from shocks within stellar winds, are used as the ionizing spectra. The luminosity of the source is chosen in such a way that the observed run of the H$\alpha$ intensity is matched, as shown in Figure \[Ha-PerseusArm\].
The density structure is exponential, as derived from the observed H$\alpha$ intensity, with a scale height of 1 kpc and a midplane density of 0.2 $\rm cm^{-3}$. The density is in clumps with a filling factor of 20%, i.e. only 1/5 of the volume is filled with this plasma. The geometry of the ionized gas is chosen to be a plane parallel layer, i.e. the ratio of the depth of the cloud to the distance of the illuminated face to the ionizing source is smaller than 1/10.
We are matching the observed H$\alpha$ intensity of the observations (Figure \[Ha-PerseusArm\]) by placing the illuminated face of the cloud at a z–height of 1 scale height of the Lockman–layer (300pc) assuming a density law of $n = 0.1 \exp(-z/0.3)$ as in Miller & Cox [@miller2]. This is done after the model is calculated as otherwise the condition of plane parallel illumination of the cloud cannot be fulfilled. The information of the actual position of the ionizing stars are effectively removed, consistent with the picture of having the DIG being ionized by radiation leaking out of the Lockman–layer. The intensity gradient of H$\alpha$ is matched as well as the estimate of the hydrogen ionizing photon flux by Reynolds [@reynolds90]): $\phi_{DIG} \ge 5 \times 10^6$ hydrogen ionizing photons $\rm cm^{-2}\,\,s^{-1}$. In order to compare the models on a common basis, which is important as extra heating is dominating for large z–heights, we choose to let the ionization structure be the same for all models (see Figure \[ionfrac\]). This is in accordance with the idea that the extra heating affects only the temperature of the gas. The forbidden levels of nitrogen and sulphur can then be more easily excited by collisions with the electrons which in turn elevates the line ratios \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$. Figure \[Ha-PerseusArm\] shows that the intensity gradient of H$\alpha$ varies only a little for the different models. The scale heights for the models are slightly different ($\ll$ 10%) for the models with an extra heating, which can be explained by assuming that the DIG–layer is in pressure equilibrium. This effect was also noted by Wood & Mathis [@wood3]. The ionization parameters (U) of all models lie in the narrow range between log(U) = -3.0 and -3.1.
The ISM composition of CLOUDY is used with N/H and S/H set to the values in Haffner99 (N/H = $7.5 \times 10^{-5}$, S/H = $1.86 \times 10^{-5}$). Graphite and silicate grains with the size distribution used for the ISM (see HAZY, Ferland [@ferland3]) are present in the gas and account for less than 10 % of the global heating. The inclusion of PAHs give only variations $<$ 3% for the line ratios and have only a small effect on the heating balance. The interaction with cosmic rays is taken into account as described in Ferland & Mushotsky [@ferland84].\
A crucial factor is the consideration of the line of sight. The observations give information about the line ratios at different positions orthogonal to the source of the ionizing radiation. The correct local line ratios are represented by the ratios of the volume emission coefficients $\epsilon_V$: $ \frac{I_1}{I_2} = \frac{\int \; {\rm d}z \; \epsilon_{V_1}}{\int \; {\rm d}z \; \epsilon_{V_2}} \approx \frac{\epsilon_{V_1}}{\epsilon_{V_2}}$. An important issue is the treatment of the line of sight to the Perseus Arm due to our position in the Milky Way as shown in Fig. \[geometry-PerseusArm\]. Each pointing of the WHAM survey contains contributions from different $|z|$ heights above the midplane ($\epsilon(z1)$ to $\epsilon(z2)$), this effect is taken into account by integrating over the particular sight line in the models. The distance to the Perseus Arm is assumed to be 2.5 kpc and the thickness of the arm to be 1 kpc as quoted in Haffner99. Moreover, the observations are the average of a beam size of $1^\circ$ which means that we have to take into account an additional integration over different z heights, on average 45 pc. These effects of geometry and observational smearing are taken into consideration here for the first time. Fig \[geometry-beam\_smearing\] shows the effect of the beam smearing and line of sight geometry on the calculated line ratios for z–heights up to 2.5 kpc, which have to be considered for the observed heights of 1.8 kpc. These effect increase the ’corrected’ line ratios by at most 15% and their trend is altered for high z–heights. The ’corrected’ model without an extra heating seems to suggest that \[NII\]/H$\alpha$ is not increasing after about 1.6 kpc, but the ’uncorrected’ model shows a further increase due to the hardening of the radiation. The result of the gas temperature (section \[comp\], Fig. \[gastemp\]) is even more influenced by this issue. The discussion of the line ratios therefore need to include an appropriate treatment of the line of sight geometry and the beam smearing to take these effects into account.
Extra heating
-------------
Additional heating to photoionization is included which is proportional to $n^1$ and $n^0$ in accordance with Reynolds et al. [@reynolds99], their factors $G_1$ and $G_2$ respectively. We choose rates in the same range as in their paper: $G_1 = 1\times10^{-25} \rm erg \, cm^{-3} \, s^{-1}$ and $G_2 = 5\times10^{-27} \rm erg \, cm^{-3} \, s^{-1}$. The heating–cooling balance can then be written as either $G_0 + G_1/n_e = \Lambda$ or $G_0 + G_2/n^2_e = \Lambda$. The heating due to photoionization is given by $G_0 n^2_e$ and the cooling by $\Lambda n^2_e$. The inclusion of an extra heating source rises the gas temperature of the models, at the same time the ionization structure varies only slightly. The temperature of the CLOUDY models are calculated by heating–cooling balance. The extra heating therefore increases the temperature, having more pronounced effects at larger z heights as the photoionization heating rate decreases like $\rm n^2$ as shown in Fig. \[gastemp\]. The graph is explained in more detail in the next section.
In our models the ionization structure is nearly unaffected by the inclusion of an extra heating source, sulphur is slightly more effected than nitrogen. This behavior is the basic assumption for the extra heating in Reynolds et al. [@reynolds99]. However they are assuming a constant ratio of $\rm N^+/N$, whereas the models show a dependence on z height, this is expected as the radiation gets progressively absorbed. Figure \[ionfrac\] shows ionization structure and the change dependent on the different heating rates. Hydrogen is nearly fully ionized throughout all models which is a basic characteristic of the DIG.
Comparison with Observations {#comp}
----------------------------
In Fig. \[n-h-z-PerseusArm\] and Fig. \[sii-nii-PerseusArm\] the models differ by the type of extra heating, ranging between models without extra heating and a rate of $G_1 = 1\times10^{-25} \rm erg \, cm^{-3} \, s^{-1}$ and $G_2 = 5\times10^{-27} \rm erg \, cm^{-3} \, s^{-1}$. There was no fitting done in order to match the models with the observations. No ’best–fit’ exists and therefore the models are independent of the observational quality, individual spectral features, or small scale variations which cannot be reproduced with a smooth density distribution. As the data for small $|z|$ heights are contaminated with radiation from the midplane and dust absorption it is convenient to consider line ratios for $|z|$–heights above 0.8 kpc to be ’pure’ DIG. This is also the range for which the H$\alpha$ scale height was determined. The models show lower values of \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$ for the lower z heights as doubly ionized nitrogen contributes 35% and doubly ionized sulphur even 80%, as seen in Fig. \[ionfrac\]. As a consequence \[NII\] and \[SII\] are weaker. Figure \[n-h-z-PerseusArm\] shows the development of the line ratios with $|z|$ height, both the data and the models show an increase with $|z|$. The trend in the line ratios is matched by the models even without an extra heating source.
The \[NII\]/H$\alpha$ line ratio above 1 kpc can be explained with the models including an extra heating rate. The modeled \[SII\]/H$\alpha$ ratio is up to a factor of two below the observations, however the theoretical uncertainty for the \[SII\] line is very high as dielectronic recombination is an important process in the DIG. As the corresponding recombination coefficients are not known (see discussion in Ferland et al. [@ferland98]), the results of the models have to be handled with care. Our models use the KLUDGE approximation (Ferland [@ferland3]). Models without dielectronic recombination have \[SII\]/H$\alpha$ decreased by 50%. If the rate is doubled then \[SII\]/H$\alpha$ is increased by 50%.
Figure \[gastemp\] shows the gas temperature of the models, the increase of the line ratios with z height is due to the progressive hardening of the radiation as the photons go through the gas layer. In order to explain the observed line ratios an extra heating source is however needed which does not alter the general shape of the predicted line ratios but elevates the line ratios. As \[NII\] and \[SII\] are forbidden lines which get collisional excited by electrons, an increase in gas temperature increases the amount of electrons capable to excite the singly ionized nitrogen and sulphur ions which can then decay by emitting the emission lines in questions. The gas temperature is deducted from the observations through the relation - following Reynolds et al. [@reynolds99]: $\frac{I_{[NII]}}{I_{\rm H\alpha}} = 1.84 \times 10^5 \left(\frac{\rm N^+}{N}\right) \left(\frac{\rm H^+}{H}\right)^{-1} T_4^{0.39} \exp(-2.18/T_4)$. Assuming that $N^+/N = H^+/H$ and the abundance as stated in section \[parameters\] gives $\frac{I_{[NII]}}{I_{\rm H\alpha}} = 13.75 \, T_4^{0.39} \exp(-2.18/T_4)$, $T_4 = T / 10^4 \, \rm K$. We are not using the collision strength of singly ionized nitrogen from Reynolds et al. [@reynolds99]: ${\rm N^+}: \Omega\left(^3P,^1D\right) = 2.28\cdot T^{0.026}_4$, Aller [@aller], but the data from Stafford et al. [@stafford]: ${\rm N^+}: \Omega\left(^3P,^1D\right) = 3.02\cdot T^{-0.01}_4$. This leads to the different coefficients and temperatures on average 250 K below the values of Reynolds et al. [@reynolds99]. For singly ionized sulfur we use the data from Lanzafame et al. [@lanzafame93] instead of Aller [@aller] as in Reynolds et al. [@reynolds99]. The plot also shows the impact of the line of sight geometry and beam smearing as well as the consideration of the ionization structure of the models. As the line of sight leads to higher \[NII\]/H$\alpha$ ratios (up to 15%,) this effects, when accounted for, lower temperatures. The ionization structure, i.e. $N^+/N \approx H^+/H$ is only valid for z $>$ 800 pc, leads to higher estimates of the temperature. The combined effect is the elevation of the derived temperature for z $<$ 800 pc and above that lower values by about 200 K. The models without an extra heating source have temperatures too low, an extra heating source independent of density shows very good agreement with the observed temperatures. The extra heating $\propto n^1$ seems to best fit the data up to 800 pc, then the extra heating $\propto n^0$ gives the best agreement for higher z values. The interpretation for the z heights $<$ 800 pc have to be handled with care as this region is contaminated by radiation from the midplane and therefore the part responsible for the DIG emission is difficult to estimate. Magnetic reconnection (e.g. Birk et al. [@birk] ) or heating by cosmic rays through linear Landau–damping (Lerche & Schlickeiser [@lerche] ) are possible processes producing a heating independent of density. As photoelectric heating from dust grains is $\propto n^1$, the data suggest that there is more dust present at z–heights below 800 pc than present in the models if this mechanism is responsible for the elevated temperature.
In Fig. \[sii-nii-PerseusArm\] the line ratios \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$ are plotted against each other. Values of \[NII\]/H$\alpha$ greater than 1, which cannot be explained by classical HII–region calculations, are reached with models using an extra heating source. Also in this case the extra heating independent of density is able to produce higher \[NII\]/H$\alpha$ ratios than the other therefore matching better with the data. The limits of Haffner99 for two constant $\rm S^+/S$ ratios (0.5 and 0.25) are also given. The ionization fractions of the models (see Fig. \[ionfrac\]) are within these two limits for z $>$ 0.6$\,$kpc. Together with the temperature plot (Fig. \[gastemp\]) the models match in all these direct and derived quantities with the data and agrees with the estimates of Haffner99 and Reynolds et al. [@reynolds99]. The application to other galaxies shows that this diagram is also a valuable diagnostic for the chemical evolution (Elwert et al. [@elwert]).
Summary
=======
We have shown that the observed trend of the line ratios \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$ above the Galactic plane can successfully be explained by photoionization models including extra heating and considering the line of sight geometry. The observed values need an extra heating source which strength lies at the lower end of the predicted values of Reynolds et al. [@reynolds99] due to our models. At high z heights (z $>$ 800 pc) an extra heating independent of density gives the best agreement with the data, whereas for smaller z heights an extra heating term $\propto n^1$ gives better results concerning the temperature. There is an intrinsically increase in the line rations and the gas temperature due to the progressive hardening of the radiation. The extra heating terms are enhancing this trend and elevate the line ratios to the observed values. It is important to incorporate the observing geometry to the Perseus Arm into the models when comparing with the data. A discussion concerning models of observed line ratios in edge–on galaxies is given in Elwert [@elwert03] and Elwert & Dettmar [@elwert05].
This work was supported by DFG through SFB 591 and through Deutsches Zentrum für Luft- und Raumfahrt through grant 50 OR 9707. TE wants to thank Kenneth Wood and Ron Reynolds for helpful comments and enlightening discussions while writing the paper. We also want to thank the anonymous referee for making many very useful suggestions and comments which helped to improve the publication.
Aller, L.H. 1984, Physics of Thermal Gaseous Nebulae (Dordrecht: Reidel) Birk, G.T., Lesch, H., Neukirch, T. 1998, MNRAS 296, 165 Bland–Hawthorn, J., Freeman, K.C., Quinn, P.J. 1997, ApJ 490, 143 Ciardi, B., Bianchi, S., Ferrara, A. 2002, MNRAS 331, 463 Collins, J.A., Rand, R.J. 2001, ApJ 551, 57 Domgörgen, H., Mathis, J.S. 1994, ApJ 428, 647 Dove, J.B., Shull, J.M. 1994, ApJ 423, 196 Elwert, T. 2003, Ph.D. Thesis, Ruhr–Universität Bochum Elwert, T., Dettmar, R.–J. 2005, ASPC 331, 203, editor: Braun, R. Elwert, T., Dettmar, R.–J., Tüllmann, R. 2003, AAS, 203, \#111.05 Ferland, G.J. 2002, Hazy, A Brief Introduction to CLOUDY, Dept. Phys. Astron. Internal Rep. 96.00 (Lexington: Univ. Kentucky) Ferland, G.J. 2000, ASP Conference Proc. Vol. 216 Ferland, G.J., Korista, K.T., Verner, D.A., Ferguson, J.W., Kingdon, J.B., Verner, E.M. 1998, PASP 110, 761 Ferland, G.J., Korista, K.T., Verner, D.A., Ferguson, J.W., Kingdon J.B., Verner, E.M. 1998, PASP 110, 761 Ferland, G.J., Mushotzky, R.F. 1984, ApJ 286, 42 Haffner, L.M., Reynolds, R.J., Madsen, G.J., Tufte, S.L., Jaehning, K.P., Percival, J.P., Hausen, N.R. 2003, ApJS 149, 405 Haffner, L.M., Reynolds, R.J., Tufte, S.L. 1999, ApJ 523, 223 Hoopes, C.G., Walterbos, R.A.M. 2003, ApJ 586, 902 Lanzafame, A.C., Tully, J.A., Berrington, K.A., Dufton, P.L., Byrne, P.B., Burgess, A. 1993, MNRAS 264, 402 Lerche, I., Schlickeiser, R. 2001, A&A 366, 1008 Miller, S.T., Veilleux, S. 2003, ApJ 592, 79 Miller, W.W. III, Cox, D.P. 1993, ApJ 417, 579 Mathis, J.S. 2000, ApJ 544, 347 Otte, B., Reynolds, R.J., Gallagher III J.S., Ferguson, A.M.N. 2001, ApJ 560, 207 Pauldrach, A.W.A., Hoffmann, T.L., Lennon, M. 2001, A&A 375, 161 Reynolds, R.J. 1993, in AIP Conf. Proc. 278, Back to the Galaxy, ed. S.S. Holt & F. Verter, 156 Reynolds, R.J. 1990, ApJ 349, L17 Reynolds, R.J., Haffner, L.M., Tufte S.L. 1999, ApJ 525, L21 Sembach, K.R., Howk, J.C., Ryans, R.S.I., and Keenan, F.P. 2000, ApJ 528, 310 Stafford, R.P., Bell, K.L., Hibbert, A., Wijesundera, W.P. 1994, MNRAS, 268, 816 Tüllmann, R., Dettmar, R.–J. 2000, A&A 362, 119 Wood, K., Mathis, J.S. MNRAS 353, 1126 Wood, K., Mathis, J.S., Ercolano, B. 2004, MNRAS 348, 1337 Wood, K., Loeb, A. 2000, ApJ 545, 86
![H$\alpha$ intensity of the Perseus Arm as observed by WHAM, overplotted is the fit of Haffner99 and the photoionization models with and without extra heating. []{data-label="Ha-PerseusArm"}](f1.ps){width="5.5cm"}
![Sketch of the line of sight geometry and the beam smearing, which have to be considered when comparing the models with the data. Deriving temperatures from the data without taking this into account lead to temperatures that are on average 250 K higher. []{data-label="geometry-PerseusArm"}](f2.ps){width="5.5cm"}
![The effect of the beam smearing and line of sight geometry on \[NII\]/H$\alpha$ specifically. The enhancement of this line ratio due to this effect is at most 15% and leads to higher temperatures derived from the observations if the effects is not taken into account. []{data-label="geometry-beam_smearing"}](f7.ps){width="5.5cm"}
![Temperature structure of the DIG, estimates from the observations with atomic data used by Stafford et al. [@stafford]. The line of sight geometry and beam smearing lead to temperatures on average 250 K lower, whereas the consideration of the ionization structure of $N^+$ elevates the derived temperatures again. The figure shows that a combination of the two different heating terms gives the best agreement with the data. []{data-label="gastemp"}](f3.ps){width="5.5cm"}
![Ionization fractions of $\rm N^+$ and $\rm S^+$ dependent on z, the assumption of Haffner99: $\rm N^+/N = H^+/H \rightarrow 1$ is approximately met for heights above 1300 pc. []{data-label="ionfrac"}](f4.ps){width="5.5cm"}
![Observed \[NII\]/H$\alpha$ and \[SII\]/H$\alpha$ in the Perseus Arm, showing the characteristic rise. The models with and without an extra heating source are able to explain this trend, the explicit values are better reproduced with an extra heating. []{data-label="n-h-z-PerseusArm"}](f5.ps){width="6cm"}
![\[NII\]/H$\alpha$ – \[SII\]/H$\alpha$ in the Perseus Arm as observed by WHAM, plotted are the models, the data, and the limits given by Haffner99 for two constant $\rm S^+/S$ ratios. []{data-label="sii-nii-PerseusArm"}](f6.ps){width="6cm"}
[^1]: 1 R = $10^6/4\pi$ photons $\rm cm^{-2}$ $\rm s^{-1}$ $\rm sr^{-1}$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the rheological characteristics of human blood plasma in shear and elongational flows. While we can confirm a Newtonian behavior in shear flow within experimental resolution, we find a viscoelastic behavior of blood plasma in the pure extensional flow of a capillary break-up rheometer. The influence of the viscoelasticity of blood plasma on capillary blood flow is tested in a microfluidic device with a contraction-expansion geometry. Differential pressure measurements revealed that the plasma has a pronounced flow resistance compared to that of pure water. Supplementary measurements indicate that the viscoelasticity of the plasma might even lead to viscoelastic instabilities under certain conditions. Our findings show that the viscoelastic properties of plasma should not be ignored in future studies on blood flow.'
author:
- 'M. Brust'
- 'C. Schaefer'
- 'R. Doerr'
- 'L. Pan'
- 'M. Garcia'
- 'P. E. Arratia'
- 'C. Wagner'
title: 'Rheology of human blood plasma: Viscoelastic versus Newtonian behavior'
---
=1
Blood is a complex fluid that consists of a suspension of blood cells in a liquid plasma which contains mostly water as well as proteins, mineral ions, hormones, and glucose. In humans, red blood cells (RBC) are the most abundant type of cells in whole blood with a concentration of approximately 45% by volume. Because of this high RBC concentration, it is often believed that rheological behavior of whole blood is mostly determined by the presence of the RBCs. Blood exhibits shear thinning and at low shear rates the red blood cells form reversible aggregates (rouleaux) that are broken up at high shear rates [@Baskurt2003]. The rouleaux formation is caused by the plasma proteins (most likely due to a depletion effect) but the plasma solution, which is approximately 92% water, is currently believed to be Newtonian [@Baskurt2003; @Sousa2011; @Wells1961]. For example, experiments by Copley and King [@Copley1972] and Jaishankar, Sharma and McKinley [@Jaishankar2011] show that the non-Newtonian and viscoelastic effects found in plasma and BSA (Bovine Serum Albumin) solutions in *shear* flows can be attributed to the surface layer of plasma proteins present at the liquid-air interface; that is, these non-Newtonian effects are surface rather than bulk effects. These surface effects can be suppressed by the addition of a small amount of surfactants to the protein solutions [@Jaishankar2011]. Much work has been devoted to the understanding of laminar blood flow as well as the appearance of flow instabilities that may appear under pathological conditions, e.g. near an aneurysm or blockage [@Sforza2009; @Baek2010]. In general, one has to distinguish between inertia and viscoelastic driven instabilities. An elegant way to separate these two kinds of instabilities is the observation of vortices in a contraction flow device. Inertial effects lead to downstream vortices whereas upstream vortices can be observed for viscoelastic fluids [@Rodd2005; @Rodd2007]. This approach has been used in a recent work by Sousa et al. [@Sousa2011] which gives an overview of the flow of two common blood analog solutions in contraction-expansion microfluidic devices. The two investigated plasma replacement solutions with the same shear but different elongational viscosity revealed a remarkably different flow behavior which shows that it is not sufficient to consider only the shear viscosity in order to build a plasma analog solution. A pronounced viscoelastic behavior in elongational flow has been found in several other bio-fluids, too, e.g. for saliva [@Bhat2010; @Haward2011] or DNA solutions.
![\[fig:thinning\] Snapshots of the capillary bridges in the CaBER experiment. Images are $ 833\ \micro\meter \times 1344\ \micro\meter$.](fig1.eps){width="0.75\columnwidth"}
![(Color online) Shear rate dependent viscosity $\eta(\dot\gamma)$ data at $T=20\celsius$ obtained by the rotational shear measurements. The grey shaded area indicates the regime below the resolution of the rheometer.[]{data-label="fig:eta_plasma"}](fig3.eps){width="\columnwidth"}
In this manuscript, we investigate the rheology and flow behavior of human blood plasma in both shear and extensional flows. We find that blood plasma shows significant viscoelastic effects in elongational flow in both (i) a capillary breakup extensional rheometer (CaBER, Figure\[fig:thinning\]) and (ii) a microfluidic contraction-extension device. In addition a synthetic (polymeric) plasma replacement solution is identified that matches the shear and extensional rheology of human plasma.
*Sample solutions*: Blood plasma used in the experiments was is taken from three healthy donors and stored in test tubes using anticoagulants (EDTA or Citra). Plasma is obtained by centrifuging the blood samples at 4000 rotations per minute for 5 minutes immediately after withdrawal. The separated transparent liquid phase is centrifuged again and finally the pure plasma phase is extracted. All measurements are performed on the day of donation. Since the main protein in the plasma is albumin, an aqueous solution of $45\ \milli\gram\per\milli\litre$ albumin (BSA, Polysciences) is prepared for comparison. The plasma replacement solutions are prepared by adding small amounts of polyethylene oxide (PEO, Sigma Aldrich, $M_w=4\cdot 10^6 \ \gram\per\mole$) to a Newtonian solvent. Three PEO solutions are prepared: (i) a 50 ppm PEO in $15/85 wt\%$ glycerol/water solution ($PEO50_{15/85}$), (ii) a 500 ppm PEO in $15/85 wt\%$ glycerol/water solution ($PEO500_{15/85}$), and (iii) a 50 ppm PEO in a $55/45 wt\%$ glycerol/water solution ($PEO50_{55/45}$). Two RBC suspension at $50 vol\%$ are also used, one in physiological buffer (PBS, Phosphate Buffered Saline, Invitrogen) with $15 wt\%$ glycerol ($PBS_{15/85-Hct50}$) and the other under the same condition but with the addition of 50 ppm of PEO ($PEO50_{15/85-Hct50}$). Finally, whole blood and distilled water are used as reference fluids.
Steady shear rheology is performed in a rheometer (Thermo Scientific, HAAKE MARS II) at shear rates ranging from $\dot\gamma=1\ s^{-1}$ to $1000\ s^{-1}$ in rate controlled (CR) mode. A double-cone DC60/1° geometry is used in order to avoid the effect of surface viscoelasticity on the measured shear viscosity $\eta(\dot\gamma)$. To reduce the instrumental noise, a special protocol is applied. For each shear rate, the torque signal is integrated over at least a full revolution of the cone. The minimal resolvable torque is $\tau_\text{min}=2\ \micro\newton\meter$ which limits the reliable minimal shear rate. Above, the accuracy is $\Delta \eta = \pm 0.1\ \milli\pascal\second$. The temperature is kept constant at $T = 20 \pm 1\celsius$.
Figure \[fig:eta\_plasma\] shows the shear viscosity of blood plasma at room temperature ($RT=20\celsius)$, along with viscosity data of whole blood (average of three donors), two dilute PEO solutions, and a reference curve of pure water. Within the reliable data range (above the grey shaded area), the plasma sample as well as the 50 ppm PEO in $15wt\%$ glycerol-water solution show Newtonian behavior, that is constant shear viscosity $\eta$. By contrast the addition of a physiologic amount of RBCs (Hct50%) significantly increases the viscosity of the PEO solution. The solution exhibits shear thinning and the flow curve can be fitted to the Carreau model [@Tanner1985]. The viscosity of full blood is higher than the viscosity of the $PEO50_{15/85-Hct50}$ solution. It is known that plasma proteins lead to a depletion induced formation of aggregates of RBCs at low shear rates, and this effect might be less pronounced for the PEO solutions [@Baskurt2003]. Table \[tab:visc\_lambda\] summarizes the estimated zero shear viscosities $\eta_0$. We should mention that our determination of $\eta_0$ for whole blood is of limited use only because blood is a yield stress fluid (yield point was found to be approximately 5 mPa).
----------------------- ---------------- ----------------------------------- -----------------------------
temperature shear viscosity relaxation time
$T (\celsius)$ $\eta_0 (\milli\pascal\;\second)$ $\lambda_C (\milli\second)$
Whole Blood 37 16.9 $7.8 \pm 0.6$
Plasma 20 1.95 $2.6 \pm 0.2$
Plasma 37 1.34 $1.5 \pm 0.2$
$PEO500_{15/85}$ 20 2.47\* $15.7 \pm 0.2$
$PEO50_{15/85}$ 20 1.40 $2.6 \pm 0.1$
$PEO50_{55/45}$ 20 7.88 $7.9 \pm 0.2$
$PEO50_{15/85-Hct50}$ 20 8.03\* $1.84 \pm 0.04$
$BSA_{H_2O}$ 20 1.24 Newt.
$PBS_{15/85-Hct50}$ 37 7.6\* Newt.
H$_2$O 20 0.97 Newt.
----------------------- ---------------- ----------------------------------- -----------------------------
: \[tab:visc\_lambda\] (Zero) shear viscosities $\eta_0$ and CaBER relaxation times $\lambda_C$ of the sample solutions. Values marked with \* are obtained by a fit based upon the Carreau model [@Tanner1985].
![\[fig:CaBER\] (Color online) Width $h(t)$ of the capillary bridge as a function of time $t-t_\text{final}$. Markers on the right vertical axis refer to the images shown in Figure \[fig:thinning\]. Straight lines are exponential fits.](fig2.eps){width="1\columnwidth"}
The relaxation time of the fluids is obtained using a custom made capillary breakup extensional rheometer (CaBER) setup [@Zell2010; @Gier2011]. The thinning of the capillary bridge is recorded using a high speed camera and 4x magnification objective (Pentax, PLAN 4x/0.10, WD=30). We find that the solutions have relatively short relaxation times, and a slow retraction method is used to allow for a quantitative evaluation of the data [@Campo-Deano2010]. For a non-Newtonian fluid, the initial inertio-capillar thinning of the capillary bridge is followed by a formation of a straight filament, which thins exponentially in time. The relation $h(t) = h_0 \cdot \exp\left(-(t-t_0)/\lambda_C\right)$ gives the characteristic CaBER relaxation time $\lambda_C$, where $h(t)$ is the diameter of the filament at the time $t$ during the exponential thinning process, which starts at time $t_0$ and width $h_0$ respectively. According to the Oldroyd-B model, a simple approach leads to an estimate for the polymer relaxation time $\lambda=\lambda_C/3$ [@Stelter2000; @Anna2001].
Figure \[fig:thinning\] shows a series of images of the thinning process of the capillary bridges of the plasma as well as the BSA and $PEO50_{15/85}$ solutions. The capillary bridge of the BSA solution shows the same dynamics as the distilled water (not shown), within experimental resolution. In this case, the capillary bridge breaks fast and leaves a well known satellite droplet. By contrast, the plasma and the $PEO50_{15/85}$ solution form a viscoelastic thread that thins exponentially in time (see Figure \[fig:CaBER\]). Both solutions also show the typical blistering instability at the very end of the thinning process characterized by the appearance of smaller droplets on the filament [@Sattler2008]. These results clearly demonstrate that plasma has viscoelastic properties.
Figure \[fig:CaBER\] shows the minimal width $h(t)$ of the capillary bridge as a function of time $t-t_\text{final}$. Starting at a width of approximately $h_1\approx0.5\ \milli\meter$, all curves initially describe a uniform thinning down to an intermediate level of $h_2\approx0.17\ \milli\meter$. At this stage, polymers in the solution do not yet affect the flow and the difference between the solutions can be only seen during further thinning of the filament. While the capillary bridge of the Newtonian sample breaks very rapidly (cp. width $h_3$), the other samples form a filament which thins exponentially over characteristic time scales $\lambda_C$. Table \[tab:visc\_lambda\] summarizes the corresponding relaxation times as averages over all performed measurements. Note that the BSA solution does not show any non-Newtonian behavior and therefore we can exclude that the protein surface layer of the plasma is responsible for the elastic thinning behavior (in agreement with ), but some other plasma proteins in the bulk that have not yet been identified are responsible for the viscoelasticity. This result is also corroborated in a cone-plate geometry with a free surface layer: adding $0.01 wt\%$ of the soluble nonionic surfactant (Tween 20, critical micellar concentration $cmc = 0.07 wt\%$ to the plasma did indeed suppress the apparent shear thinning (data not shown) in agreement with the experiments described in Ref. [@Jaishankar2011]). Please note that in principle surfactants can lead to viscoelastic filaments themselves but in our case the surfactant dissolved in pure water did not show any filament [@Craster2005; @Craster2009; @Liao2006; @Liao2004; @Kwak2001; @McGough2006; @Matar2002; @Jin2006; @Hansen2000]. However, the addition of a surfactant to the plasma did *not* alter the thinning process in the CaBER which again shows that the filament is not caused by a protein surface layer. We note that additional thinning experiments of the plasma solution were performed in which the surrounding air was replaced by silicon oil and a pronounced filament was as well observed (data not shown).
Again, we can compare our polymer solutions that should serve as model systems with the plasma samples. The data presented in Table \[tab:visc\_lambda\] shows that the $PEO50_{15/85}$ solution is a good plasma replacement solution because it matches the plasma shear and elongational properties; by simply increasing the solvent viscosity of this solution we get the $PEO50_{55/45}$ sample that matches the elongational properties of full blood but it does not match its shear rheology very well. Therefore, we compared a PBS buffer solution with $50vol\%$ hematocrit without polymer ($PBS_{15/85-Hct50}$), with polymer ($PEO50_{15/85-Hct50}$) and full blood. The polymeric hematocrit solutions reflects the elongational properties of blood to some extent. However, the $PBS_{15/85-Hct50}$ solution simply breaks-up like a Newtonian liquid. This clearly shows that the non-Newtonian elongational viscosity (i.e. the elastic properties of the macromolecules) of the plasma contributes to the non-Newtonian elongational viscosity of blood. Again, an additional clustering of the RBC’s in the whole blood due to the rouleaux formation might explain the remaining discrepancies.
Next, the effects of this apparent viscoelasticity of plasma flow are investigated in a microfluidic contraction-expansion device by measuring the pressure drop across the contraction part of the geometry [@Rodd2005; @Rodd2007]. This geometry has a significant elongation flow component and it resembles flow restricted vessels or the flow at a branching from a larger to a smaller vessel. The microchannels are made of PDMS using standard soft lithography methods. Their length, width and depth are $L=30\ \milli\meter$, $W=400\ \micro\meter$ and $h=50\ \micro\meter$ respectively. These rectangular microchannels exhibit sharp contraction of reduced width $w=25\ \micro\meter$ and length $l=100\ \micro\meter$ in the middle. Note that the ratio of the expansion width to the contraction width is 1:16 which sets the (Hencky) elongation strain $\epsilon=\ln(16)\approx2.8$. Pressure measurements are performed by directly incorporating two pressure taps (Honeywell 26PC) into the microchannel, both upstream and downstream of the contraction plane. The pressure sensors are used to cover a differential pressure range of $0 < \triangle P < 56.5$ kPa. This type of measurements have been successfully applied to investigate the flow of viscoelastic fluids in expansion-contraction microfluidic devices [@Rodd2005].
Results for the pressure drop of both human plasma and water as a function of flow rate, or equivalently Reynolds number ($Re$) are shown in Figure 4. Note that here $Re = \rho V D_h / \eta_0$ where $V$ is the mean flow velocity, $\rho$ the fluid density, $\eta_0$ the zero shear rate viscosity and $D_h=2wh/(w+h)$ the hydraulic diameter. For example, Figure 4 (inset) shows that the water pressure drop increases linearly as the flow rate is increased. On the other hand, the blood plasma pressure drop shows a deviation from this linear trend. This nonlinear behavior can also be observed by normalizing the pressure drop at a given flow rate by the pressure drop of water at the same flow rate. Figure 4 shows that the dimensionless pressure drop of plasma decreases significantly as the $Re$ increases. This indicates an extensional strain-rate thinning viscosity behavior that has been observed in other types of complex fluids such as micellar solutions [@Bhardwaj2007]. These measurements clearly show that blood plasma is a non-Newtonian fluid. We also checked by flow visualization that the increase in resistance is not caused by a viscoelastic instability, but we found that in principle the viscoelastic behavior of plasma could be responsible for viscoelastic instabilities, especially in complex flow situations and in the presence of red blood cells, i.e. $50\%$ hematocrit [@SM].
![(Color online) Dimensionless pressure drop as a function of Reynolds number. The data shows higher values of pressure drop for human plasma compared to water. Blood plasma shows a nonlinear, ’thinning’ behavior of pressure drop as a function of strain-rate [@Bhardwaj2007]. Inset: Pressure drop as a function of flow rate. Blood plasma shows nonlinear behavior.[]{data-label="fig:pressure"}](fig4_v1.eps){width="0.95\columnwidth"}
*In conclusion*, we found a pronounced viscoelastic behavior of human blood plasma in a capillary breakup extensional rheometer (CaBER). Similar to dilute polymer solution, steady shear flow or small amplitude oscillatory shear does not indicate any elasticity of the plasma because it stretches polymers less efficiently. There are several reports on viscoelasticity of plasma in shear flow but they all could be attributed to the formation of a protein surface layer which is not the case in our measurements. We could also show that the elongational properties of blood are to a large extend determined by elongational properties of the plasma proteins. In order to check the generality of our findings we prepared a plasma replacement fluid with similar elongational and shear properties. Differential pressure measurements in a microfluidic contraction-expansion geometry confirmed the non-Newtonian character of plasma. Supplementary measurements with flow visualization indicate that this viscoelasticity might even lead to viscoelastic flow instabilities. Similarly, it is known that small amounts of polymers with elongational properties as our plasma replacement solution can lead to turbulent drag reduction, i.e. turbulent instabilities due to high inertia might be suppressed by the elongational properties of the plasma. Finally, recent numerical results indicate that a slight viscoelasticity of the solvent might lead to a pronounced cell depletion layer close to the vessel boundaries [@Pranay2011]. In view of these findings, it is expected that the viscoelasticity of plasma has to be taken into account in future studies.
This work was supported by the DFG Graduate college GRK1276 “Stucture formation and transport in complex systems”. L. Pan and P. E. Arratia are supported by NSF-CBET-0932449. We thank Anke Lindner and Alexander Morozov for carefully reading the manuscript.
[32]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ]{} **, @noop [Ph.D. thesis]{}, () @noop [****, ()]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study smooth complex projective varieties $X$ of maximal Albanese dimension and of general type satisfying $\chi(X, \cO_X)=0$. We prove that the Albanese variety of $X$ has at least three simple factors. Examples were constructed by Ein and Lazarsfeld, and we prove that in dimension 3, these examples are (up to abelian étale covers) the only ones. By results of Ueno, another source of examples is provided by varieties $X$ of maximal Albanese dimension and of general type satisfying $h^0(X, K_X)=1$. Examples were constructed by Chen and Hacon, and again, we prove that in dimension 3, these examples are (up to abelian étale covers) the only ones. We also formulate a conjecture on the general structure of these varieties in all dimensions.'
address:
- |
Taita Institute for Mathematical Sciences\
National Center for Theoretical Sciences, Taipei Office\
and Department of Mathematics\
1 Sec. 4, Roosevelt Rd. Taipei 106, Taiwan
- |
Département Mathématiques et Applications\
UMR CNRS 8553\
École Normale Supérieure\
45 rue d’Ulm, 75230 Paris cedex 05, France
- |
Max-Planck-Institut für Mathematik\
Vivatsgasse 7, 53111 Bonn, Germany
author:
- Jungkai Alfred Chen
- Olivier Debarre
- Zhi Jiang
title: Varieties with vanishing holomorphic Euler characteristic
---
[^1]
Introduction
============
A smooth complex projective variety $X$ is said to have [*maximal Albanese dimension*]{} if its Albanese mapping $X\to \operatorname{Alb}(X)$ is generically finite (onto its image).
Green and Lazarsfeld showed in [@gl1] that such a variety satisfies $\chi(X, \omega_X)\ge 0$. Ein and Lazarsfeld later constructed in [@el] a smooth projective threefold $X$ of maximal Albanese dimension and of general type with $\chi(X, \omega_X)= 0$ (see Examples \[eel\] and \[chh\]).
We are interested here in describing the structure of varieties $X$ of maximal Albanese dimension (and of general type) with $\chi(X, \omega_X)=0$. This class of varieties is stable by modifications, étale covers, and products with any other variety of maximal Albanese dimension (and of general type). More generally, if $X$ is a smooth projective variety of maximal Albanese dimension with a fibration whose general fiber $F$ satisfies $\chi(F, \omega_F)= 0$, then $\chi(X, \omega_X)= 0$ ([@hp], Proposition 2.5).
So we study smooth projective varieties $X$ of general type with $\chi(X,\omega_X)=0 $ and a generically finite morphism $X\to A$ to an . In §\[strthe\], we prove a general structure theorem (Theorem \[q1\]) which implies among other things that [*$A$ has at least three simple factors.*]{} Examples where $A$ is the product of any three given non-zero factors can be constructed following Ein and Lazarsfeld, and we speculate that their construction should (more or less) describe all cases where $A$ has three simple factors but, although we prove several results in §\[3sc\] in this direction (Propositions \[l35\], \[surj\], and \[l46\]) and arrive at the rather rigid picture (\[d4\]), we are only able to get a complete description when $X$ has dimension 3: we prove that [*a smooth projective threefold $X$ of maximal Albanese dimension and of general type satisfies $\chi(X,\omega_X)=0$ if and only if it has an abelian étale cover which is an Ein-Lazarsfeld threefold*]{} (Theorem \[3cur\]).
Another source of examples is provided by varieties $X$ of maximal Albanese dimension and $h^0(X,\omega_X)=1$: it follows from work of Ueno ([@ueno]) that they satisfy $\chi(X,\omega_X)=0$. Chen and Hacon constructed examples of general type (see Example \[chh\]). We gather some properties of these varieties in §\[p11\]. However, this class of examples is not stable under étale covers and does not lend itself well to our methods of study, [*except in dimension 3,*]{} where the precise Theorem \[3cur\] allows us to give [*a complete description of all smooth projective threefolds $X$ of maximal Albanese dimension and of general type, such that $P_1(X)=1$: they are all modifications of abelian étale covers of Chen-Hacon threefolds*]{} (Theorem \[th63\]).
In §\[s7\], we propose a conjecture on the possible general structure of smooth projective varieties $X$ of maximal Albanese dimension and of general type satisfying $\chi(X, \omega_X)=0$. It seems difficult to give a complete classification, but based on the examples that we know, we conjecture that, after taking modifications and étale covers, there should exists a non-trivial fibration $X\tto Y$ which is either isotrivial, or whose general fiber $F$ satisfies $\chi(F, \omega_F)= 0$. For the converse, one does have $\chi(X, \omega_X)=0$ in the second case by [@hp], Proposition 2.5, but not necessarily in the first case, of course. Both cases do happen (Example \[eell\]).
We work over the field of complex numbers.
[**Acknowledgements.**]{} The first-named author is partially supported by NCTS and the National Science Council of Taiwan. This work started during the second-named author’s visit to Taipei under the support of the bilateral Franco-Taiwanese Project Orchid and continued during the first-named author’s visit to Institut Henri Poincaré in Paris and the third-named author’s stay at the Max Planck Institut for Mathematics in Bonn. The authors are grateful for the support they received on these occasions.
Notation and preliminaries {#notpre}
==========================
For any smooth projective variety $X$, we set $\PX=\operatorname{Pic}^0(X)$. For $\xi\in\PX$, we will denote by $P_\xi$ an algebraically trivial line bundle on $X$ that represents $\xi$.
Following standard terminology, we will say that a morphism $f:X\to A$ to an $A$ is [*minimal*]{} if the induced group morphism $\Pf:\PA\to \PX$ is injective. Equivalently, $f(X)$ generates $A$ as an algebraic group and $f$ factors through no non-trivial abelian étale covers of $A$. The Albanese mapping $a_X$ has this property. Any $f:X\to A$ factors as $f:X\xrightarrow{f'}A'\to A$, where $A'$ is an and $f'$ is minimal.
An [*algebraic fibration*]{} (or simply a fibration) is a surjective morphism between normal projective varieties, with connected fibers.
In the rest of this section, $X$ will be a smooth projective variety, of dimension $n$, with a [*generically finite*]{} morphism $f:X\to A$ to an $A$. In particular, $X$ has maximal Albanese dimension.
Cohomological loci {#21}
------------------
For each integer $i$, we define the cohomological loci $$\begin{aligned}
V_i(\omega_X,f)&=&\{\xi\in\PA\mid H^i(X,\omega_X\otimes f^*P_\xi)\ne0\}\\
&=&\{\xi\in\PA\mid H^{n-i}(X, f^*P_{-\xi})\ne0\}.\end{aligned}$$ If $Y$ is a smooth projective variety and $\eps:Y\tto X$ is birational, we have $R^j\eps_*\omega_Y=0$ for $j>0$ and $\eps_*\omega_Y\isom \omega_X$ ([@ko1], Theorem 2.1), hence $ \chi(X,\omega_X) = \chi(Y,\omega_Y)$ and $V_i(\omega_X,f)=V_i(\omega_Y,f\circ\eps)$ for all $i$. In particular, these loci do not change when $X$ is replaced with $Y$.
### {#211}
Since $R^jf_*\omega_X=0$ for $j>0$, we have for all $i$ $$V_i(\omega_X,f )= V_i(f_*\omega_X ):=\{\xi\in\PA\mid H^i(A,f_*\omega_X\otimes P_\xi)\ne0\}.$$
### {#212}
Each irreducible component of $V_i(\omega_X ,f)$ is an abelian subvariety of $\PA$ of codimension $\ge i $ ([@el], Remark 1.6 and Theorem 1.2) translated by a torsion point ([@sim]).
### {#213}
There is a chain of inclusions ([@el], Lemma 1.8) $$\label{inc}
\operatorname{Ker}(\Pf)=V_n(\omega_X,f)\subset V_{n-1}(\omega_X,f)\subset \cdots\subset V_0(\omega_X,f)\subset \PA,$$ and $\operatorname{codim}\bigl(V_n(\omega_X,f) \bigr)\ge n$.
### {#214}
If $V_0(\omega_X,f)$ has a component of codimension $i$, this component is contained in (hence is an irreducible component of) $V_i(\omega_X,f)$ ([@el], (1.10)), so that we have $i\le n$ and $f(X)$ is fibered by $i$-dimensional abelian subvarieties of $A$ ([@el], Theorem 3).
### {#215}
For $\xi\in \PA$ general, $\chi(X,\omega_X)= h^0(X,\omega_X\otimes f^*P_\xi)\ge 0$ (use \[212\]) and $$\begin{aligned}
\qquad\chi(X,\omega_X)=0 & \Longleftrightarrow&V_0(\omega_X,f)\ne \PA\\
& \Longleftrightarrow&\hbox{$V_0(\omega_X,f)$ has a component of codimension $i$}\\
&&\hskip 3cm\hbox{for some $i\in\{1,\dots,n-1\}$}\\
& \Longrightarrow&\hbox{$V_i(\omega_X,f)$ has a component of codimension $i$}\\
&&\hskip 3cm\hbox{for some $i\in\{1,\dots,n-1\}$},
\end{aligned}$$ where the last implication is not always an equivalence.
### {#221}
The variety $X$ is of general type if and only if $V_0(\omega_X,f)$ generates $\PA$ ([@CH], Theorem 2.3).
### {#222}
If $V_0(\omega_X,f)$ is finite, Ein and Lazarsfeld proved that $X$ is birational to an ([@CH], Theorem 1.3.2). In particular, if $f$ is moreover minimal, it is a birational isomorphism.
Composing $f$ with a generically finite morphism {#23}
------------------------------------------------
If $Y$ is smooth projective and $h:Y\tto X$ is surjective and generically finite, the trace map $h_*\omega_Y\tto\omega_X$ splits the natural inclusion $\omega_X\to h_*\omega_Y $, hence $H^i(X,\omega_X\otimes f^*P_\xi )
$ injects into $ H^i(X, h_*\omega_Y\otimes f^*P_\xi)$ for all $\xi\in\PA$. Thus, using also \[211\], we obtain $$V_i(\omega_X,f)\subset V_i(h_*\omega_Y,f)=V_i(\omega_Y,f\circ h ).$$ Morover, $$\label{chi}
\chi(Y,\omega_Y)\ge \chi(X,\omega_X) .$$ When $h$ is étale, we have $$\chi(Y,\omega_Y) =\deg(h) \chi(X,\omega_X).$$ Finally, when $h$ is obtained from an isogeny $\eta:B\tto A$ as in the cartesian diagram $$\xymatrix
{Y=X\times_AB\ar@{->>}[d]_h\ar[r]^-{g}\ar@{}[dr]|{\square}&B\ar@{->>}[d]^\eta\\
X\ar[r]_f&A
}$$ we have $V_i(\omega_Y,g)=\widehat\eta(V_i(\omega_X,f))$. Combining this with \[212\], we see that after making a suitable étale base change, we can always make all the components of $V_i(\omega_X,f)$ pass through 0.
Components of $V_i$ of codimension $i$ {#strthe}
======================================
Let $X$ be a smooth projective variety with a generically finite morphism $f:X\to A$ to an . If $\chi(X,\omega_X)=0 $, it follows from \[215\] that $V_i(\omega_X,f) $ has a component of codimension $i$ for some $i\in\{1,\dots,n-1\}$. We prove a structure theorem under this weaker assumption.
\[q1\] Let $X$ be a smooth projective variety of dimension $n$, let $A$ be an , and let $f:X\to A$ be a minimal generically finite morphism. Assume that for some $i\in\{0,\dots,n\}$, the locus $V_i(\omega_X,f) $ has a component $V $ of codimension $i$ in $\PA$. Let $B$ be the abelian variety $ \PV $, let $K:=\operatorname{Ker}(A\tto B)^0$, and assume $f(X)+K=f(X)$. For a suitable modification $X'$ of an abelian étale cover of $X$, the Stein factorization of the morphism $X'\to A\tto B$ induces a surjective morphism $X'\tto Y$ where $Y$ is smooth of dimension $n-i$, of general type, with $\chi(Y,\omega_Y)>0$.
\[remx\] The condition $f(X)+K=f(X)$ holds:
- when $f$ is surjective;
- when $V$ is also a component of $V_0(\omega_X,f) $ ([@el], proof of Theorem 3); this applies in particular when $\chi(X,\omega_X)=0 $ (\[215\]).
By (\[212\]) and §\[23\], we may assume, after isogeny, that $A= B \times K$ and $V= \PB $. Let $p : A \tto B$ be the projection. Considering the Stein factorization of $\pi =p \circ f:X \to B$, and replacing $X$ by a suitable modification, we may assume that $\pi $ factors as $$\pi :X \stackrel{g}{\tto} Y \stackrel{h}{\to} B,$$ where $Y$ is smooth, $h$ is generically finite, and $g$ is surjective with connected fibers. Since $f(X)+K=f(X)$, the image of $\pi$ has dimension $\dim(X)-\dim(K)$, hence general fibers of $g$ have dimension $\dim(K)=i$.
We then have ${R^ig}_* \omega_X \isom \omega_Y$ ([@ko1], Proposition 7.6). Moreover, the sheaves $R^kg_* \omega_X$ on $Y$ satisfy the generic vanishing theorem ([@hp], Theorem 2.2), hence $$V_j( R^kg_*\omega_X,h) \ne \PB\quad\hbox{ for all $j>0 $ and all $k$.}$$ For all $$\xi \in \PB\moins \bigcup_{j>0,\ k} V_j( R^kg_*\omega_X,h) ,$$ we have $$H^j(Y,R^kg_*\omega_X\otimes h^*P_\xi )=0\quad\hbox{ for all $j>0$ and all $k$.}$$ Hence, by the Leray spectral sequence, we obtain $$h^i(X,\omega_X\otimes f^* P_\xi)=h^i(X,\omega_X \otimes\pi^*P_\xi)
=h^0(Y,R^ig_*\omega_X \otimes h^*P_\xi )
=h^0(Y,\omega_Y\otimes h^*P_\xi)$$ and these numbers are non-zero because $ \PB =V\subset V_i(\omega_X,f)$. In particular, $V_0(\omega_Y,h )=\PB$. By \[221\] and \[215\], $Y$ is of general type and $\chi(Y,\omega_Y)>0$. This completes the proof.
We prove a partial converse to Theorem \[q1\]: assume that there is a generically finite morphism $f:X\to A$ and a quotient abelian variety $A\tto B$ such that $f(X)+K=f(X)$, where $K:=\operatorname{Ker}(A\tto B)^0$, and denote by $X\dra Y\to B$ a modification of the Stein factorization of $X\to A\tto B$, where $Y$ is smooth of dimension $n-i$ (we set $i:=\dim(K)$).
\[r5\] In this situation, if $Y$ is not birational to an , $V_j(\omega_X,f)$ has a component of codimension $j$ for some $j\in\{i,\dots,n-1\}$.
Replacing $X$ with a modification of an étale cover (which is allowed by §\[21\] and §\[23\]), we may assume that we have a factorization $$\label{sit}
f: X\stackrel{(g,k)}{\llraa} Y\times K\stackrel{h\times {\rm Id}_K}{\llra} B\times K,$$ where $(g,k)$ is surjective and $h:Y\to B$ is generically finite of degree $>1$. We obtain, as in the proof of Theorem \[q1\], for $\xi$ general in $\PB$, $$\label{sitt}
h^i(X,\omega_X\otimes f^*P_\xi)=h^0(Y,R^ig_* \omega_X\otimes h^*P_\xi)=h^0(Y, \omega_Y\otimes h^*P_\xi).$$
[*If $\chi(Y,\omega_Y)>0$*]{}, we have $V_0(\omega_Y,h)=\PB$, the number on the right-hand-side of (\[sitt\]) is non-zero for all $\xi$, hence $V_i(\omega_X,f)$ contains the $i$-codimensional abelian subvariety $\PB$ of $\PA$.
[*If $\chi(Y,\omega_Y)=0$*]{}, since $Y$ is not birational to an , $V_0(\omega_Y,h)$ has (by \[222\] and \[215\]) a component of codimension $l\in\{1,\dots,n-i-1\}$ in $\PB$. Thus, by Remark \[remx\], we can apply Theorem \[q1\] to $h:Y\to B$: after taking an étale cover and a modification, $h$ factors through a morphism $Y\to Z\times C$, where $C$ is an of dimension $l$, $\chi(Z,\omega_Z)>0$, and $\dim(Z)=n-i-l$. We are therefore reduced to the first case and we conclude again that $V_{i+l}(\omega_X,f)$ contains an $(i+l)$-codimensional component.
\[r33\] Under the hypotheses of Theorem \[q1\] and the assumption $A= B \times K$ made in its proof, we obtain a surjective morphism $k:X\stackrel{f}{\to} B\times K\stackrel{p_2}{\tto}K$ and, from its Stein factorization, morphisms $$k:X \stackrel{l}{\tto} Z \stackrel{m}{\tto} K,$$ where $Z$ is smooth of dimension $i$, $m$ is generically finite, and $l$ has connected (generically $(n-i)$-dimensional) fibers. We have again $R^{n-i}l_*\omega_X\isom \omega_Z$ and, for $\xi$ general in $\PK$, $$h^{n-i}(X,\omega_X\otimes f^*P_\xi)=h^0(Z, \omega_Z\otimes m^*P_\xi).$$ Then,
- either $V_0(\omega_Z,m)\subsetneq \PK$ and $\chi(Z,\omega_Z)=0$;
- or $V_0(\omega_Z,m)= \PK$ and $\chi(Z,\omega_Z)>0$, in which case $\PK$ is contained in (hence is a component of) $V_{n-i}(\omega_X,f)$. There is a surjective generically finite map $X\tto Y\times Z$, hence $\chi(X,\omega_X)\ge \chi(Y,\omega_Y)\chi(Z,\omega_Z)>0$ (this also follows from Corollary \[two\].a) below).
Finally, if $F$ is a general fiber of $l:X\tto Z$, there is a surjective generically finite map $F
\tto Y$, hence $\chi(F,\omega_F)>0$ (see (\[chi\])).
We now deduce some consequences of Theorem \[q1\] on the possible components of $V_0(\omega_X,f) $ and the number of simple factors of the $A$.
\[two\] Let $X$ be a smooth projective variety with $\chi(X,\omega_X)=0 $ and a generically finite morphism $f:X\to A$ to an .
- The locus $V_0(\omega_X,f) $ does not have complementary components.[^2]
- If $X$ is in addition of general type, $A$ has at least three simple factors.
If $V_0(\omega_X,f) $ has complementary components $V_1,\dots,V_r$, with duals $B_1,\dots,B_r$, the image $f(X)$ is stable by translation by $\prod_{j\ne i}B_j$ for each $i$ (Remark \[remx\]), hence $f$ is surjective if $r\ge 2$. We obtain from Theorem \[q1\], after passing to an étale cover and a modification of $X$, a generically finite surjective map $X\tto Y_1\times\dots\times Y_r$, with $\chi(Y_i,\omega_{Y_i})>0$ for all $i$. Since $\chi(X,\omega_X)\ge \prod_i \chi(Y_i,\omega_{Y_i}) $ (by (\[chi\])), this is absurd. This proves a). Item b) then follows from \[221\].
\[dim1\] Let $X$ be a smooth projective variety of general type, with $\chi(X,\omega_X)=0 $ and a generically finite morphism $f:X\to A$ to an . Then $V_0(\omega_X, f)$ has no 1-dimensional components.
Assume $V_0(\omega_X, f)$ has a one-dimensional component and write it as $\tau_1+ \PB_1$ for some torsion point $\tau_1\in \PA$ and some quotient elliptic curve $A\tto B_1$. By [@jia], Proposition 1.7, and since $V_0(\omega_X, f)$ generates $\PA$ (\[221\]), $\PB_1$ cannot be maximal for inclusion: more precisely, there must exist two maximal components (in the sense of [@jia], Definition 1.6) $\tau_2+\PB_2$ and $\tau_3+\PB_3$ of $V_0(\omega_X, f)$ with $\PB_1\subsetneq \PB_j\subsetneq \PA$ for each $j\in\{2,3\}$ and $ \PB_2\ne \PB_3$. There are corresponding factorizations $A\tto B_j\tto B_1$.
As in the proof of Theorem \[q1\], after passing to an étale cover and a modification of $X$, we may assume $\tau_1=\tau_2=\tau_3=0$ and that we have
- Stein factorizations $$X\stackrel{g_i}{\lraa} Y_i\xrightarrow{h_i} B_i,$$ where $h_1$, $h_2$, and $h_3$ are generically finite, $Y_1$, $Y_2$, and $Y_3$ are smooth, $Y_1$ is a curve of genus $\ge 2$, and $\chi(Y_2,\omega_{Y_2})$ and $\chi(Y_3,\omega_{Y_3})$ are both positive (Theorem \[q1\]),
- a commutative diagram $$\xymatrix{
&Y_2\ar@{->>}[dr]^{h_{21}}\ar[rr]^{h_2}&&B_2\\
X\ar@{->>}[dr]_{g_3}\ar@{->>}[ur]^{g_2} \ar@{->>}[rr]^{g_1}&&Y_1 \\
&Y_3\ar@{->>}[ur]_{h_{31}}\ar[rr]_{h_3}&&B_3 .}$$
We may further assume that the induced morphism $ X\to Y_2\times_{Y_1} Y_3$ factors as: $$\xymatrix{
&&&Y_2\ar@{->>}[dr]^{h_{21}}\ar[rr]^(.6){h_2}&&B_2\\
X\ar[r]\ar@{->>}@/_2pc/[rrrr]_(.4){g_1}&Y\ar@{->>}[r]_-{\eps}\ar@{->>}@/^1.5pc/[rrr]^(.35){q}&Y_2\times_{Y_1} Y_3\ar@{->>}[dr]|(.53){\vbox to 2mm{\hglue 1mm}}\ar@{->>}[ur]|(.43){\vbox to 2mm{\hglue 1mm}} \ar@{->>}[rr] &&Y_1 ,\\
&&&Y_3\ar@{->>}[ur]_{h_{31}}\ar[rr]_(.6){h_3}&&B_3 }$$ where $\eps$ is a resolution of singularities.
Now take $\xi_2\in \PB_2$ and $\xi_3\in \PB_3$. By [@mor], Lemma 4.10.(ii),[^3] there is an inclusion $$q_*(\omega_{Y /Y_1}\otimes \eps^*(h_2^*P_{\xi_2}\otimes h_3^*P_{\xi_3}))\subset h_{21*}(\omega_{Y_2/Y_1}\otimes h_2^*P_{\xi_2})\otimes h_{31*}(\omega_{Y_3/Y_1}\otimes h_3^*P_{\xi_3}),$$ of locally free sheaves of the same rank on the curve $Y_1$. Moreover, we saw during the proof of Theorem \[q1\] that for $j\in\{2,3\}$, we have $$0\ne h^0(Y_j, \omega_{Y_j}\otimes h_j^*P_{\xi_j})=h^0(Y_1, h_{j1*}(\omega_{Y_j}\otimes h_j^*P_{\xi_j})).$$ It follows that the sheaf $h_{j1*}(\omega_{Y_j}\otimes h_j^*P_{\xi_j})$ is non-zero, hence so is the sheaf $h_{j1*}(\omega_{Y_j/Y_1}\otimes h_j^*P_{\xi_j})$. All in all, we have obtained that the locally free sheaf $q_*(\omega_{Y /Y_1}\otimes \eps^*(h_2^*P_{\xi_2}\otimes h_3^*P_{\xi_3}))$ is non-zero.
Assume now that $\xi_2 $ and $\xi_3 $ are torsion. By [@V], Corollary 3.6,[^4] this vector bundle is nef, hence has non-negative degree. Since $Y_1$ is a curve of genus $\ge 2$, the Riemann-Roch theorem then implies $$0\ne h^0(Y_1, q_*(\omega_{Y }\otimes \eps^*(h_2^*P_{\xi_2}\otimes h_3^*P_{\xi_3})))=h^0(Y, \omega_Y\otimes \eps^*(h_2^*P_{\xi_2}\otimes h_3^*P_{\xi_3})).$$ Finally, note that both $X$ and $Y$ have maximal Albanese dimensions. This implies that $\omega_{X/Y}$ is effective, hence $h^0(X, \omega_X\otimes f^*( P_{\xi_2}\otimes P_{\xi_3}))$ is also non-zero. It follows that $\xi_2+\xi_3 $ is in $V_0(\omega_X,f)$, which therefore contains $ \PB_2+\PB_3$. This contradicts the fact that $ \PB_2$ is maximal.
Case when $A$ has three simple factors {#3sc}
======================================
Ein and Lazarsfeld constructed an example of a smooth projective threefold $X$ of maximal Albanese dimension and of general type with $\chi(X,\omega_X)=0 $, whose Albanese variety is the product of three elliptic curves. After presenting their construction (and a variant due to Chen and Hacon), we prove some general results when $\operatorname{Alb}(X)$ has three simple factors. In the next section, we will show that the Ein-Lazarsfeld example is essentially the only one in dimension 3 (Theorem \[3cur\]).
\[eel\] Let $E_1$, $E_2$, and $E_3$ be elliptic curves and let $\rho_j : C_j \tto E_j$ be double coverings, where $C_j$ is a smooth curve of genus $\ge 2$ and $\rho_{j*}\omega_{C_j}\isom\cO_{E_j}\oplus \delta_j$. Denote by $\iota_j$ the corresponding involution of $C_j$. Let $A = E_1 \times E_2 \times E_3$, and consider the quotient $Z$ of $C_1 \times C_2 \times C_3$ by the involution $\iota_1
\times \iota_2 \times \iota_3$ and the tower of Galois covers: $$C_1 \times C_2 \times C_3 \stackrel{g}{\lraa}Z \stackrel{f}{\lraa}A$$ of degrees $2$ and $4$ respectively. Observe that $Z$ has rational singularities and is minimal of general type. Let $\eps: X\tto Z$ be any desingularization. The Albanese map of $X$ is $a_X=f\circ \eps$ and $$a_{X*}\omega_X\isom \cO_A\oplus (L_1\otimes L_2)\oplus (L_3\otimes L_1)\oplus (L_2\otimes L_3),$$ where $L_j$ is the inverse image of $\delta_j$ by the projection $A\tto E_j$, hence $$\label{aX}
V_0(\omega_X, a_X)=V_1(\omega_X, a_X)
=(\widehat E_1\times \widehat E_2\times\{0\})\cup(\widehat E_1\times\{0\}\times \widehat E_3)\cup(\{0\}\times\widehat E_2\times \widehat E_3),$$ whereas $V_2(\omega_X, a_X)=V_3(\omega_X, a_X)=\{0\} $.
This provides three-dimensional examples. Obviously, the same construction works starting from double coverings $\rho_j : X_j \tto A_j$ of abelian varieties with smooth ample branch loci and provides examples in all dimensions $\ge 3$. One can also extend it to any [*odd*]{} number $2r+1$ of factors and get examples where the Albanese mapping is birationally a $(\Z/2\Z)^{2r}$-covering.
\[chh\] A variant of the construction above was given by Chen and Hacon. Keeping the same notation, choose points $\xi_j\in\PE_j$ of order 2 and consider the induced double étale covers $C'_j\tto C_j$, with associated involution $\sigma_j$, and $E'_j\tto E_j$. The involution $\iota_j$ on $C_j$ pulls back to an involution $\iota'_j$ on $C'_j$ (with quotient $E'_j$). Let $Z'$ be the quotient of $C'_1\times C'_2\times C'_3$ by the group of automorphisms generated by $\operatorname{id}_1\times \sigma_2\times\iota'_3$, $\iota'_1\times\operatorname{id}_2\times \sigma_3$, $\sigma_1\times\iota'_2\times\operatorname{id}_3$, and $\sigma_1\times\sigma_2\times\sigma_3$, and let $\eps': X'\tto Z'$ be a desingularization. There is a morphism $f':X'\tto A$ of degree 4, the Albanese map of $X'$ is $a_{X'}=f'\circ \eps'$, and $$\label{e5}
a_{X'*}\omega_{\widetilde X'}\isom \cO_A\oplus (L_1\otimes L^\xi_2\otimes P_{\xi_3})\oplus (L^\xi_1 \otimes P_{\xi_2}\otimes L_3 )\oplus (P_{\xi_1}\otimes L_2\otimes L^\xi_3),$$ where $L^\xi_j=L_j\otimes P_{\xi_j}$. In particular, $P_1(X')=1$, and $$V_0(\omega_{X'}, a_{X'})=V_1(\omega_{X'}, a_{X'})=\{0\}\cup (\PE_1\times \PE_2\times\{\xi_3\})\cup(\PE_1\times\{\xi_2\}\times \PE_3)\cup(\{\xi_1\}\times\PE_2\times \PE_3).$$ Of course, the étale cover $E'_1\times E'_2\times E'_3\tto E_1\times E_2\times E_3$ pulls back to an étale cover $X''\to X'$, where $X''$ is an Ein-Lazarsfeld threefold.
Again, this construction still works starting from double coverings of abelian varieties with smooth ample branch loci, providing examples in all dimensions $\ge 3$, and for any [*odd*]{} number $2r+1$ of factors, providing examples where the Albanese mapping is birationally a $(\Z/2\Z)^{2r}$-covering.
\[l35\] Let $X$ be a smooth projective variety of general type with $\chi(X,\omega_X)=0 $ and a generically finite morphism $f:X\to A$ to an $A$ with exactly three simple factors $A_1$, $ A_2$, $ A_3$.
1. The map $f$ is surjective.
2. After passing to an abelian étale cover, we may assume $A=A_1\times A_2\times A_3$ and that $\PA_1\times \PA_2\times \{0\}$, $\PA_1\times
\{0\}\times \PA_3$, and $ \{0\}\times \PA_2\times \PA_3$ are irreducible components of $V_0(\omega_X,f)$.
We begin with the proof of item b). As in the proof of Theorem \[q1\], we may assume, after passing to an abelian étale cover, that $A$ is $A_1\times A_2\times A_3$ and, by Corollary \[two\].a) and \[221\], that $\PA_1\times \PA_2\times \{0\}$ and $\PA_1\times
\{0\}\times \PA_3$ are irreducible components of $V_0(\omega_X,f)$.
Assume that the projection $V_0(\omega_X,f)\to \PA_2\times \PA_3$ is not surjective. For $\xi_2$ and $\xi_3$ general torsion points in $ \PA_2$ and $ \PA_3$ respectively, we then have $$(\xi_2+\xi_3+\PA_1)\cap V_0(\omega_X,f)=\vide.$$ Consider the morphism $f_1=p_1\circ f:
X\to A_1$ and the sheaf $\cE=f_{1*}(\omega_X\otimes f_2^*P_{\xi_2}\otimes f_3^*P_{\xi_3})$ on $A_1$. By [@hp], Theorem 2.2, the cohomological loci of $\cE$ satisfy the chain of inclusions \[213\]. On the other hand, for all $\xi\in \PA_1$, we have $H^0(A_1, \cE\otimes P_\xi)=0 $, hence $V_0(\cE)=\vide$. It follows that for all $\xi\in \PA_1$ and all $i\ge 0$, we have $H^i(A_1, \cE\otimes P_\xi)=0 $. This implies $\cE=0$ by Fourier-Mukai duality ([@muk2]). But the rank of $\cE$ is at least $ \chi(F_1,\omega_{F_1})$, where $F_1$ is a component of a general fiber of $f_1$ ([@hp], Corollary 2.3) and this is impossible: $F_1$ is of general type and generically finite over $A_2\times A_3$, hence $ \chi(F_1,\omega_{F_1})>0$ (Corollary \[two\].b)).
The projection $V_0(\omega_X,f)\tto \PA_2\times \PA_3$ is therefore surjective. Since $A_1$ is simple, this implies that, after passing to a (split) étale cover of $A$, there are morphisms $u_2: \PA_2\to \PA_1$ and $u_3: \PA_3\to \PA_1$ such that $$\{u_2(\xi_2)+u_3(\xi_3)+\xi_2+\xi_3\mid \xi_2\in \PA_2,\ \xi_3\in \PA_3\}$$ is a component of $V_0(\omega_X,f)$. Composing this cover with the automorphism $(a_1,a_2,a_3)\mapsto (a_1,a_2-\widehat u_2(a_1),a_3-\widehat u_3(a_1))$ of $A$, we obtain b).
Item a) then follows from the fact that $f(X)$ is stable by translation by each $A_i$ (Remark \[remx\]) hence is equal to $A$.
\[r44\] As shown by considering the product with a curve of genus $\ge 2$ of any variety $X$ of general type and maximal Albanese dimension with $\chi(X,\omega_X)=0 $, the conclusion of Proposition \[l35\].a) does not hold in general as soon as $A$ has at least four simple factors.
\[surj\] Let $X$ be a smooth projective variety of general type of dimension $n$ with $\chi(X,\omega_X)=0 $ and a generically finite morphism $ X\to A$ to an $A$ with exactly three simple factors. We have:
1. $q(X)=n$;
2. the general fiber $F$ of any non-constant fibration $X\tto Y$ satisfies $\chi(F,\omega_F)>0 $;
3. any morphism from $X$ to a curve of genus $\ge 2$ is constant;
4. $V_{n-1}(\omega_X,a_X)=\{0\}$.
Let us prove b) first. The fiber $F$ is of general type and is generically finite but not surjective over $A$, hence $\chi(F,\omega_F)>0 $ by Proposition \[l35\].a). The other items then follow from the lemma below.
\[le46\] Let $X$ be a smooth projective variety of dimension $n$, of maximal Albanese dimension, of general type, with $\chi(X, \omega_X)=0$. Assume that the general fiber $F$ of any non-constant fibration $X\tto Y$ satisfies $\chi(F,\omega_F)>0 $. Then,
1. the Albanese mapping $a_X $ is surjective ($q(X)=n$);
2. any morphism from $X$ to a curve of genus $\ge 2$ is constant;
3. $V_{n-1}(\omega_X,a_X)=\{0\}$.
Item a) follows from [@ch2], Theorem 4.2, and item b) from [@hp], Theorem 2.4. Let us prove c).
If $V_{n-1}(\omega_X,a_X)\moins\{0\} $ is non-empty, it contains a torsion point by \[212\], which defines a connected étale cover $\pi: \widetilde{X}\tto X$ such that $q(\widetilde{X})>q(X)=n$. By [@ch2], Theorem 4.2, again, there exists a non-constant fibration $ \widetilde{X}\tto Y$ with general fiber $F$ of maximal Albanese dimension, of general type, with $\chi(F, \omega_F)=0$, such that $a_{\widetilde{X}}(F)$ is a translate of a fixed abelian subvariety $\widetilde K$ of $\operatorname{Alb}(\widetilde{X})$ and $\dim (\widetilde K)=\dim (F)<\dim (X)$. We consider the image $K$ of $\widetilde K$ by the induced map $\operatorname{Alb}(\pi):\operatorname{Alb}(\widetilde X)\tto \operatorname{Alb}(X)$ and the commutative diagram: $$\xymatrix@C=40pt@M=6pt{
F\ar@{->>}[r]^{\pi\vert_F}\ar@{_{(}->}[d]&\pi(F)\ar@{_{(}->}[d]\ar@{->>}[r]^-{ a_{\widetilde{X}}\vert_{\pi(F)}}& K +x_F \ar@{_{(}->}[d]\\
\widetilde{X}\ar@{->>}[d]_{f}\ar@{->>}[r]^{\pi}& X\ar@{->>}[r]^-{a_X}\ar@{->>}[dr]_-h& \operatorname{Alb}(X)\ar@{->>}[d]\\
Y&& \operatorname{Alb}(X)/ K .}$$ The map $a_{\widetilde{X}}\vert_{\pi(F)}$ is generically finite, hence $\dim (K)=\dim\pi(F)=\dim (F)$, and $$\dim (\operatorname{Alb}(X)/ K ) =n-\dim(K)=n-\dim(F).$$ It follows that $\pi(F)$ is a general fiber of the Stein factorization of $h$. Since $\chi(\pi(F), \omega_{\pi(F)})=0$, this contradicts our hypothesis on $X$.
Assume now that we are in the situation of Proposition \[l35\].b) and consider, as in Remark \[r33\], the maps $f_i:=p_i\circ f:X\tto A_i$. Since $ \chi(X,\omega_X) =0$, we are in case a) of that remark, hence, with the notation therefrom, $V_0(\omega_Z,m)$ must be finite, because $A_i$ is simple. If $f$ is minimal, so is $f_i$ and it follows from \[222\] that $Z$ is birational to $A_i$, hence $f_i $ is a fibration. A general fiber $F_i$ satisfies $ \chi(F_i,\omega_{F_i})>0$ by Proposition \[surj\].b).
\[l46\] Let $X$ be a smooth projective variety of general type with $\chi(X,\omega_X)=0 $ and a minimal generically finite morphism $f:X\to A$ to an $A$ product of three simple factors $A_1$, $ A_2$, and $ A_3$.
Let $\{i,j,k\}=\{1,2,3\}$. For $\xi_j$ and $\xi_k$ general torsion points in $ \PA_j$ and $ \PA_k$ respectively, the sheaf $f_{i*}(\omega_X\otimes f_j^*P_{\xi_j}\otimes f_k^*P_{\xi_k})$ on $A_i$ is locally free, homogeneous, of positive rank $ \chi(F_i,\omega_{F_i})$.
We follow the proof of [@ch2], Corollary 2.3. As in the proof of Proposition \[l35\], since $\xi_j$ and $\xi_k$ are torsion, the cohomological loci of $\cE:=f_{i*}(\omega_X\otimes f_j^*P_{\xi_j}\otimes f_k^*P_{\xi_k})$ satisfy the chain of inclusions \[213\]. On the other hand, since $\xi_j$ and $\xi_k$ are general and $A_i$ is simple, the intersection $$(\xi_j+\xi_k+\PA_i)\cap V_0(\omega_X,f)$$ is finite by \[212\], hence so is $V_0(\cE)$. It follows that all $V_l(\cE)$ are finite, hence $ \cE $ is locally free and homogeneous ([@muk2], Example 3.2). Its rank is $h^0(F_i, \omega_{F_i}\otimes f_j^*P_{\xi_j}\otimes f_k^*P_{\xi_k})=\chi(F_i,\omega_{F_i})
$.
Recall that a homogeneous vector bundle is a direct sum of twists of unipotent vector bundles (successive extensions of trivial line bundles) by algebraically trivial line bundles which, in our case, are torsion by Simpson’s theorem (or rather its extension [@hp], Theorem 2.2.b)).
[*When $A_i$ is an elliptic curve,*]{} the sheaf of Proposition \[l46\] is actually a direct sum of torsion line bundles (this is explained at the bottom of page 362 of [@ko3] when $\xi_j=\xi_k=0$ and holds in general by the étale covering trick).
Finally, from the proof of Theorem \[q1\], we have (after replacing $X$ with a suitable modification), for each $\{i,j,k\}=\{1,2,3\}$, Stein factorizations $$p_{jk}\circ f:X\stackrel{g_i}{\llraa} S_i\stackrel{(h_{ij}, h_{ik})}{\lllraa} A_j\times A_k,$$ where $S_i$ is smooth of general type with $\chi(S_i,\omega_{S_i})>0$ (this follows also from Corollary \[two\].b)). Since $f_j$ has connected fibers, so does $h_{ij}:S_i\tto A_j$.
All in all, we have for each $\{i,j,k\}=\{1,2,3\}$ a commutative diagram: $$\label{d4}
\xymatrix@C=40pt{
&&X\ar@{->>}[dl]^{g_i}\ar@{->>}@/_1.5pc/[ddll]_{f_j}\ar@{->>}[dr]_{g_j}\ar@{->>}@/^1.5pc/[ddrr]^{f_i}\ar@{->>}[dd]_{f_k} \\
&S_i\ar@{->>}[dr]_{h_{ik}}\ar@{->>}[dl]^{h_{ij}}&&S_j\ar@{->>}[dl]^{h_{jk}} \ar@{->>}[dr]_{h_{ji}}\\
A_j&&A_k &&
A_i, }$$ where all the morphisms are fibrations.
Are the $h_{ij}$ isotrivial? Are the $f_i$ isotrivial? Is $X$ rationally dominated by a product $X_1\times X_2\times X_3$, where $X_i$ dominates and is generically finite over $A_i$? We are inclined to think that the answers to all these questions should be affirmative, but we were only able to go further in the case where the $A_i$ are all elliptic curves.
The 3-dimensional case {#s3}
======================
We now come to our main result, which completely describes all smooth projective threefolds $X$ of maximal Albanese dimension and of general type, with $\chi(X,\omega_X)=0$.
\[3cur\] Let $X$ be a smooth projective threefold of maximal Albanese dimension and of general type, with $\chi(X,\omega_X)=0$.
There exist elliptic curves $E_1$, $ E_2$, and $ E_3$, double coverings $C_j\tto E_j$ with associated involutions $\iota_j$, and a commutative diagram $$\xymatrix@C=15pt
{&(C_1\times C_2\times C_3)/\iota_1\times \iota_2\times\iota_3 \ar@{->>}[dr]\\
\widetilde X\ar@{->>}[rr]^{a_{\widetilde X}}\ar@{}[drr]|\square\ar@{->>}[ur]^-\eps\ar@{->>}[d]&& E_1\times E_2\times E_3\ar@{->>}[d]^\eta\\
X\ar@{->>}[rr]_{a_X} &&\operatorname{Alb}(X), }$$ where $\eta$ is an isogeny and $\eps$ a desingularization.
In other words, up to abelian étale covers, the Ein-Lazarsfeld examples (Example \[eel\]) are the only ones (in dimension 3)! Note also that $a_X$ is not finite, but that it is finite on the canonical model of $X$.
\[c52\] Under the hypotheses of the theorem, the Albanese mapping of $X$ is birationally a $(\Z/2\Z)^2$-covering.
With the notation of Example \[eel\], we set $ A:= E_1\times E_2\times E_3$ and $\widetilde X_i:=\operatorname{Spec}(\cO_A\oplus L^\vee_i)$, so that $f$ factors through the double coverings $\widetilde f_i:=\widetilde X_i\tto A$. Since the action of $ \operatorname{Ker}(\eta)$ on $ A$ by translations lifts to $\widetilde X$, it leaves $$a_{\widetilde X*}\cO_{\widetilde X}\isom \cO_A\oplus (L^\vee_1\otimes L^\vee_2)\oplus (L^\vee_3\otimes L^\vee_1)\oplus (L^\vee_2\otimes L^\vee_3)$$ invariant, hence also each $L^\vee_i$. It follows that this action lifts to each $\widetilde X_i$, hence $\widetilde f_i$ descends to a double covering $X_i\tto \operatorname{Alb}(X)$ through which $a_X$ factors.
By Proposition \[surj\].a), $a_X$ is surjective and $q(X)=3$ (see also [@ch2], Corollary 4.3). Moreover, by Corollary \[two\].b), $\operatorname{Alb}(X)$ is isogeneous to the product of three elliptic curves and, after passing to étale covers, we may assume that $a_X$ can be written as $$a_X: X\stackrel{(f_1,f_2,f_3)}{\lllraa} E_1\times E_2\times E_3,$$ where each $f_i: X\tto E_i$ is a fibration, and that $\PE_1\times \PE_2\times \{0\}$, $\PE_1\times
\{0\}\times \PE_3$, and $ \{0\}\times \PE_2\times \PE_3$ are irreducible components of $V_0(\omega_X,a_X)$ (Proposition \[l35\].b)).
Let $\{i,j,k\}=\{1,2,3\}$. As in §\[3sc\], we have a commutative diagram (\[d4\]), where each $S_i$ is a smooth minimal surface of general type and $A_i=E_i$. The proof of the theorem is very long, so we will divide it in several steps. The general scheme of proof goes as follows:
- In the diagram (\[d4\]), the fibrations $h_{ij}:S_i\tto E_j$ are all isotrivial (Step 1); we let $C_{ij}$ be a (constant) general fiber.
- There exist finite groups $G_i$ acting on $C_{ij}$ such that $C_{ij}/G_i\isom E_j$ and $C_{ik}/G_i\isom E_k$, the surface $S_i$ is birational to $(C_{ij}\times C_{ik})/G_i$, and $h_{ij}$ and $h_{ik}$ are the two projections (Step 2).
- At this point, it is quite easy to show that $X$ dominates a threefold $Y$ which is dominated by a product of 3 curves (Step 3).
- Taking an étale cover of $X$, we may assume that all the irreducible components of $V_0(\omega_X, a_X)$ pass through $0$. We then show (Step 4) that $V_0(\omega_X, a_X)$ has the same form as the corresponding locus of an Ein-Lazarsfeld threefold (see (\[aX\])), from which we deduce that $X$ is birationally isomorphic to $Y$ (Step 5) hence is also dominated by a product of 3 curves.
- Using the fact that $V_0(\omega_X, a_X)$ has no “extra” components, we finish the proof by showing that the groups $G_i$ all have order 2 (Step 6).
[**Step 1.**]{} [*The fibrations $h_{ij}:S_i\tto E_j$ are all isotrivial.*]{}
We will denote by $C_{ij}$ a general (constant) fiber of $h_{ij}$.
By the semi-stable reduction theorem ([@KKMS], Chapter II), there exist a finite cover $h: C\tto E_j$, where $C$ is a smooth curve and commutative diagrams (for each $\alpha\in\{i,k\}$) $$\xymatrix{S'_\alpha\ar@{->>}[dr]_{h_\alpha}\ar@{->>}[r]^-{\eps_\alpha}
& C\times_{E_j}S_\alpha\ar@{->>}[d]\ar@{->>}[r] &S_\alpha\ar@{->>}[d]^{h_{\alpha j}}\\
& C\ar@{->>}[r]^{h} &E_j, }$$ where $\eps_\alpha$ is a modification and the fibers of $h_\alpha$ are all reduced connected curves, with non-singular components crossing transversally.
We also make a modification $\tau : X'\to
C\times_{E_j}X$ such that there exists a commutative diagram of morphisms between smooth varieties: $$\xymatrix@C=40pt{
&&X' \ar@{->>}[dl]^{g'_i}\ar@{->>}@/_1.5pc/[ddll]_{f'_k}\ar@{->>}[dr]_{g'_k}\ar@{->>}[dd]_(0.6){f'}\ar@{->>}@/^1.5pc/[ddrr]^{f'_i} \\
&S'_i\ar@{->>}[dl]^{h'_{ik}} \ar@{->>}[dr]_{h_i}&&S'_k\ar@{->>}[dl]^{h_k}\ar@{->>}[dr]_{h'_{ki}}\\
E_k&&C&&E_i.}$$
Let $\xi_\alpha\in \PE_\alpha$. By [@V], Lemma 3.1,[^5] we have an inclusion $$\label{4}
f'_*(\omega_{X'/C}\otimes {f'_i}^*P_{\xi_i}\otimes
{f'_k}^*P_{\xi_k})\subset h^*f_{j*}(\omega_X\otimes f_i^*P_{\xi_i}\otimes
f_k^*P_{\xi_k})$$ of locally free sheaves on $C$. Since $h_\alpha$ is flat with irreducible general fibers, $S'_i\times_C S'_k$ is irreducible, and we have a surjective morphism $$g'_{ik}: X'\stackrel{(g'_i, g'_k)}{\lllraa} S'_i\times_C S'_k.$$ Moreover, $h_i$ and $h_k$ are semistable, hence by [@AK], Proposition 6.4, $S'_i\times_CS'_k$ has only rational Gorenstein singularities. After further modification of $X'$, we may assume that $g'_{ik}$ factors through a desingularization of $S'_i\times_CS'_k$: $$g'_{ik}: X'\tto Y'_{ik}\stackrel{\eps}{\tto} S'_i\times_CS'_k.$$ By definition of rational Gorenstein singularities, we have $\eps_*\omega_{Y'_{ik}}=\omega_{S'_i\times_CS'_k}$. Since $\omega_{X'/Y'_{ik}}$ is effective, we obtain an inclusion $$p_i^*(\omega_{S'_i/C}\otimes {h'_{ik}}^*P_{\xi_k})\otimes p_k^*(\omega_{S'_k/C}\otimes {h'_{ki}}^*P_{\xi_i})\subset
{g'_{ik*}}(\omega_{X'/C}\otimes {f'_i}^*P_{\xi_i}\otimes
{f'_k}^*P_{\xi_k})$$ of sheaves on $S'_i\times_CS'_k $. Pushing forward these sheaves to $C$, we obtain $$\begin{aligned}
\label{nn}
h_{i*}(\omega_{S'_i/C}\otimes
{h'_{ik}}^*P_{\xi_k})\otimes h_{k*}(\omega_{S'_k/C}\otimes
{h'_{ki}}^*P_{\xi_i})
&\subset& f'_*(\omega_{X'/C}\otimes
{f'_i}^*P_{\xi_i}\otimes {f'_k}^*P_{\xi_k}) \nonumber\\
&\subset& h^*f_{j*}(\omega_X\otimes
f_i^*P_{\xi_i}\otimes f_k^*P_{\xi_k}),\end{aligned}$$ where the second inclusion comes from (\[4\]). Let $\{\alpha,\beta\}=\{i,k\}$. Both sheaves $h_{\alpha*}(\omega_{S'_ \alpha/C}\otimes
{h'_{\alpha \beta}}^*P_{\xi_\beta})$ are nef ([@V], Corollary 3.6). On the other hand, for $\xi_i$ and $\xi_k$ general and torsion, the sheaf in (\[nn\]) has degree 0 by Proposition \[l46\], hence $$\deg(h_{\alpha*}(\omega_{S'_ \alpha/C}\otimes
{h'_{\alpha \beta}}^*P_{\xi_\beta}))=0.$$
By [@K1], Corollary 10.15, both sheaves $R^1h_{\alpha*}(\omega_{S'_ \alpha/C}\otimes
{h'_{\alpha \beta}}^*P_{\xi_\beta})$ are torsion-free and generically 0, hence 0.[^6] On the other hand, we have $
R^1h_{\alpha*}\omega_{S'_\alpha/C} =\cO_C
$ ([@ko1], Proposition 7.6) hence, by the Grothendieck-Riemann-Roch theorem, $$[\operatorname{ch}(h_{\alpha*}(\omega_{S'_\alpha/C}))-\operatorname{ch}(\cO_C)]\operatorname{Td}(C)=\operatorname{ch}(h_{\alpha*}(\omega_{S'_ \alpha/C}\otimes
{h'_{\alpha \beta}}^*P_{\xi_\beta}))\operatorname{Td}(C)$$ in the ring of cycles modulo numerical equivalence on $C$. This implies $\deg(h_{\alpha*}\omega_{S'_\alpha/C}) =0$, and $h_\alpha$ is locally trivial (see [@BPV], Theorem III.17.3). Hence $h_{\alpha j}$ is isotrivial (for each $\alpha\in\{i,k\}$).
[**Step 2.**]{} [*There exist finite groups $G_i$ acting on $C_{ij}$ such that $C_{ij}/G_i\isom E_j$ and $C_{ik}/G_i\isom E_k$, the surface $S_i$ is birational to the quotient $(C_{ij}\times C_{ik})/G_i$ for the diagonal action of $G_i$, and $h_{ij}$ and $h_{ik}$ are identified with the two projections.* ]{}
This is a consequence of the following (probably classical) result.
\[l53\] Let $S$ be a smooth projective surface with an isotrivial fibration $h_1:S\tto \Gamma_1$ onto an irrational curve with (constant) irrational general fiber $F_1$.
[a)]{} There exist a smooth curve $F_2$ and a finite group $H$ acting faithfully on $F_1$ and $F_2$ such that $\Gamma_1$ is isomorphic to $F_2/H$, the surface $S$ is birationally isomorphic to the diagonal quotient $(F_1\times F_2)/H$, and $h_1$ is the composition $S\operatorname{\stackrel{{}_{\scriptstyle\sim}}{\dra}}(F_1\times F_2)/H \tto F_2/H\isom \Gamma_1$. Let $h_2$ be the composition $S\operatorname{\stackrel{{}_{\scriptstyle\sim}}{\dra}}(F_1\times F_2)/H \tto F_1/H $.
[b)]{} Assume $S$ is of general type. Any isotrivial fibration $h:S\tto \Gamma$ onto an irrational curve $\Gamma$ is either $h_1$ or $h_2$ followed by an isomorphism between $F_1/H$ or $ F_2/H$ with $\Gamma$.
Item a) is well-known and can be found in [@Ser]. Let us prove b). Since $\Gamma$ is irrational, $h$ induces an isotrivial fibration $h':(F_1\times F_2)/H \tto \Gamma$. Let $D_2$ be a general (constant irrational) fiber of $h'$. The quotient map $\pi: F_1\times F_2\tto (F_1\times F_2)/H$ is étale outside a finite set. Hence the Stein factorization $g$ of $h'\circ \pi$ in the diagram $$\xymatrix{F_1\times F_2\ar@{->>}[dr]_g\ar@{->>}[r]^-{\pi} &(F_1\times F_2)/H\ar@{->>}[r]^-{h'} & \Gamma\\
&D_1,\ar@{->>}[ur]}$$ is also isotrivial, with general fiber $D_2'$ a (fixed) étale cover of $D_2$. By a), there is a base change $D_1'\tto D_1$ and a surjective morphism $t=(t_1, t_2): D_1'\times D_2'\tto F_1\times F_2$. Since $S$ is of general type, $F_1$ and $F_2$ are each of genus $\ge 2$, hence each $t_i$ must factor through one of the projections $p_j:D_1'\times D_2'\tto D'_j$.
If $h$ factors through neither $h_1$ nor $h_2$, the curve $D_2'$ dominates both $F_1$ and $F_2$, hence $t_1$ and $t_2$ cannot factor through $p_1$. Thus they must factor through $p_2$, which contradicts the fact that $t$ is surjective.
Let now $Y_j$ be a resolution of singularities of the irreducible threefold $S_i\times_{E_j} S_k$ and let $Y$ be a resolution of singularities of the component of $Y_1\times_{E_2\times E_3 }S_1$ that dominates both $Y_1$ and $Y_3$. After modification of $X$, we obtain a diagram $$\label{bigd}
\xymatrix@R=10pt{
&X\ar@{->>}[rr]^{g}&&Y\ar@{->>}[dr] \ar@{->>}[dl]\\
&&Y_1\ar@{->>}[dl]_{g_{13}}\ar@{->>}[dr]^{g_{12}}&&Y_3\ar@{->>}[dl]_{g_{32}}\ar@{->>}[dr]^{g_{31}}\\
&S_3\ar@{->>}[dr]_{h_{31}}\ar@{->>}[dl]^{h_{32}}&\square&S_2\ar@{->>}[dl]^{h_{21}} \ar@{->>}[dr]_{h_{23}}& \square&S_1\ar@{->>}[dr]_{h_{13}} \ar@{->>}[dl]^{h_{12}}\\
E_2&&E_1 &&
E_3&&E_2, }$$ where the squares are birationally cartesian and the isotrivial morphisms $h_{ij}:S_i\tto E_j $ fit into diagrams $$\xymatrix{
&C_{ik}\times C_{ij}\ar[r]^-{p_2}\ar@{-->}[d]^{/G_i}&C_{ij}\ar@{->>}[d]^{/G_i}_{\rho_{ij}}\\
C_{ik}\ar@{^{(}->}[r]^{\rm fiber}&S_i\ar@{->>}[r]_{h_{ij}}&E_j.}$$
[**Step 3.**]{} [*The threefold $Y$ is dominated by a product of three curves.* ]{}
The dominant maps $C_{31}\times C_{32}\dra S_3$ and $C_{21}\times C_{23}\dra S_2$ induce a factorization $$((\rho_{31}, \rho_{21}), \rho_{32}, \rho_{23}): (C_{31} \times_{E_1}C_{21} ) \times C_{32} \times C_{23}\dra S_3\times_{E_1} S_2\tto E_1\times E_2\times E_3.$$ The (Stein factorization of the) morphism $Y_1\tto E_2\times E_3$ is therefore isotrivial (its fibers are dominated by the curve $C_{31} \times_{E_1}C_{21} $). Thus, $Y$ is dominated by the product $$\label{stu}
(C_{31} \times_{E_1}C_{21})\times(C_{12} \times_{E_2}C_{32})\times (C_{23} \times_{E_3}C_{13})$$ of three (possibily reducible) curves.
Going back to the proof of Theorem \[3cur\], after passing to an étale cover, we may and will assume, from now on, that the following holds (§\[23\]): $$\label{propp}
\hbox{\em All the irreducible components of $V_0(\omega_X, a_X)$ pass through $0$.}$$ One checks, using §\[21\], §\[23\], and Proposition \[surj\].d), that if we want $X$ to be birationally covered by a product $\Gamma_1\times \Gamma_2\times \Gamma_3$, with morphisms $ \Gamma_i\tto E_i$, as in the conclusion of the theorem, we must have the following.
[**Step 4.**]{} [*We have $$V_0(\omega_X, a_X)=(\PE_1\times \PE_2\times \{0\})\cup
(\PE_1\times
\{0\}\times \PE_3)\cup( \{0\}\times \PE_2\times \PE_3).$$* ]{} We already know that $V_0(\omega_X, a_X)$ contains the right-hand-side (Proposition \[l35\].b)) and we must prove that it has no other components.
Assume $V_0(\omega_X, a_X)$ has another component $\PT$. It has dimension 2 (Corollary \[two\].a)) and, after possibly permuting the indices, we may assume the neutral component $\PE'_1$ of $\PT \cap (\PE_1\times\PE_2\times \{0\})$ is neither $\PE_1\times\{0\}\times\{0\}$ nor $\{0\}\times \PE_2\times\{0\}$. This yields an elliptic curve $E'_1$ which is a quotient of $E_1\times E_2$ which does not factor through either projection. As we saw right before Proposition \[l46\], the induced map $f_4: X\tto E'_1$ is a fibration. It factors as $$f_4: X \lraa S_3\stackrel{h_{34}}{\lraa} E'_1$$ where, by Step 1, $h_{34}$ is isotrivial. By Lemma \[l53\].b), $h_{34}$ must factor through one of the projections $h_{31}:S_3\tto E_1$ or $h_{32}:S_3\tto E_2$ so we reach a contradiction.
[**Step 5.**]{} [*The morphism $g:X\tto Y$ is birational.* ]{}
Consider, in the diagram (\[bigd\]), the generically finite morphism $v_1: X\tto Y_1$ and the three fibrations $f'_{\alpha }:Y_1\tto E_\alpha $, for $\alpha\in\{1,2,3\}$. Since $X$, $Y_1$, and $S_3$ are all of maximal Albanese dimensions, $\omega_{X/Y_1}$ and $\omega_{Y_1/S_3}$ are effective, hence $$h^0(X, \omega_X\otimes a_X^*P_\xi )\ge h^0(Y_1, \omega_{Y_1}\otimes (f'_1,f'_2)^*P_\xi)\ge h^0(S_3, \omega_{S_3}\otimes (h_{31},h_{32})^*P_\xi)$$ for all $\xi \in \PE_1\times \PE_2$. Moreover, for $\xi $ non-zero, we have by Proposition \[surj\].d) $$h^2(X, \omega_{X}\otimes a_X^*P_\xi)=h^3(X, \omega_{X}\otimes a_X^*P_\xi)=0$$ hence, since $\chi(X,\cO_X)=0$, $$h^0(X, \omega_{X}\otimes a_X^*P_\xi)=h^1(X, \omega_{X}\otimes a_X^*P_\xi).$$ Finally, for $\xi $ general in $\PE_1\times \PE_2$, we have, as in the proof of Theorem \[q1\], since $g_3:X\tto S_3$ has connected fibers, $$h^1(X, \omega_X\otimes a_X^*P_\xi )= h^0(S_3, \omega_{S_3}\otimes (h_{31},h_{32})^*P_\xi).$$ Therefore, for $\xi\in \PE_1\times \PE_2$ general, we obtain $$h^0(X, \omega_X\otimes a_X^*P_\xi)=h^0(Y_1, \omega_{Y_1}\otimes (f'_1,f'_2)^*P_\xi).$$ The induced morphism $ Y\tto E_1\times E_2\times E_3$ is the Albanese mapping of $Y$. Since in any event, we always have $$h^0(X, \omega_X\otimes a_X^*P_\xi)\ge h^0(Y , \omega_Y\otimes a_Y^*P_\xi)\ge h^0(Y_1 , \omega_{Y_1}\otimes (f'_1,f'_2)^*P_\xi)$$ for all $\xi\in \PE_1\times \PE_2 $, we obtain $$\label{fin}
h^0(X, \omega_X\otimes a_X^*P_\xi)= h^0(Y , \omega_Y\otimes a_Y^*P_\xi)$$ for $\xi $ general in $\PE_1\times \PE_2$, hence also, by Step 3, for $\xi $ general in $V_0(\omega_X, a_X)$. But for $\xi\notin V_0(\omega_X, a_X)$, both sides of (\[fin\]) vanish. By Lemma \[3-fm\] below, we conclude that $g$ is a birational morphism.
The following lemma (used in the proof above) is in the spirit of [@hp], Theorem 3.1.
\[3-fm\]Let $X\stackrel{g}{\tto} Y\stackrel{f}{\tto} A$ be generically finite morphisms between smooth projective threefolds, where $A$ is an abelian threefold, such that $ f$ and $f\circ g$ are both minimal. Assume that $X$ is of general type with $\chi(X, \omega_X)=0$ and that there exists an open subset $U\subset \PA$ with $\operatorname{codim}_{\PA}(\PA\moins U)\ge 2$ such that $$h^0(X, \omega_X\otimes g^*f^*P_\xi)=h^0(Y, \omega_Y\otimes f^*P_\xi)$$ for all $\xi\in U $. Then $g$ is birational.
By §\[23\], we can write $g_*\omega_X\isom \omega_Y\oplus \cE$, and we need to show that the sheaf $\cE$ is zero. Since $\cE$ is torsion-free and $f$ is generically finite, it is sufficient to prove $f_* \cE =0$.
As we saw at the beginning of §\[s3\], we have $q(X)=3$, hence $f\circ g$ is the Albanese mapping of $X$. By Proposition \[surj\].d), for each $i\in\{2,3\}$, we then have $\{0\}=V_i(\omega_X, f\circ g)=V_i(g_*\omega_X, f)$, hence $V_i(f_* \cE)\subset \{0\}$.
Since $q(X)=3$, we also have $q(Y)=3$, hence $h^i(Y, g_*\omega_X)=h^i(Y, \omega_Y)$. It follows that $V_i(f_* \cE)$ is empty.
The assumption $\chi(X, \omega_X)=0$ implies $\chi(Y, \omega_Y)=0$ by (\[chi\]). Thus, $$V_0(f_* \cE)=V_1(f_* \cE)\subset \PA\moins U.$$ Since $\operatorname{codim}_{\PA}(\PA\moins U)>1$, the sheaf $f_* \cE $ is therefore M-regular in the sense of [@pp1], Definition 2.1 (see also Remark 2.3), hence continuously ([@pp1], Definition 2.10 and Proposition 2.13). Since $H^0(A,f_* \cE\otimes P_\xi)=0$ for all $\xi\in U$, we obtain $f_* \cE =0$.
Let us summarize what we know. Let $\{i, j, k\}=\{1,2,3\}$. The curve $C_{ij}$ is the (constant) general fiber of the isotrivial fibration $S_i\tto E_k$; it is acted on by a group $G_i$ and $C_{ij}/G_i\isom E_j$ (Step 2). The fibration $g_i: X\tto S_i$ is also isotrivial; as we saw in Step 3, its general fiber $C_i$ is dominated by the curve $C_{ji}\times_{E_i}C_{ki}$ but also maps onto $C_{ji}$ and $C_{ki}$. Finally, a general fiber $F_k$ of the isotrivial fibration $f_k: X\tto E_k$ is an isotrivial fibration over $C_{ij}$ with (constant) general fiber $C_i$. The situation is summarized in the following diagram: $$\xymatrix@C=20pt@R=20pt
{
& C_{ij}\ar@{^{(}->}[rr]^-{\rm fiber} &&S_i\ar@{->>}[dr]^{h_{ik}}\\
F_k\ar@{^{(}->}[rr]^-{\rm fiber}\ar@{->>}[ur]^-{{\rm fiber}\ C_i}\ar@{->>}[dr]_-{{\rm fiber}\ C_j} & &X\ar@{->>}[ur]^(.4){g_i}\ar@{->>}[rr]^{f_k}\ar@{->>}[dr]_(.4){g_j}&& E_k\\
& C_{ji}\ar@{^{(}->}[rr]^-{\rm fiber} && S_j.\ar@{->>}[ur]_{h_{jk}}
}$$ By Lemma \[l53\], there exists a finite group $H_k$ acting faithfully on $C_i$ and $C_j$ such that $C_{ij}\isom C_j/H_k$, $C_{ji}\isom C_i/H_k$, and $F_k$ is isomorphic to the diagonal quotient $(C_i\times C_j)/H_k$. Moreover, the maps to $C_{ij}$ and $C_{ji}$ are the natural projections. So we have diagrams $$\label{cu}
\xymatrix@C=30pt@R=8pt
{& C_{ji}\ar@{->>}[dr]^{/G_j}\\
C_i\ar@{->>}[ur]^{/H_k}\ar@{->>}[dr]_{/H_j}&& E_i.\\
& C_{ki}\ar@{->>}[ur]_{/G_k}}$$
Let $D_1$ be the Galois closure of $C_1$ over $E_1$ and set $G=\operatorname{Gal}(D_1/E_1)$. Let $\{j, k\}=\{2, 3\}$. There is a normal subgroup $N_j\lhd G$ such that $G_j=G/N_j$ and $G$ acts on $C_{jk}$ via this quotient. By Step 2, the surface $S_j$ is birationally isomorphic to $(C_{j1}\times C_{jk})/G_j$, hence to $(D_1\times C_{jk})/G$. Therefore, the modification $Y_1$ of $S_2\times_{E_1}S_3$ (see (\[bigd\])) is birationally isomorphic to $(D_1\times C_{23}\times C_{32})/G$.
[**Step 6.**]{} [*The group $G$ is isomorphic to $ \Z/2\Z$.* ]{}
We begin with a lemma which is probably well-known. We denote by $\operatorname{Irr}(G)$ the set of isomorphism classes of irreducible representations of $G$.
\[l55\] Let $E$ be an elliptic curve and let $\pi:D\tto E$ be a Galois cover with group $G$. We can write $$\pi_*\cO_D= \bigoplus_{\chi\in \operatorname{Irr}(G)}\bigoplus_{i}\cV_{\chi, i} ,$$ where each vector bundle $\cV_{\chi, i}$ is semistable and $G$-invariant, and the representation of $G$ on the general fiber of each $\cV_{\chi, i}$ is a direct sum of $\chi$. Moreover, for each $\chi\ne 1$, the dual vector bundle $\cV^\vee_{\chi, i}$ is either ample or a direct sum of non-zero torsion line bundles.
The groups $G$ acts on $\pi_*\cO_D$. Identifying each representation $\chi$ with its character, we consider the endomorphism $\sum_{g\in G}\chi(g)g$ of $\pi_*\cO_D$ and we denote by $\cV^\vee_\chi$ its image. We then have $$\label{piod}
\pi_*\cO_D=\bigoplus_{\chi} \cV_\chi,$$ where the general fiber of $\cV_\chi$ is a (nonzero) direct sum of $\chi$ as a $G$-module.
The Harder-Narasimhan filtration $$0=\cV^\ell_\chi\subset \cV_\chi^{\ell-1}\subset \cdots \subset \cV_\chi^0=\cV_\chi$$ is preserved by the $G$-action. Serre duality shows that since we are on an elliptic curve, all the corresponding extensions are trivial, hence $\cV_\chi$ is the direct sum of the $G$-invariant semistable bundles $\cV_{\chi, i}=\cV_\chi^i/\cV_\chi^{i+1}$, for $0\le i< \ell$.
As a direct summand of $\pi_*\omega_D=(\pi_*\cO_D)^\vee$, each vector bundle $\cV^\vee_{\chi, i}$ is nef ([@V], Corollary 3.6). Moreover, it is ample if it has positive degree. Consider the maximal degree-$0$ subsheaf $\cF$ of $\pi_*\cO_D$, the direct sum of all $\cV_{\chi,i}$ that have degree 0. By [@KP], Lemma 3.2 and 3.4, $\cF$ is a $G$-invariant subalgebra and induces an étale cover of $E $, hence is a direct sum of torsion line bundles.
Let us continue with the Galois cover $\pi:D\tto E$ with group $G$ as in the lemma and assume moreover that for each $j\in\{2,3\}$, we have a Galois cover $\pi_j: D_j\tto E_j$ with Galois group $G_j=G/N_j$, where $g(D_j)\geq 2$ and $E_j$ is an elliptic curve.
Then $G$ acts on $D_1\times D_2\times D_3$ diagonally. Let $Z$ be the quotient, let $\eps:Y\tto Z$ be a resolution, and consider $$t: Y\stackrel{\varepsilon}{\tto} Z \tto E_1\times E_2\times E_3.$$
\[2elements\]Assume $$V_0(\omega_Y, t)\subset (\PE_1\times \PE_2\times \{0\})\cup
(\PE_1\times
\{0\}\times \PE_3)\cup( \{0\}\times \PE_2\times \PE_3)$$ and $V_2(\omega_Y, t)\cup V_3(\omega_Y, t)\subset \{0\}$. Then $N_2=N_3$ and $G_2\isom G_3\isom \Z/2\Z$.
We decompose $\pi_*\cO_D$ as in (\[piod\]) and write similarly $$\pi_{j*}\cO_{D_j}= \bigoplus_{\mu\in \operatorname{Irr}(G_j)}\bigoplus_i\cV_{\mu, i}^j.$$ Since quotient singularities are rational, we have as in Example \[eel\] $$t_*\omega_Y\isom (q_*\cO_Z)^\vee \isom \big((\pi_*\cO_D)^ \vee\boxtimes ((\pi_{2*}\cO_{D_2})^ \vee\boxtimes ((\pi_{3*}\cO_{D_3})^ \vee\big)^G.$$ Let $\mu$ be an non-trivial element of $\operatorname{Irr}(G_2)$. Since $G_2$ is a quotient of $G$, the representation $\mu$ and its complex conjugate $\overline{\mu}$ are also in $\operatorname{Irr}(G)$. Then, the vector bundle $$\cG:=(\cV^\vee_{\overline{\mu},1}\boxtimes \cV^{2\vee}_{\mu,1}\boxtimes\cO_{E_3})^G$$ on $E_1\times E_2\times E_3 $ is a non-zero direct summand of both $\cV^\vee_{\overline{\mu},1}\boxtimes \cV^{2\vee}_{\mu,1}\boxtimes\cO_{E_3}$ and $t_*\omega_Y$.
Assume that $\cV^{2}_{\mu,1}$ has degree 0, hence is a direct sum of non-trivial torsion line bundles.
1. If $\deg(\cV^\vee_{\overline{\mu},1})=0$, the sheaf $\cG$ is a direct sum of non-trivial torsion line bundles, which is impossible since $V_3(\cG)\subset V_3(\omega_Y, t)=\{0\}$.
2. If $\cV^\vee_{\overline{\mu},1}$ is ample, we can write $$\cG=\bigoplus_k (\cG_k\boxtimes P_{\xi_k}\boxtimes\cO_{E_3}),$$ where $\cG_k$ is a direct summand of $\cV^\vee_{\overline{\mu},1}$, hence ample, and the $\xi_k$ are non-zero torsion points in $\PE_2$. This is again impossible, because $V_2(\cG)\subset V_2(\omega_Y, t)=\{0\}$.
Therefore, $\cV^{j\vee}_{\mu,1}$ is ample for all $\mu$ non-trivial in $\operatorname{Irr}(G_j)$.
If $\operatorname{Card}(G_2)>2$, or if $N_2\ne N_3$, we may take non-trivial $\chi\in\operatorname{Irr}(G)$, $\mu\in\operatorname{Irr}(G_2)$, and $\nu\in\operatorname{Irr}(G_3)$ such that $\chi$ is a subrepresentation of $\mu\otimes \nu$. The vector bundle $$\cH:=(\cV^\vee_{\overline{\chi}, 1}\boxtimes \cV^{2\vee}_{\mu, 1} \boxtimes \cV^{3\vee}_{\nu, 1})^G$$ is then non-zero and a direct summand of $t_*\omega_Y$ (and $\cV^{2\vee}_{\mu, 1} $ and $ \cV^{3\vee}_{\nu, 1}$ are ample).
If $\cV^\vee_{\overline{\chi}, 1}$ is ample, since $\cH$ is a direct summand of $\cV^\vee_{\overline{\chi}, 1}\boxtimes \cV^{2\vee}_{\mu, 1} \boxtimes \cV^{3\vee}_{\nu, 1}$, we have $V_m(\cH)=\vide$ for all $m\in\{1,2,3\}$. Hence $h^0(E_1\times E_2\times E_3, \cH\otimes P_\xi)$ is a non-zero constant for all $\xi\in \PE_1\times \PE_2\times \PE_3$ and $V_0(\cH)=\PE_1\times \PE_2\times \PE_3$, which contradicts our assumptions.
If $\cV^\vee_{\overline{\chi}, 1}$ is a direct sum of non-trivial torsion line bundles, we may write $$\cH=\bigoplus_k (P_{\xi_k}\boxtimes \cH_k),$$ where the $\xi_k$ are non-zero torsion points in $\PE_1$ and $\cH_k$ is a direct summand of $\cV^{2\vee}_{\mu, 1} \boxtimes \cV^{3\vee}_{\nu, 1}$. Then $V_0(\cH)$, hence also $V_0(\omega_Y, t)$, contains $\{-\xi_1\}\times\PE_2\times\PE_3$, which contradicts our assumptions.
We now apply this second lemma to the Galois covers $\pi:D_1\tto E_1$, $\pi_2:C_{23}\tto E_2$, and $\pi_3:C_{32}\tto E_3$. The variety $Y$ of the lemma is the variety $Y_1$ of the proof, and since $V_0(\omega_{Y_1}, t)\subset V_0(\omega_X, a_X)$ (see §\[23\]), the hypotheses of the lemma are satisfied (Step 4).
We obtain $N_2=N_3$, hence the coverings $C_{ji}\tto E_i$ and $C_{ki}\tto E_i$ are the same (see (\[cu\])), and also $G/N_j \isom \Z/2\Z$, so that they are double covers. Denote them by $C'_i\tto E_i$. By the proof of Step 3 (see (\[stu\])), $X$ is birational to $ (C'_1\times C'_2\times C'_3)/( \Z/2\Z)$. Since the latter variety contains no rational curves, there is a birational [*morphism*]{} from $X$ to it. This finishes the proof of Theorem \[3cur\].
Varieties with $P_1=1$ {#p11}
======================
It follows from [@ueno] and \[215\] that varieties $X$ of maximal Albanese dimension and $P_1(X)=1$ satisfy $\chi(X,\omega_X)=0$. We presented in Example \[chh\] a construction of Chen and Hacon of such a variety which is in addition of general type. We gather here some properties of these varieties (most of them taken from [@ueno]).
Let $X$ be a smooth projective variety of maximal Albanese dimension $n$, with $P_1(X)=1$.
1. We have an isomorphism $$a_X^*:\bigwedge^\bullet H^0(A,\Omega_A)\isom H^0(X,\Omega^\bullet_X).$$ In particular, $h^j(X,\cO_X)=\binom{n}{j}$ for all $j$, hence $\chi(X,\omega_X)=0$, and the Albanese mapping $a_X:X\tto \operatorname{Alb}(X)$ is surjective.
2. The point 0 is isolated in $V_0(\omega_X,a_X)$.
Replacing $X$ with a modification, we may assume that there is a factorization $a_X:X \tto Z\to \operatorname{Alb}(X)$, where $ Z$ is a desingularization of $ a_X(X)$, so that $ P_1( Z)\le P_1(X)=1$. It follows from [@ueno] (or [@mor], Corollary (3.5)) that $a_X(X)$ is a translate of an abelian subvariety of $\operatorname{Alb}(X)$, hence $a_X$ is surjective. Item a) then follows from another result of Ueno ([@ueno], or [@mor], Corollary (3.4)).
By §\[23\], we can write $a_{X*}\omega_X \isom \omega_A \oplus \cE\isom \cO_A \oplus \cE $. The sheaf $\cE$ then satisfies $V_i(\cE)\moins\{0\}=V_i(\omega_X,a_X)\moins\{0\}$ for all $i$. Since $1=P_1(X)=1+h^0(A, \cE)$, the point $0$ is not in the closed set $ V_0(\cE)$, hence is isolated in $V_0(\omega_X,a_X)$. This proves b).
Regarding item b), to be more precise, a smooth projective variety $X$ of maximal Albanese dimension satisfies $P_1(X)=1$ [*if and only if*]{} 0 is isolated in $V_0(\omega_X, a_X)$.
\[th63\] Let $X$ be a smooth projective threefold of maximal Albanese dimension and of general type. If $P_1(X)=1$, the variety $X$ is a modification of an abelian étale cover of a Chen-Hacon threefold.
It is then very easy to describe all smooth projective threefolds $X$ of maximal Albanese dimension and of general type, with $P_1(X)=1$. Start from a Chen-Hacon threefold $Y$ as in Example \[chh\], with Albanese mapping $a_Y:Y\tto E_1\times E_2\times E_3$. It satisfies $$V_0(\omega_Y, a_Y)=\{0\}\cup (\PE_1\times \PE_2\times\{\xi_3\})\cup(\PE_1\times\{\xi_2\}\times \PE_3)\cup(\{\xi_1\}\times\PE_2\times \PE_3),$$ where each $\xi_j\in\PE_j$ has order 2. Take an isogeny $A\tto E_1\times E_2\times E_3$ corresponding to a (finite) subgroup of $\PE_1\times\PE_2\times \PE_3$ which contains none of the points $\xi_1,\xi_2,\xi_3$. Finally, take for $X$ a modification of $Y\times_{E_1\times E_2\times E_3}A$.
Replacing $a_X$ with it Stein factorization, we will assume that $X$ is normal and $a_X$ is finite. By Theorem \[3cur\], there exist elliptic curves $E_1$, $E_2$, and $E_3$, double coverings $\rho_i: C_i\tto E_i$, with involution $\iota_i$, and a commutative diagram $$\xymatrix@C=30pt
{C_1\times C_2\times C_3 \ar@{->>}[dr]^{(\rho_1,\rho_2,\rho_3)}\ar@{->>}[d]\\
\widetilde X\ar@{->>}[d]\ar@{->>}[r]^-{a_{\widetilde X}}\ar@{}[dr]|\square & E_1\times E_2\times E_3\ar@{->>}[d]^\eta\\
X\ar@{->>}[r]_{a_X} &\operatorname{Alb}(X),}$$ where $\eta$ is an isogeny (the variety $\widetilde X$ is the variety $Z$ of Example \[eel\]) and both $a_X$ and $a_{\widetilde X}$ are $( \Z/2\Z)^2 $-Galois coverings. In particular, $X$ has rational singularities.
We denote by $K$ the (finite) kernel of $\eta$ and by $K_i$ the image of the projection $K\to E_i$, so that $K$ is a subgroup of $ \widetilde K:=K_1\times K_2\times K_3$. The elliptic curve $F_i:=E_i/K_i$ embeds in $\operatorname{Alb}(X)$; let $\pi_i: \operatorname{Alb}(X)\tto A_i$ be the quotient. The natural morphism $h_i: X\tto A_i$ is an isotrivial fibration and we denote by $D_i$ its general (constant) fiber. Now we consider the square restricted to fibers: $$\begin{aligned}
\xymatrix@C=40pt{
C_i\ar@{->>}[r]_{\rho_i}^{2:1} \ar@{->>}[d]_-{\rm{\acute etale}} & E_i\ar@{->>}[d]_-{\lambda_i}^-{\rm{\acute etale}}\\
D_i\ar@{->>}[r]_{t_i}^{4:1} & F_i,
}\end{aligned}$$ where $t_i$ is a $(\Z/2\Z)^2$-cover. Since $D_i\times_{F_i}E_i$ is disconnected and factors as $C_i\sqcup C_i\tto C_i \stackrel{\rho_i}{\tto}E_i$, there is a non-zero 2-torsion point $\xi_i\in \PF_i$ such that $\xi_i\in \operatorname{Ker}(t_i^*)\cap\operatorname{Ker}(\lambda_i^*)$, the morphism $t_i$ factors as $D_i\stackrel{s_i}{\tto}D_i'\tto F_i$, where $s_i$ is a double étale cover, and $C_i\isom D_i'\times_{F_i}E_i$. It follows that the group $K_i$ acts on $C_i$, and the involution $\iota_i$ and the $K_i$-action commute. Therefore, $\widetilde K$ acts on $C_1\times C_2\times C_3$ and this action commutes with the involution $(\iota_1, \iota_2, \iota_3)$. It follows that $\widetilde K$ acts on $\widetilde{X}$ and the Albanese mapping $a_{\widetilde X}$ is $\widetilde K$-equivariant. Set $Y:=\widetilde{X}/\widetilde K$.
The quotient morphism $\widetilde{X}\tto Y $ is étale and factors as $\widetilde{X}\tto X\tto Y$. The variety $Y$ has maximal Albanese dimension, is of general type, $P_1(Y)=1$, and its Albanese variety is $F_1\times F_2\times F_3$.
We have a cartesian diagram $$\begin{aligned}
\label{d1}
\xymatrix@C=50pt
{
\widetilde{X}\ar@{}[dr]|\square\ar@{->>}[d]^{/\widetilde K}\ar@{->>}[r]^-{a_{\widetilde X}} & E_1\times E_2\times E_3\ar@{->>}[d]^{/\widetilde K}_\lambda\\
Y\ar@{->>}[r]^-{a_Y} & F_1\times F_2\times F_3.
}
\end{aligned}$$ Since the rightmost quotient morphism is étale, so is the leftmost quotient morphism. Since the top morphism is the Albanese mapping of $\widetilde X$, the bottom morphism is the Albanese mapping of $Y$. Furthermore, since $K$ is a subgroup of $\widetilde K$, the leftmost quotient morphism factors as $\widetilde{X}\tto X\tto Y$. Therefore, $P_1(Y)=1$.
We now claim that $Y$ is a Chen-Hacon threefold. By Theorem \[3cur\] (or the diagram (\[d1\])), $a_Y$ is a $(\Z/2\Z)^2$-Galois covering and by [@P], we can write $$a_{Y*}\cO_Y=\bigoplus_{\chi\in ( \Z/2\Z)^{2*}}L_{\chi}^\vee=\cO_{\operatorname{Alb}(Y)}\oplus L_{\chi_1}^\vee\oplus L_{\chi_2}^\vee\oplus L_{\chi_3}^\vee,$$ or equivalently, since $Y$ has rational singularities, $$\label{e15}
a_{Y*}\omega_Y=\cO_{F_1\times F_2\times F_3}\oplus L_{\chi_1}\oplus L_{\chi_2}\oplus L_{\chi_3},$$ where $L_{\chi_1}$, $ L_{\chi_2}$, and $ L_{\chi_3}$ are line bundles on $F_1\times F_2\times F_3$. Moreover, by [@P], Theorem 2.1, we have the following “building data”: there are effective divisors $D_1$, $D_2$, and $D_3$ on $F_1\times F_2\times F_3$ satisfying: $$L_{\chi_i}+L_{\chi_j}\siml L_{\chi_k}+D_k\quad{\text{and}}\quad L_{\chi_i}^2\siml D_j+D_k$$ for any $\{i, j, k\}=\{1, 2, 3\}$. These data pull back to the analogous building data on $\widetilde X$, hence $\lambda^*D_i$ is the pull-back on $E_1\times E_2\times E_3$ of the branch divisor $\Delta_i\siml 2\delta_i$ of $\rho_i$. It follows that there exists an ample line bundle $\delta'_i$ on $F_i$ which pulls back to $\delta_i$ on $E_i$ and such that $D_i$ is also the pull-back on $F_1\times F_2\times F_3$ of a divisor $\Delta'_i\siml 2\delta'_i$ on $F_i$. Let $L'_i$ be the pull-back on $F_1\times F_2\times F_3$ of $\delta'_i$. Because of the relations $L_{\chi_i}^2\siml D_j+D_k$, we can write $$L_{\chi_i}\isom P_{\xi_i}\otimes (L'_j\otimes P_{\xi_{i,j}})\otimes (L'_k\otimes P_{\xi_{i,k}}),$$ where $\xi_i\in \PE_i$, $\xi_{i,j}\in \PE_j$, and $\xi_{i,k}\in \PE_k$ are 2-torsion points. From (\[e15\]) and the fact that $P_1(Y)=1$, we deduce $H^0(E_i,L_{\chi_i})=0$, hence each $\xi_i$ has order 2 and is in the kernel of $\widehat \lambda_i$. From the relations $L_{\chi_i}+L_{\chi_j}\siml L_{\chi_k}+D_k$, we deduce $$\xi_{i,k}+\xi_{j,k}=\xi_k\quad{\text{and}}\quad \xi_{i,j}+\xi_j=\xi_{k,j}.$$ Since $\lambda_1^*\xi_1=0$, we may always change $L'_1$ to $L'_1\otimes P_{\xi_1}$, so we may assume $\xi_{3,1}=0$ and similarly, $\xi_{1,2}=0$ and $\xi_{2,3}=0$. The $\cO_{F_1\times F_2\times F_3}$ algebra $a_{Y*}\cO_Y$ is then the algebra associated to a Chen-Hacon threefold (see (\[e5\])). We conclude that $Y$ is a Chen-Hacon threefold.
A conjecture {#s7}
============
As mentioned in the introduction, we end this article with a conjecture on the possible general structure of smooth projective varieties $X$ of maximal Albanese dimension, of general type, with $\chi(X, \omega_X)=0$.
Let $X$ be a smooth projective variety of maximal Albanese dimension, of general type, with $\chi(X, \omega_X)=0$. Then there exist a smooth projective variety $X'$, a morphism $X'\tto X$ which is a composition of modifications and abelian étale covers, and a fibration $g:X'\tto Y$ with general fiber $F$, such that $0<\dim(Y)<\dim(X)$ and
- either $g$ is isotrivial;
- or $\chi(F, \omega_F)= 0$.
\[remm\] 1) Conversely, in the situation b) above, $\chi(X, \omega_X)= 0$ ([@hp], Proposition 2.5). Moreover, $\operatorname{Alb}(X)$ has at least 4 simple factors by Corollary \[two\].b) and Proposition \[surj\].b). Of course, in case a), without further constraints, one might have $\chi(X, \omega_X)>0$, but we were unable to find necessary and sufficient conditions on the isotorivial fibration $g$ (assuming $X$ does not fall into case b)) to ensure $\chi(X, \omega_X)=0$.
3\) If we are [*not*]{} in case b), it follows from Lemma \[le46\] that if $X'\tto X$ is any composition of modifications and abelian étale covers, we have $q( X')=\dim(X)$ and any morphism from $X'$ to a curve of genus $\ge 2$ is constant.
The Ein-Lazarsfeld example (Example \[eel\]) falls into case a) of the conjecture, and not into case b) by Remark \[remm\].1) above. We present an example that falls into case b), but not into case a). It is basically a non-isotrivial fibration whose general fibers are Ein-Lazarsfeld threefolds.
\[eell\] Consider a smooth projective curve $C$ of genus $\geq 2$, elliptic curves $E_1$, $E_2$, and $E_3$, and smooth double coverings $ S_j\tto C\times E_j$ ramified along ample divisors. Denote by $\iota_j$ the corresponding involution of $S_j$. We may moreover assume that the fibrations $f_j: S_j\tto C$ are all semistable and not isotrivial.
The fourfold $T:=S_1\times_C S_2\times_C S_3$ has only rational Gorenstein singularities, and so does its quotient $Z$ by the involution $\iota_1\times\iota_2\times\iota_3$. Let $\eps:X\tto Z$ be a desingularization. We have a diagram $$\xymatrix@C=15pt
{&&& T\ar@{->>}[dlll]\ar@{->>}[dll]\ar@{->>}[dl]\ar@{->>}[d]_g^{2:1}\\
S_1\ar@{->>}[d]^{2:1}&S_2\ar@{->>}[d]^{2:1}&S_3\ar@{->>}[d]^{2:1} &Z\ar@{->>}[d]_-f^{4:1}&X\ar@{->>}[l]_\eps\\
C\times E_1\ar@{->>}[drrr]&C\times E_2\ar@{->>}[drr]&C\times E_3\ar@{->>}[dr]&C\times E_1\times E_2\times E_3\ar@{->>}[d] \\
&&&C.}$$ The variety $X$ is of general type and has maximal Albanese dimension because $C\times E_1\times E_2\times E_3$ does. A general fiber of the fibration $ X\tto C$ is one of the examples constructed in Example \[eel\], hence $\chi(X, \omega_X)= 0$ by [@hp], Proposition 2.5, and $X$ falls into case b) of the conjecture. One can prove that it does not fall into case a).
[KKMS]{}
Abramovich, D., Karu, K., Weak semistable reduction in characteristic 0, [*Invent. Math.*]{} [**139**]{} (2000), 241–273.
Barth, W., Peters, C., Van de Ven, A., [*Compact complex surfaces,*]{} Ergebnisse der Mathematik und ihrer Grenzgebiete [**4**]{} (1984), Springer Verlag, Berlin-Heidelberg-New York.
Chen, J.A., Hacon, C.D., Pluricanonical maps of varieties of maximal Albanese dimension, [*Math. Ann.*]{} [**320**]{} (2001), 367–380.
Chen, J.A., Hacon, C.D., On the irregularity of the image of the Iitaka fibration, [*Comm. Algebra*]{} [**32**]{} (2004), 203–215.
Ein, L., Lazarsfeld, R., Singularities of theta divisors and the birational geometry of irregular varieties, [*J. Amer. Math. Soc.*]{} [**10**]{} (1997), 243–258.
Green, M., Lazarsfeld, R., Deformation theory, generic vanishing theorems, and some conjectures of Enriques, Catanese and Beauville, [*Invent. Math.*]{} [**90**]{} (1987), 389–407.
Hacon, C.D., Pardini, R., Birational characterization of products of curves of genus $2$, [*Math. Research Letters*]{} [**12**]{} (2005), 129–140.
Jiang, Z., Varieties with $q(X)=\dim(X)$ and $P_2(X)=2$, eprint [arXiv:1010.4709v1]{}, [*Ann. Scuola Norm. Sup. Pisa,*]{} to appear.
Kebekus, S., Peternell, T., A refinement of Stein factorization and deformations of surjective morphisms, [*Asian J. Math.*]{} [**12**]{} (2008), 365–389.
Kempf, G., Knudsen, F., Mumford, D., Saint-Donat, B., [*Toroidal Embeddings I,*]{} Springer Lecture Notes in Mathematics [**339**]{}, 1973.
Kollár, J., [*Shafarevich Maps and Automorphic Forms,*]{} Princeton University Press, 1995.
Kollár, J., Higher direct images of dualizing sheaves. I, [*Ann. of Math.*]{}Ê [**123**]{} (1986), 11–42.
Kollár, J., Subadditivity of the Kodaira dimension: fibers of general type, in [*Algebraic geometry, Sendai, 1985,*]{} 361–398, Adv. Stud. Pure Math. [**10**]{}, North-Holland, Amsterdam, 1987.
Mori, S., Classification of higher-dimensional varieties, in [*Algebraic geometry, Bowdoin, 1985*]{} (Brunswick, Maine, 1985), 269–331, Proc. Sympos. Pure Math. [**46**]{}, Part 1, Amer. Math. Soc., Providence, RI, 1987.
Mukai, S., Duality between $D(X)$ and $D(\hat X)$ with its application to Picard sheaves, [*Nagoya Math. J.*]{} [**81**]{} (1981), 153–175.
Pardini, R., Abelian covers of algebraic varieties, [*J. reine angew. Math.*]{} [**417**]{} (1991), 191–213.
Pareschi, G., Popa, M., Regularity on abelian varieties I, [*J. Amer. Math. Soc.*]{}Ê [**16**]{} (2003), 285–302.
Serrano, F., Isotrivial fibred surfaces, [*Ann. Mat. Pura Appl.*]{} (4) [**171**]{} (1996), 63–81.
Simpson, C., Subspaces of moduli spaces of rank one local systems, [*Ann. Sci. École Norm. Sup.*]{} [**26**]{} (1993), 361–401.
Uneo, K., [*Classification theory of algebraic varieties and compact complex spaces,*]{} notes written in collaboration with P. Cherenack, Springer Lecture Notes in Mathematics [**439**]{}, Berlin-New York, 1975.
Viehweg, E., [*Positivity of direct image sheaves and applications to famillies of higher dimensional manifolds,*]{} ICTP Lecture Notes [**6**]{} (2001), 249-284.
[^1]: O. Debarre is part of the project VSHMOD-2009 ANR-09- BLAN-0104-01.
[^2]: By that, we mean components such that the sum morphism induces an isogeny from their product onto $\PA$.
[^3]: This is stated in [@mor] for $\xi_2=\xi_3=0$, but the same proof works in general.
[^4]: This is stated there for $\xi_2=\xi_3=0$, but the general case follows by the étale covering trick.
[^5]: This is proved there for $\xi_i=\xi_k=0$, but the same proof works in general.
[^6]: Note that up to this point, the proof works in the more general situation where $a_X$ is surjective and $\operatorname{Alb}(X)$ has a 1-dimensional simple factor $E_1$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have recently applied the second law to an isolated system, consisting of a system $\Sigma$ such as a glass surrounded by an extremely large medium $\widetilde{\Sigma}$, to show that the instantaneous absolute temperature $T(t),$ thermodynamic entropy $S(T_{0},t)$ and enthalpy $H(T_{0},t)$ of $\Sigma$ decrease in any isothermal relaxation towards their respective equilibrium values $T_{0},S_{\text{eq}}(T_{0})$ and $H_{\text{eq}}(T_{0})$ under isobaric cooling. The decrease of the thermodynamic entropy and enthalpy during relaxation in vitrification is consistent with non-negative temperature $T(t)$. The Gibbs statistical entropy also conforms to the above relaxation behavior in a glass, which however is contrary to the conjecture by Gupta, Mauro and coworkers that the glass transition and the structural relaxation during vitrification are inverse to each other; this is then supported by computation in which their statistical entropy $\widehat{S}(T_{0},t)$ drops below $S_{\text{eq}}(T_{0})$ during the glass transition and then increases towards it during isothermal relaxation. However, they do not establish that the entropy loss during the glass transition is accompanied by a concomitant entropy gain of the medium to maintain the second law. These authors use a novel statistical formulation $\widehat{S}(T_{0},t)$ of entropy based on several conjectures such as it being zero for a microstate, but do not compare its behavior with the thermodynamic entropy $S(T_{0},t)$. The formulation is designed to show the entropy loss. Its subsequent rise not only contradicts our result but also implies that the glass during relaxation must have a negative absolute temperature. To understand these puzzling results and the above conjecture, we have carried out a critical evaluation of their unconventional approach. We find that the inverse conjecture is neither supported by their approach nor by the second law. The zero-entropy microstate conjecture is only consistent with $\widehat{S}(T_{0},t)\equiv0$ at all temperatures, not just at absolute zero and is found to have no scientific merit. We show that the maximum entropy gain of the medium during the glass transition invalidates the entropy loss conjecture. After pointing out other misleading, confusing and highly exaggerated statements in their work, we finally conclude that their unconventional statistical approach and computational scheme are not appropriate for glasses.'
author:
- 'P. D. Gujrati'
title: 'Is Structural Relaxation During Vitrification the Inverse of the Glass Transition?'
---
Introduction
============
Conventional Approach (CA) to Glass Transition
----------------------------------------------
In a recent paper [@Guj-NE-I], we have studied a homogeneous non-equilibrium system $\Sigma$ surrounded by an extremely large medium $\widetilde{\Sigma}$. The work has been extended to also cover inhomogeneous systems and internal variables [@Guj-NE-II]. The combined system $\Sigma_{0}$ forms an isolated system; see Fig. \[Fig\_Systems\]. We apply the *second law* to the isolated system to describe the behavior of the non-equilibrium system $\Sigma$. According to the second law [@note1]$$\frac{dS_{0}(t)}{dt}=\frac{dS(t)}{dt}+\frac{d\widetilde{S}(t)}{dt}\geq0,\label{Second_Law}$$ where $S_{0}(t),S(t)$ and $\widetilde{S}(t)$ denote the entropy of $\Sigma_{0},\Sigma$ and $\widetilde{\Sigma}$, respectively, at time $t$. In this work, quantities pertaining to $\Sigma_{0}$ have the suffix $0$, the system $\Sigma$ have no suffix, and $\widetilde{\Sigma}$ have a tilde. For $\Sigma_{0}$, all of its (additive) observables, variables that can be controlled by the observer, such as its energy $E_{0}$, volume $V_{0}$, number of particles $N_{0}$, etc. are *constant* in time. These observables also identify the macrostate of $\Sigma_{0}$. It is clear that for some homogeneous $\Sigma_{0}$, the variation of its instantaneous entropy $S_{0}(t)$ cannot be explained by the dependence of the latter on its constant observables. The variation can only be explained by assuming the dependence of $S_{0}(t)$ on (additive) internal variables, variables that cannot be controlled by the observer, that vary in time as $S_{0}(t)$ approaches its maximum value. For an inhomogeneous $\Sigma_{0}$, the variation of its entropy $S_{0}(t)$ can be explained by the way the inhomogeneity disappears [@Guj-NE-II] as it approaches equilibrium. The inhomogeneity gives rise to induced internal variables; see below. One may not require any additional internal variable. Thus, if we consider $\Sigma_{0}$ to consist of $\Sigma$ and $\widetilde
{\Sigma}$, as we do in this work, it is no longer a homogeneous system as long as $\Sigma$ is not in equilibrium with the medium $\widetilde{\Sigma}$. Thus, it is possible to consider $\Sigma_{0}$ without any internal variable, as was the case studied in [@Guj-NE-I]. We discover that the *instantaneous* values of its *fields* (temperature $T(t)$, pressure $P(t)$, etc.) are in general different from those of the medium ($T_{0},P_{0},$ etc.); see also Bouchbinder and Langer [@Langer]. But the most surprising result of the mathematical analysis is that the instantaneous conjugate variables, the entropy $S(t),$ the volume $V(t)$, etc. play the role of inhomogeneity-induced internal variables with the corresponding “affinity” [@Donder; @deGroot; @Prigogine; @Guj-NE-I; @Guj-NE-II; @Nemilov-Book], respectively, related to the deviation $T(t)-T_{0}$,$P(t)-P_{0}$, etc; see Eqs. (\[Gibbs\_Fundamental(t)\]) and (\[First\_Law(t)\]).
\[ptb\]
[System\_Modified\_1.eps]{}
As a non-equilibrium system at fixed $T_{0},P_{0}$ of the medium strives to come to equilibrium, it undergoes relaxation during which its instantaneous fields $T(t),P(t)$, etc. continue to change. At the completion of relaxation, the “affinities” vanish so that $$T(t)\rightarrow T_{0},P(t)\rightarrow P_{0},\ \ \ \ \text{etc.}$$ as expected. In an isobaric process, which is of central interest to us here, we will assume that the system is always in mechanical equilibrium so that its pressure $P(t)$ is always equal to $P_{0}$ at all temperatures and all times; however, there is normally no thermal equilibrium so that the instantaneous temperature $T(t)$ of the system is different from $T_{0}$ [@Guj-NE-I; @Guj-NE-II; @Langer]. It is found that during relaxation, the instantaneous entropy $S(t)$ of the system continues to decrease in an isobaric cooling experiment such as vitrification. The entropy of Glass1, see Fig. \[Fig\_entropyglass\],
\[ptb\]
[Fig\_EntropyGlass\_Mauro.eps]{}
approaches (see the downward arrows) that of the equilibrated supercooled liquid entropy $S_{\text{eq}}$ (shown by the solid curve) from above during relaxation $$S(T_{0},P_{0},t)\overset{\text{CA}}{\rightarrow}S_{\text{eq}}^{+}(T_{0},P_{0}),\label{Entropy_Behavior_CA}$$ where we have also exhibited the temperature of the medium; being in equilibrium (we do not consider possible crystallization here), the supercooled liquid cannot have any relaxation. It will be our practice to not exhibit $T_{0},P_{0}$, unless clarity is needee, in which case we will exhibit them. As we will mostly consider an isobaric process, we will not exhibit $P_{0}$ in the argument. In an isobaric heating experiment, the entropy will increase during relaxation. Such a behavior of the entropy during relaxation will be called the *conventional* behavior in this work and our approach the conventional approach (CA) as it follows from classical thermodynamics [@Donder; @deGroot; @Prigogine; @Guj-NE-I; @Guj-NE-II]. This then explains the symbol over the arrow in Eq. (\[Entropy\_Behavior\_CA\]) and the header of this subsection. The behavior appears to be the right behavior for the entropy in an isobaric process as the entropy must be an increasing function of the enthalpy with the slope given by$$\left( \frac{\partial S(t)}{\partial H(t)}\right) _{P}\approx\frac{1}{T(t)}\geq0,\label{Temperature(t)}$$ see Eq. (\[dS/dH\_relation\]); it is well known that the enthalpy falls (rises) during isobaric cooling (heating) [@Nemilov-Book]. The analysis in our previous work [@Guj-NE-I; @Guj-NE-II] is carried out for any non-equilibrium system. It is therefore also applicable to glasses where one must, in addition, make a distinction between fast and slow processes. We have considered the issue of the fictive temperature in glasses in our work [@Guj-NE-I], where such a distinction has been made.
Unconventional Approach (UCA) to Glass Transitions\[Sect\_UCA\]
---------------------------------------------------------------
In a series of papers [@Gupta; @Mauro0; @Mauro], Gupta, Mauro and coworkers, to be collectively denoted in short by GMc here, have developed a description of glasses without using any internal variables or any fictive temperature. Even the fictive temperature or pressure, which some people treat as an internal variable, is not considered in their theoretical development, notwithstanding the fact that they are used when GMc consider experimental data. Their conclusion is that their statistical entropy $\widehat{S}(T_{0},t)$ for the glass *increases* during relaxation towards $S_{\text{eq}}$$$\widehat{S}(T_{0},t)\overset{\text{UCA}}{\rightarrow}S_{\text{eq}}^{-}(T_{0});
\label{Entropy_Behavior_UCA}$$ see Glass2 and the portion of the upward thick arrow in Fig. \[Fig\_entropyglass\] above the horizontal dash; the latter is located at the value of the Glass2 entropy. In the following, we will use $\widehat{S}$ for the statistical entropy used by GMc to distinguish it from the thermodynamic or the Gibbs statistical entropy $S$ used in our approach. During the glass transition, the entropy $\widehat{S}(t)$ *falls* below $S_{\text{eq}}$ of the supercooled liquid in this approach. Thus, they suggest that the glass transition and relaxation are inverse processes. They also take their glass as homogeneous, just as we have done in our earlier work [@Guj-NE-I]; the difference is that their glass is identified by its temperature $T_{0}$, pressure $P_{0}$ and its history. Thus, they make no distinction between the instantaneous temperature of the glass and the external temperature of the medium. As the above-mentioned “affinities” vanish, $\widehat{S}(t)\ $and $V(t)$, etc. do not play the role of internal variables. Internal variables are normally used in traditional non-equilibrium thermodynamics [@Donder; @deGroot; @Prigogine; @Guj-NE-I; @Guj-NE-II], although they are very hard to identify, and even harder (really impossible) to control. For the application of internal variables to glasses, we refer the reader to Nemilov [@Nemilov-Book]. In the absence of any internal variable, the approximation in Eq. (\[Temperature(t)\]) turns into an identity [@Guj-NE-I]; see also Eq. (\[dS/dH\_relation\_0\]). As the behavior of the entropy during relaxation does not obey Eq. (\[dS/dH\_relation\_0\]), we will call their approach the *unconventional* approach (UCA) in this work; this explains the symbol over the arrow in Eq. (\[Entropy\_Behavior\_UCA\]) and the header of this subsection.
One can understand the drop $\Delta_{\text{GT}}\widehat{S}(t)$ in the entropy of Glass2 during a glass transition, provided the entropy of the medium goes up by $\Delta_{\text{GT}}\widetilde{S}(t)$ to compensate the loss and some more so that the entropy of $\Sigma_{0}$ does not decrease$$\Delta_{\text{GT}}S_{0}(T_{0},t)=\Delta_{\text{GT}}\widehat{S}(T_{0},t)+\Delta_{\text{GT}}\widetilde{S}(T_{0},t)\geq0\label{GT_Entropies}$$ during the glass transition. However, the authors have not discussed this issue at all. This is not surprising as they neither include any medium in their discussion nor do they consider an isolated system for which Eq. (\[Second\_Law\]) holds. We will always consider $\Sigma$ as a part of the isolated system $\Sigma_{0}$ and apply Eq. (\[Second\_Law\]) to the latter. This will then allow us to evaluate $\Delta_{\text{GT}}\widetilde{S}(t)$. It is the magnitude of $\Delta_{\text{GT}}\widetilde{S}(t)$ that would determine whether the entropy of the glass remains above or below $S_{\text{eq}}$ of the supercooled liquid. *As the entropy change of the medium is completely reversible, its evaluation will not suffer from any irreversibility going on in* $\Sigma_{0}$*.* This is the major benefit of investigating a glass as part of $\Sigma_{0}$.
Both glasses begin to deviate from the equilibrium supercooled liquid at $T_{0\text{g}}$, but their structures are not yet “frozen;” they freeze over a long period of time ($t>>\tau_{\text{obs}}$) at a lower temperature $T_{0\text{G}}$ to form an amorphous solid, to be identified as a glass (Glass1 and Glass2). The location of $T_{0\text{g}}$ is determined by the choice of $\tau_{\text{obs}}$; indeed, $T_{0\text{g}}$ decreases with increasing $\tau_{\text{obs}}$. Over the transition region between these two temperatures, the internal variables gradually change form their equilibrium values at $T_{0\text{g}}$ to their frozen values at $T_{0\text{G}}$.
In the absence of any internal variable, fixing the temperature and pressure fixes the instantaneous state. As the UCA glass (Glass2) is homogeneous and has the same constant temperature and pressure as the medium. it appears then that there cannot be any heat transfer between $\Sigma$ and $\widetilde
{\Sigma}$. If true, then during an isothermal relaxation (constant $T_{0}$ of the medium under isobaric condition), the first law with $dQ(T_{0},t)\overset{\text{UCA}}{=}0$ and no internal variables yields$$dE(T_{0},t)\overset{\text{UCA}}{=}-P_{0}dV(T_{0},t)\leq0;$$ This implies that $dV(T_{0},t)\geq0$ during relaxation, which is most certainly not a rule in glasses.
Another important aspect of glass transition, as has become apparent from recent work, is the violation of the principle of detailed balance and of the fluctuation disspation theorem [@Ritort], such as the equivalence of the heat capacity with enthalpy fluctuations. But the theorem is shown to be valid in UCA [@Mauro], which is quite surprising.
Faced with these hard to understand consequences and the conflict with our own results, we decided to examine the basic assumptions in the unconventional approach. These assumptions, to the best of our reading of their work, are not properly and adequately justified so far by GMc. In many cases, they are simply stated as facts alongside several statements that are either exaggerations or are outright false. Therefore, we will treat them as conjectures and investigate whether we can justify them either rigorously or on physical grounds. We defer to the next section these conjectures and the role they play in the logical development of UCA.
We should mention at this point that some aspects of UCA have already been criticized by other authors [@Goldstein; @Nemilov; @GujratiResidualentropy; @Guj-Comment1; @Guj-Comment2; @Gujrati-Symmetry; @Johari]. In particular, Goldstein [@Goldstein], see also [@Guj-Comment1; @Guj-Comment2], demonstrated that the entropy loss during the glass transition violates the second law. To this GMc responded by suggesting that the process of glass formation is not governed by the second law [@Gupta0]. This is a surprising response (as the second law is supposed to govern all processes), but quite understandable as GMc have a very unconventional view of the second law. We do not get into this debate by avoiding the issue altogether. We focus on an isolated system $\Sigma_{0}$, where there cannot be any dispute about the second law; see Eq. (\[Second\_Law\]). In that sense, our work differs from other attempts [@Goldstein; @Nemilov; @Johari]. In our investigation, which is at a fundamental level, we look at all the underlying assumptions of UCA to see if they can be justified so that UCA could become an acceptable theory. We only consider the thermodynamic entropy during this part of our investigation, so we do not get confused by which statistical entropy formulation is appropriate to study vitrification. Once, we settle the issues by using the thermodynamic entropy, we turn to the statistical formulation of entropy to assess the notion of statistical entropy GMc have advocated. It is our belief that the unconventional view of the second law and of the statistical entropy form the basis of UCA, which has been justified in various publications by following the logical steps listed below:
1. The use of equilibrium thermodynamics using calorimetric data cannot determine the entropy of the glass.
2. Thus, there is no reason for the residual entropy to exist at absolute zero.
3. The entropy of a single microstate is zero. As a glass is in one microstate at absolute zero, its entropy must be zero in accordance with the third law.
4. The glass undergoes spontaneous relaxation during which the entropy increases and reaches that of the equilibrated supercooled liquid given by the solid curve in Fig. \[Fig\_entropyglass\] from below.
5. The entropy drop due to the loss of ergodicity and the spontaneous relaxation with entropy increase are, therefore, inverse processes.
6. A calculation method for the entropy is developed to show drop in the entropy during the glass transition region, see Glass2 in Fig. \[Fig\_entropyglass\] so that the calculated entropy shows no residual entropy at absolute zero.
It is important to understand their final conclusion, the so-called inverse relationship (UCA5) and to see if it, and all of its underlying assumptions (UCA1-UCA4), are consistent with the second law, the only fundamental law of Nature that is accepted by all including GMc. Various conjectures leading to UCA5 seem not to be adequately answered so far by GMc. This deficiency by itself does not mean that UCA is unfounded, but it does mean that it requires closer scrutiny, which forms the basis of this investigation. These authors invariably consider their system $\Sigma$ (the glass) at fixed $T_{0}$ and $P_{0}$, which means that it is *not* an isolated system; rather, it is surrounded by $\widetilde{\Sigma}$; see Fig. \[Fig\_Systems\]. (In the following, we will call their system $\Sigma$ an open system, knowing very well that this is not the customary usage. We believe that this will not cause any confusion.) As GMc constantly appeal to the second law in terms of the entropy of the system, the most convenient way to examine their approach is to focus on the isolated system $\Sigma_{0}$ in which the glass will be a possible state of the system $\Sigma$. This allows us to examine their approach at the most fundamental level.
Summary of Results
------------------
We summarize our conclusions that follow from the application of the second law in terms of entropy to an isolated system. We only consider vitrification in the rest of the work. We agree with UCA1, but we find UCA2 unsubstantiated. Indeed, we find the calorimetrically obtained $S_{\text{expt}}(T_{0})$ forms a *lower bound* to the entropy $S(T_{0})$ so that the residual entropy has a lower bound $S_{\text{expt}}(T_{0})$ at absolute zero. The latter entropy is usually non-negative, so that the residual entropy must be even larger than this. Such a glass cannot satisfy the third law, which leads to UCA3 being invalid. Indeed, if there is ever any conflict between the second and the third law, it is the former that supersedes. In an isothermal relaxation, the entropy actually decreases towards the equilibrated supercooled liquid entropy, thus invalidating UCA4. However, the irreversible entropy generation remains non-negative in accordance with the second law. GMc do not recognize the importance of the irreversible entropy generation for the second law and mistakenly ascribe its *universal* non-negative ** aspect to the entropy of the system. Following UCA, we find that $\widehat{S}(t)=0$ all times including $t=0$, when the external condition (such as the temperature) of the system is changed. This is inconsistent with UCA5. Even if we follow UCA Conjecture \[Marker\_Microstate\], see Sect. \[Sect\_UCA\_Conjectures\], although it is inconsistent with the first part, we find that the entropy now increases at $t=0$ and reaches that of Glass2 at $t=t_{\text{obs}}$; the glass transition occurs at this instant if the external condition of the system is disturbed somehow. If it is not disturbed, the entropy would continue to increase. Thus the entropy is always increasing for $t\geq0$ in UCA. In both cases, we do not find that there is any justification in calling the glass transition and relaxation to be inverse processes in UCA. Thus, UCA5 is not a consequence of the previous steps UCA1-UCA4. Just because GMc have provided a computational scheme to support their invalid conclusion cannot be considered a proof of the validity of UCA. We find that UCA misses out many important aspects of non-equilibrium systems such as their temperature, pressure, etc. being different from those of the medium, absence of any internal variables to capture additional irreversible entropy generation, the failure of the fluctuation-dissipation theorem etc.
Important Conjectures in UCA\[Sect\_UCA\_Conjectures\]
======================================================
As the recent work from GMc is expected to represent their most up-to-date and current state of understanding of various issues, we will mostly focus on their recent work [@Mauro] for an understanding of the technical aspect of their approach, the cornerstone of which is that it treats structural relaxations as the *inverse* of the glass transition [@Mauro]:
\[Marker\_Inverse\]The inverse of the glass transition is structural relaxation, which involves a restoration of ergodicity as a glass spontaneously approaches the liquid state. This spontaneous relaxation process is called a unifying process and must entail an increase in entropy as the observation time constraint is lifted.
The decrease in entropy during glass transition to that of Glass2 is justified on the basis of a seemingly innocuous conjecture about the effect of confinement to an ergodic component stated as a *fact* [@Mauro]:
\[Marker\_entropy\_loss\]The loss of ergodicity at the glass transition necessarily involves a loss of configurational entropy, since this causes the system to be confined to a subset of the overall phase space. At absolute zero, any glass is confined to one and only one microstate, so the configurational entropy of a glass is necessarily zero, in accordance with the Third Law and the principle of causality.
In vitrification, the entropy does decrease. This is most clearly seen by quenching the supercooled liquid from a temperature just above $T_{0\text{g}}$ to A below $T_{0\text{g}}$; see Fig. \[Fig\_entropyglass\]. The decrease is shown by the dashed downward arrow to B, which represents the entropy of Glass1 that stays above $S_{\text{eq}}$. A similar drop of much higher value occurs for Glass2. Both glasses seem to conform to the first part of the conjecture. Therefore, this part cannot be the defining characteristics of ergodicity loss or UCA. It is merely a consequence of a positive heat capacity and nothing more and has nothing to do with ergodicity loss. Whether the loss is big enough to satisfy Eq. (\[Entropy\_Behavior\_UCA\]) is never demonstrated as they have not calculated the entropy gain $\Delta_{\text{GT}}\widetilde{S}(T_{0},t)$ of the medium. The second half of the conjecture requires the entropy to vanish at absolute zero, so the entropy of the UCA glass is given by the dashed curve Glass2 in Fig. \[Fig\_entropyglass\] and not by the dotted curve Glass1. Thus, UCA requires the entropy to drop below $S_{\text{eq}}$ of the equilibrated supercooled liquid and requires evaluating $\Delta_{\text{GT}}\widetilde{S}(T_{0},t)$ to substantiate it. This part of the conjecture is based on the following conjecture [@Mauro0]:
\[Marker\_Microstate\] …an instantaneous measurement …causes the system to collapse into a single microstate $i$ with probability $p_{i}(t)$. In the limit of zero observation time, the system is confined to one and only one microstate and the observed entropy is necessarily zero. However, the entropy becomes positive for any finite observation time $\tau_{\text{obs}}$ since transitions between microstates are not strictly forbidden except at absolute zero, barring quantum tunneling.
The conjecture refers to a system “collapsing” into one out of *many* microstates and asserts as a fact that *the entropy of a microstate is identically zero* without any supporting justification. This creates some conceptual problems. At each instant of time $t$, any system, not necessarily a glass only, is going to be in some microstate $i_{t}$; of course, we do not know which microstate it would be in at that instance. If we make “an instantaneous measurement,” the system will remain in that microstate; there is no “collapse” of the microstate. Even the macrostate, which is by definition the collection of all relevant microstates along with their probabilities, does not have time to change, because the probability distribution $p_{i}(t)$ does not change. Indeed, one does not need to make any measurement on the system to conclude “…*the system to collapse into a single microstate…*.” At each instance, the system is going to be in some microstate. If we interpret a measurement as something that instantaneously alters the external condition such as the fields of the medium, then such an instantaneous move (measurement) will probe the instantaneous microstate of the system. The statistical entropy of the system $\widehat{S}(t)=0$ at each instant, and therefore at all times, in accordance with Conjecture \[Marker\_Microstate\]. This is true even if the system is an isolated system not in equilibrium. This concept of statistical entropy in UCA is in direct contradiction with the second law in Eq. (\[Second\_Law\]) according to which the thermodynamic entropy $S_{0}(t)$ is not constant in time. Thus, the conjecture needs some justification, which GMc have not provided so far. In particular, it allows us to make the following
\[Marker\_Remark\_GMc\_entropy\]*The statistical concept of entropy used by GMc in the above conjecture has nothing to do with the thermodynamic entropy used in Eq. (\[Second\_Law\]).*
Another conceptual problem is that the statistical entropy is defined for a macrostate as an *average* quantity over all microstates; see Eq. (\[Gibbs\_Entropy\]). If it happens that a certain macrostate consists of a single microstate whose probability must be $p(t)=1$, then the statistical entropy of that macrostate such as a completely ordered crystal is certainly zero. But what does one mean by the statistical entropy, an average quantity, of a microstate? What averaging does one perform for a microstate? This issue is never addressed by GMc except by the above conjecture and by an appeal to the Boltzmann entropy formulation $\widehat{S}(t)=\ln W(t)$ (we set $k_{\text{B}}=1$), which requires the number of microstates $W(t)$ forming the macrostate. This is evident from [@Mauro]
\[Marker\_Boltzmann\]…Boltzmann’s definition of entropy is the only one valid and consistent with the Second Law for non-equilibrium system.
They take $W(t)=1$ for a single microstate at all temperatures, even if there may be other possible microstates, and argue for zero entropy. Thus, the instantaneous statistical entropy will always be zero at all times and at all temperatures since the system is in a single microstate at each instant. The idea of introducing an instantaneous measurement is highly appropriate as we need to measure instantaneous values of the observables. No measuring apparatus will ever measure the instantaneous entropy; its value can only be inferred indirectly. Thus, entropy is not an observable in the same sense as the mechanical variables such as energy, volume, etc. are; it is a thermodynamic quantity, which has been given a statistical interpretation in statistical mechanics. Despite this, GMc argue that, when the measurement takes some non-zero time, the statistical entropy increases with the duration of measurement. However, GMc never clarify if the measurement gives an accumulated value or the average value of any quantity. The first option is counter-intuitive as this suggests that the value of the energy by such a measurement will increase with the duration of the measurement. The second option seems reasonable as the value of the measurement will give an average energy. This then suggests that, since at each instant during the measurement, the system is in a single microstate so that its entropy is zero, the measurement will still result in a zero entropy. Why does it increase? Even if we adopt the first option, then the “measured” entropy would still be zero as accumulating zero always gives zero. No explanation is offered by GMc for this part of the conjecture. Recall that one cannot appeal to the second law, which uses the thermodynamic entropy, while GMc use their statistical entropy whose equivalence with the thermodynamic entropy is never shown by them. We have addressed this issue elsewhere [@GujratiResidualentropy; @Gujrati-Symmetry; @Guj-Recurrence; @Guj-Irreversibility] with a very different conclusion. We find, see Sect. \[Sect\_Statistical\_Entropy\], that the statistical entropy contribution of a microstate $i$ is $-\ln p_{i}(t)$, which is inconsistent with the above conjecture. Thus, we need to understand the basis of their conjecture. This conjecture is amplified by the following two inter-connected Conjectures [@Gupta]:
\[Marker\_Boltzmann\_Spontaneous\]With Boltzmann’s definition, the entropy increases during a spontaneous process.
\[Marker\_Glass\_Nonspontaneous\]If entropy increases during the relaxation process and the glass transition is the inverse of relaxation, then entropy must decrease during the glass transition. This is consistent with our previous conclusion that the glass transition is nonspontaneous. While this conclusion may appear as inconsistent with the second law, there is no violation since the second law is a statement about spontaneous processes, i.e., processes in which a system relaxes toward an equilibrium or less constrained state. The glass transition is not such a process since here an equilibrium system becomes a constrained equilibrium state.
Conjecture \[Marker\_Boltzmann\_Spontaneous\] is not only inconsistent with Conjecture \[Marker\_Microstate\], it is inconsistent with classical thermodynamics when we consider a system which is not isolated. For such a system, its relevant free energy decreases in any spontaneous process. It appears that GMc confuse isothermal relaxation occurring in a glass with spontaneous processes occurring at fixed observables such as energy, volume, etc. The latter processes occur for isolated systems, not for a glass at fixed temperature and pressure. If GMc insist on focusing on the entropy, then the statement should be in terms of the irreversible entropy generation $\Delta_{\text{i}}S(t)\geq0$, not in terms of the entropy change $\Delta
S(t)$. Thus, for the conjecture to make sense, the reversible entropy drop $\Delta_{\text{e}}S(t)$ during vitrification must not be too negative to ensure $\Delta S(t)>0$. This requires a justification that$$\left\vert \Delta_{\text{e}}S(t)\right\vert <\Delta_{\text{i}}S(t)$$ for the conjecture to be valid. However, no such justification is offered by GMc in their work. As given, it gives the impression that UCA treats entropy to increase during relaxation in all kinds of systems, isolated or not. This is unsettling. The same problem occurs with the last conjecture.
Therefore, to determine whether the above conjectures are justifiable, we turn to the second law for $\Sigma_{0}$.
Consequences of the Second Law for $\Sigma_{0}$\[Sect\_Second\_Law\]
====================================================================
Irreversible Entropy Generation
-------------------------------
The second law does tell us that the irreversible entropy generation in any spontaneous process is non-negative, but leaves the behavior of the entropy undetermined; the latter depends on the process. The entropy that appears in the second law in classical thermodynamics is a thermodynamic concept. *It is postulated to exist even when the system is not in equilibrium; its existence and continuity neither requires any statistical interpretation nor does it require the third law. If there is any conflict between the second law and any other laws of physics, the second law will always win.* In general, in any thermodynamic process from macrostate 1 to macrostate 2, the change in the entropy[@Donder; @deGroot; @Prigogine; @Guj-NE-I; @Guj-NE-II]$$\Delta S\equiv S_{\text{2}}-S_{\text{1}}=\Delta_{\text{e}}S+\Delta_{\text{i}}S,\ \ \Delta_{\text{i}}S\geq0,\label{Entropy_Change}$$ in which $\Delta_{\text{i}}S$ denotes the irreversible entropy generation within the system and $\Delta_{\text{e}}S$ denotes the reversible entropy change due to exchange with the medium. The actual value of $\Delta S$ will depend on the values of $\Delta_{\text{e}}S$ and $\Delta_{\text{i}}S,$ and can have any sign. In general, we have $$\Delta S\geq\Delta_{\text{e}}S,\label{Entropy_inequality}$$ which will prove extremely important below. In the following, we will only be interested in vitrification for which $\Delta_{\text{e}}S$ is negative. If it happens that $\Delta_{\text{e}}S$ is negative enough to overcome the positive contribution of $\Delta_{\text{i}}S$, then we will obtain a negative $\Delta
S$.
Second Law for $\Sigma_{0}$
---------------------------
For an isolated system such as $\Sigma_{0}$, $\Delta_{\text{e}}S=0$ so it is not surprising that$$\Delta S\overset{\text{isolated system}}{=}\Delta_{\text{i}}S\geq0,
\label{Second_Law_Isolated}$$ which explains the standard formulation [@note1] of the second law but only for an isolated system. We now consider $\Sigma_{0}$, but we allow the system to be not in equilibrium with the medium. The medium is at a fixed temperature $T_{0}$ and pressure $P_{0}$. We are thinking of the system that has been brought in contact with the medium at some instant $t=0$, which we then follow in time. The entropy of $\Sigma_{0}$ is written as a sum of the entropies of the system $S(t)$ and the medium $\widetilde{S}(t)$:$$S_{0}(t)=S(t)+\widetilde{S}(t). \label{Entropy_Sum}$$ During its approach towards the maximum, the instantaneous temperature, pressure, etc. of the system, if they can be defined, are different from those of the medium [@Guj-NE-I; @Guj-NE-II; @Langer]. The condition required for defining temperature, pressure, etc. of any non-equilibrium system is that its entropy is a function of its instantaneous observables and internal variables [@Donder; @deGroot; @Prigogine; @Guj-NE-I; @Guj-NE-II] only; it has no explicit time-dependence. The system is said to be in *internal equilibrium* [@Guj-NE-I; @Guj-NE-II], when this condition is met. Unless this condition is met, we cannot identify fields for the system, even though they exist for the medium. For simplicity, we will consider a system with a fixed number of particles with only observables $E(t)$ and $V(t)$ along with just one internal variable $\xi(t)$. Thus, $$S(t)\equiv S(E(t),V(t),\xi(t)). \label{Entropy_Function}$$ At each instance, $E(t),V(t)$ and $\xi(t)$ depend on the history of the system. The corresponding fields are now given by respective derivatives of the entropy:$$\frac{1}{T(t)}=\left( \frac{\partial S(t)}{\partial E(t)}\right)
,~\ \ \frac{P(t)}{T(t)}=\left( \frac{\partial S(t)}{\partial V(t)}\right)
,\ \frac{A(t)}{T(t)}\equiv\left( \frac{\partial S(t)}{\partial\xi(t)}\right)
;$$ the new variable $A(t)$ represents the affinity conjugate to $\xi(t).$ The Gibbs fundamental relation is given by$$dS(t)=\frac{1}{T(t)}dE(t)+\frac{P(t)}{T(t)}dV+\frac{A(t)}{T(t)}d\xi(t),
\label{Gibbs_Fundamental(t)0}$$ which can be rewritten as $$dE(t)=T(t)dS(t)-P(t)dV(t)-A(t)d\xi(t) \label{First_Law(t)0}$$ for the non-equilibrium system. In equilibrium, this relation will reduce to$$dE=T_{0}dS-P_{0}dV, \label{First_Law_Eq}$$ where all variables are independent of time and we have used the fact that the equilibrium value $A_{0}$ of $A(t)$ vanishes. We can rewrite Eqs. (\[Gibbs\_Fundamental(t)0\]) and (\[First\_Law(t)0\]) in the following form [@Guj-NE-II]$$\begin{aligned}
dS(t) & =\frac{1}{T_{0}}dE(t)+\frac{P_{0}}{T_{0}}dV+\left[ \frac{1}{T(t)}-\frac{1}{T_{0}}\right] dE(t)+\left[ \frac{P(t)}{T(t)}-\frac{P_{0}}{T_{0}}\right] dV+\frac{A(t)}{T(t)}d\xi(t),\label{Gibbs_Fundamental(t)}\\
dE(t) & =T_{0}dS(t)-P_{0}dV(t)+[T(t)-T_{0}]dS(t)-[P(t)-P_{0}]dV(t)-A(t)d\xi
(t). \label{First_Law(t)}$$ The last three terms in Eq. (\[Gibbs\_Fundamental(t)\]) each give three distinct irreversible entropy generation terms, and must be individually *non-negative* in accordance with the second law. Let us consider the middle term in this equation, which is non-negative. In a vitrification process, the energy of the system decreases so that $dE(t)$ is negative$.$ Thus, in a vitrification process, $$T(t)>T_{0} \label{Temperature_Relation}$$ during isothermal relaxation (constant $T_{0}$) and approaches $T_{0}$ from above as the relaxation ceases after equilibrium is achieved [@Guj-NE-I]. As this is a general result coming from the second law, it must be valid for all non-equilibrium systems including glasses. We need to see whether both glasses shown in Fig. \[Fig\_entropyglass\] satisfy this result.
For the enthalpy $H(t)\equiv E(t)+P_{0}V(t)$, we find$$dH(t)=T_{0}dS(t)+V(t)dP_{0}+[T(t)-T_{0}]dS(t)-[P(t)-P_{0}]dV(t)-A(t)d\xi(t).
\label{Enthalpy_Differential}$$
Let us consider the consequences of the second law. From now on, we focus on isobaric processes carried out at a fixed pressure $P_{0}$ of the medium. We will assume that in such an isobaric process, $P(t)=P_{0}$ at all times. For the time derivative of the entropy $S_{0}$ of the isolated system at fixed $T_{0}$, it can be shown [@Guj-NE-I; @Guj-NE-II] that in terms of the enthalpy and the internal variable of the system$$\frac{dS_{0}(t)}{dt}=\left( \frac{1}{T(t)}-\frac{1}{T_{0}}\right)
\frac{dH(t)}{dt}+\frac{A(t)}{T(t)}\frac{d\xi(t)}{dt}\geq0.
\label{Total_Entropy_Rate}$$ Each term on the right side of the first equation gives an irreversible entropy generation, see Eq. (\[Second\_Law\_Isolated\])), and must be non-negative. Accordingly, $$\left( \frac{1}{T(t)}-\frac{1}{T_{0}}\right) \frac{dH(t)}{dt}\geq0.
\label{Total_Entropy_Rate1}$$ which is unaffected by the number of internal variables. With Eq. (\[Temperature\_Relation\]), this shows that $$dH(t)/dt\leq0, \label{Enthalpy_rate}$$ which is found to hold in vitrification.
Determination of $\Delta_{\text{GT}}S(T_{0},t_{\text{obs}})$ and $\Delta_{\text{GT}}\widetilde{S}(T_{0},t_{\text{obs}})$
------------------------------------------------------------------------------------------------------------------------
From Eqs. (\[Gibbs\_Fundamental(t)\]) and (\[Total\_Entropy\_Rate\]), we find that$$\frac{dS(t)}{dt}=\frac{1}{T(t)}\frac{dH(t)}{dt}+\frac{A(t)}{T(t)}\frac
{d\xi(t)}{dt},\ \ \frac{d\widetilde{S}(t)}{dt}=-\frac{1}{T_{0}}\frac
{dH(t)}{dt}; \label{Entropy_variation1}$$ the two equations give the rate at which the entropy of the system and of the medium change. In vitrification, the rate for the medium is positive. The second term in the entropy rate for the system is non-negative.
The rate of the entropy drop is given by$$\left\vert \frac{dS(t)}{dt}\right\vert =\frac{1}{T(t)}\left\vert \frac
{dH(t)}{dt}\right\vert -\frac{A(t)}{T(t)}\frac{d\xi(t)}{dt}\leq\frac{1}{T(t)}\left\vert \frac{dH(t)}{dt}\right\vert \leq\frac{d\widetilde{S}(t)}{dt},$$ so that the rate of entropy drop for the system is bounded from above by the rate of entropy gain of the medium. Thus, the drop $\left\vert \Delta
S\right\vert $ for $\Sigma$ is bounded from above by the entropy gain $\Delta\widetilde{S}$ during some interval $\Delta t$:$$\left\vert \Delta S\right\vert \leq\Delta\widetilde{S},$$ where we have introduced the change in any thermodynamic quantity $\digamma$:$$\Delta\digamma(t)\equiv\digamma(t)-\digamma(t=0).$$ From Eq. (\[Entropy\_variation1\]), we find that$$\Delta_{\text{e}}S(t)=\frac{1}{T_{0}}\Delta H(t),\ \ \Delta_{\text{i}}S(t)=\int_{0}^{t}\frac{\overset{\cdot}{H}(t)}{T(t)}dt-\frac{\Delta
H(t)}{T_{0}}+\int_{0}^{t}\frac{A(t)\overset{\cdot}{\xi}(t)dt}{T(t)},\ \ \Delta_{\text{e}}\widetilde{S}(t)=-\frac{1}{T_{0}}\Delta H(t),$$ where the dot above a symbol represents the time-derivative. The maximum entropy drop $\left\vert \Delta S\right\vert $ occurs when $\Sigma\ $comes to equilibrium with the medium. In this case, $T(t)\rightarrow T_{0}$ and $A(t)\rightarrow0$. It is also equal to the maximum entropy gain of the medium at equilibration. This situation correspond to the system as the equilibrated supercooled liquid with its entropy $S_{\text{eq}}$ given by the solid curve in Fig. \[Fig\_entropyglass\]. Thus, we conclude that the entropy of the non-equilibrium system always stays above that of the supercooled liquid$$S(T_{0},t)\geq S_{\text{eq}}(T_{0})$$ during vitrification. During relaxation, $S(T_{0},t)\ $approaches $S_{\text{eq}}(T_{0})$ from above so that $$dS(t)/dt\leq0; \label{Entropy_variation0}$$ this conclusion is valid regardless of the number of internal variables. The equality occurs only when equilibrium with the medium has been achieved.
We can understand this result on physical grounds as follows. Let $T_{0}^{\prime}>T_{0\text{g}}$ denote the temperature from which the system is cooled by bringing the system in contact with the medium at $T_{0}<T_{0\text{g}}$ at $t=0$; see point A in Fig. \[Fig\_entropyglass\]. This situation corresponds to a rapid quench. Right after the contact, the system has not had any time to change its microstate and remains in the microstate it had just prior to the contact is established. Thus, at $t=0,$ the system has the older entropy $S_{\text{eq}}(T_{0}^{\prime})$ of the equilibrated supercooled liquid at the previous temperature. We will justify this fact later; see Eq. (\[Microstate\_Entropy\]). The equilibrium entropy $S_{\text{eq}}(T_{0})$ must be lower than $S_{\text{eq}}(T_{0}^{\prime})$ since $T_{0}<T_{0}^{\prime}$. Therefore, the entropy must continue to drop during relaxation at $T_{0}$ even if the contribution from $\overset{\cdot
}{\xi}(t)$ is non-negative. This gives us the entropy of Glass1 at $t=t_{\text{exp}}$. The derivation of the last equation given in [@Guj-NE-I] was simpler as we had not considered the internal variable $\xi(t)$ there. The above discussion justifies the behavior in the conventional approach described by Eq. (\[Entropy\_Behavior\_CA\]).
From Eqs. (\[Enthalpy\_rate\]), (\[Entropy\_variation1\]) and (\[Entropy\_variation0\]), we find$$\frac{dS(t)}{dH(t)}=\frac{1}{T(t)}\left[ 1+A(t)\frac{d\xi(t)}{dH(t)}\right]
\geq0,\label{dS/dH_relation}$$ which proves Eq. (\[Temperature(t)\]). This equation differs from the one given in [@Guj-NE-I], see Eq. (\[dS/dH\_relation\_0\]) because of the contribution from the internal variable, which was not considered there. However, the sign of the ratio remains the same.
From Eq. (\[Entropy\_variation1\]), we find$$\Delta\widetilde{S}(t)\equiv\widetilde{S}(t)-\widetilde{S}(t=0)=-\frac{\Delta
H(t)}{T_{0}}=-\frac{H(t)-H(t=0)}{T_{0}}$$ At the glass transition, we then have$$\Delta_{\text{GT}}\widetilde{S}(T_{0},t_{\text{obs}})\equiv-\frac
{H(T_{0},t_{\text{obs}})-H(T_{0},t=0)}{T_{0}}\leq\frac{H(T_{0},t=0)-H_{\text{eq}}(T_{0})}{T_{0}}\equiv\Delta_{\text{eq}}\widetilde{S}(T_{0}),$$ where $\Delta_{\text{eq}}\widetilde{S}(T_{0})$ is the maximum entropy gain of the medium, which occurs when the system has come to equilibrium with the medium. This gain can be easily obtained by using a reversible path from the initial temperature $T_{0}^{\prime}$ to the final temperature $T_{0}$. This means that in Fig. \[Fig\_entropyglass\], the system continues along the equilibrated supercooled liquid. Thus,$$\Delta_{\text{eq}}\widetilde{S}(T_{0})=-\Delta_{\text{eq}}S(T_{0})\equiv
S_{\text{eq}}(T_{0}^{\prime})-S_{\text{eq}}(T_{0}).$$ In such a reversible process, $\Delta S_{0}=0$, as expected.
Let us compare this entropy drop with that in the system. Because of the non-negative contribution from $\overset{\cdot}{\xi}(t)$, the entropy change $\Delta_{\text{GT}}S(T_{0},t)$ satisfies$$\left\vert \Delta_{\text{GT}}S(T_{0},t_{\text{obs}})\right\vert \leq\int
_{0}^{t_{\text{obs}}}\frac{\overset{\cdot}{H}(t)}{T(t)}dt\leq\frac{\left\vert
\Delta_{\text{GT}}H(t_{\text{obs}})\right\vert }{T_{0}}\equiv\Delta
_{\text{GT}}\widetilde{S}(T_{0},t_{\text{obs}})\leq\Delta_{\text{eq}}\widetilde{S}(T_{0}). \label{Entropy_Bound_Medium}$$ Thus, the entropy change of the medium and of the system satisfy Eq. (\[GT\_Entropies\]) at the glass transition. But most importantly, the entropy loss of the medium cannot exceed $\left\vert \Delta_{\text{eq}}S(T_{0})\right\vert $, so that $$\left\vert \Delta_{\text{GT}}S(T_{0},t_{\text{obs}})\right\vert \leq
\Delta_{\text{eq}}\widetilde{S}(T_{0})=\left\vert \Delta_{\text{eq}}S(T_{0})\right\vert . \label{Entropy_Bound_System_Medium}$$ We thus conclude [@Guj-NE-I; @Guj-NE-II] that the entropy of the glass must stay above that of the equilibrated supercooled liquid, which makes Glass1 as the physical glass, a conclusion based on the second law.
We can now extend the discussion to continuous cooling as follows. We take $T_{0}^{\prime}>T_{0\text{g}}$ and $T_{0}^{(1)}=T_{0\text{g}}-\Delta T_{0},$ and wait for $\Delta t=t_{\text{obs}}$. The entropy is that of Glass1 at $T_{0}^{(1)}$. We now decrease the temperature by $\Delta T_{0}$ to $T_{0}^{(2)}=T_{0\text{g}}-2\Delta T_{0},$ and wait for $\Delta
t=t_{\text{obs}}$. The entropy is that of Glass1 at $T_{0}^{(2)}$. We follow this cooling until the entire Glass1 curve is obtained. At each temperature, the entropy of the glass must stay above $S_{\text{eq}}$ of the equilibrated supercooled liquid. We have thus proved the following important theorem for any non-equilibrium system:
\[Theorem\_Relaxation\]The entropy of any non-equilibrium system such as a glass in isobaric cooling must stay above that of the equilibrated state.
Thermodynamic Entropy and Glasses
---------------------------------
In the above discussion, which starts with the second law behavior of the thermodynamic entropy $S_{0}(t)$ of $\Sigma_{0}$, no assumption about the form of the thermodynamic entropy $S(t)$ (such as whether $S(t)$ lies above (Glass1) or below (Glass2) the entropy of the equilibrated supercooled liquid; see Fig. \[Fig\_entropyglass\]) of or the nature of irreversibility such as loss of ergodicity, chemical reaction, chaos, phase transition, etc. in the system is made. We do not impose any statistical interpretation on these entropies either; they are assumed to exist as thermodynamic quantities in classical thermodynamics. Thus, their values are not relevant; all that is relevant is their rate. Accordingly, we do not have to even worry if the entropies needed to be treated as statistical quantities with certain particular formulation of entropy (Boltzmann versus Gibbs, or the modification in UCA). Any attempt to identify these classical entropies statistically must still conform to the consequence of the second law expressed, for example, in Eq. (\[Entropy\_variation0\]) during vitrification.
Although no assumption was made regarding $S(T_{0},t)$ lying above or below the entropy $S_{\text{eq}}(T_{0})$ of the equilibrated supercooled liquid, the second law has resulted in Glass1 as being the physically relevant glass, and not Glass2. The deviation of Glass1 entropy $S(T_{0},t)$ from $S_{\text{eq}}(T_{0})$ of the solid curve is due to the irreversible contributions. The entropy of the non-equilibrium state Glass1 approaches that of the equilibrated supercooled liquid entropy from above during isothermal relaxation. This downward approach of the entropy of Glass1 is a consequence of the second law. The entropy of the system, howsoever defined, must satisfy Eq. (\[Entropy\_variation0\]) in vitrification if the system has to obey the second law.
We assume that at time $t=0$, the system is above $T_{0\text{g}}$, so that the system is an equilibrated supercooled liquid. Its temperature and pressure are equal to those of the medium. Let the $E^{\prime},V^{\prime}$ and $S^{\prime}$ denote the energy, volume, and entropy of the equilibrated supercooled liquid at this temperature $T_{0}^{\prime}$, respectively. The equilibrium value of the internal variable is denoted by $\xi^{\prime}$. At time $t=0$, we abruptly bring this system in *contact* with another medium at temperature $T_{0}$ just below $T_{0\text{g}}$; see for example point A in Fig. \[Fig\_entropyglass\]. Immediately after the contact, the initial state of the system is characterized by its observables $E^{\prime},V^{\prime},\xi^{\prime}$ and $S^{\prime}$ at $T_{0}$. After some time $t=t_{\text{exp}}$, the system appears to be glassy as shown by B on Glass1. During further relaxation, the system eventually approaches the equilibrated supercooled liquid at $T_{0}$. During the relaxation process, the entropy $S(T_{0},t)$ decreases in accordance with Eq. (\[Entropy\_variation0\]). This is an example of a fast quench. In a continuous vitrification process carried out at a fixed rate, the resulting glass entropy is shown by Glass1. If such a glass is allowed to relax, see the two downward arrows, it also converges to the solid curve of the supercooled liquid. The resulting entropy during heating is shown by the dash-dotted curve (a) and shows the resulting hysteresis over the transition region.
With the above background about the role of the second law for the open system, we turn to UCA to see if we can justify its consequences or assumptions/conjectures.
Irreversible Contributions and Calorimetric Calculation during the Glass Transition\[Sect\_Glass\_Transition\]
==============================================================================================================
GMc assert in the abstract:[@Mauro] “*A common assumption in the glass community is that the entropy of a glass can be calculated by integration of measured heat capacity curve through the glass transition. Such integration assumes that glass is an equilibrium material and that the glass transition is a reversible process...*” This is an inaccurate and highly misleading statement, which completely overlooks the tremendous progress made by Prigogine and Defay [@Prigogine-Defay], Davies and Jones [@Davies] and others who have followed the concept of internal variables due to de Donder [@Donder; @deGroot; @Prigogine; @Guj-NE-I; @Guj-NE-II]. We refer the reader to a very nice review by Nemilov [@Nemilov] and his monogram [@Nemilov-Book]. Workers in the glass community are well aware of the fact, see Fig. \[Fig\_entropyglass\], that the glass transition neither occurs at a single temperature (it actually occurs over a range $T_{0\text{G}}-T_{0\text{g}}$) nor is it reversible (see the dotted curve and the dash-dotted curve (a) for the entropy during cooling and heating for Glass1). Any attempt to use equation 1 of GMc [@Mauro], which we slightly modify to express it in terms of the thermodynamics entropy and present below$$\Delta S\equiv S(T_{0})-S(T_{\text{M}})\overset{\text{GMc}}{=}\Delta
_{\text{e}}S\equiv{\textstyle\int\limits_{T_{\text{m}}}^{T_{0}}}
\frac{C_{P}(T_{0}^{\prime})}{T_{0}^{\prime}}dT_{0}^{\prime},\label{GM_Entropy_Diff}$$ by workers in the field merely reflects the desire to use an *approximate* description by replacing $S(T_{0})$ by its calorimetric value $S_{\text{expt}}(T_{0})$$$S_{\text{expt}}(T_{0})\equiv S(T_{\text{M}})+{\textstyle\int\limits_{T_{\text{m}}}^{T_{0}}}
\frac{C_{P}(T_{0}^{\prime})}{T_{0}^{\prime}}dT_{0}^{\prime}$$ the right hand side in the above equation [@Mauro]. Comparing with Eq. (\[Entropy\_inequality\]) shows that the approximation is simply to replace the forward inequality by a *forward* *approximate* equality (approximate equality from the greater side)$$\Delta S\gtrapprox\Delta_{\text{e}}S,\label{Entropy_Approximation}$$ and the question one should ask is: *How reliable is the forward approximation* [@Jackel; @Nemilov; @GujratiResidualentropy; @Goldstein]? This question has also been recently answered by Johari and Khouri [@Johari]. This forward approximation cannot be confused with the above-mentioned “common assumption” of equality $\Delta S\overset{\text{GMc}}{\equiv}\Delta_{\text{e}}S$ in Eq. (\[GM\_Entropy\_Diff\]). It appears that GMc confuse this forward approximation with an equality and use it (see below) to suggest that the traditional view of glasses is inapplicable [@Mauro]. This suggestion is not the right conclusion.
An important aspect of non-equilibrium systems is that their fields such as the temperature $T(t)$ are different from the constant fields such as the temperature $T_{0}$ of the surrounding medium. Such a two-field description captures the essence of non-equilibrium states and is also consistent with the violation of the fluctuation dissipation theorem [@Ritort] in non-equilibrium systems. It has become apparent that non-equilibrium systems violate the principle of detailed balance and the fluctuation dissipation theorem. The way irreversibility and the second law [@note1] are taken into account is by instantaneous fields and the introduction of internal variables; the latter is a standard practice in non-equilibrium thermodynamics [@Nemilov-Book; @Donder; @deGroot; @Prigogine; @Guj-NE-I; @Guj-NE-II], but their role is not considered in UCA. This leads them to make several inaccurate statements [@Mauro]. Not realizing that their Eq. (1) is a forward inequality [@Betsul; @Sethna-Paper; @Guj-Rigorous] $$S(T_{0})\geq S_{\text{expt}}(T_{0})$$ due to *all* irreversible contributions to the entropy, see the last three terms in Eq. (\[Gibbs\_Fundamental(t)\]), they incorrectly conclude “*… that glass is treated strictly in the framework of equilibrium thermodynamics,…*” If the correction is made and the equality is replaced by a forward inequality, it immediately rules out any contradiction with the statement “…*that glass is exempt from the Third Law due to its nonequilibrium nature.…*” since the entropy of the glass at absolute zero is bounded below by $S_{\text{expt}}(0)$$$S_{\text{res}}\equiv S(0)\geq S_{\text{expt}}(0).$$ One then discovers that there is no contradiction in logic in the traditional view, and the following statement [@Mauro] in UCA is without any scientific merit:
> “*Previous reports of a finite residual entropy of glass at absolute zero are an artifact of treating glass within the context of equilibrium thermodynamics or equilibrium statistical mechanics, assuming ergodicity and without accounting for the observation time constraint.*”
Not appreciating the important role played by instantaneous fields in non-equilibrium systems leads them to doubt the applicability of$$dS\equiv dQ(t)/T(t)$$ due to heat flow to such non-equilibrium systems, where $dQ$ is the heat transfer with the medium. That it is the correct result follows immediately from Eq. (\[First\_Law(t)0\]) by rewriting it in the form of the first law as $$dE(t)=dQ(t)-P(t)dV(t)-A(t)d\xi(t)$$ so that $$dQ(t)=T(t)dS(t).$$ The irreversible entropy generation within the system is given by $d_{\text{i}}S\equiv dQ\left\{ 1/T(t)-1/T_{0}\right\} \geq0$. In vitrification, $dQ<0$, which then requires $T(t)\geq T_{0}$. This yields $dS\geq d_{\text{e}}S\equiv dQ/T_{0}$, which results in $S(T_{0})\geq
S_{\text{expt}}(T_{0})$ as noted above. The equality occurs only in equilibrium. However, GMc [@Mauro] confuse the forward approximate equation 1 of GMc [@Mauro], reproduced here as Eq. (\[GM\_Entropy\_Diff\]), with an equality and mistakenly conclude that the classical view is inapplicable. The conclusion is without any foundation. The second law clearly establishes that residual entropy is real. As a non-zero residual entropy is in conflict with the third law can only mean that the third law is not applicable to non-equilibrium systems, a conclusion well known in theoretical physics [@Landau].
As GMc do not consider any internal variable, we must consider Eq. (\[dS/dH\_relation\]) by setting $A(t)=0$. In that case, we have $$\left( \frac{\partial S(t)}{\partial H(t)}\right) _{P}=\frac{1}{T(t)}\overset{\text{UCA}}{=}\frac{1}{T_{0}}>0,\label{dS/dH_relation_0}$$ since the temperature of the system is taken equal to $T_{0}$ in UCA. We then conclude that the increase in their statistical entropy along with the decrease in enthalpy during relaxation violates the positivity of the instantaneous temperature of the system and throws doubts that their statistical entropy can be identified with the thermodynamic entropy. As the nature of the statistical entropy is crucial to understand the reasons for the possible failure of $\widehat{S}(t)$, we turn to this issue in the following section.
Statistical Entropy for Non-equilibrium Systems\[Sect\_Statistical\_Entropy\]
=============================================================================
Statistical Entropy as an average for a Macrostate
--------------------------------------------------
The discussion so far has been about the thermodynamic entropy and its existence as used in classical thermodynamics and in the formulation of the second law for an isolated system; see Eqs. (\[Second\_Law\_Isolated\]) and (\[Second\_Law\]). All that is required for this is the fact that *there exists an entropy function* $S_{0}(t)$*, which is non-decreasing in time*. Its actual value is not relevant; all that is relevant is the change in this function. In other words, the thermodynamic entropy is not constrained by the third law in any way; the latter becomes relevant only for the statistical interpretation of the thermodynamic entropy. The latter does not even have to be non-negative, as is well known from the entropy of an ideal gas at low temperatures. The existence of $S_{0}(t)$ immediately leads to the existence of $S(t)$ and $\widetilde{S}(t)$; see Eq. (\[Entropy\_Sum\]). However, it should be emphasized that whatever value any of these entropies such as $S(t)$ has at any instance, it has this value even if no measurement is made on the system:
*Any statistical interpretation of the entropy must obey the property that its value must be unaffected by the measurement.*
This point should not be overlooked. We will explain later how this statement is justified in classical thermodynamics or non-equilibrium statistical dynamics.
We now turn to the statistical interpretation of entropy that provides a justification of the third law for equilibrium states and endow the entropy such as $S_{0}(t)$ with a definite value. Let $i$ denote a microstate of the isolated system $\Sigma_{0}$ in some macrostate. The formulation by Gibbs in terms of the probability $p_{i}(t)$ of a microstate $i$ at time $t$ is as follows:$$S_{0}(t)\equiv-{\textstyle\sum\limits_{i}}
p_{i}(t)\ln p_{i}(t)\equiv-\left\langle \eta(t)\,\right\rangle
,\label{Gibbs_Entropy}$$ where the sum is over all microstates, whose number is $W_{0}$, associated with the particular macrostate; we have set $k_{\text{B}}=1$. A microstate is called *available* at time $t$ if its probability is non-zero; otherwise, it is *unavailable* at that time [@Gujrati-Symmetry]. An available microstate does not mean that the microstate has necessarily been visited by the system during the time interval $t$. Following Gibbs [@Gibbs], we have introduced the *index of probability*$$\eta(t)\equiv\ln p(t),$$ so that the entropy becomes a statistical average of the negative index of probability *over* all microstates belonging to the macrostate. This makes entropy similar to any other average mechanical observable like the energy:$$E_{0}(t)\equiv{\textstyle\sum\limits_{i}}
p_{i}(t)E_{0i},\label{Average_Energy}$$ where $E_{0i}$ is the energy of the $i$-th microstate. The only difference is that the entropy is a thermodynamic quantity as an average of $-\eta(t)$. The index has its origin in the stochastic nature [@Guj-Recurrence; @Guj-Irreversibility; @Gujrati-Symmetry] of a statistical system. Thus, its nature is very different from the mechanical nature of observables like the energy, momentum, etc. although both averages give a statistical average. It is clear from Eq. (\[Gibbs\_Entropy\]) that the negative index $-\eta(t)$ is the contribution to the entropy from a single microstate. One may wish to think of $-\eta(t)$ as the entropy of a microstate, but this is not the conventional view as the entropy is an average quantity for the macrostate; see, however, Eq. (\[Microstate\_Entropy\]). We refer the reader to recent reviews for more details [@GujratiResidualentropy; @Gujrati-Symmetry].
The probabilities $p_{i}(t)$ can be determined by considering an ensemble or by considering the temporal evolution, as described at length elsewhere [@Guj-Irreversibility; @Guj-Recurrence; @GujratiResidualentropy; @Gujrati-Symmetry], but neither is really necessary provided the probabilities $p_{i}(t)$ are known. If it is known initially that $\Sigma_{0}$ is in some *unique* microstate $i_{0}$, then $p_{i}(0)\equiv\delta_{i,i_{0}}$ and $S_{0}(0)=0.$ As time goes on, and assuming that the dynamics is *stochastic*, the initial state will result in making various microstates available with some probabilities $p_{i}(t)$ at time $t$, and the entropy given by Eq. (\[Gibbs\_Entropy\]) will increase [@Tolman], until it reaches its maximum value. It is most certainly not a constant [@Guj-Irreversibility; @Guj-Recurrence; @Gujrati-Symmetry].
The time needed for all the microstates to be available is, in most cases, much shorter than the Poincaré recurrence time [@Guj-Irreversibility; @Guj-Recurrence; @Gujrati-Symmetry; @GujratiResidualentropy]. It may indeed be smaller than the relaxation time $\tau_{\text{relax}}$. At the shorter time, all microstates have become available, but the entropy is still not necessarily at its maximum for the macrostate, since the microstates are not yet *equiprobable*. In the latter situation, the system is in internal equilibrium to be discussed below. If and only if *all* microstates are equally probable ($p_{i}(t)\rightarrow1/W_{0}$ for all $i$), which happens after the relaxation time $\tau_{\text{relax}}$, do we have the maximum possible value of the entropy for the equilibrium macrostate:$$S_{0}(t)\rightarrow S_{0\text{,eq}}\equiv\ln W_{0}\ \ \ \ \text{for }t\gtrsim\tau_{\text{relax}}.\label{Boltzmann_Entropy}$$ This entropy is known as the Boltzmann entropy. It is the equilibrium value of the entropy and occurs because *all* microstates of the system are equiprobable. This entropy is constant in time and depends on the constant observables $E_{0},V_{0},N_{0}$, etc. The internal variables that are now constant are not independent of these observables in equilibrium [@Guj-NE-I; @Guj-NE-II], so the entropy does not depend on them anymore.
It is possible in many cases that over a period of time smaller than $\tau_{\text{relax}}$, only a part of microstates, whose number is given by $W_{0}(t)<W_{0}$ have become available. In that case, the sum in Eq. (\[Gibbs\_Entropy\]) is restricted to $W_{0}(t)$. But the entropy is strictly less than $\ln W_{0}(t)$ unless all available microstates become equally probable. In that case, the entropy is given by $$S_{0}(t)\rightarrow S_{0\text{in.eq}}\equiv\ln W_{0}(t)\ \ \ \text{\ for
}t<\tau_{\text{relax}},\label{Boltzmann_Entropy(t)}$$ which is the Boltzmann entropy at that instance. This entropy is now a function of the internal variables, which themselves depend on time. However, $S_{0\text{in.eq}}(E_{0},V_{0},N_{0},\xi_{0}(t))$ cannot have an explicit time-dependence as for fixed $E_{0},V_{0},N_{0}$, and $\xi_{0}(t)$, $S_{0\text{in.eq}}(E_{0},V_{0},N_{0},\xi_{0}(t))$ is its maximum possible value. This non-equilibrium state of the system with the entropy given by $S_{0\text{in.eq}}$ is said to be in *internal equilibrium* [@Guj-NE-I; @Guj-NE-II], introduced in Sect. \[Sect\_Second\_Law\].
The above discussion can be easily extended to an open system. It has already been shown [@Gujrati-Symmetry] that the Gibbs entropy for the open system is given by exactly the same formula as Eq. (\[Gibbs\_Entropy\]) except that $i$ now represents one of the possible microstates of the open system $\Sigma
$. Thus, everything said above applies to the entropy $S(t)$ of $\Sigma$ by removing the suffix $0$ above. The only difference is that $E_{0},V_{0},N_{0}$, etc. will be replaced by the instantaneous values $E(t),V(t),N.$ The entropy $S_{\text{in.eq}}$ is a function of $E(t),V(t),\xi(t),N_{0}$, etc. but it again cannot have an explicit dependence on time for the same reason that $S_{\text{in.eq}}$ is already maximum for *fixed* $E(t),V(t),\xi
(t),N_{0}$, etc. In the discussion below, we only consider the system $\Sigma$.
Importance of Equiprobable Microstates for Measurements and Microstate Entropy\[Sect\_Measurement\]
---------------------------------------------------------------------------------------------------
The equiprobability assumption implies that the system exhibits no bias for any particular microstate, a point already emphasized by several authors in the past including Tolman [Tolman]{}, who uses this property of a statistical system as a postulate, when he discusses the validity of statistical mechanics. This postulate should be valid even for non-equilibrium states that appear in a system as we vary macroscopic conditions. This is the main idea about the internal equilibrium in our approach. The equiprobable or unbiased sampling assumption for the application of the two Boltzmann probabilities has a very important consequence for measurements in that one does not have to wait for the system to sample all of the relevant microstates. The latter is known to take astronomically large Poincaré recurrence time [@Guj-Recurrence], as can be found in any decent textbook on statistical mechanics; see for example, Huang [@Huang]. Let us consider a non-equilibrium system in internal equilibrium. Because of the equiprobable assumption, $$p_{i}(t)=1/W(t),\ \forall i,$$ where $W(t)$ denotes the number of microstates in the macrostate at that instant, so that the average of any thermodynamic quantity like the energy or entropy is given by$$E(t)\equiv{\textstyle\sum\limits_{i}}
E_{i}/W(t),\ S(t)\equiv-{\textstyle\sum\limits_{i}}
\eta/W(t)=-\eta,$$ in which the sum is over $W(t)$ microstates. In reality, equiprobable microstates do not have to imply an exact equiprobability; they can be within statistical error. One can think of $-\eta\equiv\ln W(t)$ as the entropy per microstate or the entropy of a microstate *under the assumption of equiprobability*. As the observables in each of these microstates take values within statistical fluctuations of the average observables, even a few samples will result in a highly reliable value of the observables. The only difference is that we need to replace $W(t)$ by the number of samples. This is what makes classical thermodynamics so reproducible within statistical fluctuations. For example, let us take a *single sample*, which happens to be in some microstate of energy $E$ with probability $1/W(t)$. The value of $E$ is within statistical error to the average energy $E(t)$. There is no sum in the definition for $E(t)$ now. The result is that$$E(t)=E,\ \ S(t)\equiv-\eta=\ln W(t).\label{Microstate_Entropy}$$ A single sample, or equivalently a single microstate with probability $p(t)=1/W(t)$, provides us with the energy $E$ within statistical error and with the Boltzmann entropy. The latter is not zero and contradicts the UCA Conjecture \[Marker\_Microstate\]. The same is also true of other observables.
*There is no need to take the average over a large ensemble or over a long period of time.*
This is why only a few samples to obtain average thermodynamic quantities give rise to highly reproducible results in thermodynamics. One most certainly does not have to take a very long time average or a very large ensemble average. The above discussion shows how the measurement will not affect the thermodynamic properties, in particular, the entropy of the system in accordance with the expectation noted in Sect. \[Sect\_Statistical\_Entropy\]. The requirement that the measurement should have ample time to *sample all relevant microstates* $W(t)$ is not only unnecessary but also not physical as that time is comparable to the Poincaré cycle [@Guj-Recurrence; @Gujrati-Symmetry; @GujratiResidualentropy]. We believe that GMc have unnecessarily confused the issue by their following suggestion [@Mauro]:
> *Consequently, only the time average can correctly reproduce the measured properties of glass. The underlying reason for this is that at any instant in time a system has one and only one representative point in phase space. The properties measured during a given experiment are a result of averaging over only those microstates that are accessed by the system during the measurement time. This, in a nutshell, is the principle of causality. For short observation times, only a small number of microstates are accessed. For long observation times, a large number of microstates are accessed*.
Of course, it is possible in some rare cases that the sample we have is not a representative of internal equilibrium. In that case, we will obtain results that are not reproducible. But such a situation will be truly rare.
Gibbs vs Boltzmann Entropy Formulation
--------------------------------------
It should be clear form above that the Gibbs formulation is more general than the Boltzmann formulation [@Guj-Irreversibility; @Guj-Recurrence; @Gujrati-Symmetry; @GujratiResidualentropy]. In both cases considered above, the Boltzmann entropy is the maximum possible entropy which occurs only when the *available microstates have become equiprobable*, and the system is either in equilibrium or in internal equilibrium [@Guj-NE-I; @Guj-NE-II]. The system is said to be in equilibrium when Eq. (\[Boltzmann\_Entropy\]) determines the entropy, and in internal equilibrium when Eq. (\[Boltzmann\_Entropy(t)\]) determines the entropy. In all other cases, the Gibbs entropy is the correct entropy of the system. As the Gibbs formulation supersedes the Boltzmann formulation, it seems to be the more general one to use for non-equilibrium systems. The relevance of the Gibbs formulation of entropy for non-equilibrium systems has been discussed recently [@Guj-Irreversibility; @Guj-Recurrence; @Gujrati-Symmetry; @GujratiResidualentropy] by us, and we refer the reader to them for more details. We should, however, mention that Boltzmann’s *H*-theorm already shows that the Gibbs formulation is more general and conforms to the second law. We have also discussed [@Gujrati-Symmetry; @GujratiResidualentropy] there how the time-average is not very useful at low temperatures.
With the above discussion of the statistical entropy, we now turn to UCA. GMc [@Mauro] go on to state as a fact that “…*the Gibbs entropy is valid for canonical systems in equilibrium,…cannot be used in nonequilibrium systems since it implicitly assumes ergodicity.*” This is far from the truth; see above also. The Gibbs entropy is valid for any system (isolated or not), which need not be in equilibrium. We refer the reader to the derivation of the Gibbs entropy formulation in Eq. (40.7) for a non-equilibrium ideal gas in the famous textbook by Landau and Lifshitz [@Landau]; when this entropy is maximized, it gives the *grand canonical distribution*. But the point is that the Gibbs entropy is valid even for non-equilibrium systems. It also does not require ergodicity. It should be stressed that Gibbs never mentions ergodicity in his famous treatise [@Gibbs]. The entropy of a non-equilibrium isolated system using Gibbs formulation is considered by Tolman [@Tolman] to show that it is a non-decreasing function of time and satisfies the second law. Using this formulation for the isolated system, it is easy to show [@Gujrati-Symmetry; @GujratiResidualentropy; @Guj-NE-II] that the same formulation also applies to open systems. GMc use their above limited view of the Gibbs entropy to argue that the approach by Lebowitz and Goldstein of using the Boltzmann entropy formulation “*…is the only one valid and consistent with the Second Law of nonequilibrium thermodynamics;*” see Claim \[Marker\_Boltzmann\]. This is a very strong statement with the implication that it is the truth to be accepted by the reader. Unfortunately, the statement is not the truth as Gibbs formulation of the entropy also satisfies the second law as we have discussed above. Moreover, it is also not accepted by all workers in non-equilibrium thermodynamics. Even Ruelle [@Ruelle], who is cited by GMc [@Mauro], categorically *disagrees* with the interpretation in UCA We quote Ruelle [@Ruelle]
> "*The fact that we take seriously the expression* $S(\rho)=-\int
> dx\rho(x)\ln\rho(x)$ *for the entropy seems to be at variance with the point of view defended by Lebowitz,*$^{\text{(20)}}$ *who prefers to give physical meaning to a *Boltzmann entropy* different from* $S(\rho
> )$*. There is, however, no necessary contradiction between the two points of view, which correspond to idealizations of different physical situations. Specifically, Lebowitz discusses the entropy of states which are locally close to equilibrium, while here we analyze entropy production for certain particular steady states (which maybe far from equilibrium)."*
It is our opinion that GMc have unnecessarily confused the issue of the statistical interpretation of the entropy. While they argue for the superiority of the Boltzmann entropy for which no reasonable arguments are offered, they go back to use the Gibbs formulation, which they blame to be an equilibrium quantity, knowing well that the glass is not an equilibrium system. We find nothing wrong with the Gibbs formulation, contrary to the implications in UCA.
The suggestion by GMc that the glass confined to a component is like a canonical system at fixed temperature and volume (while it is really a non-canonical system with time-varying temperature and constant pressure) misses out the most important aspect of the glass transition. The temperature controlling the vibrations within the component and the temperature describing component hopping over a longer period of time are two distinct temperatures. As they do not include any internal variable in their approach, they miss out in capturing all non-equilibrium contributions to the problem. All they seem to be concerned with is to justify the loss of entropy using a computational approach. Proposing a computational approach that shows entropy of the glass below that of the supercooled liquid is not a proof of the conjecture of the entropy loss. We need to ensure that the resulting physics is consistent with the established laws of physics, such as the second law. We now turn to this aspect of their approach.
Ergodicity and Causality
------------------------
When the entropy is given by Eq. (\[Boltzmann\_Entropy\]), the system is said to be *ergodic*. A system is either ergodic or it is not. When the entropy is given by Eq. (\[Boltzmann\_Entropy(t)\]), one can say that the system is “ergodic with respect to the available microstates belonging to $W(t)$.” But this is not equivalent to the original concept of ergodicity, which is mathematically defined [@Gallavotti; @Patrascioiu; @Szasz] by requiring the equality of infinite time and ensemble or phase-space averages; see also Tolman [@Tolman]:$$\overline{f}=\left\langle f\right\rangle ,$$ where $$\overline{f}\equiv\lim_{t\rightarrow\infty}\frac{1}{t}\int_{0}^{t}f(t^{\prime
})dt^{\prime},\ \ \,\left\langle f\right\rangle \equiv\int_{\Gamma
}f(p,q)dpdq/\int_{\Gamma}dpdq.$$ The infinite-time average is required to ensure that the average does not depend on the initial state of the system. Thus, the Deborah number$$D_{\text{e}}(t,\tau_{\text{relax}})\equiv\frac{\tau_{\text{relax}}}{t}\overset{t\rightarrow\infty}{\rightarrow}0$$ if we wish to test whether a given system is ergodic or not. If we observe a system over a period much shorter than $\tau_{\text{relax}}$, so that $D_{\text{e}}>>1$, all we observe is a non-equilibrium system, but it tells us nothing about the system being ergodic or not. That can only be answered by observing a system much much longer than $\tau_{\text{relax}}$; indeed this time must be comparable to the Poincaré cycle. Even if we observe the system for a period $t$ comparable to $\tau_{\text{relax}}$, the system has not have enough time to visit all relevant microstate. In this case,$$\overline{f}(t\sim\tau_{\text{relax}})\equiv\frac{1}{t}\int_{0}^{t}f(t^{\prime})dt^{\prime}$$ will be dominated by microstates that the system has visited during $t\sim
\tau_{\text{relax}}$; but these microstates have a strong correlation with the initial state, which may be far from equilibrium. Thus, such a finite time-average will not be equal to the ensemble average $\left\langle
f\right\rangle $ for the system, and one would incorrectly conclude that the system is non-ergodic, even if it is ergodic. It should be clear that because of the limit $t\rightarrow\infty$, ergodicity is a property of an equilibrium state. It has no meaning for a non-equilibrium state. Therefore, any suggestion that a glass is non-ergodic requires the *phenomenological assumption* that it is a permanently frozen structure. This is most certainly not a valid assumption in the glass transition region. Thus, to speak of ergodicity breaking at or near $T_{0\text{g}}$ is a misnomer in our opinion, even though it is loosely used in the literature.
In practical terms, a system is ergodic if, after *sufficiently long* time $t>>\tau_{\text{relax}}$, it visits all possible microstates consistent with a macrostate with *equal probability*. This is no different than the above mathematical definition, as the time required to visit all microstates is comparable to the Poincaré cycle [@Guj-Recurrence; @Gujrati-Symmetry; @GujratiResidualentropy]. However, the practical definition, which uses the macrostate, causes the following problem. It depends on how the macrostate is defined. As we have seen in Sect. \[Sect\_Second\_Law\], the concept of a macrostate in non-equilibrium systems depends on time and will eventually become the equilibrium macrostate when $t\approx\tau_{\text{relax}}$. This will make every system ergodic, whether the equilibrium state is unique or not, such as an Ising ferromagnet which has two distinct equilibrium states, and for which the macrostate can be described by the magnetization along with other observables. If, however, the magnetization is not used in specifying the macrostate, then the practical definition will show that the ergodicity is broken in ferromagnets. Usually, we require the equilibrium state to be not unique for the loss of ergodicity. Therefore, we believe that the mathematical definition of ergodicity as a limiting property is the proper way to investigate ergodicity. Such a definition will surely make the liquid above the melting temperature ergodic. Now, just because we observe an ergodic system such as this liquid at some time $t<\tau_{\text{relax}}$, so that we observe a non-equilibrium state of the system, does *not* make the ergodic system non-ergodic. The system remains ergodic as it will eventually equilibrate to its unique equilibrium state for $t\approx\tau_{\text{relax}}$ in accordance with its ergodic nature. Therefore, to speak of ergodicity restoration for a glass is not proper as the glass fully relaxes. From the proper mathematical definition of ergodicity, a glass is also “ergodic” as it will eventually equilibrate to the unique equilibrated supercooled liquid. The quotation marks on ergodic here is to reflect the fact that we are taking the corresponding crystal out of the consideration. These issues have been discussed elsewhere [@Gujrati-Symmetry; @GujratiResidualentropy].
Palmer [@Palmer] does talk about the loss of ergodicity, but it is understood that the relevant part of the phase space is a union of *disjoint* components with no possible transitions between them; the union of these components determine the macrostate. However, the system will be confined *forever* to one of these components, so that we can set $\tau_{\text{relax}}\rightarrow\infty$ and $D_{\text{e}}\rightarrow\infty$. The situation is similar to that in a ferromagnet, except that there are many more macrostates considered by Palmer. The ergodicity is clearly lost in this case. However, the situation with glasses is quite different since $\tau_{\text{relax}}<\infty$ so that $D_{\text{e}}\rightarrow0$ in the limit.
We believe that GMc unnecessarily complicate the issue of glass transition by invoking the loss of ergodicity just because we happen to observe the system in its non-equilibrium state at time intervals $t<\tau_{\text{relax}}$. If we accept this rendition of ergodicity loss, even a liquid above its melting temperature will become non-ergodic at $t<\tau_{\text{relax}}$, and no useful purpose is served by introducing such a concept of ergodicity. We refer the reader to a very good discussion of ergodicity by Tolman [@Tolman] and by Gallavotti [@Gallavotti see, in particular, p. 257]. In our view, the glass transition at $T_{0\text{g}}$ is a transition from equilibrated supercooled liquid to a non-equilibrium supercooled liquid and the transition at $T_{0\text{G}}$ a transition from this non-equilibrium supercooled liquid to a glass, which is almost solid and its structure appears frozen over a long period of time ($t>>\tau_{\text{obs}}$). Thus, our interpretation is different from GMc.
GMc also refer to the concept of causality in their work; see Conjecture \[Marker\_entropy\_loss\]. This issue seems to be first raised by Kievelson and Reiss [@Kievelson; @Reiss]. It basically refers to the possible existence of a large number of degenerate microstates for a macrostate at absolute zero. According to Reiss [@Reiss]
> “*Besides the residual entropy at 0 K being an artifact resulting from apparent entropy measurements along at least partially irreversible paths, this specification is incompatible with a view of the second law which establishes entropy as a function of state. If it is a state function it depends only on its measured state, not upon the history of the system and certainly not upon its future. Since the system does not visit its alternative degenerate states during the time of measurement, it is unaware of these states, and the principle of causality forbids it to be affected by these states.*”
The entropy in non-equilibrium thermodynamics is a *generalized* state function in that it is not only a function of instantaneous observables but also internal variables. The internal variables are no longer independent of the observables only when the system has come to equilibrium. Only the observables, and not the internal variables, are measurable. Thus, Reiss proposes a very narrow concept of entropy used in non-equilibrium thermodynamics. As the instantaneous observables and internal variables clearly depend on the history, Reiss’s assessment about the history dependence is incorrect. Moreover, as the second law destroys time-reversibility, the system is very much controlled by its “future,” i.e. the equilibrium state. Every system, no matter how it is prepared, “knows” exactly where its future lies and relaxes towards it. The idea of causality in the above quote with respect to the statistical entropy is somewhat misleading. Just as each role of a die results in an independent outcome, yet their probabilities are not independent (after all, they have to add to unity), different microstates are independent, yet their probabilities are not [@Gujrati-Symmetry; @GujratiResidualentropy]. Let us clarify this by a simple example discussed by us elsewhere [@Gujrati-Symmetry; @GujratiResidualentropy]. Let us throw an six-face unloaded die. Let the outcome of the throw be $5$. Then, we have $$p_{i}=\delta_{i,5},\ \ \ i=1,2,\cdots,6,$$ where $\delta_{i,j}$ is the kronecker delta, and where $i,j$ denote the six possible outcomes. Let us assume that the outcome of the next throw is $3$. Then, $p_{5}=p_{3}=1/2$, and all other probabilities remaiin zero. Even if the two throws are independent, the probaility distribution changes depending not only on the numeber of throws, but also on the particular outcomes. As the entropy is determined by the probability distribution, it should not come as surprise that the microstates (throws here), though independent, affect the value of the entropy. Only when the number of throws has become so large that $p_{i}\rightarrow1/6$, the “equilibrium value,” can we say that additional throws will not affect the entropy. But this is precisely the property of an equilibrium state.
It should be clear that the probabilities of independent events are not independent in probability theory. As entropy is a statistical quantity (after all it is the average of the negative of the index $\eta$), its value is determined by microstate probabilities. Therefore, the entropy is a measure of the index of probability of all “independent” microstates. Causality has nothing to do with the concept of statistical entropy. Let us consider the case when microstates are equiprobable. The entropy of any sample at $t$ (a single microstate at that instance) is given by its probability, as shown in Eq. (\[Microstate\_Entropy\]). Mechanical quantities such as energy, volume, etc. are not affected by this probability. The mechanical quantities are independent for each sample, as expected. But entropy, being a statistical or thermodynamic quantity, is determined by the probabilities, which themselves are controlled by the sum rule$$\sum_{i}p_{i}(t)\equiv1$$ over all microstates or samples, so that the probability is *determined* by the number of microstates $W(t)$. This expected result has nothing to do with the temperature such as the absolute zero and remains valid at all temperatures and at all times whenever internal equilibrium is present.
Loss of Entropy in UCA and the Glass transition\[Sect\_Entropy\_Loss\]
======================================================================
Continuous vitrification results in the entropy curves (thermodynamic and statistical entropy $S(T_{0},t)$ of Glass1 in CA and statistical entropy $\widehat{S}(T_{0},t)$ of Glass2 in UCA) in Fig. \[Fig\_entropyglass\]. Let us consider our system above $T_{0\text{g}}$, where the system is either the equilibrated supercooled liquid or the equilibrated liquid. Let $E^{\prime
},V^{\prime}$ and $S^{\prime}$ denote the energy, volume, and entropy of the equilibrated state at this temperature $T_{0}^{\prime}$, respectively. The equilibrium value of the internal variable is denoted by $\xi^{\prime}$. At time $t=0$, we abruptly bring this system in *contact* with another medium at temperature $T_{0}$. Immediately prior to the instant the contact is made, the system is in some microstate $i^{\prime}$, but we do not know precisely which microstate it is in. There is a probability $p_{i^{\prime}}$ that the system is in microstate $i^{\prime}$. Let $\tau$ denote the time required for $i^{\prime}$ to evolve to another microstate at $T_{0}$. This microstate has no time to change immediately after the contact, so the system remains in the same microstate initially for $t<\tau$.
What is the entropy $S(T_{0},t)$ for $t<\tau$ after the contact?
The Unconventional Approach
---------------------------
According to Conjecture \[Marker\_Boltzmann\], the statistical entropy $\widehat{S}(T_{0},t)$ is identically *zero* (recall that we are considering the entropy and not just the configurational entropy here and in Fig. \[Fig\_entropyglass\]; similarly, our microstate refers to the system and not just to its configurational state):$$\widehat{S}(T_{0},t)\equiv0\ \ \ \text{for }t<\tau\text{.}$$ This will be true regardless of whether $T_{0}>T_{0}^{\prime}$ or $T_{0}<T_{0}^{\prime}$. We have already used this argument in Sect. \[Sect\_UCA\], which we will now formalize. This zero entropy for $t<\tau$ will result in an entropy curve similar to the entropy curve of Glass2 in Fig. \[Fig\_entropyglass\] at $T_{0}$ in that it lies *below* $S_{\text{eq}}(T_{0})$, except that it is identically zero for $t<\tau$. The argument works the same way even if $T_{0}^{\prime}$ and $T_{0}$ happen to be above the melting temperature $T_{\text{M}}$, where we have an ordinary liquid, which is not thought to lose ergodicity.
It thus follows that the argument of the entropy loss in UCA has nothing to do with any impending glass transition or any impending loss of ergodicity, both of which require temperatures near $T_{0\text{G}}$. It is merely a consequence of two distinct facts:
1. *the duration of observation* $\tau_{\text{obs}}<\tau$ (we will see below that this restriction on observation time is toally irrelevant for the conclusion)*;*
2. *the entropy of a microstate is zero per Conjecture \[Marker\_Microstate\]*.
Let us now consider the above thought experiment in time at any temperature $T_{0}$. The entropy is $\widehat{S}(T_{0},t)\equiv0\ $for $t<\tau$. We now watch the microstate $i^{\prime}$ to evolve to some other microstate $i^{\prime(1)}$ at $t=\tau$, and let $\tau^{(1)}$ be the evolution time for $i^{\prime(1)}$. Since the system is in a single microstate, it follows from Conjecture \[Marker\_Microstate\] that the entropy of the system is still zero for $t<\tau+\tau^{(1)}$. We wait till $t=\tau+\tau^{(1)}$ so that the current microstate evolves into another microstate $i^{\prime(2)}$, and let $\tau^{(2)}$ be the its evolution time. From the same reasoning, we find that $$\widehat{S}(T_{0},t)\equiv0\ \ \ \text{for }t<\tau+\tau^{(1)}+\tau^{(2)}.$$ It is easy to see that $$\widehat{S}(T_{0},t)\equiv0\ \ \ \text{for }t\leq\infty.$$ This makes the second part of Conjecture \[Marker\_Microstate\] inconsistent with its first part.
Conjecture \[Marker\_Microstate\] cannot be justified.
Even though we have rejected Conjecture \[Marker\_Microstate\], let us assume that the entropy $\widehat{S}(T_{0},t)\ $continues to increase in time from its initial value $\widehat{S}(T_{0},0)\equiv0$ for reasons not clearly specified by GMc. We should recall, see Remark \[Marker\_Remark\_GMc\_entropy\], that the statistical concept of entropy adopted by GMc cannot entertain the second law. So, its increase must be justified on some other grounds, which GMc have not done so far. Within the framework of the unconventional approach, let us ask: what would happen if $t=\tau_{\text{obs}}$? If the relaxation time $\tau_{\text{relax}}<\tau_{\text{obs}}$, the entropy $\widehat{S}(T_{0},t)$ will continue to increase and become equal to the equilibrium entropy. For $T_{0\text{g}}<T_{0}<T_{\text{M}}$, the entropy will equal $S_{\text{eq}}(T_{0})$ of the supercooled liquid. For $T_{0}<T_{0\text{g}}$, $\widehat{S}(T_{0},t)$ will continue to increase from zero and become equal to the entropy $\widehat
{S}(T_{0},\tau_{\text{obs}})$ of Glass2, see the horizontal bars on upward pointing arrows in Fig. \[Fig\_entropyglass\], at $t=\tau_{\text{obs}}$ as it tries to grow to its equilibrium value $S_{\text{eq}}(T_{0})$ for reasons not mentioned in UCA. If we disrupt the time-evolution at $t=\tau_{\text{obs}}$ such as by abruptly changing the temperature of the medium, the system will have this entropy $\widehat{S}(T_{0},\tau_{\text{obs}})$ at the moment the change is made. According to our understanding of UCA, this is the glass transition in the system. If we let the system relax at $T_{0}$, the entropy will continue to increase form $\widehat{S}(T_{0},\tau_{\text{obs}}),$ this time from above the horizontal bar on the upwards arrow, until it reaches $S_{\text{eq}}(T_{0})$ as $t$ $\rightarrow\tau_{\text{relax}}(T_{0})$. The entropy is *always increasing*, with the glass transition playing no special role in the growth of the entropy. We see no evidence of this process being inverse of the glass transition at $t=\tau_{\text{obs}}$; entropy $\widehat{S}(T_{0},t)$ is an increasing function of $t$ at all times:
We see no real difference in the way entropy behaves during the evolution of the system, which suggests that the glass transition and relaxation are not inverse processes.
Entropy Loss and the Second Law
-------------------------------
Regardless of the amount of drop (it does not even have to be comparable to $S_{\text{res}}$), the statistical entropy $\widehat{S}(t)$ of such a non-equilibrium state in UCA must approach that of the supercooled liquid from below. This will result in the increase of the entropy during relaxation, which violates Eq. (\[Entropy\_variation0\]). As this equation was a consequence of the second law, we come to the following
A conjecture of entropy drop below that of the supercooled liquid will violate the second law as is clear from Eqs. (\[Total\_Entropy\_Rate\]), (\[Entropy\_variation1\]) and (\[Entropy\_variation0\]). Thus, the statistical entropy $\widehat{S}(T_{0},t)$ and the thermodynamic entropy $S(T_{0},t)$ are two distinct quantities, with the statistical entropy $\widehat{S}(T_{0},t)$ having no relevance to the glassy state.
Our Conventional Approach
-------------------------
We believe that the core of the problem with UCA is the conjecture about the entropy of a microstate; see Conjecture \[Marker\_Microstate\]. The entropy is a property of a macrostate. However, if the system is in internal equilibrium or in equilibrium, then one can obtain the entropy of the system by simply using a microstate [@Guj-Irreversibility; @Guj-Recurrence], as seen in Eq. (\[Microstate\_Entropy\]). This entropy is not always zero; it will be zero if and only if the microstate is unique. The *macrostate* corresponding to given $E(t),V(t)$ and $\xi(t)$ is the collection of all $W(t)$ microstates with given $E(t),V(t)$ and $\xi(t)$ along with their probabilities [@Gujrati-Symmetry]. All instantaneous thermodynamic averages including the instantaneous entropy are average quantities over the macrostate at that instant. Under the assumption of internal equilibrium, the instatntaneous averages can be obtained from a single microstate or sample, as discussed in Sect. \[Sect\_Measurement\]. The dynamics within a glass for it to jump from one microstate to another in time is not necessary for determining these instantaneous averages, an issue discussed elsewhere [@GujratiResidualentropy; @Gujrati-Symmetry] to which we refer the reader for details. When we pick a glass, or when we make an instantaneous measurement, we do not know which microstate it belongs to. All we know is the probability $p_{i}$ for the microstate $i$. If the glass formation occurs under an unbiased condition, all microstates will be equally probable so that$$p_{i}\equiv1/W(t),$$ and we obtain the Boltzmann entropy $\ln W(t)$. Let $W_{\text{G}}$ denote the number of possible microstates at absolute zero. When a glass is formed, it is equally likely to be in any of the $W_{\text{G}}$ microstates at absolute zero so that the residual entropy resulting from this will be $$S_{\text{res}}=\ln W_{\text{G}}.$$ The residual entropy will be zero if and only if we know for sure that the glass is a particular microstate, which will happen only if $W_{\text{G}}\equiv1.$ This we believe will represent an ideal glass. Just because one glass sample at absolute zero is in some microstate out of $W_{\text{G}}$ $(>1)$ does not mean that the glass entropy is zero. The latter would be the case if we knew which particular microstate the glass sample happens to be. Only when $W_{\text{G}}\equiv1$ can we be sure that all glass samples would be in the same microstate, and the glass entropy would be precisely zero [@Gujrati-Symmetry]. In all other cases, all we know is that the probability that the system is in microstate $i$ is $p_{i}$, and the entropy is given by the Gibbs entropy in Eq. (\[Gibbs\_Entropy\]).
There is another way to understand this probability [@GujratiResidualentropy; @Gujrati-Symmetry]. We consider dividing the system into a large number of macroscopically large but quasi-independent parts of equal size, each of which can be in any microstate $\iota$ associated with a part with a probability $p_{\iota}$. Then the entropy $s(t)$ of each part is given$$s(t)\equiv-{\textstyle\sum\limits_{\iota}}
p_{\iota}(t)\ln p_{\iota}(t),$$ and the entropy of the system, using its additive property, is given by $$S(t)=\sum s(t)\equiv N_{\text{P}}\overline{s}(t),$$ where the sum is over all $N_{\text{P}}$ parts of the system and $\overline
{s}(t)$ is the average entropy per part.
Once we recognize that the entropy of a microstate is in general not zero identically, we have no problem understanding that when we bring the system in contact with a medium at another temperature, the entropy immediately after the contact is also unchanged. It then changes towards the new equilibrium value during its relaxation, which is shown by downward arrows in Fig. \[Fig\_entropyglass\]. If we interrupt this relaxation at $t=\tau
_{\text{obs}}$ ($T_{0}<T_{0\text{g}}$) by bringing the system in contact with a different medium at a lower temperature, the system will not completely relax. The current value of the entropy $S(T_{0},t=\tau_{\text{obs}})$ becomes the initial value of the entropy at the new temperature. A sequence of such interruptions will eventually result in a “frozen” glass below $T_{0\text{G}}$. This understanding of the microstate entropy also shows that one does not have to observe the system over a period necessary to sample *many* of or *all* of the microstates associated with the macrostate or one does not require that the entropy is maximum only when all the microstates have been visited. The latter understanding of entropy has been criticized in the past by several authors; see for example Huang [@Huang], Tolman [@Tolman] , Gallovatti [@Gallavotti], Gujrati [@Guj-Recurrence], and the argument has been revisited recently by Goldstein [@Goldstein]: the time needed for *all* the microstates of a macroscopic system to be visited so that the entropy becomes maximum is beyond the current age of the universe. We have already argued against the time-average to be relevant for any measurement [@Gujrati-Symmetry; @GujratiResidualentropy]. In this work, we have shown clearly that a single instantaneous measurement is sufficient to provide us with a thermodynamic description of the system at that instant. Any measurement that takes some finite non-zero duration will never yield any instantaneous information about the system.
Discussion and Conclusions
==========================
We have briefly described and extended the conventional approach we have developed earlier for any non-equilibrium system. We consider the system $\Sigma$ to be surrounded by a very large medium $\widetilde{\Sigma}$ to form the combined system $\Sigma_{0}$ so that the fields of the medium are not affected by the presence of processes going on inside the system, whatever they may. Thus, the approach can be applied to glasses; some authors sometimes identify them by some stretch of imagination as non-ergodic. However, whether ergodicity is lost or not plays no role in the behavior of the collection $\Sigma_{0},$ which we treat as an isolated body so that its thermodynamic entropy cannot decrease with time. This is the statement of the second law for the isolated body. We assume that both the body and the medium are separately in internal equilibrium, but not in equilibrium with each other. The internal equilibrium allows us to introduce instantaneous fields $T(t),P(t)$, etc. for the system and the constant fields $T_{0},P_{0}$, etc. for the medium. We also use a single internal variable $\xi(t)$, in addition to $S(t),V(t)$ as induced internal variables, to describe possible relaxation in the system as it approaches equilibrium. The non-equilibrium nature of the system appears in the values of instantaneous fields $T(t),P(t)$, etc. being different from $T_{0},P_{0}$, etc. of the medium. These differences in the fields cause non-negative irreversible entropy generation in accordance with the second law. In an isobaric vitrification, which is what we consider in this work, we assume that $P(t)=P_{0}$ at all times. The irreversible entropy generation requires $T(t)\geq T_{0}$; the equality occurs when there is equilibrium between the system and the medium. The instantaneous observables, internal variables and entropy at time $t$ are described by the microstate $i_{t}$ the system happens to be in at that instance along with its probability. This microstate also represents the instantaneous macrostate of the system. The effect of an instantaneous measurement is to give the values of the instantaneous observables and the entropy. The measurement does not alter the instantaneous value of the observables, internal variables, and the entropy. The entropy above refers to the *thermodynamic entropy* and its statistical interpretation is obtained by the Gibbs entropy formualtion: The *statistical entropy* is a statistical average of the negative index of microstate probabilities, just like the observables are of mechanical quantities. In our approach (CA), the two entropies behave in identical fashion. Any attempt to provide the classical entropy with a statistical interpretation must satisfy two important requirements:
1. It must decrease during relaxation in an isobaric vitrification process.
2. Its instantaneous value must not be affected by any instantaneous measurement.
In our approach, any non-equilibrium state, such as the one produced by changes in the medium by changing its temperature, relaxes towards its new equilibrium. During such a relaxation under isobaric cooling, the entropy, ehthalpy and the instantaneous temperature decrease towards their respective equilibrium values. The relaxation is complete when we wait for $t=\tau
_{\text{relax}}$. During the relaxation process, the system will undergo a glass transition below $T_{0\text{g}}$, if the system is abruptly changed at $t=\tau_{\text{obs}}<\tau_{\text{relax}}$ by bringing it in contact with a medium at a lower temperature. The *instantaneous macrostate* of the system, described in terms of its observables and internal variable, does not change when the contact is made. This also means that the entropy also does not change. If the contact is not made, the relaxation will continue to go on. Thus, the glass transition and relaxation are part of the same relaxation process in CA. There is nothing inverse about them. The gain in the thermodynamic entropy of the medium is shown to be bounded from above by the maximum change $\Delta S_{\text{eq}}(T_{0})$; the latter is the maximum possible decrease in the thermodynamic entropy of the system. From a careful analysis, we have concluded that the thermodynamic entropy $S(T_{0},t)$ of the system, such as Glass1, must always be above $S_{\text{eq}}(T_{0})$; thus, the thermodynamic entropy must decrease during relaxation; the decrease is a consequence of the second law.
The conclusion of the thermodynamic approach is summarized in Theorem \[Theorem\_Relaxation\]: the thermodynamic entropy must decrease with time during any isothermal relaxation in isobaric vitrification.
UCA developed by GMc as an attempt to describe glass transition in any material does not fulfill both requirements CA1 and CA2. Not only that, the glass transition and relaxation are described as inverse processes. Faced with these discrepancies and several other unconventional and not adequetly explained aspects of UCA, we have carefully examined it in this work. To test the validity of their inverse conjecture UCA5, we decided to treat $\Sigma$ as part of $\Sigma_{0}$. We do this to determine the entropy gain by the medium to show unambiguously whether the system can lose so much entropy at the glass transition that it lies below $S_{\text{eq}}$ in Fig. \[Fig\_entropyglass\]; see UCA5.
GMc incorrectly conclude that the use of classical thermodynamics to calculate the thermodynamic entropy is logically inconsistent (UCA1). Using this erroneous conclusion they argue that the residual entropy must vanish in accordance with the third law (UCA2). However, a careful reconsideration shows that there is nothing wrong in using the classical non-equilibrium thermodynamics. Indeed, its use clearly establishes that the calorimetric entropy $S_{\text{expt}}(0)$ at absolute zero is a *lower bound to the residual entropy*; the former entropy is normally found to be non-negative, which makes $$S_{\text{res}}\geq S_{\text{expt}}(0)>0.$$ Thus, the primary motivation of GMc to develop their unconventional approach UCA is based on an incorrect understanding of classical thermodynamics.
As discussed in Sect. \[Sect\_UCA\_Conjectures\], UCA is based on a set of conjectures, some of which are inter-related, left unproven or satisfactorily justified by GMc. In particular, as summarized in Remark \[Marker\_Remark\_GMc\_entropy\], their statistical formulation $\widehat
{S}(T_{0},t)$ of the entropy, which is based on the zero-entropy conjecture UCA3 for a microstate, has nothing to do with the thermodynamic entropy $S(T_{0},t)$ used in the second law in Eq. (\[Second\_Law\]). The entire UCA is based solely on this statistical notion of entropy and its computation. This formulation of $\widehat{S}(T_{0},t)$ in UCA has been developed with a goal to show entropy loss; yet we find that this formalism, in particular the growth of the statistical entropy with time in Conjecture \[Marker\_Microstate\], is *inconsistent* with UCA3; the latter, if accepted, only results in $\widehat{S}(T_{0},t)=0$ at all times under all conditions including all temperatures. *This is most certainly unphysical.* In our opinion, it is this conjecture that results in the entropy loss during a glass transition under cooling.
Even if we allow for the entropy to increase from zero due to some unknown reasons, not offered by GMc, we find that $\widehat{S}(T_{0},t)$ will always increase towards $S_{\text{eq}}(T_{0})$ of the equilibrated supercooled liquid. We find no justification that the relaxation and glass formation are inverse processes. Their statistical entropy $\widehat{S}(T_{0},t)$ increases in both processes. The increase of entropy scenario would hold at all temperatures, not just at and below $T_{0\text{g}}$. Thus, the entropy loss conjecture has nothing to do with any glass transition or any ergodicity loss; it is merely a consequence of the zero-entropy conjecture UCA3. The increase of the statistical entropy $\widehat{S}(T_{0},t)$ in UCA contradicts the decrease of the thermodynamic entropy found in CA. Thus, *the statistical entropy in UCA cannot represent the thermodynamic entropy of a glass.* As their computation also shows an increase of their statistical entropy, their computational scheme is not useful to understand glasses.
The statistical entropy due to Gibbs that is used in CA remains in conformity with the behavior of the thermodynamic entropy. The entropy of a microstate, see Eq. (\[Microstate\_Entropy\]), is not necessarily zero. Thus, at each instant of time, the entropy of a system, which happens to be in a microstate, is not zero in CA. This instantaneous entropy for an isolated system will always increase, but for an open system such as a glass may *decrease*. The latter behavior is in accordance with the second law. In both cases, it is the irreversible entropy generation that can never be negative. It is our belief that GMc have overlooked this distinction beteen the entropy and irreversible entropy generation in their approach, which causes them to incorrectly believe that the entropy must increase during isothermal relaxation in vitrification.
It is our opinion that GMc have unnecessarily confused the issue of the statistical concept of entropy by implying that the Gibbs entropy is not suitable to describe glasses but the Boltzmann entropy is. This is not a correct conclusion. Both formulations are appropriate, but care must be exercised to interpret them properly. Let us assume equiprobability of microstates. The number of relevant microstats $W(t)$ is most certainly *not* the number of microstates *sampled* by any measurement for a macrostate in time, as GMc suggest. The time for that is of the order of Poincaré cycle and far exceeds the age of the universe. It is really the number of microstates *available* to the system, as explained earlier, and even an instantaneous measurement will give the expected value of the observables. This is what makes thermodynamics so reliable a science. This interpretation is the same whether we use the Gibbs entropy formulation or the Boltzmann entropy formulation. There is no difference between them as long as we deal with internal equilibrium. Their continuous assertion in various publication that they are different is most probably due to their misunderstanding, and serves no purpose except to confuse the issue of the statistical entropy. There is no reason at all to doubt that the thermodynamic and statistical entropies are different in any way.
It is our opinion that they have also unnecessarily made too much of an issue of ergodicity loss and of causality in glasses. All one needs to do is to treat glasses as a non-equilibrium state and to recall that the statistical entropy is an average of a statistical quantity, the negative index of probability as discussed in the work. GMc have taken a very simplistic view of glasses by ignoring internal variables, two-temperature description and the fact that fluctuation-dissipation theorems used by them [@Mauro] fail for glasses.
We finally conclude that their current theoretical and computational attempts using UCA has no relevance for glasses.
I am thankful to M. Goldstein for introducing me to the work by GMc, and to G.P. Johari for his useful comments on an earlier version of the manuscript.
[99]{}
P.D. Gujrati, Phys. Rev. E **81**, 051130 (2010); P.D. Gujrati, arXiv:0910.0026.
P.D. Gujrati, arXiv:1101.0438.
By the second law we mean the law of increase of entropy according to which *the entropy of an isolated system can never decrease*.
E. Bouchbinder and J.S. Langer, Phys. Rev. E **80**, 031131 (2009); *ibid.* 031132 (2009); *ibid.* 031133 (2009).
Th. de Donder and P. van Rysselberghe, *Thermodynamic Theory of Affinity*, Stanford University, Stanford (1936).
S.R. de Groot and P. Mazur, *Non-Equilibrium Thermodynamics***,** First Edition, Dover, New York (1984).
D. Kondepudi and I. Prigogine, *Modern Thermodynamics*, John Wiley and Sons, West Sussex (1998).
Nemilov, S.V., *Thermodynamic and Kinetic Aspects of the Vitreous State*, CRC Press, Boca Raton (1995).
J. C. Mauro, P. K. Gupta, and R. J. Loucks, J. Chem. Phys.**126**, 184511 (2007).
P.K. Gupta and J.C. Mauro, J. Chem. Phys. **126**, 224504 (2007); J. Non-Cryst. Solids, **355**, 595 (2009).
J.C.Mauro, R.J. Loucks, and S. Sen, J. Chem. Phys. **133**, 164503 (2010).
see for example the review A. Crisanti and F. Rotort, J. Phys. A **36**, R181 (2003).
M. Goldstein, J. Chem. Phys. **128**, 154510 (2008).
S.V. Nemilov, J. Non-Cryst. Solids, **355**, 607 (2009).
P.D. Gujrati, arXiv:0908.1075.
P.D. Gujrati, Symmetry **2**, 1201 (2010).
P.D. Gujrati, arXiv:0909.0238.
P.D. Gujrati, arXiv:0909.0734.
G.P. Johari and J. Khouri, J. Chem. Phys. **134**, 034515 (2011).
P.K. Gupta and J.C. Mauro, J. Chem. Phys. **129**, 067101 (2008).
P.D. Gujrati, arXiv:0803.0983.
P.D. Gujrati, arXiv:0803.1099.
I. Progogine and R. Defay, *Chemical Thermodynamics*, Longmans, London (1954).
R.O Davies and G.O. Jones, Adv. Phys. **2**, 370 (1953).
J. Jäckle, (a) Philos. Mag. B **44**, 533 (1981); (b) Physica B **127**, 79 (1984).
A. B. Bestul and S. S. Chang, J. Chem. Phys. **43**, 4532 (1965).
S.A. Langer and J.P. Sethna, Phys. Rev. Lett. **61**, 570 (1988).
P.D. Gujrati, arXiv:1101.5623.
K. Huang, *Statistical Mechanics,* second edition, John Wiley, New York (1987).
L.D. Landau, E.M. Lifshitz, *Statistical Physics*, Vol. 1, Third Edition, Pergamon Press, Oxford (1986).
J.W. Gibbs, *Elementary Principles in Statistical Mechanics*, Ox Bow Press,Woodbridge, Conn. (1981).
R.C. Tolman, *The Principles of Statistical Mechanics*,Oxford University, London (1959).
D. Ruelle, J. Stat. Phys. **85**, 1 (1996).
G. Gallovatti, *Statistical Mechanics: A Short Treatise,* Springer, Berlin (1999).
Adrian Patrascioiu, http://library.lanl.gov/cgi-bin/getfile?00285754.pdf
D. Szàsz, Studia Sci. Math. Hung. **31**, 299 (1996); www.renyi.hu/szasz/preprints/boltz.ps
R.G. Palmer, Philos. Mag. B **44**, 533 (1981); Adv. Phys. **31**, 669 (1982).
D. Kievelson and H. Reiss, J. Phys. Chem. B **103**, 8337 (1999).
H. Reiss, J. Non-Cryst. Solids, **355**, 617 (2009).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Simultaneous observations by the large number of gamma-ray burst detectors operating in the GLAST era will provide the spectra, lightcurves and locations necessary for studying burst physics and testing the putative relations between intrinsic burst properties. The detectors’ energy band and the accumulation timescale of their trigger system affect their sensitivity to hard vs. soft and long vs. short bursts. Coordination of the Swift and GLAST observing plans consistent with Swift’s other science objectives could increase the detection rate of GLAST bursts with redshifts.'
author:
- 'David L. Band, on behalf of the GLAST collaboration'
title: 'THE SYNERGY OF GAMMA-RAY BURST DETECTORS IN THE GLAST ERA'
---
[ address=[CRESST and Code 661, NASA/Goddard Space Flight Center, Greenbelt, MD 20771]{}, altaddress=[CSST, University of Maryland, Baltimore County, 1000 Hilltop Circle, Baltimore, MD 21250]{} ]{}
Anticipated to be launched in spring, 2008, the Gamma-ray Large Area Space Telescope (GLAST) will join a large number of gamma-ray burst detectors that are already operating in space. The strengths of these different detectors complement each other, both in providing capabilities that are absent in other detectors and in allowing cross-calibration. In this work I compare the different detectors and their capabilities.
The Table lists burst detectors that will operate during the first few years of the GLAST mission. Quantitative comparisons between different missions are difficult because of the operational details. For example, the sensitivity usually varies across a detector’s field-of-view (FOV), resulting in a burst detection threshold that is not uniform. Because many detectors can provide spectra over a larger energy band than used for the burst triggers, I provide two energy bands in the Table.
--------------------- ------------------------------- ------------ --------------- -------------- ----------------- ----------------- ------------ ----------------------------------------------
GLAST-LAT 565 km, $\iota$=25.3$^\circ$ $\sim$3.5 8000 0.1$^\circ$ 25 MeV–300 GeV 25 MeV–300 GeV Variable [@Omodei:2007]
GLAST-GBM 565 km, $\iota$=25.3$^\circ$ $\sim$9 122$\times$12 8$^\circ$ 8 keV–30 MeV 50–300 keV 0.064–4 s [@Meegan:2007]
Swift-BAT 590 km $\iota$=20.1$^\circ$ $\sim$1.4 2600 4$^\prime$ 15–150 keV 15–150 keV 0.004–64 s [@Gehrels:2004; @Barthelmy:2005; @Band:2006]
Konus-Wind L1 $\sim4\pi$ 133$\times2$ — 12 keV–10 MeV 45-190 keV 0.15, 1 s [@Aptekar:1995; @Mazets:2004]
Suzaku-WAM 570 km, $\iota$=31$^\circ$ $\sim4\pi$ 800$\times$4 — 50 keV–5 MeV 110–240 keV 0.25, 1 s [@Yamaoka:2006]
RHESSI 580 km, $\iota$=38$^\circ$ $\sim4\pi$ $\sim150$ — $>$50 keV $>$50 keV Variable [@Wigger:2007]
Super-AGILE $\sim$580 km, $\iota<3^\circ$ 1.4 312$\times$4 1.5$^\prime$ 15–45 keV 15–45 keV Variable [@Costa:2001; @Tavani:2006]
AGILE Mini-Cal $\sim$580 km, $\iota<3^\circ$ $\sim$2.5 $\sim$1400 — 300 keV–100 MeV 300 keV–100 MeV Variable [@Labanti:2006; @Tavani:2006]
AGILE TKR $\sim$580 km, $\iota<3^\circ$ $\sim$2.5 $\sim$1000 15$^\prime$ 30 MeV–50 GeV 30 MeV–50 GeV Variable [@Tavani:2006]
INTEGRAL ISGRI/IBIS Ecc. 0.1 1300 2$^\prime$ 15 keV–1 MeV 15 keV–1 MeV 8ms–40 s [@Lebrun:2003; @Mereghetti:2005]
INTEGRAL SPI Ecc. 0.1 250 10$^\prime$ 20 keV–8 MeV — — [@vonKienlin:2004]
INTEGRAL SPI ACS Ecc. $\sim4\pi$ $\sim3000$ — — — $>50$ ms [@Lebrun:2003; @Mereghetti:2005]
--------------------- ------------------------------- ------------ --------------- -------------- ----------------- ----------------- ------------ ----------------------------------------------
: Burst Detectors in GLAST Era[]{data-label="tab:a"}
Burst triggers ultimately compare an increase in the number of detected counts in an energy band $\Delta E$ and accumulation time $\Delta t$ to the expected background fluctuations; the burst threshold is derived from the signal-to-noise ratio for a $\Delta E - \Delta t$ bin. The burst detection sensitivity is the threshold flux $F_T$ (here over the 1–1000 keV band) as a function of the burst spectrum (here over 1 s). Burst spectra can be parameterized by the ‘Band’ function,[@Band:1993] characterized by low and high energy spectral indices $\alpha$ and $\beta$, and a characteristic energy $E_p$, the photon energy of the peak of the $E^2N(E)\propto \nu
f_\nu$. The Figure’s left hand panel presents $F_T$ as a function of $E_p$, fixing $\alpha=-1/2$ and $\beta=-2$, for different burst detectors. Note that this figure does not show a detector’s sensitivity at a given energy but instead the detector sensitivity to a burst with a given $E_p$. Here I show the sensitivity for $\Delta t=1$, but detector triggers operate with a variety of $\Delta t$ values, and differ in their sensitivity to bursts with different durations.[@Band:2003; @Band:2006]
{height=".25\textheight"} {height=".25\textheight"}
In many cases we are not interested in whether a detector detects a burst—spectral data may be available regardless of whether the detector triggered—but in the spectra the detector accumulates. The Figure’s right hand panel shows the detectors’ spectral sensitivity, the continuum sensitivity over a 1 s accumulation time.
The synergy between missions will be maximized by simultaneous burst observations. Particularly important is the overlap between detectors with spectral capability (GLAST, Konus-Wind, Suzaku-WAM) and localization capability (Swift-BAT and INTEGRAL-ISGRI). Konus-Wind and INTEGRAL SPI-ACS essentially see the entire sky, while GLAST-GBM, Suzaku-WAM and RHESSI see down to the horizon for a low-Earth orbit. Although ISGRI is sensitive, it has a (relatively) small FOV. The GLAST instruments—the LAT ($<$20 MeV–$>$300 GeV) and the GBM (8 keV–30 MeV)—have large FOVs and the GLAST observatory will operate in a fixed survey mode. Swift-BAT also has a large FOV and Swift has a very flexible observing timeline.
Because of their large FOVs and complementary strengths—localizing bursts and following afterglows for Swift, accumulating spectra over 7 energy decades for GLAST—increasing the overlap between these two missions will have major scientific gains. During most of its mission, the LAT’s pointing will follow a fixed pattern to execute an all-sky survey. On the other hand, Swift observes a number of targets each orbit with its Narrow Field Instruments (NFIs)—X-ray Telescope (XRT) and Ultraviolet-Optical Telescope (UVOT); the wide FOV Burst Alert Telescope (BAT), which observes bursts’ prompt emission, is centered on the NFIs. The NFI targets are burst afterglows and other astrophysically interesting sources. Semi-analytic calculations show that if Swift does not coordinate its pointings with GLAST, $\sim$13% of GLAST-LAT bursts will be in the Swift-BAT FOV, and $\sim$27% of Swift-BAT bursts in the GLAST-LAT FOV. If Swift points as close as possible to the LAT pointing direction these overlap numbers could increase by $\sim$3$\times$! However, this estimate neglects Swift’s observational constraints, and sacrifices many of Swift’s scientific objectives. Nonetheless, the judicious choice of Swift NFI targets could increase the LAT-BAT overlap by $\sim$2$\times$. For example, Swift’s timeline could include two sets of targets, one observed when the LAT is pointed towards the northern hemisphere, and the second for the southern hemisphere. Procedures to increase this overlap with little impact on Swift’s science objectives are under development. An increase in the LAT-BAT overlap would of necessity increase the GBM-BAT overlap. Note that a burst in a detector’s FOV may nonetheless be too faint to be detected.
Coordination with GLAST of the timelines of most of the other burst missions would yield meagre increases in the detector overlaps because the detectors are nearly all-sky (or are occulted by the Earth), or their operations do not lend themselves to such coordination. Because of the different FOVs and the varying detection sensitivies, I forsee that most bursts detected by one of the constellation of burst detectors will have at best upper limits from other detectors. A few bursts a year will be well-observed by a varying assortment of missions; for example, Swift might provide a location, Konus-Wind spectra, and Swift and GLAST-LAT afterglow observations. If the Swift timeline is coordinated with GLAST, then the LAT and BAT will observe simultaneously $\sim$50 bursts a year, although the LAT will probably detect less than half of these. The localizations, spectra and lightcurves of this last set of bursts will advance the study of gamma-ray bursts.
I thank members of the BAT team at GSFC and of the GLAST GRB working group for insightful discussions.
[17]{} natexlab\#1[\#1]{}\[1\][“\#1”]{} url \#1[`#1`]{}urlprefix\[2\]\[\][[\#2](#2)]{}
D. [Band]{}, et al., *Ap.J.* **413**, 281–292 (1993).
D. Band, *Ap.J.* **588**, 945 (2003).
N. [Omodei]{}, and J. [Norris]{}, “LAT observation of GRBs: Simulations and Sensitivity studies”, in *AIPC 921*, edited by S. [Ritz]{}, P. [Michelson]{}, and C. A. [Meegan]{}, pp. 472–475 (2007).
C. [Meegan]{}, et al., “The GLAST Burst Monitor”, in *AIPC 921*, edited by S. [Ritz]{}, P. [Michelson]{}, and C. A. [Meegan]{}, pp. 13–18 (2007).
N. [Gehrels]{}, et al., *Ap.J.* **611**, 1005–1020 (2004).
S. D. [Barthelmy]{}, et al., *Space Science Reviews* **120**, 143–164 (2005).
D. L. [Band]{}, *Ap.J.* **644**, 378–384 (2006).
R. L. [Aptekar]{}, et al., *Space Science Reviews* **71**, 265–272 (1995).
E. P. [Mazets]{}, et al., “Konus catalog of short GRBs”, in *ASPC 312*, edited by M. [Feroci]{}, F. [Frontera]{}, N. [Masetti]{}, and L. [Piro]{}, pp. 102–105 (2004).
K. [Yamaoka]{}, et al., “In-orbit performance of the Suzaku wideband all-sky monitor”, in *Space Telescopes and Instrumentation II: Ultraviolet to Gamma Ray.* Edited by Turner, Martin J. L.; Hasinger, G[ü]{}nther. Proceedings of the SPIE, Volume 6266, pp. 626643 (2006).
C. [Wigger]{}, et al., *ArXiv e-prints* **704** (2007), .
E. [Costa]{}, et al., “Super-agile-The X-ray detector for the gamma-ray mission agile”, in *AIPC 599*, edited by N. E. White, G. Malaguti, and G. G. C. Palumbo, pp. 582–585 (2001).
M. [Tavani]{}, et al., “The AGILE mission and its scientific instrument”, in *Space Telescopes and Instrumentation II: Ultraviolet to Gamma Ray.* Edited by Turner, Martin J. L.; Hasinger, G[ü]{}nther. Proceedings of the SPIE, Volume 6266, pp. 626603 (2006).
C. [Labanti]{}, et al., “The mini-calorimeter of the AGILE satellite”, in *Space Telescopes and Instrumentation II: Ultraviolet to Gamma Ray.* Edited by Turner, Martin J. L.; Hasinger, G[ü]{}nther. Proceedings of the SPIE, Volume 6266, pp. 62663Q (2006).
F. [Lebrun]{}, et al., *A&A* **411**, L141–L148 (2003).
S. [Mereghetti]{}, and D. [G[ö]{}tz]{}, *Nuovo Cimento C Geophysics Space Physics C* **28**, 259 (2005).
A. V. [Kienlin]{}, N. [Arend]{}, G. [Lichti]{}, A. [Strong]{}, and P. [Connell]{}, “[Gamma-Ray Burst Detection with INTEGRAL/SPI]{},” in *ASPC 312*, edited by M. [Feroci]{}, F. [Frontera]{}, N. [Masetti]{}, and L. [Piro]{}, 2004, p. 551.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Leonard W. T. Ng'
- Xiaoxi Zhu
- Guohua Hu
- Nasiruddin Macadam
- Dooseung Um
- 'Tien-Chun Wu'
- Frederic Le Moal
- Chris Jones
- Tawfique Hasan
title: 'Conformal printing of graphene for single and multi-layered devices on to arbitrarily shaped 3D surfaces'
---
Introduction
============
Although a number of methods have been explored for the deposition of 2D materials for functional devices, printing holds specific promise for high-volume, low-cost and large-area manufacturing [@Cui2016; @Ng2019; @Hu2018; @Torrisi2012]. These include inkjet [@Hu2018; @Torrisi2012; @Mcmanus2017; @Howe2015], screen [@Joseph2016; @Secor2014] and roll-to-roll (R2R) gravure and flexographic printing which are already widely used in the large-scale, ultra-low-cost production of packaging materials, everyday documents, magazines and newspapers [@Ng2019a]. Among these, screen printing has emerged as an effective method in depositing high-viscosity graphene inks directly on to the substrates, typically *via* a stencil made of a fine synthetic fibre mesh that is capable of depositing $\sim$10-16 $\mu$m thick layers of ink [@R.H.Leach1993; @Ng2019a]. Recent investigations have shown customised screen printing can achieve high resolution ($<\sim$40 $\mu$m) patterns of 2D materials on a variety of substrates, including glass, paper, textiles, polyester film and even solid graphite [@Zhang2017; @Rowley-Neale2017; @Hyun2015; @Arapov2016; @Xu2013; @Yeates2017]. However, it is usually capable of printing only on to flat surfaces [@Hu2018; @Ng2019a].
Conformal printing is an attractive and emerging field, which can enable specific electronic features such as sensors and circuits directly on to 3D objects. This is especially important in the present-day internet of things (IoT) era where 3D objects are increasingly interconnected. However, functional ink deposition on 3D structures is not a trivial process due to uneven and sometimes, complex topography. To date, several techniques have been explored to achieve conformal functional printing of 2D material, carbon nanotube and metal nanoparticle-based inks. These include aerosol jetting [@Jabari2015] or specially adapted inkjet or screen printing that manipulate arbitrarily-shaped substrates *via* rotary systems around ink nozzle heads or printing mesh[@Krebs2009]. However, these usually require highly complex and expensive printer designs that are capable of only depositing functional patterns on to one object at a time. Thermoforming has also been used in conjunction with other printing techniques such as flexography and gravure where inks are first deposited on planar polymer substrates which are then made pliable *via* the application of heat and formed into a complex 3D shape in a multi-stage, offline process. [@Krebs2009]. Other more esoteric methods such as direct-write deposition [@Ahn2009] have also been explored. However, these demonstrations have been conducted only on a small scale, are capable of depositing only one material at a time and are highly time consuming.
We note that the general concept we present here has been previously demonstrated for the deposition or transfer printing of nanomaterials and functional structures on to flat and 3D objects. In 2008, transfer deposition of 0D and 1D nanostructures [@Jiao2008] as well as atomically thin sheets of 2D materials [@Reina2008] on to flat substrates were demonstrated using poly(methyl methacrylate) (PMMA) as the sacrificial layer (SF). Subsequently, this strategy was adapted for printing of nanostructures on to 3D objects using PMMA [@Lee2011] and other SFs such as polyvinyl alcohol (PVA) [@Aziz2011] and silk-fibroin [@Kim2010]. Very recently, Saada *et al.* used inkjet printed commercial silver inks to demonstrate patterns on to 3D objects by transfer printing using a PVA film as the SF [@Saada2017]. However, this nanoparticle-based ink requires sintering to provide the desired functionality and mechanical integrity for the transfer process, potentially restricting the technique to single layer patterns only.
Among the various strategies discussed above, we propose that water or solvent-assisted printing is potentially the most promising method for large-area deposition of functional 2D material inks directly on to arbitrarily shaped objects. This is due to the ability of this technique to conform to the complex topographies of 3D objects. To date, such a method has not yet been developed for graphene and other 2D materials that is robust for large-area, single and multilayer devices for electronics and sensing.
Here, we introduce a method of conformal printing of a specifically formulated conductive graphene ink, exploiting the advantages of water-assisted conformal printing. This method first involves the formulation of a water-insoluble ink that can be printed *via* screen printing. For this, we use a combination of One Factor At A Time (OFAT) experiments (where we vary one factor in ink formulation and measure the corresponding outcome) and Design of Experiments (DOE), a method that we use to determine the inter-relationship between multiple components of the final ink formulation. OFAT experiments involve increasing the graphene loading until the ink is unprintable ($>$ 10 Pa s), while determining the minimum amount of binder required for the ink to be shear thinning during printing and to maintain its mechanical integrity on the substrate after curing. The results of OFAT are used to qualitatively define sensible upper and lower bounds for the DOE, which is then used to quantitatively optimise the electrical conductivity of the ink after printing and curing.
After formulation, the ink (and additional layers for the case of multi-layered devices) is printed on an ultrathin, water-soluble PVA as the SF. The composite structure, comprising the SF and printed layer(s) is then floated in a water bath to allow the SF to dissolve. The floating patterns are then ‘fished’ out from the water-bath using the target substrates, allowing the pattern to adhere directly on to the substrate. This process is highly versatile, capable of printing on to different substrates to fabricate functional devices without the loss of the conductive properties of the printed graphene patterns after the transfer.
We demonstrate this by conformally printing conductive graphene patterns to fabricate single and multi-layered devices on to multiple hard (glass, 3D printed thermoplastics), flexible and porous (bandages) and non-porous (nitrile gloves), and even on soft substrates such as marshmallows and gummy candies. These structures are then used to demonstrate applications of conformal printing, including joule heaters on glassware, resistive sensors on off-the-shelf candies and electrodes on non-conductive objects, making them capable of capacitive proximity sensing. We further demonstrate the capability of this method in the fabrication of multi-layered devices by printing graphene/polyurethane(PU)/graphene parallel-plate capacitors on arbitrarily shaped objects.
Experimental section
====================
Graphene ink formulation
------------------------
![Graphene ink formulation **a** Photographs of the graphene ink as it is passed through the three-roll mill and corresponding SEM images. Scale bar 5 $\mu$m. **b** Effect of three-roll milling on the R$_s$ of the ink. **c** Graph showing the influence of graphene ink on the final ink’s viscosity after 6 passes through the three-roll mill. **d** Rheogram of the graphene screen printable ink, showing shear thinning behaviour. **e** Log-log plot of viscosity and shear rate of the power law region of a flow curve. **f** Ternary plot of the interaction of different loadings of IPA, EC binder and graphene as determined by DOE. Experiments are indicated with black dots. []{data-label="panel1"}](panel1){width="90.00000%"}
Like the majority of the graphics inks [@R.H.Leach1993; @Goldschmidt2003; @Flick1999], the main components of the functional graphene ink is typically made up of a functional material (in this case, graphene), a binder system (a polymeric film former to provide mechanical integrity after curing) and appropriate solvents (ink carrier). Due to the need for the graphene ink to be insoluble during the water-assisted conformal printing, a water insoluble binder, ethyl cellulose (EC), is chosen. Terpineol is chosen as the primary solvent due to its compatibility with EC. The high boiling point of the primary solvent terpineol ($\sim$218 $^{\circ}$ C) allows for a long ‘dwell time’ on the screen. This prevents clogging of the screens during or immediately after printing[@Ng2019a]. Isopropyl alcohol (IPA) (boiling point $\sim82.3 ^{\circ}$ C), is chosen as the secondary solvent for on press ink viscosity control. As the inks are printed using a hand-operated press, the inks do not have to travel from surface to surface within a printer and is hence subject to less agitation. This in turn prevents common printing issues such as foaming (the formation of air bubbles in the ink due to mechanical agitation usually caused by printing machines) and orange peeling (a mottle-like print defect characterised by dimples in the finished print, giving the print the texture of an orange peel. Usually caused by a failure of the ink to flow out in a smooth, uniformly thick film) [@Goldschmidt2003]. Therefore, an additive package (i.e. defoamers, surface tension modifiers) that are commonly used in commercial formulations is not included in our ink formulation.
The high viscosity of our ink means that a suitable processing method is required to sufficiently disperse the agglomerates and aggregates of the graphene flakes into primary particles [@Ng2019e]. Hence, three-roll milling is selected as the main method of dispersion. A three-roll mill consists of a feed roll, a center roll and an apron roll that rotate in opposing directions and at different speeds. It is commonly used in ink, paint and coating industries to disperse pigment particles within high viscosity liquids [@Goldschmidt2003]. As a dispersion method, it utilises the internal shear and impact forces generated from the movement of the three rollers to break down agglomerates and aggregates into primary particles.
In our formulation, the EC is first mixed with graphene powder (CamGraph G3, Cambridge Nanosystems, avg. flake thickness $\sim$12.4 nm, avg. flake size $\sim$94.5 nm) until the premix becomes a coarse slurry; Fig. \[panel1\]a. We note that the dimensions of the graphene powder fall out of the scope of nomenclature postulated by academic literatures which define 2D graphitic materials with a thickness and/or lateral dimension less than 100 nm as “Graphite nanoplates”, “graphite nanosheets” or “graphite nano-flakes” [@Bianco2013]. This is to distinguish from finely milled graphite powders whose thickness is typically $>$100 nm and Few Layer Graphene (FLG) with layer numbers of 2 to about 5 and Multi-Layer Graphene (MLG) with layer numbers of between 2 and 10. We use the term “graphene powder” to describe the starting 2D graphitic material as it is a term used by the manufacturer. The slurry is then processed by the three-roll mill six times until it assumes a glossy, paste-like appearance. Figure \[panel1\] depicts photographs and scanning electron microscope (SEM) images of the graphene ink as it is processed through the three-roll mill from pass 1 to 6. With subsequent passes, improved dispersion of the graphene flakes into primary particles within the ink gradually lowers the sheet resistance (R$_s$) and measurement variability from different areas of the print; Fig. \[panel1\]b. The R$_s$ saturates at 6 passes and does not improve further with increasing three-roll milling; Supplementary Fig. 6. In our experiments, it is able to achieve separation of the graphene agglomerates from an average flake diameter of $\sim$90.5 nm to $\sim$40.6 nm and an average thickness from 12.4 nm to 2.5 nm; supplementary Fig. 2. Raman measurements of the graphene powders before and after ink formulation do not show any noticeable changes; supplementary Fig 3.
We observe no major variations and shifts in either the $G$ peak, $2D$ peak or I($D$)/I($G$) ratio indicating that three-roll milling during the ink formulation does not cause any significant morphological or chemical change of graphene. On the other hand, AFM data shows a reduction in both the flake lateral dimension and thickness despite having no observed changes in the Raman spectrum; supplementary Fig. 2. Hence the combination of AFM data and Raman data indicates three-roll milling provides a good method of mechanically dispersing the aggregates of the starting graphene powder within the EC polymeric binder to improve percolation.
In considering the screen printing of the water-insoluble graphene ink, the formulation must be designed to be homogenized (well dispersed graphene flakes), stable (against aggregation or precipitation) and have thixotropic (shear thinning) rheological properties [@Ng2019e]. This is shown by previous demonstrations such as those in references [@Arapov2016] and [@Karagiannidis2016]. As the ink is drawn along the screen *via* the squeegee during printing, the shear stress rate ($\dot{\gamma}$) acting upon the ink changes from very low ($\sim$1$ s^{-1}$) to very high ($\sim$1000$ s^{-1}$) [@Ng2019a; @Dybowska-Sarapuk2015] while the viscosity ($\eta$) decreases [@Ng2019a; @Rosu1999]. Upon deposition, $\dot{\gamma}$ decreases as shear rate from the squeegee is removed while $\eta$ increases to form a solid, unbroken image on the substrate surface with a rectangular cross sectional profile. The narrowest printed lines we have achieved on PET are of 200$\mu$m width with a line edge roughness of 4.04 $\mu$ m.; Supplementary Fig. 5. To ensure our formulated ink has these properties, a combination of OFAT, for qualitative screening and DOE for quantitative optimisation are utilised. A detailed explanation on the use of OFAT and DOE is given in supplementary note 1.
We use OFAT to firstly optimise the rheology of the ink by varying the graphene content, thereby determining the maximum loading of graphene permissible in the EC binder before the ink becomes unprintable. OFAT is a method of experimentation that varies one factor at a time in order to find an optimum response and is highly useful for bi-variate analysis [@Ng2019e; @Jacquez1998]. We define the screen printability of the ink as having starting viscosity under zero shear being in between 2 to 10 Pa s [@Ng2019e; @Hu2018]. To derive the maximum loading of graphene, first 20 wt.% of EC powder is fully dissolved in terpineol *via* stator-rotor mill, creating the EC binder. The graphene powder is then dispersed in this EC binder *via* three-roll mill to ascertain the maximum loading of graphene before the ink becomes unprintable. We find that the maximum amount of graphene that can be introduced to the EC binder before the viscosity of the ink exceeds the upper limit of the screen printable viscosity range ($\sim$2-10 Pa s) is $\sim$23 wt.%; Fig. \[panel1\]c. In addition, rheological characterisation is also carried out on the ink at this stage to ensure it has the necessary shear-thinning properties as depicted in \[panel1\]d for screen printing. To further characterise shear thinning behaviour, the measured viscosity and shear rate were plotted in a log-log plot to observe power law behaviour; Fig. \[panel1\]e. The power law model, also known as the Ostwald de Waele model is a rheological model that can be applied to shear thinning fluids to observe their behaviour under shear [@Bjrn2012]. The model can be expressed mathematically using the following equation:
$$\eta = K ^* \dot{\gamma}^{n-1}
\label{powerlaw}$$
where $K$ is the consistency factor and $n$ is the power law index; $n <$1 represents shear thinning behaviour, $n$ = 1 represents Newtonian behaviour and $n >$1 = shear thickening behaviour [@Bjrn2012]. The resultant curve with a power law model, yield a good fit to the flow curve (R$^2$ = 0.9998) between the values 5 to 10 $^2$ s$^1$ s$^{-1}$. We apply a curve fitting based on equation \[powerlaw\] to the log-log plot which determine $K$= 9530 m Pa.s$^n$ and $n$ = 0.403, indicating that the graphene ink behaves as a viscous fluid that is very shear thinning.
Using the results of the first set of OFAT experiments, we establish a realistic range of parameter levels, based on the three basic ink components- graphene (10 – 23 wt.%), EC binder (75 – 88 wt.%) and IPA (0 – 2 wt.%) as a solvent. These ranges are used as the upper and lower limits for the DOE to optimise the conductivity of the ink. We note that the above ranges may differ should graphene from a different source be used. DOE is a widely-used statistical approach to experimental formulation in screening and optimising parameters (also known as factors) [@Ng2019e; @Jacquez1998] and multivariate analysis. Inks are produced according to the parameter ranges indicated in Fig. \[panel1\]e and plotted in a manner reflecting the interaction between graphene, IPA and EC binder. The experiments and results of the DOE are given in supplementary information and have also been plotted in Fig. \[panel1\]f in a ternary plot reflecting the interaction between the three ink components. An optimum formulation for the screen printing graphene ink is established as graphene (23 wt.%), EC binder (75 wt.%) and IPA (2 wt.%) with a R$_s$ response of 64 $\Omega$ $\square^{-1}$ while on PET. The formulated ink is stable and not prone to sedimentation for more than 1 month; Supplementary Fig. 4. The lowest measured value of R$_s$ at 25 $\mu$m is 24 $\Omega$ $\square^{-1}$ when deposited on glass (bulk conductivity = 0.17 $\times$ 10$^4$ S m $^{-1}$). We stress that this value represents the bulk conductivity of the screen printed ink without further annealing, compression, and partial or full decomposition of the non-conductive polymeric binder. As will be discussed later, this variation in R$_s$ is attributed to differences in surface roughness. We further note that the above process optimises the electrical conductivity while maintaining the ink printability, and to our knowledge, have not been methodically applied to formulation of graphene or other 2D material inks in published literature. Although the final conductivity value will vary depending on the source of graphene and type of binder used, the above discussion provides a systematic and universal approach to achieve optimised ink formulation.
Water-assisted conformal printing method
----------------------------------------
![Water assisted conformal printing. **a** Schematic of screen printing of graphene ink on to the SF. **b** Screen printed conductive pattern array on water-soluble SF. **c** Schematic of the water assisted conformal printing process starting with the floating of the printed SF on water followed by the dissolution of the surrounding SF, leaving behind residual wet SF underneath to act as an adhesive to the substrate. Immersion of the substrate into the water bath to ‘fish’ the conductive pattern out which eventually adheres to the substrate by drying of the SF. **d** Interaction of the printed graphene ink with water soluble SF over 120 s. The SF is coloured using a food dye. **e** Photograph of the graphene pattern on a glass rod with representative SEM views of the top and cross section. Scale bar 5 $\mu$m **f** Comparison of R$_s$ before and after water assisted conformal printing across a variety of printed graphene ink thicknesses on PET.[]{data-label="panel2"}](panel2){width="\textwidth"}
Figure \[panel2\]a-g schematically illustrates the key steps in the water-assisted conformal printing process. This involves firstly screen printing the formulated graphene ink directly on to a thin ($\sim$30 $\mu$m), water soluble SF of PVA that has been previously cast on a polytetrafluoroethylene (PTFE) plate using screen printing; Fig. \[panel2\]a. For this, the water-insoluble graphene ink is first flooded over the screen. A squeegee is then drawn across it as shown in figure \[panel2\]a, applying shear forces to push the ink through the open pores of the screen [@Ng2019a] directly on to the SF, producing an array of conductive patterns (length, 200 mm; height, 50 mm per pattern); Fig. \[panel2\]b. When printing multi-layered structures, the first deposited layer of ink is allowed to dry before the next layer is printed over the first print. The SF containing the single and multi-layered patterns are then detached from the PTFE plate *via* simple peeling.
The printed patterns on the SF are then floated on to the surface of a de-ionised water bath (25 $^{\circ}$C); Fig. \[panel2\]c. This allows the dissolution of PVA ($\sim$120 s PVA dissolution time), while retaining the single and multi-layered patterns which are water insoluble with the exception of an ultrathin layer of still wet PVA SF ($\sim$20 $\mu$m); Fig. \[panel2\]d. The object to be printed on is then immersed into the water bath to ‘fish’ out the conductive pattern, allowing the pattern to adhere to it with the wet SF acting as the main interface between the ink patterns and the substrate; Fig. \[panel2\]e, f. The printed pattern conforms directly to the topography of the 3D object. Upon desiccation, the residual SF under the conductive patterns dries, reduces in thickness to $\sim$15 $\mu$m and acts as an adhesive for the graphene patterns directly on to the substrate; Fig. \[panel2\]e. We note that this process is still in the proof-of-concept stage and will require further optimisation before it can be scaled-up.
The success of this method of conformal printing relies on the interaction between the water-insoluble graphene ink previously printed in a pattern with the water soluble SF film. To illustrate this in figure \[panel2\]d, the water-soluble PVA is artificially coloured brown with food dye and a water-insoluble graphene pattern printed on to it *via* screen printing. This composite structure is then exposed to water and observed for 120 s which is the time needed for the water-soluble SF to fully dissolve around the printed pattern at 25 $^{\circ}$ C, while leaving the graphene pattern fully intact (Fig. \[panel2\]d), with the ultrathin PVA layer underneath acting as the ink-substrate interface.
Figure \[panel2\]e shows the final printed pattern on a glass rod and corresponding SEM images of the top and cross-sectional view of the printed conductive pattern, revealing the dried SF film ($\sim$15 $\mu$m) acting as the adhesive layer between the conductive pattern and the glass substrate. We also carry out a pull off adhesion test of a transferred pattern on a glass substrate using a BGD 500 adhesion pull off tester. From three separate measurements, we find that it requires an average of 8.8 $\times$ 10${^5}$ N m${^{-2}}$ of force to detach the pattern from the substrate. We observe a $\sim$40 % improvement in adhesion over time as the same set of three adhesion tests carried out three months later on the same sample requires an average of 1.25 $\times$ 10${^6}$ N m${^{-2}}$ of force to pull off. This is comparable to our measurements on commercial carbon inks printed on glass substrates (1.1 $\times$ 10${^6}$ N m${^{-2}}$). Representative sheet resistance (R$_s$) data recorded for the patterns show no significant increase in R$_s$ before and after transfer; Fig. \[panel2\]f. At thinner prints of graphene ink ($\sim$4 $\mu$m), there is an observed improvement in R$_s$, presumably due to polymer shrinkage after the dissolution of the SF. As the printed graphene ink thickness increases to 25 $\mu$m, the electrical response becomes increasingly monotonic as R$_s$ drops to 71$\sim$88 $\Omega$ $\square^{-1}$ after transfer. We note that this is achieved without any post-processing other than drying at ambient temperature ($\sim$25 $^{\circ}$C). The predictable change in R$_s$ after the transfer process underscores the potential of this technique to fabricate electrical devices with well-defined performance parameters on to arbitrarily-shaped objects without the need for any post deposition treatment.
![Conformal printing on to various substrates. **a** In a spiral pattern on glassware. **b** On a 3D printed thermoplastic dome. **c** A 2 mm graphene strip on to a latex glove on a finger. **d** On a 3m textile bandage in a sensor pattern. **e** On to pink and white marshmallows, green heart-shaped gummy candies and two-toned white/pink candies. Scale bar 1 cm. Top view SEM images of the **f** section highlighted in red in **c**, **g** section highlighted in green in **d** and **h** section highlighted in orange in **e**. Scale bar 5 $\mu$m. **i** R$_s$ of patterns after water assisted printing on different substrates correlated with R$_a$.[]{data-label="panel3"}](panel3){width="85.00000%"}
To highlight the versatility of this technique, we print single and multi-layered patterns on to various arbitrarily shaped objects: (1) hard, conformal, (2) flexible and (3) soft, conformal substrates, even with sharp angles; Supplementary Fig. 8. We note that the process is most proficient when printing on surfaces with limited curvature. Certain shapes, especially 3D concave surfaces (e.g. the inside of a dome), are difficult to print on using this method. In demonstrating the ability to print on hard conformal substrates, we create spiral conductive graphene patterns directly on to the surface of a glass beaker (figure \[panel3\]a) and on a 3D-printed dome made of polylactic acid (PLA), a thermoplastic; Fig. \[panel3\]b. In both cases, the resulting prints yield continuous lines that are electrically conductive in their entirety without damaging the original dimensions of the pattern (R$_s$ $\sim$120 $\pm$10 $\Omega$ $\square^{-1}$ at $\sim$25 $\mu$m before and after transfer). Overall, the process is suitable for most hard surfaces and there is no significant change in the dimensions of the printed lines at $\sim$25 $\mu$m thickness or structures before and after printing. Large-device prints (130 mm $\times$ 130 mm) can also be achieved, especially on the glass beaker. The resolution of the deposited pattern (i.e. edge definition) is also highly dependant on the printing method employed in the process. We therefore propose that this water-assisted conformal printing process will work just as well with other high-throughput forms of printing, including flexography, rotogravure and inkjet printing [@Ng2019a; @R.H.Leach1993]. We also note that thin lines like those in Fig. \[panel3\]b run the risk of deformation during the dissolution phase due to the natural, uneven dissolution of the SF film in water at 25 $^{\circ}$C. As such, the prints in Fig. \[panel3\](a, b) are reliably achieved only with a precise 120 s dissolution time in our experimental setup.
Adhesive bandages are used daily to protect small wounds and adhere securely to the body and conform to every contortion of the skin. Similarly, nitrile gloves adapt and conform well to the human body. To date, research has gone into developing sensors that are fabricated on to bandages [@Yamada2011] and developing other forms of human motion detection devices on to gloves [@Zhao2017] for wearable applications. To demonstrate the ability of this strategy to conformally print on to flexible substrates, we therefore use nitrile gloves and bandages as target substrates. Using the method described above, a thin 1 mm width strip is conformally printed directly on the tip of the finger of a latex glove. A serpentine pattern which could potentially be used for strain sensing is also conformally printed on to an off-the-shelf 3M Durapore^TM^ textile bandage. SEM images at the interface of the conductive graphene patterns and the substrates (nitrile glove: Fig. \[panel3\]f, Durapore bandage: Fig. \[panel3\]g, candies: Fig. \[panel3\]h) reveal the nature of the water transfer process whereby the functional patterns are adhered and ‘sit’ atop the substrates as compared to being printed directly. We note that the adherence of the printed patterns to the substrate is highly dependent on a water-soluble PVA layer which renders the final product water-sensitive. Therefore the potential of the final application being exposed to water must be considered. Additionally, we recognise that many functional devices may include water-sensitive materials and their compatibility with this method must also be taken into account.
Finally, we evaluate the ability of the water-assisted conformal printing method to print on soft substrates. To demonstrate this, we print a series of patterns on assorted candies including marshmallows, green heart-shaped gummy candies and two-toned white/pink candies; Fig. \[panel3\]e. Despite mild deformation, the patterns are able to maintain their shape, electrical conductivity and adhesion to their respective substrates. Printing conductive patterns directly on to conventionally non-functionalised surfaces such as those for candies allows for the fabrication of small form factor devices which will be further demonstrated in the following section.
We note that the properties of the water-assisted conformally printed structures may be affected by the roughness of the target substrate. We take average roughness (R$_a$) measurements *via* profilometry of the various above-mentioned substrates and find that increased R$_a$ is correlated with increased R$_s$ of single-layer water-assisted conformal patterns; Fig. \[panel3\]i. This is attributed to the printed patterns taking on the different unique morphologies of the respective substrates. The process we present is a proof-of-concept demonstration of conformal printing. One particular area for further work is in effective registration of the functional patterns. To achieve this, two strategies that are compatible with our process and could potentially be used are the water-level lowering method [@Saada2017] or automatic registration using robotics [@LeBorgne2017].
Applications of water-assisted conformal printing
-------------------------------------------------
We have already shown that water-based conformal printing can be used as a simple and effective method to create functional printed patterns on to different substrates while maintaining the properties of the structure. This allows fabrication of functional devices directly on to previously-inert surfaces. To demonstrate this, we conformally-print devices, including joule heaters on glassware, conductive patterns on small gummy candies for capacitive based proximity sensing and deformation sensors on marshmallows and candies. We also demonstrate the ability of this method to conformally print multi-layered devices by demonstrating the fabrication of a graphene/PU/graphene thin film parallel plate capacitor and subsequent water transfer on to soft and hard substrates.
![Applications of water assisted conformal printing. **a** Temperature evolution curves of conformally printed spiral heater under constant voltages of 10 V, 15 V, 20 V, 25 V and 30 V. **b** Photograph of a spiral pattern functioning as a joule heater on a round bottom flask with the corresponding IR image. **c** Photograph of a spiral pattern on a glass beaker with a corresponding IR image. **d** Conformally printed square pattern functionalising green heart-shaped candies for capacitive touch sensing. **e** Circuit discharge time before and after proximity capacitive sensing of the candy. **f** Conductivity response of the deformation of a marshmallow’s surface. **g** Deformation sensor printed on two-toned white/pink candies with corresponding $\Delta R/R$ responses when applied with increasing levels of force. **h** Average $\Delta R/R$ response in relation to increasing levels of force.[]{data-label="panel4"}](panel4){width="80.00000%"}
Joule heaters fabricated *via* printing nanomaterials [@Kim2013] such as graphene [@Kang2011; @Sui2011] are of great research interest. Joule heating, also known as ohmic heating is the process by which the passage of an electric current through a conductor produces heat and is currently the primary method by which many thin-film heaters function. Printed graphene heaters are have huge potential for disposable devices such as wearable heaters for therapeutic purposes [@Jang2017]. Demonstrations to date on flat surfaces has seen screen printing [@He2017] emerging as an attractive method for the large-scale fabrication of joule heaters. This is due to its ability to print thicker layers of functional materials accurately compared to other conventional printing techniques. We demonstrate the potential of our water-assisted conformal printing method to fabricate joule heaters on arbitrary substrates.
To realise joule heaters, we conformally print an unbroken spiral pattern (130 mm $\times$ 130 mm) on to a beaker and a round bottom flask; Fig. \[panel3\]a, b. The prints are allowed to dry in ambient temperature ($\sim$25 $^{\circ}$C). We then apply varying voltages of 10 V, 15 V, 25 V and 30 V using a DC power supply producing a maximum of 54.5 $^{\circ}$C of heating in 200 s; Fig. \[panel4\]a. Representative IR images of heating are depicted in Fig. \[panel4\](b, c). The heater functions consistently throughout all voltage inputs, demonstrating how water-assisted conformal printing could be used to create small devices on arbitrarily shaped, previously non-conductive surfaces.
We next demonstrate the use of water-assisted conformal printing for proximity capacitive sensing. A proximity capacitive sensing device consists of two conductive electrodes separated by a dielectric. To realise this, we first conformally print a conductive graphene pattern on a candy (*T* electrode). The other conductive graphene pattern (*S* electrode)is screen printed on to the reverse side of PET where PET acts as the dielectric. We use a 16-bit microcontroller that continuously charges the *S* electrode every 250 ms intervals to 3.3 V, generating a surrounding electric field. After a fixed time period at 3.3 V, the microcontroller starts to discharge the *S* electrode through a 512 k*$\Omega$* resistor. This voltage is brought to 0 V through a Schmitt trigger if the *S* electrode discharges to below 1 V or if the overall discharge time exceeds 40 $\mu$s. A threshold of 35 $\mu$s (denoted by a vertical dashed line; Fig. \[panel4\]e) is set to trigger a response from the microcontroller. As seen in Fig. \[panel4\]e, the discharge time for the *S* electrode alone (i.e. before sensing) is $\sim$ 30 $\mu$s. When the *T* electrode is brought in close proximity to the *S* electrode, it causes a disruption in the electric field, resulting in an increase in overall capacitance. [@Yao2014; @Hayashi2005; @Cotton2009]. This is manifested by an increase in the RC constant, and therefore, discharge time of the *S* electrode beyond the set threshold of 35 $\mu$s; Fig. \[panel4\]e, triggering a response from a speaker connected to the microcontroller. The sensing process is shown in the supplementary video. This demonstrates how conformally printed conductive patterns could enable any object for proximity capacitive sensing in a simple, cost-effective manner.
![Conformal printing of parallel plate capacitors. Photographs of a capacitor on **a** white/pink candy, **b** acrylic and **c** glass substrate. Frequency response of **d** the capacitance **e** Impedance value and **f** phase angle of the capacitor before and after conformal printing. Inset: SEM image of cross-section of printed graphene/pu/graphene parallel plate capacitor. The graphene layers have been false-coloured for clarity. Scale bar: 30 $\mu$m[]{data-label="panel5"}](panel5){width="90.00000%"}
Graphene based sensors capable of detecting motion *via* resistive changes responding to force have already been widely researched [@Amjadi2016; @Tian2014] and have been used in a variety of applications including human-motion detection. We demonstrate the versatility of our printing method in the fabrication of sensors functioning in a similar manner capable of detecting deformation on soft substrates. For this, we conformally print a graphene pattern directly on to the surface of a marshmallow to measure its surface deformation when heated at a constant temperature. The marshmallow is placed on a hotplate pre-heated to 300 $^{\circ}$C and resistance readings of the deformation sensor are taken every 5 s for a period of 360 s; Fig. \[panel4\]f(inset). The melt profile allows us to monitor the corresponding structural deformation in real time. In a second resistive deformation sensor demonstration, we conformally print a conductive square of graphene ink directly on a white/pink candy; Fig. \[panel4\]h. Increasing levels of force of 0.05 N, 0.06 N, 0.07 N, 0.08 N and 0.09 N are applied at 5 second intervals to the square of graphene as shown in Fig. \[panel4\]g. Two strips of copper tape are attached to the edges of the printed graphene pattern as terminals and the resistance measured using a multimeter as force is applied. This surface deformation sensor can achieve a relative change in resistance ($\Delta R/R$) upon application of force in a repeatable fashion due to the elasticity of the substrate. The $\Delta R/R$ response for the different forces applied is linear up 0.08 N; see Fig. \[panel4\]h. At 0.08 N of force, the trend appears to become non-linear, indicating that the graphene pattern is approaching its elastic limit. Additional force applied past 0.09 N no longer generates an effective resistive response as resistances do not return to baseline levels ($\Delta R/R$ = 0 %). This indicates that the possible effective range of the deformation sensor on this substrate is between 0.05 N to 0.08 N. The ability to conformally print directly on to small, soft and delicate surfaces again highlights the versatility of water-based conformal printing to produce small item sensors that are able to provide comprehensive information for a wide variety of applications. For instance, in our demonstration, the ability to print directly on to food items and derive physical information such as deformation rate could be applied to other fields in food and agriculture.
The above examples showcase the ability of water-assisted conformal printing in the fabrication of single layered devices only. However, the use of conventional printing techniques in tandem with conformal printing also means that this method could be easily adapted to printing multi-layered devices such as parallel-plate capacitors. The capacitors fabricated here consist of printed, stacked layers of graphene/PU/graphene. The bottom electrode is first deposited *via* screen printing on to the SF. This is followed by blade applicator coating of pre-polymerised, liquid PU which is then allowed to cure at room temperature, producing a $\sim$100 $\mu$m thick dielectric layer. Finally, we screen-print the top graphene electrode in a manner similar to the deposition of the bottom electrode over the cured PU film, forming a graphene/PU/graphene capacitive structure (10 mm $\times$ 10 mm overlap area). The capacitor is then conformally printed on to a variety of substrates including glass, acrylic and white/pink candies; Fig. \[panel5\]a-c.
To characterise the properties of the capacitor printed on to the glass substrate, impedance analysis is carried out before and after the transfer process. Figures \[panel5\]d-f present the frequency response of the capacitor with a 10 mm $\times$ 10 mm electrode overlap area (as shown in Fig. \[panel5\]c) and a dielectric thickness of $\sim$100 $\mu$m (measured from the cross section SEM image in Fig. \[panel5\]f: inset). The capacitance ($C$) ($\sim$40pF) and impedance ($|Z|$) across the measured range (10$^4$ $\sim$ 10$^7$ Hz) before and after water-assisted printing does not show any significant change; Fig. \[panel5\]d, e. The data recorded for the phase angle indicates a dominating capacitive component with a corresponding phase angle of -90 $^{\circ}$ from 10$^4$ Hz to 10$^7$ Hz, consistent with the operations of a typical printed capacitor with an R-C equivalent behaviour; Fig. \[panel5\]f. The relative dielectric constant ($\epsilon{_r}$) extracted from the measurements gives a value of $\sim$4.6 through the frequency range 10$^4$ $\sim$ 10$^7$ Hz. This is in agreement with previous literature stating the $\epsilon{_r}$ of a pure PU matrix is $\sim$3.2-8.0 [@Li2016; @Huang2004]. The transfer of the capacitor using waster-assisted conformal printing directly on to arbitrary substrates demonstrates exciting opportunities in using our printing process in the fabrication of more complex multi-layer, multi-component devices and systems from functional inks of 2D crystals.
Conclusion
==========
In this work, we demonstrate a room-temperature method of conformally printing functional structures on an arbitrarily-shaped 3D objects using a water-assisted technique. This is achieved by carefully designing a screen-printable, water insoluble conductive graphene ink *via* OFAT and DOE methods of optimisation. The printing method is highly versatile in its ability to fabricate functional structures on to arbitrarily-shaped surfaces and substrates (glass, 3D printed objects, textiles, nitrile and gummy candies) without the loss of the electrical properties. To demonstrate the potential applications of this conformal printing, we fabricate joule heaters directly on to glasswares that are capable of achieving 54.5 $^{\circ}$C in 200 s. In another demonstration, we show how small items such as a gummy candy can also be rendered functional to make it capable of proximity capacitive sensing. Small-item sensing is also realised in another demonstration where we print versatile and highly sensitive deformation sensors on delicate substrates such as marshmallows and gummy candies. Finally, we demonstrate multi-layered device fabrication by conformally printing a graphene/PU/graphene parallel-plate capacitor on both soft and hard substrates without any noticeable change in device parameters. Our versatile conformal printing method gives the ability to print devices on different arbitrarily-shaped 3D objects without any harsh chemical or temperature treatment, lending new functionalities to previously-inert surfaces.
Methods {#methods .unnumbered}
=======
Ink preparation, printing and characterisation {#ink-preparation-printing-and-characterisation .unnumbered}
----------------------------------------------
The ink binder is first prepared by mixing EC (Sigma Aldrich, 200646, viscosity 4 cP, 5% in toluene/ethanol) with terpineol (Sigma Aldrich, W304506, $\alpha$-Terpineol) into a glass jar at a concentration of EC:terpineol 20:80 wt.%. The mixture is then mixed in a Silverson L5m laboratory high shear stator-rotor mill at 4,000 rpm for 45 minutes. The glass jar sat in a water-bath during dissolution and homogenisation to dissipate the heat generated from the mixing process. Commercially-sourced graphene powder (Cambridge Nanosystems) and IPA is then added to the homogenised binder solution in a 23:2:75 wt.% graphene:IPA:EC and mixed until a coarse slurry is formed. The slurry is then put through a three-roll mill (EXAKT 50I) 6 times with roller gaps set to $\sim$30 $\mu$m until a glossy appearance is achieved. Prints are carried out using a hand operated screen printing press (80 mesh/inch). R$_s$ measurements are taken using a four-point probe system. A parallel plate rotational rheometer (DHR rheometer, TA instruments) is used to evaluate the viscosity of the inks as a function of the shear rate.
SF preparation and characterisation {#sf-preparation-and-characterisation .unnumbered}
-----------------------------------
For water soluble SF, PVA (MP Biomedicals, MW 15,000 Da) is mixed with de-ionised (DI) water into a sealed container at a ratio of 20:80 wt.% and stirred with a magnetic stir bar for 12 h at 400 rpm under 100 $^{\circ}$C. To cast the SF film, the obtained solution is bar coated at a 50 $\mu$m wet thickness on to a PTFE plate and is allowed to dry at ambient temperature ($\sim$25 $^{\circ}$C). The dry thickness thickness is measured *via* SEM. After screen printing and curing of the single or multi-layered functional structures on to the SF, the resultant composite stack is detached from the PTFE plate *via* simple peeling. For exposure studies, graphene ink is printed in a square pattern on to the SF colored with food dye, peeled and exposed to water using an enclosed glass observation jar and timed.
Joule heater {#joule-heater .unnumbered}
------------
IR imaging is carried out using a FLIR ONE thermal imaging camera. Temperature curve measurements are carried out using an arduino UNO unit with a TMP-36 temperature sensor attached directly on to the graphene heating element. Readings are sampled every 1 ms and averaged over 1 s.
Proximity sensor {#proximity-sensor .unnumbered}
----------------
The graphene $S$ electrode is fabricated *via* screen printing of the ink in to a single strip and connected to the microcontroller using copper tape. Water-assisted conformal printing is used to print 1 cm $\times$ 1 cm squares ($T$ electrode) of graphene directly on to the gummy candies.
Deformation sensor on marshmallow and white and pink candy {#deformation-sensor-on-marshmallow-and-white-and-pink-candy .unnumbered}
----------------------------------------------------------
A copper tape is attached to the edges of the printed graphene sensors to be used as contacts. Resistance measurements are carried out using a Keithley 2100 Digital Multimeter with probes attached directly to the copper tape *via* crocodile clips.
Parallel plate capacitor {#parallel-plate-capacitor .unnumbered}
------------------------
Commercial pre-polymerised PU is obtained as a 2-component resin and hardener. The two components are mixed (resin to hardener ratio 10:1) and then degassed in a vacuum jar for 10 minutes before application *via* blade applicator as the dielectric layer. A Newtons4th Impedance Analysis Interface 2 with Newtons4th PSM3750 is used to measure the frequency response of the capacitor.
Scanning electron microscopy {#scanning-electron-microscopy .unnumbered}
----------------------------
SEM images are taken with a high-resolution LEO 1530 VP Ultra-high performance VP SEM system. The field emission gun is operated at an accelerating voltage of 2.4 keV.
Atomic force microscopy {#atomic-force-microscopy .unnumbered}
-----------------------
Measurements of graphene flakes are taken with a Bruker Dimension Icon in scan assist mode. Samples are diluted 100 times and then drop cast on to pre-cleaned (isopropanol) Si/SiO$_2$ substrate.
Profilometry measurements {#profilometry-measurements .unnumbered}
-------------------------
Profilometry measurements of the substrates are taken with a Bruker DektakXT stylus profilomter with a scan length of 4 mm, a resolution of 0.66 $\mu$m, stylus force of 3 mg and a stylus tip radius of 12.5 $\mu$m.
Acknowledgements {#acknowledgements .unnumbered}
================
We acknowledge support from EP/L016087/1, National Research Fellowship (NRF) of Korea and the Graphene Flagship.
[10]{} url \#1[`#1`]{} urlprefix \[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
** (, , ).
*et al.* . In **, (, ).
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , & . In **, vol. , ().
, , & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . In **, (, ).
& ** (, ).
*et al.* . ** ****, ().
, & . ** ****, ().
, , , & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
& . ** ****, ().
. ** ****, ().
*et al.* . ** ****, ().
, *et al.* . ** ****, ().
, *et al.* . ** ****, ().
, *et al.* . ** ****, ().
, *et al.* . ** ****, ().
, *et al.* . ** ****, ().
, , & . ** ****, ().
& ** (, ).
** (, ).
*et al.* . In **, (, , ).
, , . ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ****, ().
, & . ** ****, ().
. ** ****, ().
*et al.* ** (, ).
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ().
*et al.* . ** ****, ().
& . ** ****, ().
, , & . ** ****, ().
, & . ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[**Abstract.**]{} We study first-order model checking, by which we refer to the problem of deciding whether or not a given first-order sentence is satisfied by a given finite structure. In particular, we aim to understand on which sets of sentences this problem is tractable, in the sense of parameterized complexity theory. To this end, we define the notion of a graph-like sentence set, which definition is inspired by previous work on first-order model checking wherein the permitted connectives and quantifiers were restricted. Our main theorem is the complete tractability classification of such graph-like sentence sets, which is (to our knowledge) the first complexity classification theorem concerning a class of sentences that has no restriction on the connectives and quantifiers. To present and prove our classification, we introduce and develop a novel complexity-theoretic framework which is built on parameterized complexity and includes new notions of reduction.'
author:
- HUBIE CHEN
bibliography:
- '/Users/hubiec/Dropbox/active/writing/hubiebib.bib'
title: 'The Tractability Frontier of Graph-Like First-Order Query Sets'
---
\[theorem\][Proposition]{} \[theorem\][Assumption]{}
Introduction {#section:introduction}
============
Model checking, the problem of deciding if a logical sentence holds on a structure, is a fundamental computational task that appears in many guises throughout computer science. In this article, we study *first-order model checking*, by which we refer to the case of this problem where one wishes to evaluate a first-order sentence on a finite structure. This case is of principal interest in database theory, where first-order sentences form a basic, heavily studied class of *database queries*, and where it is well-recognized that the problem of evaluating such a query on a database can be taken as a formulation of first-order model checking. Indeed, the investigation of model checking in first-order logic entails an examination of one of the simplest, most basic logics, and it can be expected that understanding of the first-order case should provide a well-founded basis for studying model checking in other logics, such as those typically considered in verification and database theory. First-order model checking is well-known to be intractable in general: it is PSPACE-complete.
As has been articulated [@PapadimitriouYannakakis99-database; @FlumGrohe06-parameterizedcomplexity], the typical model-checking situation in the database and verification settings is the evaluation of a relatively short sentence on a relatively large structure. Consequently, it has been argued that, in measuring the time complexity of model checking, one could reasonably allow a slow (non-polynomial-time) preprocessing of the sentence, so long as the desired evaluation can be performed in polynomial time following the preprocessing. Relaxing polynomial-time computation to allow arbitrary preprocessing of a *parameter* of a problem instance yields, in essence, the notion of *fixed-parameter tractability*. This notion of tractability is the base of *parameterized complexity theory*, which provides a taxonomy for reasoning about and classifying problems where each instance has an associated parameter. We utilize this paradigm, and focus the discussion on this form of tractability (here, the sentence is the parameter).
A typical way to understand which types of sentences are well-behaved and exhibit desirable, tractable behavior is to simply consider model checking relative to a set $\Phi$ of sentences, and to attempt to understand on which sets one has tractable model checking. We restrict attention to sets of sentences having bounded arity.[^1] Here, there have been successes in understanding which sets of sentences are tractable (and which are not) in fragments of first-order logic described by restricting the connectives and quantifiers that may be used: there are systematic classification results for so-called conjunctive queries (formed using the connectives and quantifiers in $\{ \wedge, \exists \}$) [@GroheSchwentickSegoufin01-conjunctivequeries; @Grohe07-otherside; @ChenMueller13-fineclassification], existential positive queries ($\{ \wedge, \vee, \exists \}$) [@Chen14-existentialpositive], and quantified conjunctive queries ($\{ \wedge, \exists, \forall \}$) [@ChenDalmau12-decomposingquantified; @ChenMarx13-blocksorted]. However, to the best of our knowledge, there has been no classification theorem for general first-order logic, without any restriction on the connectives and quantifiers. In this article, we present the first such classification.
[**Our approach.**]{} In the fragments of first-order logic where the only connective permitted is one of the binary connectives ($\{ \wedge, \vee \}$)—such as those of conjunctive queries and quantified conjunctive queries—a heavily studied approach to describing sets $\Phi$ of sentences is a *graphical approach*. In this graphical approach, one studies a sentence set $\Phi$ if it is *graphical* in the following sense: if one prenex sentence $\phi$ is contained in $\Phi$ and a second prenex sentence $\psi$ has the same prefix as $\phi$ and also has the same graph as $\phi$, then $\psi$ is also in $\Phi$. (By the graph of a prenex sentence $\phi$, we mean the graph whose vertices are the variables of $\phi$ and where two vertices are adjacent if they occur together in an atomic formula.) In the fragments where it was considered, this approach of studying graphical sentence sets is not only a natural way to coarsen the project of classifying all sets $\Phi$ of sentences, but in fact, can be used cleanly as a key module in obtaining general classifications of sentence sets: such general classifications have recently been proved by using the respective graphical classifications as black boxes [@ChenMueller13-fineclassification; @ChenMarx13-blocksorted].
In this article, we adapt this graphical approach to the full first-order setting. To explain how this is done, consider that a graphical set $\Phi$ of sentences satisfies certain syntactic closure properties. For instance, if one takes a sentence from such a graphical set $\Phi$ and replaces the relation symbol of an atomic formula, the resulting sentence will have the same graph, and will hence continue to be in $\Phi$; we refer to this property of $\Phi$ as *replacement closure*. As another example, if one rewrites a sentence in $\Phi$ by invoking associativity or commutativity of the connective $\wedge$, the resulting sentence will likewise still have the same graph and will hence be contained in $\Phi$ also. Inspired by these observations, we define a sentence set $\Phi$ to be *graph-like* if it is replacement closed and also closed under certain well-known syntactic transformations, such as associativity and commutativity of the binary connectives (see Section \[sect:preliminaries\] for the full definition).
Our principal result is the complete tractability characterization of graph-like sentence sets (see Theorem \[thm:main\] and the corollaries that follow). In particular, we introduce a measure on first-order formulas which we call *thickness*, and show that a set of graph-like sentences is tractable if and only if it has bounded thickness (under standard complexity-theoretic assumptions). In studying unrestricted first-order logic, we believe that our building on the syntactic, graphical approach—which has an established, fruitful tradition—will facilitate the formulation and obtention of future, more general results.
As evidence of our result’s generality and of its faithfulness to the graphical approach, we note that the graphical classification of quantified conjunctive queries [@ChenDalmau12-decomposingquantified] can be readily derived from our main classification result (this is discussed in Section \[sect:discussion\]); it follows readily that the dual graphical classification of quantified disjunctive queries (also previously derived [@ChenDalmau12-decomposingquantified]) can be dually derived from our main classification result. We therefore give a single classification theorem that naturally unifies together these two previous classifications. Indeed, we believe that the technology that we introduce to derive our classification yields a cleaner, deeper and more general understanding of these previous classifications. Observe that, for each of those classifications, since only one binary connective is present, the two quantifiers behave asymmetrically; this is in contrast to the present situation, where in building formulas, wherever a formula may be constructed, its dual may be as well.
[**Parameterized complexity.**]{} An increasing literature investigates the following general situation:
Given a parameterized problem $P$\
whose instances consist of two parts, where the first part is the parameter, and a set $S$,\
define $P \llbracket S \rrbracket$ to be the restricted version of $P$\
where $(x, y)$ is admitted as an instance iff $x \in S$.
Then, attempt to classify and understand, over all sets $S$,\
the complexity of the problem $P \llbracket S \rrbracket$.
Examples of classifications and studies that can be cast in this situation include [@GroheSchwentickSegoufin01-conjunctivequeries; @Grohe07-otherside; @ChenThurleyWeyer08-inducedsubgraph; @ChenGrohe10-succinct; @ChenDalmau12-decomposingquantified; @DurandMengel13-structuralcounting; @ChenMueller13-fineclassification; @JonssonLagerkvistNordh13-blowingholes]. It is our view that this literature suffers from the defect that there is no complexity-theoretic framework for discussing the families of problems obtained thusly. As a consequence, different authors and different articles used divergent language and notions to present hardness results on and reductions between such problems, and applied different computability assumptions on the sets $S$ considered. We attempt to make a foundational contribution and to ameliorate this state of affairs by presenting a complexity-theoretic framework for handling and classifying problems of the described form. In particular, we introduce notions such as reductions and complexity classes for problems of the above type, which we formalize as *case problems* (Section \[sect:case\]). Although we do not carry out this exercise here, we believe that most of the results in the mentioned literature can be shown to be naturally and transparently expressible within our framework.
In order to derive our classification, we present (within our complexity-theoretic framework) a new notion of reduction which we call *accordion reduction* and which is crucial for the proof of our hardness result. (See Sections \[sect:accordion\] and \[sect:hardness\] for further discussion.) We believe that this notion of reduction may play a basic role in future classification projects of the form undertaken here.
Let us emphasize that, while the establishment of our main classification theorem makes use of the complexity-theoretic framework and accompanying machinery that was just discussed (and is presented in Sections \[sect:case\] and \[sect:accordion\]), this framework and machinery is fully generic in that it does not make any reference to and is not specialized to the model checking problem. We believe and hope that the future will find this framework to be a suitable basis for presenting, developing and discussing complexity classification results.
Preliminaries {#sect:preliminaries}
=============
When $g: A \to B$ and $h: B \to C$ are mappings, we will typically use $h(g)$ to denote their composition. When $f$ is a partial mapping, we use ${\mathsf{dom}}(f)$ to denote its domain, and we use $f {\upharpoonright}S$ to denote its restriction to the set $S$. We will use $\pi_i$ to denote the $i$th projection, that is, the mapping that, given a tuple, returns the value in the tuple’s $i$th coordinate. For a natural number $k$, we use ${\underline{k}}$ to denote the set $\{ 1, \ldots, k \}$.
First-order logic
-----------------
We use the syntax and semantics of first-order logic as given by a standard treatment of the subject. In this article, we restrict to relational first-order logic, so the only symbols in signatures are relation symbols. We use letters such as ${\mathbf{A}}$, ${\mathbf{B}}$ to denote structures and $A$, $B$ to denote their respective universes. The reader may assume for concreteness that relations of structures are represented using lists of tuples, although in general, we will deal with the setting of bounded arity, and natural representations of relations will be (for the complexity questions at hand) equivalent to this one. We assume that equality is not built-in to first-order logic, so a *formula* is created from *atoms*, the usual connectives ($\neg$, $\wedge$, $\vee$) and quantification ($\exists$, $\forall$); by an *atom*, we mean a formula $R(v_1, \ldots, v_k)$ where a relation symbol is applied to a tuple of variables (of the arity of the symbol). A formula is *positive* if it does not contain negation ($\neg$). We use ${\mathsf{free}}(\phi)$ to denote the set of free variables of a formula $\phi$. The *width* of a formula $\phi$, denoted by ${\mathsf{width}}(\phi)$, is defined as the maximum of $|{\mathsf{free}}(\psi)|$ over all subformulas $\psi$ of $\phi$. The *arity* of a formula is the maximum arity over all relation symbols that occur in the formula.
We use $\phi_1 \wedge \cdots \wedge \phi_n$ as notation for $(\cdots ((\phi_1 \wedge \phi_2) \wedge \phi_3) \cdots)$, and $\phi_1 \vee \cdots \vee \phi_n$ is defined dually. We refer to a formula of the shape $\phi_1 \wedge \cdots \wedge \phi_n$ as a *conjunction*, and respectively refer to a formula of the shape $\phi_1 \vee \cdots \vee \phi_n$ as a *disjunction*. Let $\Phi$ be a set of formulas. A *positive combination* of formulas from $\Phi$ is a formula in the closure of $\Phi$ under conjunction and disjunction. A *CNF* of formulas from $\Phi$ is a conjunction of disjunctions of formulas from $\Phi$, and a *DNF* of formulas from $\Phi$ is a disjunction of conjunctions of formulas from $\Phi$.
We will use the following terminology which is particular to this article. A subformula $\psi$ of a formula $\phi$ is a *positively combined subformula* if, in viewing $\phi$ as a tree, all nodes on the unique path from the root of $\phi$ to the parent of the root of $\psi$ (inclusive) are conjunctions or disjunctions. We say that a formula $\phi$ is *variable-loose* if no variable is quantified twice, and no variable is both quantified and a free variable of $\phi$. We say that a formula is *symbol-loose* if no relation symbol appears more than once in the formula. We say that a formula is *loose* if it is both variable-loose and symbol-loose.
We now present a number of syntactic transformations; that each preserves logical equivalence is well-known.[^2]
- Associativity and commutativity of $\wedge$ and $\vee$
- $\exists x (\bigvee_{i = 1}^n \phi_i) \equiv \bigvee_{i=1}^n (\exists
x \phi_i)$,
$\forall y (\bigwedge_{i = 1}^n \phi_i) \equiv \bigwedge_{i=1}^n (\forall y \phi_i)$
- $\exists x (\phi \wedge \psi) \equiv (\exists x \phi) \wedge \psi
\textup{ if $x \notin {\mathsf{free}}(\psi)$}$,
$\forall y (\phi \vee \psi) \equiv (\forall y \phi) \vee \psi
\textup{ if $y \notin {\mathsf{free}}(\psi)$}$
- (Distributivity for $\wedge$ and $\vee$)
$\phi \wedge (\psi \vee \psi') \equiv
(\phi \wedge \psi) \vee (\phi \wedge \psi')$,
$\phi \vee (\psi \wedge \psi') \equiv
(\phi \vee \psi) \wedge (\phi \vee \psi')$
- (DeMorgan’s laws)
$\neg \exists v \phi \equiv \forall v \neg \phi$, $\neg \forall v \phi \equiv \exists v \neg \phi$
$\neg (\phi \wedge \psi) \equiv \neg \phi \vee \neg \psi$, $\neg (\phi \vee \psi) \equiv \neg \phi \wedge \neg \psi$
We say that a set $\Phi$ of formulas is *syntactically closed* if, for each $\phi \in \Phi$, when a formula $\phi'$ can be obtained from $\phi$ by applying one of the syntactic transformations $(\alpha)$, $(\beta)$, $(\gamma)$, $(\delta)$, $(\epsilon)$ to a subformula of $\phi$, it holds that $\phi' \in \Phi$. The *syntactic closure* of a formula $\phi$ is the intersection of all syntactically closed sets that contain $\phi$.
Let us say that a formula $\phi'$ on signature $\sigma'$ is obtainable from a formula $\phi$ on signature $\sigma$ by *replacement* if $\phi'$ can be obtained from $\phi$ by replacing instances of relation symbols in $\phi$ with instances of relation symbols from $\sigma'$ (without making any other changes to $\phi$).
Let $\phi$ be the formula $\forall y \exists x \exists x' (E(y, x) \wedge E(x, x'))$. Let $\tau$ be the signature $\{ E, F \}$ where $E$ and $F$ are relation symbols of binary arity. Each of the four formulas $\phi$, $\forall y \exists x \exists x' (F(y, x) \wedge E(x, x'))$, $\forall y \exists x \exists x' (E(y, x) \wedge F(x, x'))$, and $\forall y \exists x \exists x' (F(y, x) \wedge F(x, x'))$ is obtainable from $\phi$ by replacement; moreover, these are the only four formulas over signature $\tau$ that are obtainable from $\phi$ by replacement.
Let us say that a set of formulas $\Phi$ is *replacement closed* if, for each $\phi \in \Phi$, when $\phi'$ is obtainable from $\phi$ by replacement, it holds that $\phi' \in \Phi$.
A set of formulas $\Phi$ is *graph-like* if it is syntactically closed and replacement closed.
Graphs and hypergraphs
----------------------
When $S$ is a set, we use $K(S)$ to denote the set containing all size $2$ subsets of $S$, that is, $K(S) = \{ \{ s, s' \} ~|~ s, s' \in S, s \neq s' \}$. For us, a *graph* is a pair $(V, E)$ where $V$ is a set and $E \subseteq K(V)$.
Here, a *hypergraph* $H$ is a pair $(V(H), E(H))$ consisting of a vertex set $V(H)$ and an edge set $E(H)$ which is a subset of the power set ${\wp}(V(H))$. We will sometimes specify a hypergraph just by specifying the edge set $E$, in which case the vertex set is understood to be $\bigcup_{e \in E} e$. We associate a hypergraph $(V(H), E(H))$ with the graph $(V(H), \bigcup_{e \in E(H)} K(e))$, and thereby refer to (for example) the treewidth of or an elimination ordering of a hypergraph.
An *elimination ordering* of a graph $(V, E)$ is a pair $$((v_1, \ldots, v_n), E')$$ that consists of a superset $E'$ of $E$ and an ordering $v_1, \ldots, v_n$ of the elements of $V$ such that the following property holds: for each vertex $v_k$, any two distinct lower neighbors $v, v'$ of $v_k$ are adjacent in $E'$, that is, $\{ v, v' \} \in E'$; here, a *lower neighbor* of a vertex $v_k$ is a vertex $v_i$ such that $i < k$ and $\{ v_i, v_k \} \in E'$. Relative to an elimination ordering $e$, we define the *lower degree* of a vertex $v$, denoted by ${\mathsf{lower}\textup{-}\mathsf{deg}}(e, v)$, to be the number of lower neighbors that it has; we define ${\mathsf{lower}\textup{-}\mathsf{deg}}(e)$ to be the maximum of ${\mathsf{lower}\textup{-}\mathsf{deg}}(e, v)$ over all vertices $v$. We assume basic familiarity with the theory of treewidth [@Bodlaender98-arboretum]. The following is a key property of treewidth that we will utilize; here, we use ${\mathsf{tw}}(H)$ to denote the treewidth of $H$.
\[prop:tw-gives-elimination-ordering\] For each $k \geq 2$, there exists a polynomial-time algorithm that, given as input a hypergraph $H$ with a distinguished edge $f$, will return the following whenever ${\mathsf{tw}}(H) < k$: an elimination ordering $e = ((v_1, \ldots, v_m), E)$ with ${\mathsf{lower}\textup{-}\mathsf{deg}}(e) = {\mathsf{tw}}(H)$ and where $\{ v_1, \ldots, v_{|f|} \} = f$.
The treewidth of (for example) a set of graphs ${\mathcal{G}}$ is the set $\{ {\mathsf{tw}}(G) ~|~ G \in {\mathcal{G}}\}$; it is said to be *unbounded* if this set is infinite, and *bounded* otherwise. We employ similar terminology, in general, when dealing with a complexity measure defined on a class of objects.
Parameterized complexity
========================
In this section, we specify the framework of parameterized complexity to be used in this article.
Throughout, we use $\Sigma$ to denote an alphabet over which languages are defined. As is standard, we will sometimes view elements of $\Sigma^* \times
\Sigma^*$ as elements of $\Sigma^*$. A *parameterization* is a mapping from $\Sigma^*$ to $\Sigma^*$. A *parameterized problem* is a pair $(Q, \kappa)$ consisting of a language $Q \subseteq \Sigma^*$ and a parameterization $\kappa: \Sigma^* \to \Sigma^*$. A *parameterized class* is a set of parameterized problems.
\[assumption:non-trivial-languages\] We assume that each parameterized problem $(Q, \kappa)$ has a non-trivial language $Q$, that is, that neither $Q = \Sigma^*$ nor $Q = \emptyset$.
Let us remark on a difference between our setup and that of other treatments. Elsewhere, a *parameterization* is often defined to be a mapping from $\Sigma^*$ to ${\mathbb{N}}$, and in the context of query evaluation, the parameterization studied is typically the size of the query. In contrast, this article takes the parameterization to be the query itself. Since there are finitely many queries of any fixed size, model checking on a set of queries will be fixed-parameter tractable (that is, in the class ${\textup{\small $\mathsf{FPT}$}}$, defined below) under one of these parameterizations if and only if it is under the other. However, we find that—as concerns the theory in this article—taking the query itself to be the parameter allows for a significantly cleaner presentation. One example reason is that the reductions we present will generally be “slice-to-slice”, that is, they will send all instances with the same query to instances that share another query. Indeed, to understand the complexity of a set of queries, we will apply a closure operator to pass to a larger set of queries having the same complexity, using the notion of *accordion reduction* (see Sections \[sect:accordion\] and \[sect:hardness\]); we believe that the theory justifying this passage is most cleanly expressed under the used parameterization.
Let us also remark that we do not put in effect any background assumption on the computability/complexity of parameterizations; this is because our theory does not require such an assumption. ${\ensuremath{\footnotesize \textup{$\Box$}}}$
We now define what it means for a partial mapping $r$ to be *FPT-computable*; we actually first define a non-uniform version of this notion. This definition coincides with typical definitions in the case that $r$ is a total mapping; in the case that $r$ is a partial mapping, we use a “promise” convention, that is, we do not impose any mandate on the behavior of the respective algorithm in the case that the input $x$ is not in the domain of $r$.
\[def:fpt-computable\] Let $\kappa: \Sigma^* \to \Sigma^*$ be a parameterization.
A partial mapping $r: \Sigma^* \to \Sigma^*$ is *nu-FPT-computable with respect to $\kappa$* if there exist a function $f: \Sigma^* \to {\mathbb{N}}$ and a polynomial $p: {\mathbb{N}}\to {\mathbb{N}}$ such that for each $k \in \Sigma^*$, there exists an algorithm $A_k$ satisfying the following condition: on each string $x \in {\mathsf{dom}}(r)$ such that $\kappa(x) = k$, the algorithm $A_k$ computes $r(x)$ within time $f(\kappa(x))p(|x|)$.
A partial mapping $r: \Sigma^* \to \Sigma^*$ is *FPT-computable with respect to $\kappa$* if, in the just-given definition, the function $f$ can be chosen to be computable and there exists a single algorithm $A$ that can play the role of each algorithm $A_k$; formally, if there exist a computable function $f: \Sigma^* \to {\mathbb{N}}$, a polynomial $p: {\mathbb{N}}\to {\mathbb{N}}$, and an algorithm $A$ such that for each $k \in \Sigma^*$, the above condition is satisfied when $A_k$ is set equal to $A$. ${\ensuremath{\footnotesize \textup{$\Box$}}}$
We define ${\textup{\small $\mathsf{FPT}$}}$ to be the class that contains a parameterized problem $(Q, \kappa)$ if and only if the characteristic function of $Q$ is FPT-computable with respect to $\kappa$.
We now introduce the notion of a reduction between parameterized problems. When $A$ is a set, we use ${{\wp}_{\mathsf{fin}}}(A)$ to denote the set containing all finite subsets of $A$.
Let $(Q, \kappa)$ and $(Q', \kappa')$ be parameterized problems. A *FPT-reduction* (respectively, *nu-FPT-reduction*) from $(Q, \kappa)$ to $(Q', \kappa')$ is a total mapping $g: \Sigma^* \to \Sigma^*$ that is FPT-computable (respectively, nu-FPT-computable) with respect to $\kappa$ and a computable (respectively, not necessarily computable) mapping $h: \Sigma^* \to {{\wp}_{\mathsf{fin}}}(\Sigma^*)$ such that:
- for each $x \in \Sigma^*$, it holds that $x \in Q$ if and only if $g(x) \in Q'$; and
- for each $x \in \Sigma^*$, it holds that $\kappa'(g(x)) \in h(\kappa(x))$.
\[assumption:closure-under-reduction\] We assume that each parameterized class $C$ is closed under FPT-reductions, that is, if $(Q', \kappa')$ is in $C$ and $(Q, \kappa)$ FPT-reduces to $(Q', \kappa')$, then $(Q, \kappa)$ is in $C$.
\[prop:fpt-reductions-compose\] Let $(Q, \kappa)$, $(Q', \kappa')$, and $(Q'', \kappa'')$ be parameterized problems.
- If $g$ is a FPT-reduction from $(Q, \kappa)$ to $(Q', \kappa')$ and $h$ is a FPT-reduction from $(Q', \kappa')$ to $(Q'', \kappa'')$, then their composition $h(g)$ is a FPT-reduction from $(Q, \kappa)$ to $(Q'', \kappa'')$.
- Similarly, if $g$ is a nu-FPT-reduction from $(Q, \kappa)$ to $(Q', \kappa')$ and $h$ is a nu-FPT-reduction from $(Q', \kappa')$ to $(Q'', \kappa'')$, then their composition $h(g)$ is a nu-FPT-reduction from $(Q, \kappa)$ to $(Q'', \kappa'')$.
Also, each FPT-reduction from $(Q, \kappa)$ to $(Q', \kappa')$ is an nu-FPT-reduction from $(Q, \kappa)$ to $(Q', \kappa')$.
We may now define, for each parameterized class $C$, a non-uniform version of the class, denoted by ${\textup{{\small $\mathsf{nu}$}-{$ C $}}}$.
(non-uniform classes) When $C$ is a parameterized class, we define ${\textup{{\small $\mathsf{nu}$}-{$ C $}}}$ to be the set that contains each parameterized problem that has a nu-FPT-reduction to a problem in $C$.
It is straightforwardly verified that a parameterized problem $(Q, \kappa)$ is in the class ${\textup{{\small $\mathsf{nu}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$ (under the above definitions) if and only if the characteristic function of $Q$ is nu-FPT-computable with respect to $\kappa$.
We next present two notions of hardness for parameterized classes.
(hardness) Let $C$ be a parameterized class. We say that a problem $(Q, \kappa)$ is $C$-hard if every problem in $C$ has a FPT-reduction to $(Q, \kappa)$. We say that a problem $(Q, \kappa)$ is non-uniformly $C$-hard if every problem in $C$ has a nu-FPT-reduction to $(Q, \kappa)$.
We end this section by observing some basic closure properties of what we call *degree-bounded functions*, which, roughly speaking, are the functions which can serve as the running time of algorithms in the definition of *FPT-computable* (Definition \[def:fpt-computable\]). Let $\kappa: \Sigma^* \to \Sigma^*$ be a parameterization. A partial function $T: \Sigma^* \to {\mathbb{N}}$ is *degree-bounded* with respect to $\kappa$ if there exist a computable function $f: \Sigma^* \to {\mathbb{N}}$ and a polynomial $p: {\mathbb{N}}\to {\mathbb{N}}$ such that, for each $x \in {\mathsf{dom}}(T)$, it holds that $T(x) \leq f(\kappa(x)) p(|x|)$. We will make use of the observation that a partial mapping $h: \Sigma^* \to \Sigma^*$ is FPT-computable with respect to $\kappa$ if and only if there is a degree-bounded function $T$ (with ${\mathsf{dom}}(T) \supseteq {\mathsf{dom}}(h)$) and an algorithm that, for all $x \in {\mathsf{dom}}(h)$, computes $h(x)$ within time $T(x)$.
The following proposition will be of use in establishing that functions are degree-bounded.
\[prop:db-closure\] Let $\kappa$ be a parameterization, and let $T_1, \ldots, T_m: \Sigma^* \to {\mathbb{N}}$ be partial functions sharing the same domain.
1. \[closure:sum\] If each of $T_1, \ldots, T_m$ is degree-bounded with respect to $\kappa$, then $T_1 + \cdots + T_m$ is as well.
2. \[closure:product\] If each of $T_1, \ldots, T_m$ is degree-bounded with respect to $\kappa$, then the product $T_1 \cdots T_m$ is as well.
3. \[closure:polynomial\] Let $q: {\mathbb{N}}\to {\mathbb{N}}$ be a polynomial; if a partial function $T: \Sigma^* \to {\mathbb{N}}$ is degree-bounded with respect to $\kappa$, then $q(T)$ is as well.
For (\[closure:sum\]), one can use the fact that $f_1(k) p_1(n) + \cdots + f_m(k) p_m(n)$ is bounded above by $(f_1(k) + \cdots + f_m(k))(p_1(n) + \cdots + p_m(n))$.
For (\[closure:product\]), it suffices to observe that the product $(f_1(k) p_1(n)) \cdots (f_m(k) p_m(n))$ can be grouped as $(f_1(k) \cdots f_m(k))(p_1(n) \cdots p_m(n))$.
For (\[closure:polynomial\]), it suffices to observe that $q(f(k)p(n))$ is bounded above by $q(f(k)) q(p(n))$; note that $q(p)$ is the composition of two polynomials, and hence itself a polynomial.
Case complexity {#sect:case}
===============
A number of previous works focus on a decision problem $Q$ where each instance consists of two parts, and obtain restricted versions of the problem by taking sets $S \subseteq \Sigma^*$ and, for each such set, considering the restricted version where one allows only instances where (say) the first of the parts falls into $S$. This is precisely the type of restriction that we will consider here. We are interested in first-order model checking, which we view as the problem of deciding, given a first-order sentence and a structure, whether or not the sentence holds on the structure; our particular interest is to study restricted versions of this problem where the allowed sentences come from a set $S$.
It has been useful (see for instance the articles [@ChenDalmau12-decomposingquantified; @ChenMueller13-fineclassification]) and is useful in the present article to present reductions between such restricted versions of problems. In order to facilitate our doing this, we present a framework wherein we formalize this type of restricted version of problem as a *case problem*, and then present a notion of reduction for comparing case problems. We believe that our notion of reduction, called *slice reduction*, faithfully abstracts out precisely the key useful properties that are typically present in such reductions in the literature. Note that, in the existing literature, different articles imposed different computability assumptions on the sets $S$ considered (assumptions used include that of computable enumerability, of computability, and of no computability assumption). One feature of our framework is that such reductions can be carried out and discussed independently of whether or not any such computability assumption is placed on the sets $S$; a general theorem (Theorem \[thm:slice-reduction-gives-fpt-reduction\]) allows one to derive normal parameterized reductions from slice reductions, where the exact computability of the reduction derivable depends on the computability assumption placed on the sets $S$. We now introduce our framework. Suppose that $Q \subseteq \Sigma^* \times \Sigma^*$ is a language of pairs; for a set $T \subseteq \Sigma^*$, we use $Q_T$ to denote the language $Q \cap (T \times \Sigma^*)$ and for a single string $t \in \Sigma^*$, we use $Q_t$ to denote the language $Q \cap (\{ t \} \times \Sigma^*)$.
A *case problem* consists of a language of pairs $Q \subseteq \Sigma^* \times \Sigma^*$ and a subset $S \subseteq \Sigma^*$, and is denoted $Q[S]$. When $Q[S]$ is a case problem, we use ${\mathsf{param}\textup{-}Q[S]}$ to denote the parameterized problem $(Q_S, \pi_1)$.
Ultimately, our purpose in discussing a case problem $Q[S]$ is to understand the complexity of the associated parameterized problem ${\mathsf{param}\textup{-}Q[S]}$. As mentioned, formalizing the notion of a case problem allows us to cleanly present reductions between such problems.
Let $(Q, \kappa)$ be a parameterized problem. Under the assumption that $\kappa$ is FPT-computable with respect to itself, the parameterized problem $(Q, \kappa)$ is straightforwardly verified to be equivalent, under FPT-reduction, to the parameterized problem $(Q', \pi_1)$ where $Q' = \{ (\kappa(x), x) ~|~ x \in \Sigma^* \}$. Hence, any such given parameterized problem $(Q, \kappa)$ may be canonically associated to the case problem $Q'[\Sigma^*]$, as one has ${\mathsf{param}\textup{-}Q'[\Sigma^*]} = (Q', \pi_1)$.
We define ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ to be the case problem $Q[\Sigma^*]$ where $Q$ contains a pair $(k, G)$ if and only if $G$ is a graph that contains a clique of size $k$ (we assume that both $G$ and $k$ are encoded as strings over $\Sigma$); we define ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ to be the problem $\overline{Q}[\Sigma^*]$.
We now present the notion of *slice reduction*, which allows us to compare case problems.
A case problem $Q[S]$ *slice reduces* to a second case problem $Q'[S']$ if there exist:
- a computably enumerable language $U \subseteq \Sigma^* \times \Sigma^*$ and
- a partial function $r: \Sigma^* \times \Sigma^* \times \Sigma^* \to \Sigma^*$ that has ${\mathsf{dom}}(r) = U \times \Sigma^*$ and is FPT-computable with respect to the parameterization $(\pi_1, \pi_2)$
such that the following conditions hold:
- *(coverage)* for each $s \in S$, there exists $s' \in S'$ such that $(s, s') \in U$, and
- *(correctness)* for each $(t, t') \in U$, it holds (for each $y \in \Sigma^*$) that $$(t, y) \in Q \Leftrightarrow (t', r(t, t', y)) \in Q'.$$
We call the pair $(U, r)$ a *slice reduction* from $Q[S]$ to $Q'[S']$.
In this definition, we understand the parameterization $(\pi_1, \pi_2)$ to be the mapping that, for all $(s, s') \in U$ and $y \in \Sigma^*$, returns the pair $(s, s')$ given the triple $(s, s', y)$.
\[thm:slice-reduction-transitive\] (Transitivity of slice reducibility) Suppose that $Q_1[S_1]$ slice reduces to $Q_2[S_2]$ and that $Q_2[S_2]$ slice reduces to $Q_3[S_3]$. Then $Q_1[S_1]$ slice reduces to $Q_3[S_3]$.
The following theorem allows one to derive an FPT-reduction or an nu-FPT-reduction from a slice reduction.
\[thm:slice-reduction-gives-fpt-reduction\] Suppose that a case problem $Q[S]$ slice reduces to another case problem $Q'[S']$. Then, it holds that ${\mathsf{param}\textup{-}Q[S]}$ nu-FPT-reduces to ${\mathsf{param}\textup{-}Q'[S']}$; if in addition $S$ and $S'$ are computable, then ${\mathsf{param}\textup{-}Q[S]}$ FPT-reduces to ${\mathsf{param}\textup{-}Q'[S']}$.
Let $(U, r)$ be the slice reduction. First, consider the case where both $S$ and $S'$ are computable. Since $S$ and $S'$ are both computable and $U$ is computably enumerable, there exists a computable function $f: \Sigma^* \to \Sigma^*$ such that, for all $s \in S$, it holds that $(s, f(s)) \in U$. We define the reduction $g$ so that $g(s, x)$ is equal to $(f(s), r(s, f(s), x))$ for all $(s, x) \in S \times \Sigma^*$, and is otherwise defined to be a fixed string outside of $Q'$ (such a string exists by Assumption \[assumption:non-trivial-languages\]). The mapping $g$ has a natural algorithm, namely, given $(s, x)$, check if $s \in S$ (via an algorithm for $S$); if $s \notin S$, return a fixed string outside of $Q'$, otherwise, compute the value of $g$ using an algorithm for $f$ and an algorithm witnessing FPT-computability of $r$. Suppose that $s \in S$; the value $f(s)$, viewed as a function of $(s, x)$, is FPT-computable with respect to $\pi_1$; since $r$ is FPT-computable with respect to $(\pi_1, \pi_2)$, the value $r(s, f(s), x)$, viewed as a function of $(s, x)$, is also FPT-computable with respect to $\pi_1$. Thus, the mapping $g$ is FPT-computable with respect to $\pi_1$.
In the case that no computability assumptions are placed on $S$ and $S'$, there exists a function $f: \Sigma^* \to \Sigma^*$ such that, for all $s \in S$, it holds that $(s, f(s)) \in U$. The reduction $g$ defined as above is readily verified to be nu-FPT-computable with respect to $\pi_1$, via an ensemble of algorithms where $A_s$ contains as hard-coded information whether or not $s \in S$ and (if so) the value of $f(s)$.
For each parameterized class $C$, we define ${\textup{{\small $\mathsf{case}$}-{$ C $}}}$ to be the set of case problems that contains a case problem $Q[S]$ if and only if there exists a case problem $Q'[S']$ such that
- ${\mathsf{param}\textup{-}Q'[S']}$ is in $C$,
- $Q[S]$ slice reduces to $Q'[S']$, and
- $S'$ is computable.
\[prop:in-casec\] Suppose that $C$ is a parameterized class, and that $Q[S]$ is a case problem in ${\textup{{\small $\mathsf{case}$}-{$ C $}}}$. Then, the problem ${\mathsf{param}\textup{-}Q[S]}$ is in ${\textup{{\small $\mathsf{nu}$}-{$ C $}}}$; if it is assumed additionally that $S$ is computable, then the problem ${\mathsf{param}\textup{-}Q[S]}$ is in $C$.
There exists a case problem $Q'[S']$ satisfying the conditions given in the definition of ${\textup{{\small $\mathsf{case}$}-{$ C $}}}$. Since $Q[S]$ slice reduces to $Q'[S']$, by Theorem \[thm:slice-reduction-gives-fpt-reduction\], it holds that ${\mathsf{param}\textup{-}Q[S]}$ nu-FPT-reduces to ${\mathsf{param}\textup{-}Q'[S']}$. Since ${\mathsf{param}\textup{-}Q'[S']}$ is in $C$, it follows that ${\mathsf{param}\textup{-}Q[S]}$ is in ${\textup{{\small $\mathsf{nu}$}-{$ C $}}}$. If in addition it is assumed that $S$ is computable, by Theorem \[thm:slice-reduction-gives-fpt-reduction\], we obtain that ${\mathsf{param}\textup{-}Q[S]}$ FPT-reduces to ${\mathsf{param}\textup{-}Q'[S']}$, and hence that ${\mathsf{param}\textup{-}Q[S]}$ is in $C$ (by appeal to Assumption \[assumption:closure-under-reduction\]).
Let $Q[S]$ be a case problem and let $C$ be a parameterized class. We say that $Q[S]$ is ${\textup{{\small $\mathsf{case}$}-{$ C $}}}$-hard if there exists a case problem $Q^-[S^-]$ such that
- ${\mathsf{param}\textup{-}Q^-[S^-]}$ is $C$-hard,
- $Q^-[S^-]$ slice reduces to $Q[S]$, and
- $S^-$ is computable.
\[prop:casec-hard\] Suppose that $C$ is a parameterized class, and that $Q[S]$ is a case problem that is ${\textup{{\small $\mathsf{case}$}-{$ C $}}}$-hard. Then, the problem ${\mathsf{param}\textup{-}Q[S]}$ is non-uniformly $C$-hard; if it is assumed additionally that $S$ is computable, then the problem ${\mathsf{param}\textup{-}Q[S]}$ is $C$-hard.
There exists a case problem $Q^-[S^-]$ satisfying the conditions given in the definition of ${\textup{{\small $\mathsf{case}$}-{$ C $}}}$-hard. Since $Q^-[S^-]$ slice reduces to $Q[S]$, by Theorem \[thm:slice-reduction-gives-fpt-reduction\], it holds that ${\mathsf{param}\textup{-}Q^-[S^-]}$ nu-FPT-reduces to ${\mathsf{param}\textup{-}Q[S]}$. By the $C$-hardness of ${\mathsf{param}\textup{-}Q^-[S^-]}$, each problem in $C$ FPT-reduces to ${\mathsf{param}\textup{-}Q^-[S^-]}$, and hence, by Proposition \[prop:fpt-reductions-compose\], each problem in $C$ nu-FPT-reduces to ${\mathsf{param}\textup{-}Q[S]}$. If in addition it is assumed that $S$ is computable, by Theorem \[thm:slice-reduction-gives-fpt-reduction\], the problem ${\mathsf{param}\textup{-}Q^-[S^-]}$ FPT-reduces to the problem ${\mathsf{param}\textup{-}Q[S]}$; from the $C$-hardness of ${\mathsf{param}\textup{-}Q^-[S^-]}$ and from Proposition \[prop:fpt-reductions-compose\], we obtain that ${\mathsf{param}\textup{-}Q[S]}$ is $C$-hard.
Thickness {#sect:thickness}
=========
In this section, we define a measure of first-order formulas that we call *thickness*, which we will show is the crucial measure that determines whether or not a graph-like set of sentences is tractable. From a high-level viewpoint, the measure is defined in the following way. We first define a notion of *organized formula* and show that each formula is logically equivalent to a positive combination of organized formulas; we then define a notion of *layered formula* and show that each organized formula is logically equivalent to a layered formula. We then, for each layered formula $\phi$, define its thickness (denoted by ${{\mathsf{thick}}_l}(\phi)$), and then naturally extend this definition to positive combinations of layered formulas, and hence to all formulas. A key property of thickness, which we prove (Theorem \[thm:thick-many-variables\]), is that there exists an algorithm that, given a formula $\phi$, outputs an equivalent formula that uses at most ${\mathsf{thick}}(\phi)$ many variables.
We define the set of organized formulas inductively, as follows.
- Each atom and each negated atom is an organized formula.
- If each of $\phi_1, \ldots, \phi_n$ is an organized formula and $v \in {\mathsf{free}}(\phi_1) \cap \cdots \cap {\mathsf{free}}(\phi_n)$, then $\exists v (\bigwedge_{i=1}^n \phi_i)$ and $\forall v (\bigvee_{i=1}^n \phi_i)$ are organized formulas.
\[thm:formula-to-pos-comb-organized\] There exists an algorithm ${\mathrm{org}^+}$ that, given a formula $\phi$ as input, outputs a positive combination ${\mathrm{org}^+}(\phi)$ of organized formulas that is logically equivalent to $\phi$ and that is in the syntactic closure of $\phi$.
The algorithm of this theorem is defined recursively with respect to formula structure.
To give the algorithm, we first recursively define a procedure that, given a formula $\phi$ where negations appear only in front of atoms, outputs both a CNF of organized formulas and a DNF of organized formulas, each of which is equivalent to $\phi$ and in the syntactic closure of $\phi$. We will make tacit use of transformation $(\alpha)$. Observe that it suffices to show that either such a CNF or such a DNF can be computed, since one can convert between such a CNF and such a DNF by use of transformation $(\delta)$. The procedure is defined as follows.
- When $\phi$ is an atom or a negated atom, the procedure returns $\phi$.
- When $\phi$ has the form $\psi_1 \wedge \psi_2$, the procedure computes a CNF for $\phi$ by taking the conjunction of CNFs for $\psi_1$ and $\psi_2$, which can be computed recursively. The case where $\phi$ has the form $\psi_1 \vee \psi_2$ is defined dually.
- When $\phi$ has the form $\exists x \psi$, the procedure performs the following. (The case where $\phi$ has the form $\forall y \psi$ is dual.) First, it recursively computes a DNF for $\psi$; denote the DNF by $\bigvee_i \psi_i$. By appeal to transformation $(\beta)$, the formula $\phi$ is logically equivalent to $\bigvee_i (\exists x \psi_i)$. To obtain a DNF for $\phi$, it thus suffices to show that each formula of the form $\exists x \psi_i$ can be transformed to an equivalent conjunction of organized formulas. Each formula $\psi_i$ is a conjunction $\psi_i^1 \wedge \cdots \wedge \psi_i^k$ of organized formulas; write $\psi_i$ as $\psi_i^x \wedge \psi_i^{-x}$, where $\psi_i^x$ is the conjunction of the formulas $\psi_i^j$ such that $x \in {\mathsf{free}}(\psi_i^j)$, and $\psi_i^{-x}$ is the conjunction of the remaining formulas $\psi_i^j$. We have, by transformation $(\gamma)$, that $\exists x \psi_i$ is equivalent to $(\exists x \psi_i^x) \wedge \psi_i^{-x}$.
The algorithm ${\mathrm{org}^+}$, given a formula $\phi$, computes an equivalent formula $\phi'$ where negations appear only in front of atoms (using transformations $(\epsilon)$), and then invokes the described procedure on $\phi'$ and outputs either the CNF or the DNF returned by the procedure.
In what follows, when $V$ is a set of variables and $Q \in \{ \exists, \forall \}$ is a quantifier, we will use $Q V$ as shorthand for $Q v_1 \ldots Q v_n$, where $v_1, \ldots, v_n$ is a list of the elements of $V$. Our discussion will always be independent of the particular ordering chosen. Relative to a hypergraph $H$ and a subset $S \subseteq V(H)$, we consider a set of edges $\{ e_1, \ldots, e_k \}$ to be *$S$-connected* if one has connectedness of the graph with vertices $\{ e_1, \ldots, e_k
\}$ and having an edge between $e_i$ and $e_j$ if and only if $S \cap e_i \cap e_j \neq \emptyset$; we say that the hypergraph $H$ is itself *$S$-connected* if $E(H)$ is $S$-connected.
We define the sets of *$\exists$-layered formulas* and of *$\forall$-layered formulas* to be the variable-loose formulas that can be constructed inductively, as follows:
- Each atom and each negated atom is both an $\exists$-layered formula and a $\forall$-layered formula.
- If each of $\phi_1, \ldots, \phi_n$ is a $\forall$-layered formula, and\
$X \subseteq {\mathsf{free}}(\phi_1) \cup \cdots \cup {\mathsf{free}}(\phi_n)$ is such that the hypergraph $\{ {\mathsf{free}}(\phi_1), \ldots, {\mathsf{free}}(\phi_n) \}$ is $X$-connected, then $\exists X (\bigwedge_{i=1}^n \phi_i)$ is an $\exists$-layered formula.
- If each of $\phi_1, \ldots, \phi_n$ is an $\exists$-layered formula, and\
$Y \subseteq {\mathsf{free}}(\phi_1) \cup \cdots \cup {\mathsf{free}}(\phi_n)$ is such that the hypergraph $\{ {\mathsf{free}}(\phi_1), \ldots, {\mathsf{free}}(\phi_n) \}$ is $Y$-connected, then $\forall Y (\bigvee_{i=1}^n \phi_i)$ is a $\forall$-layered formula.
\[thm:organized-to-layered\] There exists an algorithm that, given an organized formula $\phi$ as input, outputs a layered formula that is logically equivalent to $\phi$.
The intuitive idea behind the algorithm of Theorem \[thm:organized-to-layered\] is to combine together (into a set quantification $Q V$) quantifiers of the same type that occur adjacently in $\phi$.
We define the algorithm recursively.
- If $\phi$ is an atom or a negated atom, the algorithm returns $\phi$.
- If $\phi$ is an organized formula of the form $\exists v (\bigwedge_{i=1}^{\ell} \theta_i)$ with $v \in {\mathsf{free}}(\theta_1) \cap \cdots \cap {\mathsf{free}}(\theta_{\ell})$, the algorithm proceeds as follows. By renaming variables if necessary, the algorithm ensures that no variable is quantified twice in $\phi$, and also that no variable is both free and quantified in $\phi$. For each $i$, the algorithm recursively computes, from $\theta_i$, a logically equivalent layered formula $\theta'_i$. The algorithm then writes $\phi$ as the formula $$\exists v ((\exists X_1 \phi_1) \wedge \cdots \wedge (\exists X_m
\phi_m)) \wedge (\psi_1 \wedge \cdots \wedge \psi_n)$$ where $\exists X_1 \phi_1, \ldots, \exists X_m \phi_m$ is a list of the formulas $\theta'_i$ that begin with existential quantification, and where $\psi_1 \wedge \cdots \wedge \psi_n$ is a list of the remaining formulas $\theta'_i$. Note that each $\phi_j$ is the conjunction of $\forall$-layered formulas, and each $\psi_j$ is a $\forall$-layered formula.
For each $j \in {\underline{m}}$, let $H_j$ denote the hypergraph of $\phi_j$, that is, the hypergraph where an edge is present if it is the set of free variables of a conjunct of $\phi_j$. We have that the hypergraph $H_j$ is $X_j$-connected. Since each $H_i$ contains $v$ in an edge and it holds that $v \in {\mathsf{free}}(\psi_1) \cap \cdots \cap {\mathsf{free}}(\psi_n)$, we have that the hypergraph $H$ with edge set $$E(H_1) \cup \cdots \cup E(H_m) \cup
\{ {\mathsf{free}}(\psi_1) \} \cup \cdots \cup \{ {\mathsf{free}}(\psi_n) \}$$ is $(X_1 \cup \cdots \cup X_m \cup \{ v \})$-connected. Consider the conjunction of the $\phi_j$ and of the $\psi_j$. This conjunction can be viewed as the conjunction of $\forall$-layered formulas whose free variable sets are exactly the edges of $H$; denote this conjunction by $\chi$. The $\exists$-layered formula $\exists (X_1 \cup \cdots \cup X_m \cup \{ v \}) \chi$ is logically equivalent to $\phi$; this is because, by the variable renaming done initially, the variables in a set $X_i$ do not appear in a formula $\phi_j$ when $j \neq i$, nor do they appear in a formula $\psi_j$.
- If $\phi$ is an organized formula of the form $\forall v (\bigvee_{i=1}^{\ell} \theta_i)$ with $v \in {\mathsf{free}}(\theta_1) \cap \cdots \cap {\mathsf{free}}(\theta_{\ell})$, the algorithm proceeds dually to the previous case.
\[thm:formula-to-pos-comb-layered\] There exists an algorithm ${\mathrm{lay}^+}$ that, given a formula $\phi$ as input, outputs a positive combination ${\mathrm{lay}^+}(\phi)$ of layered formulas that is logically equivalent to $\phi$.
Immediate from Theorems \[thm:formula-to-pos-comb-organized\] and \[thm:organized-to-layered\].
We define the following measures on layered formulas.
- When $\phi$ is an atom or a negated atom, we define ${{\mathsf{thick}}_l}(\phi) = |{\mathsf{free}}(\phi)|$.
- Suppose that $\phi$ is a layered formula of the form $\exists U (\bigwedge_{i=1}^n \phi_i)$ or $\forall U (\bigvee_{i=1}^n \phi_i)$.
We define the *local thickness* of $\phi$ as $${\mathsf{local}\textup{-}{{\mathsf{thick}}_l}}(\phi) = 1 + {\mathsf{tw}}(\{ {\mathsf{free}}(\phi_i) ~|~ i \in {\underline{n}} \} \cup \{ {\mathsf{free}}(\phi) \})$$ where the object to which the treewidth is applied is a hypergraph specified by its edge set, which hypergraph has vertex set $\cup_{i=1}^n {\mathsf{free}}(\phi_i)$.
We define the *thickness* of $\phi$ inductively as $${{\mathsf{thick}}_l}(\phi) = \max (\{{\mathsf{local}\textup{-}{{\mathsf{thick}}_l}}(\phi) \} \cup \{ {{\mathsf{thick}}_l}(\phi_i)
~|~ i \in {\underline{n}} \}).$$ We define the *quantified thickness* of $\phi$ as $${\mathsf{quant}\textup{-}{{\mathsf{thick}}_l}}(\phi) = 1 + {\mathsf{tw}}(\{ {\mathsf{free}}(\phi_i) \cap U ~|~ i \in
{\underline{n}} \}).$$
The *thickness* of an arbitrary formula $\phi$ is defined as follows: let $\Psi$ be the set of layered formulas that are positively combined subformulas of ${\mathrm{lay}^+}(\phi)$ such that ${\mathrm{lay}^+}(\phi)$ is a positive combination over $\Psi$; then, $${\mathsf{thick}}(\phi) = \max_{\psi \in \Psi} {{\mathsf{thick}}_l}(\psi).$$
\[prop:thickness-computable\] The function ${\mathsf{thick}}(\cdot)$ is computable.
This follows from the computability of ${\mathrm{lay}^+}(\phi)$ from $\phi$ (Theorem \[thm:formula-to-pos-comb-layered\]) and the definition of thickness.
We indeed now demonstrate a principal property of thickness, namely, that this measure provides an upper bound on the number of variables needed to express a formula; this upper bound is effective in that there is an algorithm that computes an equivalent formula using a bounded number of variables. When $k \geq 1$, let us say that a formula *uses $k$ many variables* if the set containing all variables that occur in the formula has size less than or equal to $k$.
\[thm:thick-many-variables\] There exists an algorithm that, given as input a formula $\phi$, outputs an equivalent formula that uses ${\mathsf{thick}}(\phi)$ many variables.
\[def:elim-ordering-of-formula\] Let $\phi$ be a formula of the form $\exists V (\bigwedge_{i=1}^n \phi_i)$ or $\forall V (\bigvee_{i=1}^n \phi_i)$. Let us say that a pair $e = ((v_1, \ldots, v_m), E)$ consisting of an ordering $v_1, \ldots, v_m$ of the elements in $\bigcup_{i=1}^n {\mathsf{free}}(\phi_i)$ and a subset $E \subseteq K( \{ v_1, \ldots, v_m \} )$ is an *elimination ordering of $\phi$* if:
- the variables in ${\mathsf{free}}(\phi)$ occur first in the ordering, that is, $\{ v_1, \ldots, v_{|{\mathsf{free}}(\phi)|} \} = {\mathsf{free}}(\phi)$; and,
- $e$ is an elimination ordering of the hypergraph $\{ {\mathsf{free}}(\phi) \} \cup \{ {\mathsf{free}}(\phi_i) ~|~ i \in {\underline{n}} \}$.
The following lemma can be taken as a variation of a lemma of Kolaitis and Vardi [@KolaitisVardi00-containment Lemma 5.2].
\[lemma:formula-and-elimination-ord\] There exists a polynomial-time algorithm that, given a formula $\phi$ of the form $\exists V (\bigwedge_{i=1}^n \phi_i)$ or $\forall V (\bigvee_{i=1}^n \phi_i)$ and an elimination ordering $e$ of $\phi$, outputs a formula $\phi'$ that is logically equivalent to $\phi$ such that ${\mathsf{width}}(\phi') \leq
\max (\{ 1 + {\mathsf{lower}\textup{-}\mathsf{deg}}(e) \} \cup \{ {\mathsf{width}}(\phi_i) ~|~ i \in {\underline{n}} \})$.
Given a formula $\phi$ of the described form having $|V| \geq 1$ and an elimination ordering $e = ((v_1, \ldots, v_m), E)$ of it, we explain how eliminate the last variable; precisely speaking, we explain how to compute a logically equivalent $\psi$ of the form $\exists \{ v_1, \ldots, v_{m-1} \} \bigwedge \psi_j$ and having elimination ordering $((v_1, \ldots, v_{m-1}), E \cap K(\{ v_1, \ldots, v_{m-1}\}))$. The desired algorithm iterates this variable elimination.
Let $\phi^{v_m}$ denote the conjunction of the formulas of the form $\phi_i$ having $v_m \in {\mathsf{free}}(\phi_i)$, and let $I$ denote the set containing the indices of the remaining formulas $\phi_i$. The formula $\psi$ is defined as $\exists \{ v_1, \ldots, v_{m-1} \}$ followed by the conjunction of all formulas in $\Psi = \{ \phi_i ~|~ i \in I \} \cup \{ \exists v_m \phi^{v_m}\}$. It is clear that $\psi$ and $\phi$ are logically equivalent, and we have $|{\mathsf{free}}(\phi^{v_m})| = 1 + {\mathsf{lower}\textup{-}\mathsf{deg}}(e, v_m)$. We verify that $e^{\psi} = ((v_1, \ldots, v_{m-1}), E \cap K(\{ v_1, \ldots, v_{m-1}\}))$ is an elimination ordering of $\psi$ as follows. Since ${\mathsf{free}}(\psi) = {\mathsf{free}}(\phi)$, it is clear that the variables in ${\mathsf{free}}(\psi)$ occur first in the ordering $(v_1, \ldots, v_{m-1})$. To verify that $e^{\psi}$ is an elimination ordering of the hypergraph named in Definition \[def:elim-ordering-of-formula\], we need to verify that, for each edge $f$ of that hypergraph, one has the containment $K(f) \subseteq E$. For each such edge $f$, other than the edge ${\mathsf{free}}(\exists v_m \phi^{v_m})$, this containment holds because it held for the original elimination ordering for $\phi$. For the edge ${\mathsf{free}}(\exists v_m \phi^{v_m})$, the containment holds due to the lower neighbor property and due to $v_m$ being a neighbor of each element of ${\mathsf{free}}(\phi^{v_m}) \setminus \{ v_m \}$ in $E$.
We now confirm that we can apply the algorithm to $\psi$ to obtain the desired formula $\phi'$. We have ${\mathsf{lower}\textup{-}\mathsf{deg}}(e^{\psi}) \leq {\mathsf{lower}\textup{-}\mathsf{deg}}(e)$ and ${\mathsf{width}}(\phi^{v_m}) \leq
\max( \{ {\mathsf{width}}(\phi_i) ~|~ i \in {\underline{n}} \setminus I \} \cup
\{ 1 + {\mathsf{lower}\textup{-}\mathsf{deg}}(e) \})$. From this, it follows that $$\max (\{ 1 + {\mathsf{lower}\textup{-}\mathsf{deg}}(e^{\psi}) \} \cup \{ {\mathsf{width}}(\psi_j) ~|~
\psi_j \in \Psi \})$$ $$\leq
\max (\{ 1 + {\mathsf{lower}\textup{-}\mathsf{deg}}(e) \} \cup \{ {\mathsf{width}}(\phi_i) ~|~ i \in {\underline{n}} \}).$$
(Theorem \[thm:thick-many-variables\]) By definition of ${\mathsf{thick}}(\phi)$ and by the computability of ${\mathrm{lay}^+}(\phi)$ from $\phi$ (Theorem \[thm:formula-to-pos-comb-layered\]), it suffices to prove that there is an algorithm that, given a layered formula $\phi$, returns an equivalent formula that uses ${{\mathsf{thick}}_l}(\phi)$ many variables.
Consider the following algorithm $A$ defined recursively on layered formulas. When $\phi$ is an atom or a negated atom, $A(\phi) = \phi$. When $\phi$ is of the form $\exists V (\bigwedge_{i=1}^n \phi_i)$ or $\forall V (\bigvee_{i=1}^n \phi_i)$, the algorithm proceeds as follows. Set $H$ to be the hypergraph $$\{ {\mathsf{free}}(\phi_i) ~|~ i \in
{\underline{n}} \} \cup \{ {\mathsf{free}}(\phi) \}.$$ We have $1 + {\mathsf{tw}}(H) = {\mathsf{local}\textup{-}{{\mathsf{thick}}_l}}(\phi) \leq {{\mathsf{thick}}_l}(\phi)$. By calling the algorithm of Proposition \[prop:tw-gives-elimination-ordering\] on the hypergraph $H$ and the distinguished edge ${\mathsf{free}}(\phi)$, we can obtain an elimination ordering $e$ of $\phi$ having $1 + {\mathsf{lower}\textup{-}\mathsf{deg}}(e) = 1 + {\mathsf{tw}}(H) \leq {{\mathsf{thick}}_l}(\phi)$. Let $\phi'$ be the formula obtained from $\phi$ by replacing each formula $\phi_i$ by $A(\phi_i)$; for each $i \in {\underline{n}}$, we have ${\mathsf{width}}(\phi_i) \leq {{\mathsf{thick}}_l}(\phi_i)$. By applying the algorithm of Lemma \[lemma:formula-and-elimination-ord\] to $\phi'$ and $e$, we obtain a formula $\phi''$ having ${\mathsf{width}}(\phi'') \leq {{\mathsf{thick}}_l}(\phi)$. By renaming variables in $\phi''$, we can obtain an equivalent formula where the number of variables used is equal to ${\mathsf{width}}(\phi'')$. The output $A(\phi)$ of the algorithm is this equivalent formula.
Graph-like queries {#sect:main}
==================
In this section, we state our main theorem and two corollaries thereof, which describe the tractable graph-like sentence sets. In the following two sections, we prove the hardness portion of the main theorem.
Define ${\mathsf{MC}}$ to be the language of pairs $$\{ (\phi, {\mathbf{B}}) ~|~ {\mathbf{B}}\models \phi \}$$ where $\phi$ denotes a first-order sentence, and ${\mathbf{B}}$ denotes a relational structure.
(Main theorem) \[thm:main\] Let $\Phi$ be a graph-like set of sentences having bounded arity. If $\Phi$ has bounded thickness, then the case problem ${\mathsf{MC}}[\Phi]$ is in ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$; otherwise, the case problem ${\mathsf{MC}}[\Phi]$ is ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{W[1]}$}}$}}}$-hard or ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{co}$-$\mathsf{W[1]}$}}$}}}$-hard.
We give a proof of this theorem that makes a single forward reference to the main theorem of Section \[sect:hardness\].
Suppose $\Phi$ has thickness bounded above by $k$. Define $S'$ to be the set of sentences having thickness less than or equal to $k$. The set $S'$ is computable by Proposition \[prop:thickness-computable\]. We have that ${\mathsf{param}\textup{-}{\mathsf{MC}}[S']}$ is in ${\textup{\small $\mathsf{FPT}$}}$ via the algorithm that, given an instance $(\phi, {\mathbf{B}})$, first checks if $\phi \in S'$, and if so, invokes Theorem \[thm:thick-many-variables\] to obtain $\phi'$, and then performs the natural bottom-up, polynomial-time evaluation of $\phi'$ on ${\mathbf{B}}$ (à la Vardi [@Vardi95-boundedvariable]). The case problem ${\mathsf{MC}}[\Phi]$ slice reduces to ${\mathsf{MC}}[S']$ via the slice reduction $(\{ (s', s') ~|~ s' \in S' \}, \pi_3)$.
Suppose that $\Phi$ has unbounded thickness. Theorem \[thm:hardness\] yields that either ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ or ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ slice reduces to ${\mathsf{MC}}[\Phi]$. It then follows by definition that ${\mathsf{MC}}[\Phi]$ is ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{W[1]}$}}$}}}$-hard or ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{co}$-$\mathsf{W[1]}$}}$}}}$-hard, since the problems ${\textup{{\small $\mathsf{CLIQUE}$}}}$ and ${\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$ are ${\textup{\small $\mathsf{W[1]}$}}$-hard and ${\textup{\small $\mathsf{co}$-$\mathsf{W[1]}$}}$-hard, respectively.
We now provide two corollaries that describe the complexity of the problems ${\mathsf{param}\textup{-}{\mathsf{MC}}[\Phi]}$ addressed by the main theorem; the first corollary assumes that $\Phi$ is computable, while the second corollary makes no computability assumption on $\Phi$. Both of these corollaries follow directly from Theorem \[thm:main\] via use of Propositions \[prop:in-casec\] and \[prop:casec-hard\].
\[cor:main-computable\] Let $\Phi$ be a computable, graph-like set of sentences having bounded arity. If $\Phi$ has bounded thickness, then the problem ${\mathsf{param}\textup{-}{\mathsf{MC}}[\Phi]}$ is in ${\textup{\small $\mathsf{FPT}$}}$; otherwise, the problem ${\mathsf{param}\textup{-}{\mathsf{MC}}[\Phi]}$ is not in ${\textup{\small $\mathsf{FPT}$}}$, unless ${\textup{\small $\mathsf{W[1]}$}}\subseteq {\textup{\small $\mathsf{FPT}$}}$.
When $\Phi$ has bounded thickness, the claim follows directly from Theorem \[thm:main\] and Proposition \[prop:in-casec\]. When $\Phi$ does not have bounded thickness, Theorem \[thm:main\] and Proposition \[prop:casec-hard\] imply that ${\mathsf{param}\textup{-}{\mathsf{MC}}[\Phi]}$ is either ${\textup{\small $\mathsf{W[1]}$}}$-hard or ${\textup{\small $\mathsf{co}$-$\mathsf{W[1]}$}}$-hard, from which the claim follows.
\[cor:main-noncomputable\] Let $\Phi$ be a graph-like set of sentences having bounded arity. If $\Phi$ has bounded thickness, then the problem ${\mathsf{param}\textup{-}{\mathsf{MC}}[\Phi]}$ is in ${\textup{{\small $\mathsf{nu}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$; otherwise, the problem ${\mathsf{param}\textup{-}{\mathsf{MC}}[\Phi]}$ is not in ${\textup{{\small $\mathsf{nu}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$, unless ${\textup{\small $\mathsf{W[1]}$}}\subseteq {\textup{{\small $\mathsf{nu}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$.
When $\Phi$ has bounded thickness, the claim follows directly from Theorem \[thm:main\] and Proposition \[prop:in-casec\]. When $\Phi$ does not have bounded thickness, Theorem \[thm:main\] and Proposition \[prop:casec-hard\] imply that ${\mathsf{param}\textup{-}{\mathsf{MC}}[\Phi]}$ is either non-uniformly ${\textup{\small $\mathsf{W[1]}$}}$-hard or non-uniformly ${\textup{\small $\mathsf{co}$-$\mathsf{W[1]}$}}$-hard, so, invoking the fact that ${\textup{{\small $\mathsf{nu}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$ is closed under nu-FPT-reductions, the claim follows.
In Section \[sect:discussion\], we provide a discussion of how the main theorem of the present paper can be used to readily derive the dichotomy theorem of graphical sets of quantified conjunctive queries [@ChenDalmau12-decomposingquantified]; a dual argument yields the corresponding theorem on graphical sets of quantified disjunctive queries. Note that it follows immediately from this discussion and [@ChenDalmau12-decomposingquantified Example 3.5] that there exists a set of graph-like sentences having bounded arity that is tractable, but not contained in one of the tractable classes identified by Adler and Weyer [@AdlerWeyer12-treewidth].
Let us also note here that, as pointed out by Adler and Weyer, it is undecidable, given a first-order sentence $\phi$ and a value $k \geq 1$, whether or not $\phi$ is logically equivalent to a $k$-variable sentence, and hence one cannot expect an algorithm that takes a first-order sentence and outputs an equivalent one that minimizes the number of variables. (This undecidability result holds even for positive first-order logic [@BovaChen14-width-ep].)
Accordion reductions {#sect:accordion}
====================
In this section, we introduce a notion that we call *accordion reduction*. When $C \subseteq \Sigma^* \times \Sigma^*$ is a set of string pairs and $S \subseteq \Sigma^*$ is a set of strings, we use ${\mathsf{closure}}_C(S)$ to denote the intersection of all sets $T$ containing $S$ and having the closure property that, if $(u, u') \in C$ and $u' \in T$, then $u \in T$. When an accordion reduction exists for such a $C$ (with respect to a language $Q$), the theorem of this section (Theorem \[thm:accordion-reduction-gives-slice-reduction\]) yields that the problem $Q[{\mathsf{closure}}_C(S)]$ slice reduces to $Q[S]$. Hence, an accordion reduction is not itself a slice reduction, but its existence provides a sufficient condition for the existence of a class of slice reductions.
How is this section’s theorem proved? One component of an accordion reduction is an FPT-computable mapping $r$ that, for each $(u, u') \in C$, maps a $Q$-instance $(u, y)$ to a $Q$-instance $(u', y')$. Intuitively, to give a slice reduction from $Q[{\mathsf{closure}}_C(S)]$ to $Q[S]$, one needs to reduce, for any $s \in S$ and $s_1 \in {\mathsf{closure}}_C(S)$, instances of the form $(s_1, \cdot)$ to instances of the form $(s, \cdot)$. The containment $s_1 \in {\mathsf{closure}}_C(S)$ implies the existence of a sequence $s_1, \ldots, s_k = s$ such that every pair $(s_i, s_{i+1})$ is in $C$. This naturally suggests applying the map $r$ repeatedly, but note that there is no constant bound on the length $k$ of the sequence. Hence, $r$ needs to be sufficiently well-behaved so that, when composed with itself arbitrarily many times (in the described way), the end effect is that of an FPT-computable function that may serve as the map in the definition of slice reduction. (The author is mentally reminded of the closing of an accordion in thinking that this potentially long sequence of compositions yields a single well-behaved map.) To ensure this well-behavedness, we impose a condition that we call *measure-linearity*.
In the context of accordion reductions, a *measure* is a mapping $m: \Sigma^* \times \Sigma^* \to
{\mathbb{N}}$ such that there exist a computable function $f: \Sigma^* \to {\mathbb{N}}$ and a polynomial $p: {\mathbb{N}}\to {\mathbb{N}}$ whereby, for all pairs $(s, y) \in \Sigma^* \times \Sigma^*$, it holds that $m(s,y) \leq |y| \leq f(s) p(m(s,y))$.
Let $Q \subseteq \Sigma^* \times \Sigma^*$ be a language of pairs. With respect to $Q$, an *accordion reduction* consists of:
- a computably enumerable language $C \subseteq \Sigma^* \times \Sigma^*$,
- a measure $m: \Sigma^* \times \Sigma^* \to {\mathbb{N}}$,
- a partial computable function $B: \Sigma^* \times \Sigma^* \to {\mathbb{N}}$ with ${\mathsf{dom}}(B) = C$, and
- a mapping $r: \Sigma^* \times \Sigma^* \times \Sigma^* \to
\Sigma^*$ that has ${\mathsf{dom}}(r) = C \times \Sigma^*$, that is FPT-computable with respect to the parameterization $(\pi_1, \pi_2)$, and that is *measure-linear* in that, for each pair $(u, u') \in C$ and for each $y \in \Sigma^*$, it holds that $m(u', r(u, u', y)) \leq B(u, u') m(u, y)$,
such that the following condition holds:
- *(correctness)* for each $(u, u') \in C$, it holds (for each $y \in \Sigma^*$) that $$(u, y) \in Q \Leftrightarrow (u', r(u, u', y)) \in Q.$$
\[thm:accordion-reduction-gives-slice-reduction\] Suppose that $Q[S]$ is a case problem, and that $(C, m, r)$ is an accordion reduction with respect to $Q$. Then, the case problem $Q[{\mathsf{closure}}_C(S)]$ slice reduces to the case problem $Q[S]$.
We define a slice reduction $(U, r^+)$ from $Q[{\mathsf{closure}}_C(S)]$ to $Q[S]$. Set $U$ to be the set containing the pairs $(u, u')$ such that there exists $k \geq 1$ and a sequence $u_1, \ldots, u_k$ such that $u = u_1$, $u_k = u'$, and $(u_i, u_{i+1}) \in C$ when $1 \leq i < k$. It is straightforwardly verified that $U$ is computably enumerable and that the coverage criterion is satisfied, that is, if $u \in {\mathsf{closure}}_C(S)$, then there exists $u' \in S$ such that $(u, u') \in U$. Fix $A_U$ to be an algorithm that, given a pair $(u, u')$, returns a sequence $u = u_1, u_2, \ldots, u_k = u'$ of the just-described form whenever $(u, u') \in U$. Consider the algorithm $A_{r^+}$ that does the following: given a triple $(u, u', y)$, it invokes $A_U(u, u')$; if this computation halts with output $u = u_1, u_2, \ldots, u_k = u'$, the algorithm sets $y_1 = y$; the algorithm computes
$y_2 = r(u_1, u_2, y_1)$, $y_3 = r(u_2, u_3, y_2)$, $\ldots$, $y_k = r(u_{k-1}, u_k, y_{k-1})$,
and outputs $y_k$. We define $r^+$ as the partial mapping computed by this algorithm $A_{r^+}$. In order to compute each of the strings $y_2, \ldots, y_k$, the algorithm $A_{r^+}$ uses the algorithm $A_r$ for $r$ provided by the definition of an accordion reduction; let $f_r$ and $p_r$ be a computable function and a polynomial, respectively, such that the running time of $A_r$, on an input $(u, u', y)$, is bounded above by $f_r(u, u') p_r(|(u, u', y)|)$.
On an input $(u, u', y)$ where $r^+$ is defined, the running time of $A_{r^+}$ can be bounded above by the running time of $A_U(u, u')$ plus $$f_r(u_1, u_2) p_r(|(u_1, u_2, y_1)|) + \cdots
+ f_r(u_{k-1}, u_k) p_r(|(u_{k-1}, u_k, y_{k-1})|).$$ Note that the running time of $A_U(u, u')$ and $k$ are both functions of $(u, u')$. So, in order to show that the running time of $A_{r^+}(u, u', y)$ is FPT-computable with respect to $(\pi_1,
\pi_2)$, it suffices to show that, for each $i \geq 1$, the term $f_r(u_i, u_{i+1}) p_r(|(u_i, u_{i+1}, y_i)|)$ is degree-bounded, when defined (hereon, when discussing degree-boundedness, this is with respect to $(\pi_1, \pi_2)$). So consider such a term; the strings $u_i$ and $u_{i+1}$ are computable functions of $(u, u')$, and so by appeal to Proposition \[prop:db-closure\], it suffices to show that the string length $|y_i|$ is degree-bounded. This is clear in the case that $i = 1$, that is, we have that $|y_1|$ is degree-bounded. Let us consider the case that $i > 1$. By the measure-linearity of $r$, we have that $$m(u_i, y_i) \leq B(u_1, u_2) \cdots B(u_{i-1}, u_i) m(u_1,y_1).$$ By the definition of measure, it follows that $$m(u_i, y_i) \leq B(u_1, u_2) \cdots B(u_{i-1}, u_i) |y_1|.$$ Since the constants $B(u_j, u_{j+1})$ depend only on $(u, u')$, we obtain that the value $m(u_i, y_i)$ is degree-bounded. Letting $f_m$ and $p_m$ be the computable function and polynomial (respectively) provided by the definition of measure, we then have $|y_i| \leq f_m(u_i) p_m(m(u_i, y_i))$, and conclude that $|y_i|$ is degree-bounded.
Hardness {#sect:hardness}
========
In this section, we establish the main intractability result of the paper, namely, that the case problem ${\mathsf{MC}}[\Phi]$ is hard when $\Phi$ is graph-like and has unbounded thickness.
\[thm:hardness\] Suppose that $\Phi$ is a set of graph-like sentences of bounded arity such that ${\mathsf{thick}}(\Phi)$ is unbounded. Then, either ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ or ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ slice reduces to ${\mathsf{MC}}[\Phi]$.
Immediate from Lemmas \[lemma:red-friendly-to-graphlike\] and \[lemma:red-fullysorted-to-normal\], and Theorem \[thm:red-clique-to-sorted-friendly\].
This intractability result is obtained by composing three slice reductions. Define a formula to be *friendly* if it is loose, positive, and layered. We first show (Lemma \[lemma:red-friendly-to-graphlike\]) that there exists a set of friendly sentences $\Psi$, with ${{\mathsf{thick}}_l}(\Psi)$ unbounded, such that ${\mathsf{MC}}[\Psi]$ slice reduces to ${\mathsf{MC}}[\Phi]$. We next show (Lemma \[lemma:red-fullysorted-to-normal\]) that a multi-sorted version ${\mathsf{full}}(\Psi)$ of $\Psi$ has the property that ${\mathsf{MC}_s}[{\mathsf{full}}(\Psi)]$ slice reduces to ${\mathsf{MC}}[\Psi]$; here, ${\mathsf{MC}_s}$ denotes the multi-sorted generalization of ${\mathsf{MC}}$. Finally, we directly slice reduce either ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ or ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ to ${\mathsf{MC}_s}[\Psi]$ (Theorem \[thm:red-clique-to-sorted-friendly\]); this third reduction is obtained via an accordion reduction.
\[lemma:red-friendly-to-graphlike\] Suppose that $\Phi$ is a set of graph-like sentences of bounded arity such that ${\mathsf{thick}}(\Phi)$ is unbounded. There exists a set of friendly sentences $\Psi$ such that ${\mathsf{MC}}[\Psi]$ slice reduces to ${\mathsf{MC}}[\Phi]$ and such that ${{\mathsf{thick}}_l}(\Psi)$ is unbounded.
Define $\Phi'$ to be the set that contains a loose layered sentence if it occurs as a positively combined subformula of a loose sentence in ${\mathrm{lay}^+}(\{ \phi \in \Phi ~|~ \textup{$\phi$ is loose} \})$. We have that ${{\mathsf{thick}}_l}(\Phi')$ is unbounded; this is because, for each sentence $\phi \in \Phi$, if we define $\phi^L$ to be equal to ${\mathrm{org}^+}(\phi)$ but with symbols renamed (if necessary) so that $\phi^L$ is loose, then $\phi^L \in \Phi$ and it is straightforwardly verified that ${\mathsf{thick}}(\phi) = {\mathsf{thick}}(\phi^L)$.
We claim that ${\mathsf{MC}}[\Phi']$ slice reduces to ${\mathsf{MC}}[\Phi]$. We define a slice reduction as follows. The set $U$ contains a pair $(\phi', \phi)$ if $\phi$ is a loose sentence and $\phi'$ is a loose layered sentence that is a positively combined subformula of ${\mathrm{lay}^+}(\phi)$; we have that $U$ is computable. For pairs $(\phi', \phi) \in U$, we set $r(\phi', \phi, {\mathbf{B}}') = {\mathbf{B}}$, where ${\mathbf{B}}$ is defined as follows. For each symbol $R$ (of arity $k$) that appears in $\phi$ but not in $\phi'$, we define $R^{{\mathbf{B}}} = B^k$ or $R^{{\mathbf{B}}} = \emptyset$ as appropriate so that ${\mathbf{B}}' \models \phi'$ if and only if ${\mathbf{B}}\models \phi$; this is possible since $\phi'$ is a positively combined subformula of ${\mathrm{lay}^+}(\phi)$ and because ${\mathrm{lay}^+}(\phi)$ and $\phi$ are logically equivalent (by Theorem \[thm:formula-to-pos-comb-layered\]).
Define $\Psi$ to be the set that contains a sentence $\psi$ if it can be obtained from a sentence $\phi' \in \Phi'$ by removing all negations that appear immediately in front of atoms. Since each sentence in $\Phi'$ is loose and layered, we obtain that each sentence in $\Psi$ is friendly. We give a slice reduction from ${\mathsf{MC}}[\Psi]$ to ${\mathsf{MC}}[\Phi']$, as follows. Define $U$ to be the set that contains a pair $(\psi, \phi')$ if $\phi'$ is a loose layered sentence and $\psi$ can be obtained from $\phi'$ by removing all negations that appear in front of atoms. When $(\psi, \phi') \in U$, define $r(\psi, \phi', {\mathbf{B}}) = {\mathbf{B}}'$ where $B' = B$ and, for each symbol $R$ of arity $k$, it holds that $R^{{\mathbf{B}}'} = B^k \setminus R^{{\mathbf{B}}}$ when $R$ appears in $\phi'$ with a negation before it, and $R^{{\mathbf{B}}'} = R^{{\mathbf{B}}}$ otherwise. (Note that $B^k \setminus R^{{\mathbf{B}}}$ can always be computed in polynomial time from $R^{{\mathbf{B}}}$ due to our assumption of bounded arity.) From the definition of ${{\mathsf{thick}}_l}(\cdot)$, we have that $(\psi, \phi') \in U$ implies ${{\mathsf{thick}}_l}(\psi) =
{{\mathsf{thick}}_l}(\phi')$, so ${{\mathsf{thick}}_l}(\Psi)$ is unbounded.
By Theorem \[thm:slice-reduction-transitive\], there is a slice reduction from ${\mathsf{MC}}[\Psi]$ to ${\mathsf{MC}}[\Phi]$.
In what follows, we will work with multi-sorted relational first-order logic, formalized as follows. (For differentiation, we will refer to formulas in the usual first-order logic considered thus far as *one-sorted*.) A *signature* is a pair $(\sigma, {\mathcal{S}})$ where ${\mathcal{S}}$ is a set of *sorts* and $\sigma$ is a set of relation symbols; each relation symbol $R \in \sigma$ has an associated arity ${\mathrm{ar}}(R)$ which is an element of ${\mathcal{S}}^*$. In a formula over signature $(\sigma, {\mathcal{S}})$, each variable $v$ has associated with it a sort $s(v)$ from ${\mathcal{S}}$; an atom is a formula $R(v_1, \ldots, v_k)$ where $R \in \sigma$ and $s(v_1) \ldots s(v_k) = {\mathrm{ar}}(R)$. A structure ${\mathbf{B}}$ on signature $(\sigma, {\mathcal{S}})$ consists of an ${\mathcal{S}}$-sorted family $\{ B_s ~|~ s \in {\mathcal{S}}\}$ of sets called the *universe* of ${\mathbf{B}}$, and for each symbol $R \in \sigma$, an interpretation $R^{{\mathbf{B}}} \subseteq B_{{\mathrm{ar}}(R)}$, where for a word $w = w_1 \ldots w_k \in {\mathcal{S}}^*$, we use $B_w$ to denote the product $B_{w_1} \times \cdots \times B_{w_k}$. We use ${\mathsf{MC}}_s$ to denote the multi-sorted version of ${\mathsf{MC}}$, that is, it is the language of pairs $(\phi, {\mathbf{B}})$ where $\phi$ is a sentence and ${\mathbf{B}}$ is a structure both having the same signature $(\sigma, {\mathcal{S}})$, and ${\mathbf{B}}\models \phi$.
Suppose that $\phi$ is a multi-sorted friendly formula on signature $(\sigma, {\mathcal{S}})$, and let $V$ be the set of variables occurring in $\phi$; we say that $\phi$ is *fully-sorted* if $V \subseteq {\mathcal{S}}$ and for each $v \in V$, the sort of $v$ is $v$ itself (that is, $s(v) = v$). When $\psi$ is a one-sorted friendly formula, and $V$ is the set of variables that occur in $\psi$, we use ${\mathsf{full}}(\psi)$ to denote the natural fully-sorted formula induced by $\psi$, namely, the formula on signature $(\sigma, V)$ where $\sigma$ contains those symbols occurring in $\psi$, and, if $R(v_1, \ldots, v_k)$ appears in $\psi$, then ${\mathrm{ar}}(R) = v_1 \ldots v_k$ (this is well-defined since $\phi$ is symbol-loose).
\[lemma:red-fullysorted-to-normal\] Let $\Psi$ be a set of one-sorted friendly sentences. The set ${\mathsf{full}}(\Psi)$ of fully-sorted friendly sentences has the property that ${\mathsf{MC}_s}[{\mathsf{full}}(\Psi)]$ slice reduces to ${\mathsf{MC}}[\Psi]$.
We give a slice reduction. Define $U$ to be the set that contains each pair of the form $({\mathsf{full}}(\psi), \psi)$. Suppose that $\psi$ and ${\mathbf{B}}$ are over signature $(\sigma, {\mathcal{S}})$. Define $r({\mathsf{full}}(\psi), \psi, {\mathbf{B}}) = {\mathbf{B}}'$, where ${\mathbf{B}}'$ is defined as follows. Let $B'$ be a set whose cardinality is $\max_{s \in {\mathcal{S}}} |B_s|$. For each sort $s \in {\mathcal{S}}$, fix a map $f_s: B' \to B_s$ that is surjective. For each relation symbol $R \in \sigma$ of arity $s_1 \ldots s_k$, define $R^{{\mathbf{B}}'} = \{ (b'_1, \ldots, b'_k) \in B'^k ~|~
(f_{s_1}(b'_1), \ldots, f_{s_k}(b'_k)) \in R^{{\mathbf{B}}} \}$. It is straightforward to prove by induction that, for all subformulas $\phi$ of $\psi$, and each assignment $g$ from the set of variables to $B'$, that ${\mathbf{B}}', g \models \phi$ if and only if ${\mathbf{B}}, f {\circ^s}g \models {\mathsf{full}}(\phi)$. Here, $f {\circ^s}g$ denotes the mapping that sends each variable $v$ to $f_{s(v)}(g(v))$ and we view ${\mathsf{full}}(\phi)$ as a formula over the signature $(\sigma, {\mathcal{S}})$ of $\psi$. The correctness of the reduction follows.
*In the remainder of this section, we assume that all formulas and structures under discussion are multi-sorted.*
Let us say that a friendly formula is a *simple formula* if it is of the form $\exists X (\bigwedge_{i=1}^n \alpha_i)$ or of the form $\forall Y (\bigvee_{i=1}^n \alpha_i)$ where each $\alpha_i$ is an atom. When $\phi$ is a simple formula where the variables $V$ are those that are quantified initially, we say that $\psi$ is *a sentence based on $\phi$* if $\psi$ is a simple sentence derivable from $\phi$ by replacing each atom $R(w_1, \ldots, w_k)$ with an atom whose variables are the elements in $\{ w_1, \ldots, w_k \} \cap V$.
We now present an accordion reduction that will be used to derive our hardness result. When $\Psi$ is a set of friendly sentences, this accordion reduction will allow a simple subformula $\phi' = \exists V \chi$ to simulate a disjunction of atoms on ${\mathsf{free}}(\phi')$, and likewise for a simple subformula $\phi' = \forall V \chi$ to simulate a corresponding conjunction; this is made precise as follows. The set $C$ is defined to contain a pair $(\psi, \phi)$ of friendly sentences if there exists a simple subformula $\phi' = Q V \chi$ of $\phi$ such that one of the following conditions holds:
1. $\psi$ is a sentence based on $\phi'$.
2. $Q = \exists$ and $\psi$ is a friendly sentence obtained from $\phi$ by replacing $\phi'$ with $\bigvee_{i=1}^m E_i(v_{i1}, v_{i2})$, where the tuples $(v_{i1}, v_{i2})$ are such that $\{ v_{11}, v_{12} \}, \ldots, \{ v_{m1}, v_{m2} \}$ is a list of the elements in $K({\mathsf{free}}(\phi'))$ and the $E_i$ are relation symbols (each of which is fresh in that it does not appear elsewhere in $\psi$).
3. $Q = \forall$ and $\psi$ is a friendly sentence obtained from $\phi$ by replacing $\phi'$ with $\bigwedge_{i=1}^m E_i(v_{i1}, v_{i2})$ where the tuples $(v_{i1}, v_{i2})$ and the symbols $E_i$ are as described in the previous case.
In order to present the accordion reduction, we view, without loss of generality, each pair of strings as a pair $(\phi, {\mathbf{B}})$ where $\phi$ is a sentence whose encoding includes the signature $(\sigma, {\mathcal{S}})$ over which it is defined; and ${\mathbf{B}}$ is a structure over this signature.
\[thm:formula-accordion-reduction\] Let $M$ be the measure such that $M(\phi, {\mathbf{B}})$ is equal to $\max_{s \in {\mathcal{S}}} |B_s|$ when ${\mathbf{B}}$ is a multi-sorted structure defined on the signature $(\sigma, {\mathcal{S}})$ of $\phi$. There exists a mapping $r: \Sigma^* \times \Sigma^* \times \Sigma^* \to \Sigma^*$ such that the triple $(C, M, r)$ is an accordion reduction with respect to ${\mathsf{MC}_s}$. (Here, $C$ is the set defined above.)
Suppose $(\psi, \phi, {\mathbf{A}})$ is a triple with $(\psi, \phi) \in C$ and where ${\mathbf{A}}$ is a structure over the signature of $\psi$. We define $r(\psi, \phi, {\mathbf{A}})$ to be the structure ${\mathbf{B}}$, defined as follows.
It is straightforward to treat the case where there exists a simple subformula $\phi'$ of $\phi$ such that $\psi$ is the sentence based on $\phi'$ (here, we omit discussion of this case).
In the remainder of this proof, we consider the case that there exist a simple subformula $\phi' = \exists V \chi$ of $\phi$ such that $\psi$ is a friendly sentence obtained from $\phi$ by replacing $\phi'$ with a disjunction $\psi'$ as in the definition of $C$. (Dual to this case is the remaining case where $\phi' = \forall V \chi$ and $\psi$ is obtained from $\phi$ by replacing $\phi'$ with a conjunction.) We will denote the disjunction $\psi'$ by $E_1(w_{11}, w_{12}) \vee \cdots \vee E_m(w_{m1}, w_{m2})$. We define the universe of ${\mathbf{B}}$ as follows.
- For each sort $u$ of ${\mathbf{A}}$, we define $B_u = A_u$.
- For each $v \in V$, define $B_v = \{ (E_{\ell}, u, a) ~|~
\ell \in {\underline{m}}, u \in \{ w_{\ell 1}, w_{\ell 2} \}, a \in A_u
\}$
As $m \leq |V|^2$ and a variable $u$ appearing as the second coordinate in an element of a set $B_v$ must be an element of ${\mathsf{free}}(\phi')$, We have $M(\phi, {\mathbf{B}}) \leq |V|^2 \cdot |{\mathsf{free}}(\phi')| \cdot M(\psi, {\mathbf{A}})$; this confirms measure-linearity of $M$, since $|V|$ and $|{\mathsf{free}}(\phi')|$ are computable functions of the pair $(\psi, \phi)$.
For each atom $R(u_1, \ldots, u_k)$ in $\phi$ that occurs outside of $\phi'$ (equivalently, that also appears in $\psi$), define $R^{{\mathbf{B}}} = R^{{\mathbf{A}}}$.
For each atom $R(u_1, \ldots, u_k)$ that occurs in $\phi'$, define $R^{{\mathbf{B}}}$ to contain a tuple $(b_1, \ldots, b_k) \in B_{u_1 \ldots u_k}$ if and only if the following two conditions hold:
- For all $i, j \in {\underline{k}}$, if $u_i \in V$ and $u_j \in V$, then $b_i = b_j$.
- For all $i, j \in {\underline{k}}$, if $u_i \in V$, $b_i = (E_{\ell}, u, a)$, and $u_j \notin V$ (equivalently, $u_j \in {\mathsf{free}}(\phi')$), then
- $u_j = u$ implies $b_j = a$, and
- $\{ u_j, u \} = \{ w_{\ell 1}, w_{\ell 2} \}$ implies ${\mathbf{A}}, \{ (u, a), (u_j, b_j) \} \models
E_{\ell}(w_{\ell 1}, w_{\ell 2})$. (Here, we use a set of pairs to denote a partial map.)
To verify that ${\mathbf{A}}\models \psi$ if and only if ${\mathbf{B}}\models
\phi$, it suffices to verify that, for any assignment $f$ defined on ${\mathsf{free}}(\phi') = {\mathsf{free}}(\psi')$ taking each variable $u$ to an element of $A_u$, that $${\mathbf{A}}, f \models \psi' \Leftrightarrow {\mathbf{B}}, f \models \phi'.$$
$(\Rightarrow)$: There exists $\ell \in {\underline{m}}$ such that ${\mathbf{A}}, f \models E_{\ell}(w_{\ell 1}, w_{\ell 2})$. Pick $w$ to be a variable in $\{ w_{ \ell 1 }, w_{ \ell 2 } \}$, and consider the extension $f^+$ of $f$ that sends each variable in $V$ to $(E_{\ell}, w, f(w))$. It is straightforward to verify that ${\mathbf{B}}, f^+$ satisfies the conjunction of $\phi'$; we do so as follows. Suppose that $R(u_1, \ldots, u_k)$ is an atom in this conjunction. We claim that $(f^+(u_1), \ldots, f^+(u_k)) \in R^{{\mathbf{B}}}$. This tuple clearly satisfies the first condition in the definition of $R^{{\mathbf{B}}}$; to check the second condition, suppose that $i, j \in {\underline{k}}$ are such that $u_i \in V$ and $u_j \notin V$. We have $f^+(u_i) = (E_{\ell}, w, f(w))$. If $u_j = w$, then indeed $f^+(u_j) = f(w)$. If $\{ u_j, w \} = \{ w_{ \ell 1 }, w_{ \ell 2 } \}$, then $\{ (w, f(w)), (u_j, f^+(u_j)) \}$ is equal to $f {\upharpoonright}\{ w_{ \ell 1}, w_{\ell 2} \}$, and we have ${\mathbf{A}}, f {\upharpoonright}\{ w_{ \ell 1}, w_{\ell 2} \} \models
E_{\ell}(w_{\ell 1}, w_{\ell 2})$ by our choice of $\ell$.
$(\Leftarrow)$: Suppose that ${\mathbf{B}}, f^+$ satisfies the conjunction of $\phi'$. By the definition of ${\mathbf{B}}$ and since $\phi'$ is a layered formula, $f^+$ maps all variables in $V$ to the same value $(E_{\ell}, u, a)$. There exists an atom $R(u_1, \ldots, u_k)$ in $\phi'$ such that one of its variables $u_j$ has the property that $\{ u_j, u \} = \{ w_{\ell 1}, w_{\ell 2} \}$. It follows, from the definition of $R^{{\mathbf{B}}}$, that ${\mathbf{A}}, f^+ {\upharpoonright}\{ u, u_j \} \models E_{\ell}(w_{\ell 1}, w_{\ell 2})$.
\[thm:red-clique-to-sorted-friendly\] Let $\Psi$ be a set of fully-sorted, friendly sentences such that ${{\mathsf{thick}}_l}(\Psi)$ is unbounded. Then, either ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ or ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ slice reduces to ${\mathsf{MC}_s}[\Psi]$.
Suppose that $\psi$ is a simple friendly formula that occurs as a subformula of a formula $\phi$. We say that $\psi$ is an *existential $k$-clique* if it is of the form $\exists X (\bigwedge_{i=1}^n \alpha_i)$ and there exists a set $V$ of variables, with $|V| \geq k$, that are existentially quantified (in $\phi$) such that for each set $\{ v, v' \} \in K(V)$, there exists an atom $\alpha_i$ with $\{ v, v' \} \subseteq {\mathsf{free}}(\alpha_i)$. We define a *universal $k$-clique* dually.
Let $C$ be as defined above in the discussion. By appeal to Theorem \[thm:formula-accordion-reduction\], it suffices to prove that either ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ or ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ slice reduces to ${\mathsf{MC}_s}[{\mathsf{closure}}_C(\Psi)]$. When $\Theta$ is a set of layered sentences, let us use the term *$\Theta$-relevant subformula* to refer to a layered subformula $\phi$ of a sentence $\theta$ in $\Theta$. For each $\Psi$-relevant subformula $\phi$, it holds that ${\mathsf{local}\textup{-}{{\mathsf{thick}}_l}}(\phi) \leq {\mathsf{quant}\textup{-}{{\mathsf{thick}}_l}}(\phi) + |{\mathsf{free}}(\phi)|$ (this is since the hypergraph from which ${\mathsf{quant}\textup{-}{{\mathsf{thick}}_l}}(\phi)$ is defined is the hypergraph from which ${\mathsf{local}\textup{-}{{\mathsf{thick}}_l}}(\phi)$ is defined, but with the vertices in ${\mathsf{free}}(\phi)$ removed). By the definition of ${{\mathsf{thick}}_l}(\cdot)$, the quantity ${\mathsf{local}\textup{-}{{\mathsf{thick}}_l}}(\phi)$ over $\Psi$-relevant subformulas $\phi$ is unbounded. Thus, over $\Psi$-relevant subformulas $\phi$, either the quantity ${\mathsf{quant}\textup{-}{{\mathsf{thick}}_l}}(\phi)$ or the quantity $|{\mathsf{free}}(\phi)|$ is unbounded. We may therefore consider two cases.
Assume that $|{\mathsf{free}}(\phi)|$ is unbounded over $\Psi$-relevant subformulas $\phi$. By conditions (2) and (3) in the definition of $C$, we obtain that $|{\mathsf{free}}(\phi)|$ is unbounded over simple ${\mathsf{closure}}_C(\Psi)$-relevant subformulas $\phi$. We assume that $|{\mathsf{free}}(\phi)|$ is unbounded over such simple subformulas $\phi$ that use existential quantification (if not, then $|{\mathsf{free}}(\phi)|$ is unbounded over such simple subformulas $\phi$ that use universal quantification, and the argumentation is dual). We now consider two cases. If the number of universally quantified variables in ${\mathsf{free}}(\phi)$ is unbounded over such subformulas $\phi$, then by condition (2) of the definition of $C$ applied to each such subformula (and the sentence in which it appears), we obtain that, for each $k \geq 1$, the set ${\mathsf{closure}}_C(\Psi)$ contains a sentence that contains a universal $k$-clique. Otherwise, the number of existentially quantified variables in ${\mathsf{free}}(\phi)$ is unbounded over such subformulas $\phi$; in this case, by an application of condition (3) followed by an application of condition (2) (again, to each such subformula $\phi$ and the sentence in which it appears), we obtain that, for each $k \geq 1$, the set ${\mathsf{closure}}_C(\Psi)$ contains a sentence that contains an existential $k$-clique. In this latter case, let us explain how to exhibit a slice reduction from ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ to ${\mathsf{MC}_s}[{\mathsf{closure}}_C(\Psi)]$ (in the former case, one dually obtains a reduction from ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ to ${\mathsf{MC}_s}[{\mathsf{closure}}_C(\Psi)]$). We define $U$ to be the set of pairs $(k, \theta)$ such that $k \geq 1$ and $\theta$ is a sentence that contains an existential $k$-clique as a subformula. Let us define $r(k, \theta, (V, E))$ to be the structure ${\mathbf{B}}$ described as follows. Let $W$ be the set of variables that witnesses the existential $k$-clique. For each relation symbol $R$ of an atom in $\theta$ that does not witness the existential $k$-clique, we may set $R^{{\mathbf{B}}}$ to either $\emptyset$ or $B^r$ (here, $r$ is the arity of $R$) in such a way that ${\mathbf{B}}\models \theta$ if and only if ${\mathbf{B}}\models \exists W (\wedge_j \beta_j)$, where the $\beta_j$ are the atoms witnessing the existential $k$-clique. By defining, for each remaining relation symbol $F$ (that is, for each relation symbol $F$ of an atom $\beta_j$), the relation $F^{{\mathbf{B}}}$ to be $E$, we obtain that $(V, E)$ contains a $k$-clique if and only if ${\mathbf{B}}\models \theta$.
Now assume that ${\mathsf{quant}\textup{-}{{\mathsf{thick}}_l}}(\phi)$ is unbounded over $\Psi$-relevant subformulas $\phi$. By conditions (2) and (3) in the definition of $C$, we obtain that ${\mathsf{quant}\textup{-}{{\mathsf{thick}}_l}}(\phi)$ is unbounded over simple ${\mathsf{closure}}_C(\Psi)$-relevant subformulas $\phi$. By considering the sentences based on these subformulas (and by invoking condition (1) in the definition of $C$), we obtain that ${\mathsf{closure}}_C(\Psi)$ contains simple sentences $\exists X (\bigwedge \alpha_i)$ or contains simple sentences $\exists Y (\bigvee \alpha_i)$ such that, over these sentences, ${\mathsf{quant}\textup{-}{{\mathsf{thick}}_l}}(\cdot)$ is unbounded. In the former case, the intractability result of Grohe, Schwentick and Segoufin [@GroheSchwentickSegoufin01-conjunctivequeries] directly yields a slice reduction from ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ to ${\mathsf{MC}_s}[{\mathsf{closure}}_C(\Psi)]$ where the set $U$ contains a pair $(k, \theta)$ when $k \geq 1$ and $\theta$ is a sentence $\exists X (\bigwedge \alpha_i)$ whose corresponding graph admits a $k$-by-$k$ grid as a minor. The latter case is dual (one obtains a slice reduction from ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{co}$}-{\small $\mathsf{CLIQUE}$}}}$}}}}$ to ${\mathsf{MC}_s}[{\mathsf{closure}}_C(\Psi)]$).
Proof of Theorem \[thm:slice-reduction-transitive\]
===================================================
Let $(U_1, r_1)$ be a slice reduction from $Q_1[S_1]$ to $Q_2[S_2]$, and let $(U_2, r_2)$ be a slice reduction from $Q_2[S_2]$ to $Q_3[S_3]$. Define $U$ to be the set
$\{ (t_1, t_3) ~|~ \textup{ there exists $t_2 \in \Sigma^*$ such that}$
$\textup{$(t_1, t_2)
\in U_1$ and $(t_2, t_3) \in U_2$} \}.$
It is straightforward to verify that $U$ is computably enumerable and that there exists an algorithm $A_U$ that, given a pair $(t_1, t_3) \in U$, outputs a value $t_2 \in \Sigma^*$ such that $(t_1, t_2) \in U_1$ and $(t_2, t_3) \in U_2$. Define $r$ to be a partial function such that, when $(t_1, t_3) \in U$, for each $y \in \Sigma^*$ it holds that $r(t_1, t_3, y) = r_2(t_2, t_3, r_1((t_1, t_2), y))$; here, we use $t_2$ to denote $A_U(t_1, t_3)$.
We verify that $(U, r)$ is a slice reduction from $Q_1[S_1]$ to $Q_3[S_3]$. For each $t_1 \in S_1$, there exists $t_2 \in S_2$ such that $(t_1, t_2) \in U_1$; and, for each $t_2 \in S_2$, there exists $t_3 \in S_3$ such that $(t_2, t_3) \in U_2$. Hence, for each $t_1 \in S_1$, there exists $t_3 \in S_3$ such that $(t_1, t_3) \in U$. This confirms the coverage condition; we now check correctness. Suppose that $(t_1, t_3) \in U$, and let $y \in \Sigma^*$. Set $y' = r_1(t_1, t_2, y)$. We have that $$(t_1, y) \in Q_1 \Leftrightarrow (t_2, y') \in Q_2$$ and $$(t_2, y') \in Q_2 \Leftrightarrow (t_3, r_2(t_2, t_3, y')) \in
Q_3.$$ It follows that $$(t_1, y) \in Q_1 \Leftrightarrow (t_3, r(t_1, t_3, y)) \in Q_3.$$
It remains to verify that the function $r$ is FPT-computable with respect to the parameterization $(\pi_1, \pi_2)$. We verify this by considering the natural algorithm at this point, namely, the following algorithm: given $(t_1, t_3, y)$, invoke $A_U$ on $(t_1, t_3)$ and, if this computation halts, set $t_2$ to the result; then, compute $y' = r_1(t_1, t_2, y)$ using the algorithm witnessing FPT-computability, and finally, compute $r_2(t_2, t_3, y')$ using the algorithm witnessing FPT-computability. On an input $x = (t_1, t_3, y) \in U \times \Sigma^*$, the running time of this algorithm can be upper bounded by $$F(t_1, t_3) +
f_1(t_1, t_2) p_1(|(t_1, t_2, y)|) +
f_2(t_2, t_3) p_2(|(t_2, t_3, y')|)$$ where $F(t_1, t_3)$ denotes the running time of $A_U$ on input $(t_1, t_3)$; and $(f_1, p_1)$ and $(f_2, p_2)$ witness the FPT-computability of $r_1$ and $r_2$, respectively. We need to show that this running time is degree-bounded with respect to $(\pi_1, \pi_2)$; for the remainder of the proof, let us simply use *degree-bounded* to mean degree-bounded with respect to this parameterization. We view the various quantities under discussion as functions of $x = (t_1, t_3, y)$. Since $F(t_1, t_3)$, $f_1(t_1, t_2)$ and $f_2(t_2, t_3)$ can be viewed as computable functions of $(t_1, t_3)$, by Proposition \[prop:db-closure\] (\[closure:sum\], \[closure:product\]), it suffices to verify that each of $p_1(|(t_1, t_2, y)|)$, $p_2(|(t_2, t_3, y')|)$ is degree-bounded. By appeal to Proposition \[prop:db-closure\](\[closure:polynomial\]), to verify that $p_1(|(t_1, t_2, y)|)$ is degree-bounded, it suffices to observe that each of $|t_1|$, $|t_2|$, $|y|$ is degree-bounded. Similarly, to verify that $p_2(|(t_2, t_3, y')|)$ is degree-bounded, it suffices to verify that each of $|t_2|$, $|t_3|$, and $|y'|$ are degree-bounded; this is clear for $|t_2|$ and $|t_3|$, and the size $|y'|$ is bounded above by $f_1(t_1, t_2) p_1(|(t_1, t_2, y)|)$, which is degree-bounded since we just verified that $p_1(|(t_1, t_2, y)|)$ is degree-bounded.
Discussion: Derivation of the Classification of Prefixed Graphs {#sect:discussion}
===============================================================
Let us say that a sentence is *quantified conjunctive* if conjunction ($\wedge$) is the only connective that occurs therein. A *prefixed graph* is a pair $(P, G)$ where $P$ is a quantifier prefix and $G$ is a graph whose vertices are the variables appearing in $P$. Let ${\mathcal{G}}$ be a set of prefixed graphs. Let ${\mathrm{QC}}({\mathcal{G}})$ denote the set that contains a prenex quantified conjunctive sentence $P \phi$ if there exists a prefixed graph $(P, G) \in {\mathcal{G}}$ such that $\phi$ is a conjunction of atoms, where the variable set of each atom forms a clique in $G$. We will assume that no variable occurs more than once in an atom; we make this assumption without loss of interestingness, since given a sentence $\Phi \in {\mathrm{QC}}({\mathcal{G}})$ and a structure ${\mathbf{B}}$, one can efficiently compute a sentence $\Phi' \in {\mathrm{QC}}({\mathcal{G}})$ and a structure ${\mathbf{B}}'$ such that (1) each atom of $\Phi$ containing more than one variable occurrence is replaced, in $\Phi'$, with an atom with the same variables but not having multiple variable occurrences, and (2) ${\mathbf{B}}\models \Phi$ iff ${\mathbf{B}}' \models \Phi'$.
The previous work [@ChenDalmau12-decomposingquantified] studied the complexity of model checking on sentence sets ${\mathrm{QC}}({\mathcal{G}})$, proving a comprehensive classification. Here, we show how this classification can be readily derived using the main theorem of the present article. This witnesses the strength and generality of our main theorem (Theorem \[thm:main\]). We first present two lemmas, then proceed to the derivation.
When ${\mathcal{G}}$ is a set of prefixed graphs, let ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$ denote the subset of ${\mathrm{QC}}({\mathcal{G}})$ that contains a sentence if
- in each atom, the latest occurring variable (in the quantifier prefix) is existentially quantified, and
- it is symbol-loose.
Let ${\mathcal{G}}$ be any set of prefixed graphs. The case problems ${\mathsf{MC}}[{\mathrm{QC}}({\mathcal{G}})]$ and ${\mathsf{MC}}[{\mathrm{norm\mbox{-}QC}}({\mathcal{G}})]$ slice reduce to each other.
It is straightforward to verify that ${\mathsf{MC}}[{\mathrm{norm\mbox{-}QC}}({\mathcal{G}})]$ slice reduces to ${\mathsf{MC}}[{\mathrm{QC}}({\mathcal{G}})]$ (by definition, ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$ is a subset of ${\mathrm{QC}}({\mathcal{G}})$). We slice reduce from ${\mathsf{MC}}[{\mathrm{QC}}({\mathcal{G}})]$ to ${\mathsf{MC}}[{\mathrm{norm\mbox{-}QC}}({\mathcal{G}})]$ as follows. Define $U$ to contain a pair $(\phi, \phi')$ of prenex quantified conjunctive sentences if they share the same quantifier prefix, $\phi'$ is symbol-loose, and each atom $\alpha$ of $\phi$ can be placed in bijective correspondence with an atom $\alpha'$ of $\phi'$ in such a way that the variables of $\alpha'$ is the set obtained by taking the variables of $\alpha$ and iteratively eliminating the latest occurring variable when it is universally quantified, until the latest occurring variable is existentially quantified. The partial function $r$ is defined in a natural way, namely, such that $r(\phi, \phi', {\mathbf{B}})$ is a structure ${\mathbf{B}}'$ having the same universe $B$ as ${\mathbf{B}}$ and having the following property: for each atom $\alpha = R(u_1, \ldots, u_m)$ of $\phi$, it holds that the satisfying assignments for $\alpha'$ over ${\mathbf{B}}'$ are precisely the satisfying assignments for $\forall D \alpha$ over ${\mathbf{B}}$, where $D$ denotes the variables in $\alpha$ that do not occur in $\alpha'$.
Let $\Phi$ be a set of symbol-loose quantified conjunctive sentences. It holds that ${\mathsf{MC}}[\Phi]$ and ${\mathsf{MC}}[\Phi']$ are slice reducible to each other, where $\Phi'$ is the graph-like closure of $\Phi$.
It is straightforward to verify that ${\mathsf{MC}}[\Phi]$ slice reduces to ${\mathsf{MC}}[\Phi']$, as $\Phi \subseteq \Phi'$. We thus show that ${\mathsf{MC}}[\Phi']$ slice reduces to ${\mathsf{MC}}[\Phi]$. Let $U$ be the set of pairs $(\phi', \phi)$ such that $\phi$ is a symbol-loose quantified conjunctive sentence, and $\phi'$ is in the graph-like closure of $\phi$. We use $\sigma$ and $\sigma'$ to denote the signatures of $\phi$ and $\phi'$, respectively. Since disjunction and negation do not occur in $\phi$, we have that $\phi'$ is obtained from $\phi$ via applications of the syntactic transformations $(\alpha)$, $(\beta)$, and $(\gamma)$, and replacement. Since each of these syntactic transformations preserves logical equivalence as well as the number of atom occurrences, there is a mapping $S: \sigma \to \sigma'$ such that when each symbol in $\phi$ is mapped under $S$, the resulting sentence is logically equivalent to $\phi'$. The partial function $r$ is defined by $r(\phi', \phi, {\mathbf{B}}') = {\mathbf{B}}$ where for each symbol $R \in \sigma$, the relation $R^{{\mathbf{B}}}$ is defined as $S(R)^{{\mathbf{B}}'}$.
We now explain how to obtain the main dichotomy of the previous work [@ChenDalmau12-decomposingquantified], in particular, we show how to classify precisely the sets of prefixed graphs ${\mathcal{G}}$ such that ${\mathsf{MC}}[{\mathrm{QC}}({\mathcal{G}})]$ is in ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$.
First, consider the case where atoms in ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$ may have unboundedly many variables. If for each $k \geq 1$ there exists an atom in ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$ with at least $k$ existentially quantified variables, then there is a direct reduction from ${{\textup{{\small $\mathsf{case}$}-{$ {\textup{{\small $\mathsf{CLIQUE}$}}}$}}}}$ and one has ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{W[1]}$}}$}}}$-hardness of ${\mathsf{MC}}[{\mathrm{QC}}({\mathcal{G}})]$. Otherwise, define $F_k$ to be the sentence $\forall y_1 \ldots \forall y_k \exists x \bigwedge_{i=1}^k
E_i(y_i,x)$; up to the insertion of additional variables that do not appear in atoms and up to renaming of variables, we have that the sentences $F_k$ are instances of ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$. By the above two lemmas, it suffices to prove that the graph-like closure of $\{ F_k \}$ is hard. It is readily verified that ${\mathrm{lay}^+}( F_k )$ is $\forall \{ y_1, \ldots, y_k \} (\exists x \bigwedge_{i=1}^k
E_i(y_i, x))$, where the formula in parentheses is viewed as the single disjunct of a disjunction; we have that ${\mathsf{thick}}(F_k) = k+1$, and by the main theorem, we obtain hardness (either ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{W[1]}$}}$}}}$-hardness or ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{co}$-$\mathsf{W[1]}$}}$}}}$-hardness) of the graph-like closure of $\{ F_k \}$.
Next, consider the case where there is a constant upper bound on the number of variables that occur in atoms in ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$. By our assumption that no variable occurs more than once in an atom, the set of sentences ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$ has bounded arity. In this case, by the two presented lemmas, we have that ${\mathsf{MC}}[{\mathrm{QC}}({\mathcal{G}})]$ is equivalent, under slice reduction, to ${\mathsf{MC}}[\Phi']$, where $\Phi'$ is the graph-like closure of ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$. Since ${\mathrm{norm\mbox{-}QC}}({\mathcal{G}})$ has bounded arity, $\Phi'$ does as well. Hence, our main theorem (Theorem \[thm:main\]) then can be applied to infer that ${\mathsf{MC}}[\Phi']$ is either in ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{FPT}$}}$}}}$, is ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{W[1]}$}}$}}}$-hard, or is ${\textup{{\small $\mathsf{case}$}-{$ {\textup{\small $\mathsf{co}$-$\mathsf{W[1]}$}}$}}}$-hard.
By arguing as in Corollaries \[cor:main-computable\] and \[cor:main-noncomputable\], one can derive dichotomies in the complexity of ${\mathsf{param}\textup{-}{\mathsf{MC}}[{\mathrm{QC}}({\mathcal{G}})]}$.
The author was supported by the Spanish Project FORMALISM (TIN2007-66523), by the Basque Government Project S-PE12UN050(SAI12/219), and by the University of the Basque Country under grant UFI11/45. The author thanks Montserrat Hermo and Simone Bova for useful comments.
[^1]: Note that in the case of unbounded arity, complexity may depend on the choice of representation of relations [@ChenGrohe10-succinct].
[^2]: Note that in the transformation $(\gamma)$, we permit that $\phi$ is not present.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The trace algebra $C_{nd}$ over a field of characteristic 0 is generated by all traces of products of $d$ generic $n\times n$ matrices, $n,d\geq 2$. Minimal sets of generators of $C_{nd}$ are known for $n=2$ and $n=3$ for any $d$ as well as for $n=4$ and $n=5$ and $d=2$. The defining relations between the generators are found for $n=2$ and any $d$ and for $n=3$, $d=2$ only. Starting with the generating set of $C_{3d}$ given by Abeasis and Pittaluga in 1989, we have shown that the minimal degree of the set of defining relations of $C_{3d}$ is equal to 7 for any $d\geq 3$. We have determined all relations of minimal degree. For $d=3$ we have also found the defining relations of degree 8. The proofs are based on methods of representation theory of the general linear group and easy computer calculations with standard functions of Maple.'
address:
- 'Dipartimento di Matematica ed Applicazioni, Università di Palermo, Via Archirafi 34, 90123 Palermo, Italy'
- 'Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, 1113 Sofia, Bulgaria'
author:
- Francesca Benanti and Vesselin Drensky
title: 'Defining Relations of Minimal Degree of the Trace Algebra of $3 \times 3$ Matrices'
---
[^1] [^2]
Introduction {#introduction .unnumbered}
============
Let $K$ be any field of characteristic 0. All vector spaces, tensor products, algebras considered in this paper are over $K$. Let $X_i=\left(x_{pq}^{(i)}\right)$, $p,q=1,\ldots,n$, $i=1,\ldots,d$, be $d$ generic $n\times n$ matrices. We consider the pure (or commutative) trace algebra $C_{nd}$ generated by all traces of products $\text{\rm tr}(X_{i_1}\cdots X_{i_k})$. The algebra $C_{nd}$ coincides with the algebra of invariants of the general linear group $GL_n=GL_n(K)$ acting by simultaneous conjugation on $d$ matrices of size $n\times n$. General results of invariant theory of classical groups imply that the algebra $C_{nd}$ is finitely generated. Theory of PI-algebras provides upper bounds for the generating sets of the algebras $C_{nd}$. The Nagata-Higman theorem states that the polynomial identity $x^n=0$ implies the identity $x_1\cdots x_N=0$ for some $N=N(n)$. If $N$ is minimal with this property, then $C_{nd}$ is generated by traces of products $\text{\rm tr}(X_{i_1}\cdots X_{i_k})$ of degree $k\leq N$. This estimate is sharp if $d$ is sufficiently large. A description of the defining relations of $C_{nd}$ is given by the Razmyslov-Procesi theory [@R; @P] in the language of ideals of the group algebras of symmetric groups. For a background on the algebras of matrix invariants see, e.g. [@F; @DF] and for computation aspects of the theory see [@D2].
Explicit minimal sets of generators of $C_{nd}$ and the defining relations between them are found in few cases only. It is well known that, in the Nagata-Higman theorem, $N(2)=3$, $N(3)=6$, and $N(4)=10$, which gives bounds for the degrees of the generators of the algebras $C_{2d}$, $C_{3d}$, and $C_{4d}$, respectively. Nevertheless, the defining relations of $C_{nd}$ are explicitly given for $n=2$ and any $d$, see e.g. [@DF] for details, and for $n=3$, $d=2$, see the comments below. For $n=3$, $d\geq 3$ and $n\geq 4$ and $d\geq 2$, nothing is known about the concrete form of the defining relations with respect to fixed minimal systems of generators.
Teranishi [@T1] found a system of 11 generators of $C_{32}$. It follows from his description that, with respect to these generators, $C_{32}$ can be defined by a single relation of degree 12. The explicit (but very complicated) form of the relation was found by Nakamoto [@N], over $\mathbb Z$, with respect to a slightly different system of generators. Abeasis and Pittaluga [@AP] found a system of generators of $C_{3d}$, for any $d\geq 2$, in terms of representation theory of the symmetric and general linear groups, in the spirit of its usage in theory of PI-algebras. Aslaksen, Drensky and Sadikova [@ADS] gave the defining relation of $C_{32}$ with respect to the set found in [@AP]. Their relation is much simpler than that in [@N]. For $C_{42}$, a set of generators was found by Teranishi [@T1; @T2] and a minimal set by Drensky and Sadikova [@DS], in terms of the approach in [@AP]. Djoković [@D] gave another minimal set of 32 generators of $C_{42}$ consisting of traces of products only. He found also a minimal set of 173 generators of $C_{52}$.
As usually in invariant theory, the determination of generators and defining relations is simpler, if one has some additional information about the algebras of invariants. In particular, it is very useful to know the Hilbert (or Poincaré) series of the algebra. Again, the picture is completely clear for $n=2$. The only other cases, when the Hilbert series are explicitly given, are $n=3$, $d=2$ (Teranishi [@T1]) and $d=3$ (Berele and Stembridge [@BS]), $n=4$, $d=2$ (Teranishi [@T2] (with some typos) and corrected by Berele and Stembridge [@BS]). Recently Djoković [@D] has calculated also the Hilbert series of $C_{52}$ and $C_{62}$.
The minimal generating set of $C_{3d}$ given in [@AP] consists of $$g=g(d)=\frac{1}{240}d(5d^5+19d^4-5d^3+65d^2+636)$$ homogeneous trace polynomials $u_1,\ldots,u_g$ of degree $\leq 6$. In more detail, the number of polynomials $u_i$ of degree $k$ is $g_k$, and $$g_1=d,g_2=\frac{1}{2}(d+1)d,g_3=\frac{1}{3}d(d^2+2),
g_4=\frac{1}{24}(d+1)d(d-1)(5d-6),$$ $$g_5=\frac{1}{30}d(d-1)(d-2)(3d^2+4d+6),
g_6=\frac{1}{48}(d+2)(d+1)d(d-1)(d^2-3d+4).$$ Hence $C_{3d}$ is isomorphic to the factor algebra $K[y_1,\ldots,y_g]/I$. Defining $\text{\rm deg}(y_i)=\text{\rm deg}(u_i)$, the ideal $I$ is homogeneous. For $d=3$, surprisingly, the comparison of the Hilbert series of $C_{33}\cong K[y_1,\ldots,y_g]/I$ given in [@BS], with the Hilbert series of $K[y_1,\ldots,y_g]$, gives that any homogeneous minimal system of generators of the ideal $I$ contains no elements of degree $\leq 6$, three elements of degree 7 and 30 elements of degree 8. The purpose of the present paper is to find the defining relations of minimal degree for $C_{3d}$ and any $d\geq 3$, with respect to the generating set in [@AP]. It has turned out that the minimal degree of the relations is equal to 7 for all $d\geq 3$, and there are a lot of relations of degree 7. (Compare with the single relation of degree 12 in the case $d=2$.) The dimension of the vector space of relations of degree 7 is equal to $$r_7=r_7(d)=\frac{2}{7!}(d+1)d(d-1)(d-2)(41d^3-86d^2+114d-360).$$ For $d=3$ we have computed also the homogeneous relations of degree 8. The defining relations are given in the language of representation theory of $GL_d$. There is a simple algorithm which gives the explicit form of all relations of degree 7, and of degree 8 for $d=3$. The proofs involve basic representation theory of $GL_d$ and develop further ideas of [@ADS; @DS] and our recent paper [@BD] combined with computer calculations with Maple. In the case $d=3$ we have used essentially the Hilbert series of $C_{33}$ from [@BS] which has allowed to reduce the number of computations. Our methods are quite general and we believe that they can be successfully used for further investigation of generic trace algebras and other algebras close to them.
Preliminaries
=============
In what follows, we fix $n=3$ and $d\geq 3$ and denote by $X_1,\ldots,X_d$ the $d$ generic $3\times 3$ matrices. Often, when the value of $d$ is clear from the context, we shall denote $C_{3d}$ by $C$. It is a standard trick to replace the generic matrices with generic traceless matrices. We express $X_i$ in the form $$X_i=\frac{1}{3}\text{\rm tr}(X_i)e+x_i,\quad i=1,\ldots,d,$$ where $e$ is the identity $3\times 3$ matrix and $x_i$ is a generic traceless matrix. Then $$\label{replacing with traceless matrices}
C_{3d}\cong K[\text{\rm tr}(X_1),\ldots,\text{\rm tr}(X_d)]\otimes C_0,$$ where the algebra $C_0$ is generated by the traces of products $\text{\rm tr}(x_{i_1}\cdots x_{i_k})$, $k\leq 6$. Hence the problem for the defining relations of $C$ can be replaced by a similar problem for $C_0$.
As in the case of “ordinary” generic matrices, without loss of generality we may replace $x_1$ by a generic traceless diagonal matrix. Changing the variables $x_{pp}^{(i)}$, we may assume that $$\label{first matrix}
x_1=\left( \begin{array}{ccc}
x_{11}^{(1)} & 0 & 0 \\
0 & x_{22}^{(1)} & 0 \\
0 & 0 & -(x_{11}^{(1)}+x_{22}^{(1)})\\
\end{array} \right),$$ $$\label{other matrices}
x_i=\left( \begin{array}{ccc}
x_{11}^{(i)} & x_{12}^{(i)} & x_{13}^{(i)} \\
x_{21}^{(i)} & x_{22}^{(i)} & x_{23}^{(i)} \\
x_{31}^{(i)} & x_{32}^{(i)} & -(x_{11}^{(i)}+x_{22}^{(i)})\\
\end{array} \right), \quad i=2,\ldots,d.$$ Till the end of the paper we fix these $d$ generic traceless matrices. Let $C_0^+=\omega(C_0)$ be the augmentation ideal of $C_0$. It consists of all trace polynomials $f(x_1,\ldots,x_d)\in C_0$ without constant terms, i.e., satisfying the condition $f(0,\ldots,0)=0$. Any minimal system of generators of $C_0$ lying in $C_0^+$ forms a basis of the vector space $C_0^+$ modulo $(C_0^+)^2$. Conversely, if a system of polynomials $f_1,\ldots,f_g$ forms a basis of $C_0^+$ modulo $(C_0^+)^2$, and each $f_i$ is a linear combination of traces of products $\text{\rm tr}(x_{i_1}\cdots x_{i_k})$ then it is a minimal generating system of $C_0$. The algebra $C=C_{3d}$ is $\mathbb Z$-graded assuming that the trace $\text{\rm tr}(X_{i_1}\cdots X_{i_k})$ is of degree $k$, and this grading is inherited by $C_0$. Similarly, $C$ (and also $C_0$) has a more precise ${\mathbb Z}^d$-multigrading induced by the condition that $X_1,\ldots,X_d$ are, respectively, of multidegree $(1,0,\ldots,0,0),\ldots,(0,0,\ldots,0,1)$. The considerations below, stated for the $\mathbb Z$-grading hold also for the ${\mathbb Z}^d$-multigrading. The numbers $g_1,g_2,\ldots,g_6$ of elements of degree $1,2,\ldots,6$, respectively, in any homogeneous minimal system of generators is an invariant of $C$. Any homogeneous minimal system $\{f_1,\ldots,f_h\}$ of generators of $C_0$ consists of $g_2,\ldots,g_6$ elements of degree $2,\ldots,6$, and $h=g_2+\cdots+g_6$. Hence $$C_0\cong K[z_1,\ldots,z_h]/J,$$ with isomorphism defined by $z_j+J\to f_j$, $j=1,\ldots,h$. If $u_j(z_1,\ldots,z_h)$, $j=1,\ldots,r$, is a system of generators of the ideal $J$, then $u_j(f_1,\ldots,f_h)=0$, $j=1,\ldots,r$, is a system of defining relations of $C_0$ with respect to the system of generators $\{f_1,\ldots,f_h\}$. Any homogeneous system of polynomials in $J$ (where, by definition $\text{\rm deg}(z_j)=\text{\rm deg}(f_j)$) which forms a basis of the vector space $J$ modulo the subspace $JK[z_1,\ldots,z_h]^+$, is a minimal system of generators of the ideal $J$. We denote by $r_k$ the number of elements of degree $k$ in such a system. Clearly, $r_k$ is the dimension of the homogeneous component of degree $k$ of the vector space $J/JK[z_1,\ldots,z_h]^+$.
Now we summarize the necessary background on representation theory of $GL_d$. We refer e.g. to [@M] for general facts and to [@D1] for applications in the spirit of the problems considered here. All $GL_d$-modules which appear in this paper are completely reducible and are direct sums of irreducible polynomial modules. The irreducible polynomial representations of $GL_d$ are indexed by partitions $\lambda=(\lambda_1,\ldots,\lambda_d)$, $\lambda_1\geq \cdots\geq \lambda_d\geq 0$. We denote by $W(\lambda)=W_d(\lambda)$ the corresponding irreducible $GL_d$-module, assuming that $W_d(\lambda)=0$ if $\lambda_{d+1}\not=0$. The group $GL_d$ acts in the natural way on the $d$-dimensional vector space $K\cdot x_1+\cdots+K\cdot x_d$ and this action is extended diagonally on the free associative algebra $K\langle x_1,\ldots,x_d\rangle$.
The module $W(\lambda)\subset K\langle x_1,\ldots,x_d\rangle$ is generated by a unique, up to a multiplicative constant, homogeneous element $w_{\lambda}$ of degree $\lambda_j$ with respect to $x_j$, called the highest weight vector of $W(\lambda)$. It is characterized by the following property.
\[criterion for hwv\] Let $1\leq i<j\leq d$ and let $\Delta_{ij}$ be the derivation of $K\langle x_1,\ldots,x_d\rangle$ defined by $\Delta_{ij}(x_j)=x_i$, $\Delta_{ij}(x_k)=0$, $k\not=j$. If $w(x_1,\ldots,x_d) \in K\langle x_1,\ldots,x_d\rangle$ is multihomogeneous of degree $\lambda=(\lambda_1,\ldots,\lambda_d)$, then $w(x_1,\ldots,x_d)$ is a highest weight vector for some $W(\lambda)$ if and only if $\Delta_{ij}(w(x_1,\ldots,x_d))=0$ for all $i<j$. Equivalently, $w(x_1,\ldots,x_d)$ is a highest weight vector for $W(\lambda)$ if and only if $$g_{ij}(w(x_1,\ldots,x_d))=w(x_1,\ldots,x_d),\quad 1\leq i<j\leq d,$$ where $g_{ij}$ is the linear operator of the $d$-dimensional vector space which sends $x_j$ to $x_i+x_j$ and fixes the other $x_k$.
The lemma is a partial case of a result by De Concini, Eisenbud, and Procesi [@DEP], see also Almkvist, Dicks, and Formanek [@ADF]. In the version which we need, the first part of the lemma was established by Koshlukov [@K]. The equivalence follows from the fact that the kernel of any locally nilpotent derivation $\Delta$ coincides with the fixed points of the related exponential automorphism $\exp(\Delta)=1+\Delta/1!+\Delta^2/2!+\cdots$, and $g_{ij}=\exp(\Delta_{ij})$.
If $W_i$, $i=1,\ldots,m$, are $m$ isomorphic copies of the $GL_d$-module $W(\lambda)$ and $w_i\in W_i$ are highest weight vectors, then the highest weight vector of any submodule $W(\lambda)$ of the direct sum $W_1\oplus\cdots\oplus W_m$ has the form $\xi_1w_1+\cdots+\xi_mw_m$ for some $\xi_i\in K$. Any $m$ linearly independent highest weight vectors can serve as a set of generators of the $GL_d$-module $W_1\oplus\cdots\oplus W_m$.
It is convenient to work with an explicit copy of $W(\lambda)$ in $K\langle x_1,\ldots,x_d\rangle$ obtained in the following way. Let $$s_k(x_1,\ldots,x_k)=\sum_{\sigma\in S_k}\text{\rm sign}(\sigma)x_{\sigma(1)}\cdots x_{\sigma(k)}$$ be the standard polynomial of degree $k$. (Clearly, $$s_2(x_1,x_2)=x_1x_2-x_2x_1=[x_1,x_2]$$ is the commutator of $x_1$ and $x_2$.) If the lengths of the columns of the diagram of $\lambda$ are, respectively, $k_1,\ldots,k_p$, $p=\lambda_1$, then $$\label{canonical hwv}
w_{\lambda}=w_{\lambda}(x_1,\ldots,x_{k_1})=s_{k_1}(x_1,\ldots,x_{k_1})\cdots s_{k_p}(x_1,\ldots,x_{k_p})$$ is the highest weight vector of a submodule $W(\lambda)\subset K\langle x_1,\ldots,x_d\rangle$. Sometimes we shall write $w_{\lambda}=w_{\lambda}(x_1,\ldots,x_d)$, even when $k_1<d$.
Recall that the $\lambda$-tableau $$T=(a_{ij}),\quad a_{ij}\in \{1,\ldots,d\},\quad i=1,\ldots,d,\quad j=1,\ldots,\lambda_i,$$ is semistandard if its entries do not decrease from left to right in rows and increase from top to bottom in columns. The following lemma gives a basis of the vector subspace $W(\lambda)\subset K\langle x_1,\ldots,x_d\rangle$. It also provides an algorithm to construct this basis.
\[basis of module\] Let $T=(a_{ij})$ be a semistandard $\lambda$-tableau such that its $i$-th row contains $b_{i,i}$ times $i$, $b_{i,i+1}$ times $i+1$, $\ldots$, $b_{i,d}$ times $d$. Let $w(x_1,\ldots,x_d)$ be the highest weight vector of $W(\lambda)\subset K\langle x_1,\ldots,x_d\rangle$ and let $$u_T(x_{11},x_{12},\ldots,x_{1d},x_{22},\ldots,x_{2d},\ldots,x_{dd})$$ be the multihomogeneous component of degree $b_{iq}$ in $x_{iq}$, $q=i,i+1,\ldots,d$, of the polynomial $$w(x_{11}+x_{12}+\cdots+x_{1d},x_{22}+\cdots+x_{2d},\ldots,x_{dd}).$$ When $T$ runs on the set of semistandard $\lambda$-tableaux, the polynomials $$v_T=v_T(x_1,\ldots,x_d)=u_T(x_1,x_2,\ldots,x_d,x_2,\ldots,x_d,\ldots,x_d)$$ form a basis of the vector space $W(\lambda)$.
By standard Vandermonde arguments, the polynomial $u_T(x_{11},x_{12},\ldots,x_{dd})$ is a linear combination of some $w(\sum\alpha_{1j}x_{1j},\ldots,\sum\alpha_{dj}x_{dj})$, $\alpha_{ij}\in K$. Hence $v_T(x_1,\ldots,x_d)$ is a linear combination of $w(\sum\alpha_{1j}x_j,\ldots,\sum\alpha_{dj}x_j)$ and belongs to the $GL_d$-module $W(\lambda)$ generated by $w(x_1,\ldots,x_d)$. Without loss of generality, it is sufficient to consider the case when $W(\lambda)$ is generated by the element (\[canonical hwv\]). The polynomial $w(x_{11}+x_{12}+\cdots+x_{1d},x_{22}+\cdots+x_{2d},\ldots,x_{dd})$ is a product of evaluations $$s_{k_j}(x_{11}+x_{12}+\cdots+x_{1d},x_{22}+\cdots+x_{2d},\ldots,x_{k_j,k_j}+\cdots+x_{k_j,d})$$ of standard polynomials. Hence $u_T(x_{11},x_{12},\ldots,x_{dd})$ is a linear combination of monomials starting with $x_{\sigma(1),q_1}\cdots x_{\sigma(k_1),q_{k_1}}$. We order the variables by $x_1>\cdots>x_d$ and consider the lexicographic order of $K\langle x_1,\ldots,x_d\rangle$. The first column of the semistandard tableau $T$ contains $a_{11}<\cdots<a_{k1}$ and in each row $a_{i1}\leq a_{ij}$ for all $j=2,\ldots,\lambda_i$. This easily implies that the leading monomial of $v_T$ starts with $x_{a_{11}}x_{a_{21}}\cdots x_{a_{k_11}}$. Hence the first $k_1$ variables in the leading monomial are indexed by the entries of the first column of the tableau. Continuing in the same way, we obtain that the next $k_2$ variables in the leading monomial are indexed with the entries of the second row, etc. Hence, for fixed $\lambda$, the leading monomial of $v_T$ determines completely the tableau $T$, and the polynomials $v_T$ are linearly independent. Since $W(\lambda)$ has a basis which is in 1-1 correspondence with the semistandard $\lambda$-tableaux, and the number of $v_T$ is equal to the number of semistandard tableaux, we obtain that the polynomials $v_T$ form a basis of $W(\lambda)$.
If $W$ is a $GL_d$-submodule or a factor module of $K\langle x_1,\ldots,x_d\rangle$, then $W$ inherits the ${\mathbb Z}^d$-grading of $K\langle x_1,\ldots,x_d\rangle$. Recall that the Hilbert series of $W$ with respect to its ${\mathbb Z}^d$-multigrading is defined as the formal power series $$H(W,t_1,\ldots,t_d)=\sum_{k_i\geq 0}
\dim(W^{(k_1,\ldots,k_d)})t_1^{k_1}\cdots t_d^{k_d},$$ with coefficients equal to the dimensions of the homogeneous components $W^{(k_1,\ldots,k_d)}$ of degree $(k_1,\ldots,k_d)$. It plays the role of the $GL_d$-character of $W$: If $$W\cong \sum_{\lambda}m(\lambda)W(\lambda),$$ then $$H(W,t_1,\ldots,t_d)=\sum_{\lambda}m(\lambda)S_{\lambda}(t_1,\ldots,t_d),$$ where $S_{\lambda}=S_{\lambda}(t_1,\ldots,t_d)$ is the Schur function associated with $\lambda$, and the multiplicities $m(\lambda)$ are determined by $H(W,t_1,\ldots,t_d)$. One of the possible ways to introduce Schur functions is via Vandermonde-like determinants. For a partition $\mu=(\mu_1,\ldots,\mu_d)$, define the determinant $$V(\mu_1,\ldots,\mu_d)=\left\vert\begin{matrix}
t_1^{\mu_1}&t_2^{\mu_1}&\cdots&t_d^{\mu_1}\\
t_1^{\mu_2}&t_2^{\mu_2}&\cdots&t_d^{\mu_2}\\
\vdots&\vdots&\ddots&\vdots\\
t_1^{\mu_d}&t_2^{\mu_d}&\cdots&t_d^{\mu_d}\\
\end{matrix}\right\vert.$$ Then the Schur function is $$S_{\lambda}(t_1,\ldots,t_d)=
\frac{V(\lambda_1+d-1,\lambda_2+d-2,\ldots,\lambda_{d-1}+1,\lambda_d)}
{V(d-1,d-2,\ldots,1,0)}.$$ The dimension of $W(\lambda)$ is given by the formula $$\label{dim of W}
\text{\rm dim}(W_d(\lambda))=\prod_{1\leq i<j\leq d}\frac{\lambda_i-\lambda_j+j-i}{j-i}.$$ The decomposition of the tensor product $W_d(\lambda)\otimes W_d(\mu)$ of two irreducible $GL_d$-modules $W_d(\lambda)$ and $W_d(\mu)$ is given by the Littlewood-Richardson rule. We shall need it in one case only: $$\label{usage of LRrule}
W_3(2^2)\otimes W_3(2^2)\cong W_3(4^2)\oplus W_3(4,3,1)\oplus W_3(4,2^2).$$ But even in this case we can check the equality (\[usage of LRrule\]) verifying directly the equality of symmetric functions $$H(W_3(2^2)\otimes W_3(2^2),t_1,t_2,t_3)=S_{(2^2)}^2(t_1,t_2,t_3)$$ $$=S_{(4^2)}(t_1,t_2,t_3)
+S_{(4,3,1)}(t_1,t_2,t_3)
+S_{(4,2^2)}(t_1,t_2,t_3).$$ In all other cases it will be sufficient to use the Young rule which is a partial case of the Littlewood-Richardson one: $$\label{Young rule A}
W_d(\lambda_1,\ldots,\lambda_d)\otimes W_d(p)\cong
\sum W_d(\lambda_1+p_1,\ldots,\lambda_d+p_d),$$ where the sum runs on all nonnegative integers $p_1,\ldots,p_d$ such that $p_1+\cdots+p_d=p$ and $\lambda_i\geq \lambda_{i+1}+p_{i+1}$, $i=1,\ldots,d-1$, and its dual version $$\label{Young rule B}
W_d(\lambda_1,\ldots,\lambda_d)\otimes W_d(1^p)\cong
\sum W_d(\lambda_1+\varepsilon_1,\ldots,\lambda_d+\varepsilon_d),$$ where the sum is on all partitions $(\lambda_1+\varepsilon_1,\ldots,\lambda_d+\varepsilon_d)$ such that $\varepsilon_i=0,1$, $\varepsilon_1+\cdots+\varepsilon_d=p$.
In the $q$-th symmetric tensor power $$W^{\otimes_sq}=\underbrace{W\otimes_s\cdots\otimes_sW}_{q\text{ \rm times}}$$ of the $GL_d$-module $W$, we identify the tensors $w_{\sigma(1)}\otimes\cdots\otimes w_{\sigma(q)}$ and $w_1\otimes\cdots\otimes w_q$, $\sigma\in S_q$. If $W=W_1\oplus\cdots \oplus W_k$, then $$\label{symmetric powers}
W^{\otimes_sq}
=\bigoplus W_1^{\otimes_sq_1}\otimes\cdots\otimes W_k^{\otimes_sq_k},\quad q_1+\cdots+q_k=q.$$ There is no general combinatorial rule for the decomposition of the symmetric tensor powers of $W_d(\lambda)$. We shall need the following partial results due to Thrall [@Th], see also [@M]: $$\label{first Thrall}
K[W_d(2)]=\sum_{q\geq 0}W_d(2)^{\otimes_sq}
=\sum W_d(2\lambda_1,2\lambda_2,\ldots,2\lambda_d),$$ where the sum is over all partitions $\lambda$, $$\label{second Thrall}
W_d(p)\otimes_sW_d(p)=\sum_{0\leq k\leq p/2}W_d(p+2k,p-2k),$$ $$\label{third Thrall}
W_d(1^p)\otimes_sW_d(1^p)=\sum_{0\leq k\leq p/2}W_d(2^{p-2k},1^{4k}).$$ Besides, we shall need the decomposition $$\label{symmetric square of W22}
W_3(2^2)\otimes_sW(2^2)=W_3(4^2)\oplus W_3(4,2^2).$$ The easiest way to check (\[symmetric square of W22\]) is to use (\[usage of LRrule\]), hence $$W_3(2^2)\otimes_sW(2^2)\subset W_3(4^2)\oplus W_3(4,3,1)\oplus W_3(4,2^2).$$ Therefore $$W_3(2^2)\otimes_sW(2^2)= \varepsilon_1W_3(4^2)\oplus
\varepsilon_2W_3(4,3,1)\oplus \varepsilon_3W_3(4,2^2),$$ $\varepsilon_i=0,1$. The Hilbert series of $W_3(2^2)\otimes_sW_3(2^2)$ contains the summand $t_1^4t_2^4$ and $S_{(4^2)}$ is the only Schur function among $S_{(4^2)},S_{(4,3,1)},S_{(4,2^2)}$ which contains $t_1^4t_2^4$. This implies that $W_3(4^2)$ participates in the decomposition of $W_3(2^2)\otimes_sW_3(2^2)$ and $\varepsilon_1=1$. Finally, we apply dimension arguments: $$\text{\rm dim}(W_3(2^2)\otimes_sW(2^2))=\text{\rm dim}(W_3(4^2))+
\varepsilon_2\text{\rm dim}(W_3(4,3,1))+\varepsilon_3\text{\rm dim}(W_3(4,2^2)).$$ Since $$\text{\rm dim}(W_3(2^2))=6,\quad
\text{\rm dim}(W_3(2^2)\otimes_sW(2^2))=\binom{6+1}{2}=21,$$ $$\text{\rm dim}(W_3(4^2))=\text{\rm dim}(W_3(4,3,1))=15,\quad
\text{\rm dim}(W_3(4,2^2))=6,$$ we obtain the only possibility $\varepsilon_2=0$, $\varepsilon_3=1$.
The action of $GL_d$ on $K\langle x_1,\ldots,x_d\rangle$ is inherited by the algebras $C_{3d}$ and $C_0$. Now we discuss the approach of Abeasis and Pittaluga [@AP] for the special case $n=3$. (Pay attention that the partitions in [@AP] are given in “Francophone” way, i.e., transposed to ours.) The algebra $C_{3d}$ has a system of generators of degree $\leq 6$. Without loss of generality we may assume that this system consists of traces of products $\text{\rm tr}(X_{i_1}\cdots X_{i_k})$. Let $U_k$ be the subalgebra of $C_{3d}$ generated by all traces $\text{\rm tr}(X_{i_1}\cdots X_{i_l})$ of degree $l\leq k$. Clearly, $U_k$ is also a $GL_d$-submodule of $C_{3d}$. Let $C_{3d}^{(k+1)}$ be the homogeneous component of degree $k+1$ of $C_{3d}$. Then the intersection $U_k\cap C_{3d}^{(k+1)}$ is a $GL_d$-module and has a complement $G_{k+1}$ in $C_{3d}^{(k+1)}$, which is the $GL_d$-module of the “new” generators of degree $k+1$. We may assume that $G_{k+1}$ is a submodule of the $GL_d$-module spanned by traces of products $\text{\rm tr}(X_{i_1}\cdots X_{i_{k+1}})$ of degree $k+1$. The $GL_d$-module of the generators of $C_{3d}$ is $$G=G_1\oplus G_2\oplus \cdots\oplus G_6.$$
\[generating module of C\] [(Abeasis and Pittaluga [@AP])]{} The $GL_d$-module $G$ of the generators of $C_{3d}$ decomposes as $$G=W(1)\oplus W(2)\oplus W(3)\oplus W(1^3)\oplus W(2^2)\oplus
W(2,1^2)$$ $$\oplus W(3,1^2)\oplus W(2^2,1)\oplus W(1^5)\oplus
W(3^2)\oplus W(3,1^3).$$ Each module $W(\lambda)\subset G$ is generated by the “canonical” highest weight vector $\text{\rm tr}(w_{\lambda}(X_1,\ldots,X_d))$, where $w_{\lambda}$ is given in [(\[canonical hwv\])]{}.
\[dimensions of G\] The numbers $g_k$ of generators of degree $k\leq 6$ in any homogeneous minimal system of generators of $C_{3d}$ are $$g_1=d,g_2=\frac{1}{2}(d+1)d,g_3=\frac{1}{3}d(d^2+2),
g_4=\frac{1}{24}(d+1)d(d-1)(5d-6),$$ $$g_5=\frac{1}{30}d(d-1)(d-2)(3d^2+4d+6),
g_6=\frac{1}{48}(d+2)(d+1)d(d-1)(d^2-3d+4).$$ The total number of generators is $$g=g(d)=\frac{1}{240}d(5d^5+19d^4-5d^3+65d^2+636)$$
The number $g_k$ is equal to the dimension of the $GL_d$-submodule $G_k$ of the $GL_d$-module $G$ of generators of $C_{3d}$. Applying the formula (\[dim of W\]) we obtain that the dimensions of the $GL_d$-modules $$W(1),W(2),W(3),W(1^3),W(2^2),W(2,1^2),$$ $$W(3,1^2),W(2^2,1),W(1^5),W(3^2),W(3,1^3).$$ are, respectively, $$d,\binom{d+1}{2},\binom{d+2}{3},\binom{d}{3},
\frac{d}{2}\binom{d+1}{3},3\binom{d+1}{4},$$ $$6\binom{d+2}{5},d\binom{d+1}{4},
\binom{d}{5},
3\binom{d+2}{4}\binom{d+1}{2},10\binom{d+2}{6}.$$ This easily implies the results, because $$\text{\rm dim}(G_3)=\text{\rm dim}(W(3))+\text{\rm dim}(W(1^3)),
\text{\rm dim}(G_4)=\text{\rm dim}(W(2^2))+\text{\rm dim}(W(2,1^2)),$$ $$\text{\rm dim}(G_5)=\text{\rm dim}(W(3,1^2))+\text{\rm dim}(W(2^2,1))
+\text{\rm dim}(W(1^5)),$$ $$\text{\rm dim}(G_6)=\text{\rm dim}(W(3^2))+\text{\rm dim}(W(3,1^3)),$$ and $g=g_1+g_2+\cdots+g_6$.
In the sequel we shall need the Hilbert series of $C_{33}$ calculated by Berele and Stembridge [@BS]: $$H(C_{33},t_1,t_2,t_3)= \frac{p(t_1,t_2,t_3)}{q(t_1,t_2,t_3)},$$ where $$p=1-e_2+e_3+e_1e_3+e_2^2+e_1^2e_3-e_2e_3-2e_1e_2e_3+e_3^2+e_2^2e_3$$ $$-e_1^2e_2e_3+2e_1^2e_3^2+e_1^3e_3^2+e_2^2e_3^2-e_1^2e_2e_3^2-e_1e_3^3-2e_1e_2^2e_3^2$$ $$+2e_2e_2e_3^3-e_2^3e_3^2+e_1^3e_3^3+2e_1^2e_2e_3^3-2e_1e_3^4-e_1^2e_3^4+e_1e_2^2e_3^3$$ $$+e_2e_3^4-e_2^3e_3^3-2e_2^2e_3^5+e_1e_2^2e_3^4+2e_1e_2e_3^5-e_3^6-e_2^2e_3^5$$ $$+e_1e_3^6-e_2e_3^6-e_1^2e_3^6-e_3^7+e_1e_3^7-e_3^8,$$ $$q=\!\!\left(\prod_{i=1}^3(1-t_i)(1-t_i^2)(1-t_i^3)\right)\!\!\!
\left(\prod_{1\leq i<j\leq
3}(1-t_it_j)^2(1-t_i^2t_j)(1-t_it_j^2)\right)\!\! (1-t_1t_2t_3),$$ and $$e_1=t_1+t_2+t_3,\quad e_2=t_1t_2+t_1t_3+t_2t_3,\quad e_3=t_1t_2t_3$$ are the elementary symmetric polynomials in three variables. Since the Hilbert series of the tensor product is equal to the product of the Hilbert series of the factors, and $$H(K[\text{\rm \rm tr}(X_1),\ldots,\text{\rm \rm tr}(X_d)],t_1,\ldots,t_d)=
\frac{1}{(1-t_1)\cdots(1-t_d)},$$ (\[replacing with traceless matrices\]) implies that $$H(C_{33},t_1,t_2)=\frac{H(C_0,t_1,t_2,t_3)}{(1-t_1)(1-t_2)(1-t_3)}.$$ In this way, for $d=3$, $$\label{Hilbert series of C0}
H(C_0,t_1,t_2,t_3)=
(1-t_1)(1-t_2)(1-t_3)H(C_{33},t_1,t_2).$$
The symmetric algebra of the generators
=======================================
We consider the symmetric algebra $$S=K[G_2\oplus \cdots\oplus G_6]$$ of the $GL_d$-module of the generators of the algebra $C_0$. Clearly, the grading and the $GL_d$-module structure of $G_2\oplus \cdots\oplus G_6$ induce a grading and the structure of a $GL_d$-module also on $S$. The defining relations of the algebra $C_0$ are in the square of the augmentation ideal $\omega(S)$ of $S$. Since we are interested in the defining relations of degree 7 for $C_0$ for any $d\geq 3$ and of degree 8 for $d=3$, we shall decompose the homogeneous components of degree 7, respectively, 8 of the ideal $\omega^2(S)$ into a sum of irreducible $GL_d$-, respectively, $GL_3$-modules. Then we shall find explicit generators of those irreducible components which may give rise to relations.
\[multiplicities of degree 7\] The following $GL_d$-module isomorphisms hold: $$\label{W2xW2}
W(2)\otimes_sW(2)\cong W(4)\oplus W(2^2),$$ $$\label{W31xW2}
W(3)\otimes W(2)\cong W(5)\oplus W(4,1)\oplus W(3,2),$$ $$\label{W32xW2}
W(1^3)\otimes W(2)\cong W(3,1^2)\oplus W(2,1^3),$$ $$\label{W41xW2}
W(2^2)\otimes W(2)\cong W(4,2)\oplus W(3,2,1)\oplus W(2^3),$$ $$\label{W42xW2}
W(2,1^2)\otimes W(2)\cong W(4,1^2)\oplus W(3,2,1)\oplus W(3,1^3)
\oplus W(2^2,1^2),$$ $$\label{W31xW31}
W(3)\otimes_sW(3)\cong W(6)\oplus W(4,2),$$ $$\label{W31xW32}
W(3)\otimes W(1^3)\cong W(4,1^2)\oplus W(3,1^3),$$ $$\label{W32xW32}
W(1^3)\otimes_sW(1^3)\cong W(2^3)\oplus W(2,1^4),$$ $$\label{W2xW2xW2}
W(2)\otimes_sW(2)\otimes_sW(2)\cong W(6)\oplus W(4,2)\oplus W(2^3),$$ $$\label{W51xW2}
W(3,1^2)\otimes W(2)\cong W(5,1^2)\oplus W(4,2,1)\oplus W(4,1^3)
\oplus W(3^2,1)\oplus W(3,2,1^2),$$ $$\label{W52xW2}
W(2^2,1)\otimes W(2)\cong W(4,2,1)\oplus W(3,2^2)\oplus W(3,2,1^2)
\oplus W(2^3,1),$$ $$\label{W53xW2}
W(1^5)\otimes W(2)\cong W(3,1^4)\oplus W(2,1^5),$$ $$\label{W41xW31}
W(2^2)\otimes W(3)\cong W(5,2)\oplus W(4,2,1)\oplus W(3,2^2),$$ $$\label{W41xW32}
W(2^2)\otimes W(1^3)\cong W(3^2,1)\oplus W(3,2,1^2)\oplus W(2^2,1^3),$$ $$\label{W42xW31}
W(2,1^2)\otimes W(3)\cong W(5,1^2)\oplus W(4,2,1)\oplus W(4,1^3)
\oplus W(3,2,1^2),$$ $$\label{W42xW32}
W(2,1^2)\otimes W(1^3)\cong W(3,2^2)\oplus W(3,2,1^2)$$ $$\nonumber
\oplus W(3,1^4) \oplus W(2^3,1)\oplus W(2^2,1^3)\oplus W(2,1^5),$$ $$\label{W31xW2xW2}
W(3)\otimes(W(2)\otimes_sW(2))\cong W(7)\oplus W(6,1)$$ $$\nonumber
\oplus 2W(5,2) \oplus W(4,3)\oplus W(4,2,1)\oplus W(3,2^2),$$ $$\label{W32xW2xW2}
W(1^3)\otimes(W(2)\otimes_sW(2))\cong W(5,1^2)\oplus W(4,1^3)\oplus W(3^2,1)
\oplus W(2^2,1^3).$$
Equations (\[W31xW31\]), (\[W32xW32\]) are partial cases of (\[first Thrall\]), (\[second Thrall\]), respectively; (\[W2xW2\]) and (\[W2xW2xW2\]) follow from (\[third Thrall\]). Equations (\[W31xW2\]), (\[W32xW2\]), (\[W41xW2\]), (\[W42xW2\]), (\[W31xW32\]), (\[W51xW2\]), (\[W52xW2\]), (\[W53xW2\]), (\[W41xW31\]), and (\[W42xW31\]) are obtained from (\[Young rule A\]); (\[W41xW32\]) and (\[W42xW32\]) follow from (\[Young rule B\]). Finally, (\[W31xW2xW2\]) and (\[W32xW2xW2\]) are calculated applying, respectively, (\[Young rule A\]) and (\[Young rule B\]) to (\[W2xW2\]).
\[decomposition of low degree\] The homogeneous components $(\omega^2(S))^{(k)}$ of degree $k\leq 7$ of the square $\omega^2(S)$ of the augmentation ideal of the symmetric algebra of $G_2\oplus\cdots\oplus G_6$ decomposes as $$(\omega^2(S))^{(4)}=W(4)\oplus W(2^2),$$ $$(\omega^2(S))^{(5)}=W(5)\oplus W(4,1)\oplus W(3,2)\oplus W(3,1^2)\oplus W(2,1^3),$$ $$(\omega^2(S))^{(6)}=2W(6)\oplus 3W(4,2)\oplus 2W(4,1^2)$$ $$\oplus 2W(3,2,1)
\oplus 2W(3,1^3)\oplus 3W(2^3)\oplus W(2^2,1^2)\oplus W(2,1^4),$$ $$(\omega^2(S))^{(7)}=W(7)\oplus W(6,1)\oplus 3W(5,2)\oplus 3W(5,1^2)$$ $$\oplus W(4,3)\oplus 5W(4,2,1)
\oplus 3W(4,1^3)\oplus 3W(3^2,1)\oplus 4W(3,2^2)$$ $$\oplus 6W(3,2,1^2)\oplus 2W(3,1^4)\oplus 2W(2^3,1)\oplus 3W(2^2,1^3)\oplus 2W(2,1^5).$$
By (\[symmetric powers\]), the homogeneous component of degree $k$ of $\omega^2(S)$ has the form $$(\omega^2(S))^{(k)}
=\bigoplus G_2^{\otimes_sq_2}\otimes\cdots\otimes G_6^{\otimes_sq_6},\quad 2q_2+\cdots+6q_6=k, q_2+\cdots+q_6\geq 2.$$ The decomposition of $G_2,\ldots,G_6$ is given in Proposition \[generating module of C\]. Again, the equality (\[symmetric powers\]) implies that $(\omega^2(S))^{(k)}=0$ for $k<4$. For $k=4$ we use (\[W2xW2\]) and obtain $$(\omega^2(S))^{(4)}= G_2\otimes_sG_2=W(2)\otimes_sW(2)=W(4)\oplus W(2^2).$$ For $k=5$, we derive $$(\omega^2(S))^{(5)}= G_3\otimes G_2=(W(3)\oplus W(1^3))\otimes W(2)$$ and the decomposition follows from (\[W31xW2\]) and (\[W32xW2\]). For $k=6$ we have $$(\omega^2(S))^{(6)}=G_4\otimes G_2\oplus G_3\otimes_sG_3
\oplus G_2\otimes_s G_2\otimes_sG_2$$ and we use the decompositions given in (\[W41xW2\]) – (\[W2xW2xW2\]). The decomposition of $(\omega^2(S))^{(7)}$ is obtained in a similar way and makes use of (\[W51xW2\]) – (\[W32xW2xW2\]).
\[hwv of degree up to 6\] The following elements of $S=K[G_2\oplus\cdots\oplus G_6]$ are highest weight vectors:
For $\lambda=(2,1^3)$: $$w=\sum_{\sigma\in S_4}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_{\sigma(1)},x_{\sigma(2)},x_{\sigma(3)}))
\text{\rm \rm tr}(x_{\sigma(4)}x_1);$$ For $\lambda=(3,1^3)$: $$w_1=\sum_{\sigma\in S_4}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_{\sigma(1)},x_{\sigma(2)},x_{\sigma(3)})x_1
+s_3(x_1,x_{\sigma(1)},x_{\sigma(2)})x_{\sigma(3)})
\text{\rm tr}(x_{\sigma(4)}x_1),$$ $$w_2=\sum_{\sigma\in S_4}\text{\rm sign}(\sigma)
\text{\rm tr}(x_1^2x_{\sigma(1)})
\text{\rm tr}(s_3(x_{\sigma(2)},x_{\sigma(3)},x_{\sigma(4)})),$$ For $\lambda=(2^2,1^2)$: $$w=\sum_{\sigma\in S_4,\tau\in S_2}\text{\rm sign}(\sigma\tau)
\text{\rm tr}(s_3(x_{\sigma(1)},x_{\sigma(2)},x_{\sigma(3)})x_{\tau(1)}$$ $$+s_3(x_{\tau(1)},x_{\sigma(1)},x_{\sigma(2)})x_{\sigma(3)})
\text{\rm tr}(x_{\sigma(4)}x_{\tau(2)}),$$ For $\lambda=(2,1^4)$: $$w=\sum_{\sigma\in S_5}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_{\sigma(1)},x_{\sigma(2)},x_{\sigma(3)}))
\text{\rm tr}(s_3(x_{\sigma(4)},x_{\sigma(5)},x_1)).$$ For $\lambda=(2,1^3), (2^2,1^2), (2,1^4)$, every highest weight vector $w\in W(\lambda)\subset \omega^2(S)$ is equal, up to a multiplicative constant, to the corresponding $w$. For $\lambda=(3,1^3)$ the highest weight vectors are linear combinations of $w_1$ and $w_2$.
By Proposition \[generating module of C\], the submodules $W(2), W(3),W(1^3),W(2,1^2)$ of $G_2\oplus\cdots\oplus G_6$ are generated, respectively, by $$v_1=\text{\rm tr}(x_1^2),v_2=\text{\rm tr}(x_1^3),v_3=\text{\rm tr}(s_3(x_1,x_2,x_3)),
v_4=\text{\rm tr}(s_3(x_1,x_2,x_3)x_1).$$ The trace polynomials $\text{\rm tr}(x_1x_2)$ and $\text{\rm tr}(s_3(x_1,x_2,x_3)x_4+s_3(x_2,x_3,x_4)x_1)$ are linearizations of $v_1$ and $v_4$, respectively. Hence the trace polynomials $w,w_1,w_2$ defined in the proposition belong to the ideal $\omega^2(S)$. They have the necessary skew-symmetries, as the “canonical” highest weight vectors $w_{\lambda}$ from (\[canonical hwv\]). Hence all $w$ and $w_1,w_2$ are highest weight vectors. Clearly, they are nonzero in $S$ and their number coincides with the multiplicities of $W(\lambda)$ in the decomposition of $\omega^2(S)$. Hence, every highest weight vector $w\in W(\lambda)\subset \omega^2(S)$ can be expressed as their linear combination.
\[hwv of degree 7\] The following elements of $S=K[G_2\oplus\cdots\oplus G_6]$ are highest weight vectors:
For $\lambda=(4,1^3)$: $$w_1=(\text{\rm tr}(s_3(x_1,x_2,x_3)(x_1x_4+x_4x_1))-\text{\rm tr}(s_3(x_1,x_2,x_4)(x_1x_3+x_3x_1))$$ $$+\text{\rm tr}(s_3(x_1,x_3,x_4)(x_1x_2+x_2x_1))
+3\text{\rm tr}(s_3(x_2,x_3,x_4)x_1^2))\text{\rm tr}(x_1^2)$$ $$+5(-\text{\rm tr}(s_3(x_1,x_2,x_3)x_1^2)\text{\rm tr}(x_1x_4)$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_4)x_1^2)\text{\rm tr}(x_1x_3)
-\text{\rm tr}(s_3(x_1,x_3,x_4)x_1^2)\text{\rm tr}(x_1x_2)),$$ $$w_2=(\text{\rm tr}(s_3(x_1,x_2,x_3)x_4)-\text{\rm tr}(s_3(x_1,x_2,x_4)x_3)+\text{\rm tr}(s_3(x_1,x_3,x_4)x_2)$$ $$+3\text{\rm tr}(s_3(x_2,x_3,x_4)x_1))\text{\rm tr}(x_1^3)
+4(-\text{\rm tr}(s_3(x_1,x_2,x_3)x_1)\text{\rm tr}(x_1^2x_4)$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_4)x_1)\text{\rm tr}(x_1^2x_3)
-\text{\rm tr}(s_3(x_1,x_3,x_4)x_1)\text{\rm tr}(x_1^2x_2)),$$ $$w_3=(\text{\rm tr}(s_3(x_2,x_3,x_4)))\text{\rm tr}(x_1^2)-\text{\rm tr}(s_3(x_1,x_3,x_4)))\text{\rm tr}(x_1x_2)$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_4)))\text{\rm tr}(x_1x_3)
-\text{\rm tr}(s_3(x_1,x_2,x_3)))\text{\rm tr}(x_1x_4))\text{\rm tr}(x_1^2).$$
For $\lambda=(3,2^2)$: $$w_1=\sum_{\sigma\in S_3}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_1,x_2,x_3)x_{\sigma(1)}x_{\sigma(2)})\text{\rm tr}(x_1x_{\sigma(3)}),$$ $$w_2=\text{\rm tr}(s_3(x_1,x_2,x_3)x_1)\text{\rm tr}(s_3(x_1,x_2,x_3)),$$ $$w_3=\text{\rm tr}([x_1,x_2]^2)\text{\rm tr}(x_1x_3^2)+\text{\rm tr}([x_1,x_3]^2)\text{\rm tr}(x_1x_2^2)
+\text{\rm tr}([x_2,x_3]^2)\text{\rm tr}(x_1^3)$$ $$-\text{\rm tr}([x_1,x_2][x_1,x_3])\text{\rm tr}(x_1(x_2x_3+x_3x_2))$$ $$+2\text{\rm tr}([x_1,x_2][x_2,x_3])\text{\rm tr}(x_1^2x_3)
-2\text{\rm tr}([x_1,x_3][x_2,x_3])\text{\rm tr}(x_1^2x_2),$$ $$w_4=\text{\rm tr}(x_1^3)(\text{\rm tr}(x_2^2)\text{\rm tr}(x_3^2)-\text{\rm tr}^2(x_2x_3))
+\text{\rm tr}(x_1x_2^2)(\text{\rm tr}(x_1^2)\text{\rm tr}(x_3^2)-\text{\rm tr}^2(x_1x_3))$$ $$+\text{\rm tr}(x_1x_3^2)(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)-\text{\rm tr}^2(x_1x_2))$$ $$+2\text{\rm tr}(x_1^2x_2)(-\text{\rm tr}(x_1x_2)\text{\rm tr}(x_3^2)+\text{\rm tr}(x_1x_3)\text{\rm tr}(x_2x_3))$$ $$+2\text{\rm tr}(x_1^2x_3)(-\text{\rm tr}(x_1x_3)\text{\rm tr}(x_2^2)+\text{\rm tr}(x_1x_2)\text{\rm tr}(x_2x_3))$$ $$+\text{\rm tr}(x_1(x_2x_3+x_3x_2))(-\text{\rm tr}(x_1^2)\text{\rm tr}(x_2x_3)
+\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3)).$$
For $\lambda=(3,2,1^2)$: $$w_1=(\text{\rm tr}(s_3(x_1,x_2,x_3)(x_2x_4+x_4x_2)) -\text{\rm tr}((s_3(x_1,x_2,x_4)(x_2x_3+x_3x_2))$$ $$+4\text{\rm tr}(s_3(x_2,x_3,x_4)(x_1x_2+x_2x_1))+ 2\text{\rm tr}(s_3(x_1,x_3,x_4)x_2^2))\text{\rm tr}(x_1^2)$$ $$+(-\text{\rm tr}(s_3(x_1,x_2,x_3)(x_1x_4+x_4x_1)) +\text{\rm tr}(s_3(x_1,x_2,x_4)(x_1x_3+x_3x_1))$$ $$-6\text{\rm tr}(s_3(x_1,x_3,x_4)(x_1x_2+x_2x_1)) -8\text{\rm tr}(s_3(x_2,x_3,x_4)x_1^2))\text{\rm tr}(x_1x_2)$$ $$+5(-\text{\rm tr}(s_3(x_1,x_2,x_3)(x_1x_2+x_2x_1))\text{\rm tr}(x_1x_4)$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_4)(x_1x_2+x_2x_1))\text{\rm tr}(x_1x_3))$$ $$+10(\text{\rm tr}(s_3(x_1,x_2,x_3)x_1^2)\text{\rm tr}(x_2x_4)
-\text{\rm tr}(s_3(x_1,x_2,x_4)x_1^2)\text{\rm tr}(x_2x_3)$$ $$+\text{\rm tr}(s_3(x_1,x_3,x_4)x_1^2)\text{\rm tr}(x_2^2)),$$ $$w_2=\text{\rm tr}(x_1^2)(-\text{\rm tr}(s_3(x_1,x_2,x_3)[x_2,x_4])
-\text{\rm tr}(s_3(x_2,x_3,x_4)[x_1,x_2])$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_4)[x_2,x_3])
+ 3\text{\rm tr}(s_3(x_2,x_3,x_4)[x_1,x_2]))$$ $$+\text{\rm tr}(x_1x_2)(\text{\rm tr}(s_3(x_1,x_2,x_3)[x_1,x_4])
- \text{\rm tr}(s_3(x_1,x_3,x_4)[x_1,x_2])$$ $$-\text{\rm tr}(s_3(x_1,x_2,x_4)[x_1,x_3])-\text{\rm tr}(s_3(x_1,x_3,x_4)[x_1,x_2]))$$ $$+ 3\text{\rm tr}(x_1x_3)\text{\rm tr}(s_3(x_1,x_2,x_4)[x_1,x_2])
- 3\text{\rm tr}(x_1x_4)\text{\rm tr}(s_3(x_1,x_2,x_3)[x_1,x_2]),$$ $$w_3=\text{\rm tr}([x_1,x_2]^2)\text{\rm tr}(s_3(x_1,x_3,x_4))
- \text{\rm tr}([x_1,x_2][x_1,x_3])\text{\rm tr}(s_3(x_1,x_2,x_4))$$ $$+ \text{\rm tr}([x_1,x_2][x_1,x_4])\text{\rm tr}(s_3(x_1,x_2,x_3)),$$ $$w_4=-2\text{\rm tr}(s_3(x_2,x_3,x_4)x_2)\text{\rm tr}(x_1^3)
+ 2(\text{\rm tr}(s_3(x_1,x_3,x_4)x_2)$$ $$+ \text{\rm tr}(s_3(x_2,x_3,x_4)x_1))\text{\rm tr}(x_1^2x_2)
- 2\text{\rm tr}(s_3(x_1,x_2,x_4)x_2)\text{\rm tr}(x_1^2x_3)$$ $$+2\text{\rm tr}(s_3(x_1,x_2,x_3)x_2)\text{\rm tr}(x_1^2x_4)
- 2\text{\rm tr}(s_3(x_1,x_3,x_4)x_1)\text{\rm tr}(x_1x_2^2)$$ $$+ \text{\rm tr}(s_3(x_1,x_2,x_4)x_1)\text{\rm tr}(x_1(x_2x_3+x_3x_2))
- \text{\rm tr}(s_3(x_1,x_2,x_3)x_1)\text{\rm tr}(x_1(x_2x_4+x_4x_2)),$$ $$w_5=\text{\rm tr}(s_3(x_1,x_2,x_3)x_1)\text{\rm tr}(s_3(x_1,x_2,x_4))
- \text{\rm tr}(s_3(x_1,x_2,x_4)x_1)\text{\rm tr}(s_3(x_1,x_2,x_3)),$$ $$w_6=(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)-\text{\rm tr}(x_1x_2)^2)\text{\rm tr}(s_3(x_1,x_3,x_4))$$ $$+ (-\text{\rm tr}(x_1^2)\text{\rm tr}(x_2x_3)
+\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3))\text{\rm tr}(s_3(x_1,x_2,x_4))$$ $$+ (\text{\rm tr}(x_1^2)\text{\rm tr}(x_2x_4)
-\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_4))\text{\rm tr}(s_3(x_1,x_2,x_3)).$$
For $\lambda=(3,1^4)$: $$w_1=\text{\rm tr}(s_5(x_1,x_2,x_3,x_4,x_5))\text{\rm tr}(x_1^2),$$ $$w_2=\sum_{\sigma\in S_5}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_1,x_{\sigma(2)},x_{\sigma(3)})x_1)\text{\rm tr}(s_3(x_1,x_{\sigma(4)},x_{\sigma(5)})),
\quad \sigma(1)=1.$$
For $\lambda=(2^3,1)$: $$w_1=\text{\rm tr}(s_3(x_2,x_3,x_4)[x_2,x_3])\text{\rm tr}(x_1^2)$$ $$-(\text{\rm tr}(s_3(x_1,x_3,x_4)[x_2,x_3]) +\text{\rm tr}(s_3(x_2,x_3,x_4)[x_1,x_3]))\text{\rm tr}(x_1x_2)$$ $$+(\text{\rm tr}(s_3(x_1,x_2,x_4)[x_2,x_3]) +\text{\rm tr}(s_3(x_2,x_3,x_4)[x_1,x_2]))\text{\rm tr}(x_1x_3)$$ $$-\text{\rm tr}(s_3(x_1,x_2,x_3)[x_2,x_3])\text{\rm tr}(x_1x_4)
+\text{\rm tr}(s_3(x_1,x_3,x_4)[x_1,x_3])\text{\rm tr}(x_2^2)$$ $$-(\text{\rm tr}(s_3(x_1,x_2,x_4)[x_1,x_3]) + \text{\rm tr}(s_3(x_1,x_3,x_4)[x_1,x_2]))\text{\rm tr}(x_2x_3)$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_3)[x_1,x_3])\text{\rm tr}(x_2x_4)
+\text{\rm tr}(s_3(x_1,x_2,x_4)[x_1,x_2])\text{\rm tr}(x_3^2)$$ $$-\text{\rm tr}(s_3(x_1,x_2,x_3)[x_1,x_2])\text{\rm tr}(x_3x_4),$$ $$w_2=(-3(\text{\rm tr}(s_3(x_1,x_2,x_3)x_4) + \text{\rm tr}(s_3(x_2,x_3,x_4)x_1))+\text{\rm tr}(s_3(x_1,x_3,x_4)x_2)$$ $$+ \text{\rm tr}(s_3(x_2,x_3,x_4)x_1))-\text{\rm tr}(s_3(x_1,x_2,x_4)x_3)$$ $$+ \text{\rm tr}(s_3(x_2,x_3,x_4)x_1)))\text{\rm tr}(s_3(x_1,x_2,x_3))$$ $$+4\text{\rm tr}(s_3(x_1,x_2,x_3)x_3)\text{\rm tr}(s_3(x_1,x_2,x_4))$$ $$-4\text{\rm tr}(s_3(x_1,x_2,x_3)x_2)\text{\rm tr}(s_3(x_1,x_3,x_4))$$ $$+4\text{\rm tr}(s_3(x_1,x_2,x_3)x_1)\text{\rm tr}(s_3(x_2,x_3,x_4)).$$
For $\lambda=(2^2,1^3)$: $$w_1=(\text{\rm tr}([x_1,x_4][x_2,x_5]) + \text{\rm tr}([x_1,x_5][x_2,x_4])$$ $$+2\text{\rm tr}([x_1,x_2][x_4,x_5]) - 2\text{\rm tr}([x_1,x_5][x_2,x_4]))\text{\rm tr}(s_3(x_1,x_2,x_3))$$ $$+ (-\text{\rm tr}([x_1,x_3][x_2,x_5]) - \text{\rm tr}([x_1,x_5][x_2,x_3])-2\text{\rm tr}([x_1,x_2][x_3,x_5])$$ $$+2\text{\rm tr}([x_1,x_5][x_2,x_3]))\text{\rm tr}(s_3(x_1,x_2,x_4))+ (\text{\rm tr}([x_1,x_3][x_2,x_4]) +$$ $$\text{\rm tr}([x_1,x_4][x_2,x_3])+2\text{\rm tr}([x_1,x_2][x_3,x_4]) -$$ $$2\text{\rm tr}([x_1,x_4][x_2,x_3]))\text{\rm tr}(s_3(x_1,x_2,x_5))$$ $$+ 3\text{\rm tr}([x_1,x_2][x_2,x_5])\text{\rm tr}(s_3(x_1,x_3,x_4))
- 3\text{\rm tr}([x_1,x_2][x_2,x_4])\text{\rm tr}(s_3(x_1,x_3,x_5))$$ $$+ 3\text{\rm tr}([x_1,x_2][x_2,x_3])\text{\rm tr}(s_3(x_1,x_4,x_5))
- 3\text{\rm tr}([x_1,x_2][x_1,x_5])\text{\rm tr}(s_3(x_2,x_3,x_4))$$ $$+ 3\text{\rm tr}([x_1,x_2][x_1,x_4])\text{\rm tr}(s_3(x_2,x_3,x_5))
- 3\text{\rm tr}([x_1,x_2][x_1,x_3])\text{\rm tr}(s_3(x_2,x_4,x_5))$$ $$+ 3\text{\rm tr}([x_1,x_2]^2)\text{\rm tr}(s_3(x_3,x_4,x_5)),$$ $$w_2=\sum_{\sigma\in S_5}\sum_{\tau\in S_2}\text{\rm sign}(\sigma\tau)
\text{\rm tr}(s_3(x_{\tau(1)},x_{\sigma(1)},x_{\sigma(2)}))
(\text{\rm tr}(s_3(x_{\sigma(3)},x_{\sigma(4)},x_{\sigma(5)})x_{\tau(2)})$$ $$+\sum_{\sigma\in S_5}\sum_{\tau\in S_2}\text{\rm sign}(\sigma\tau)
\text{\rm tr}(s_3(x_{\tau(1)},x_{\sigma(1)},x_{\sigma(2)}))
(\text{\rm tr}(s_3(x_{\tau(2)},x_{\sigma(3)},x_{\sigma(4)})x_{\sigma(5)}),$$ $$w_3=\text{\rm tr}(s_3(x_1,x_2,x_3))(\text{\rm tr}(x_1x_4)\text{\rm tr}(x_2x_5)
-\text{\rm tr}(x_1x_5)\text{\rm tr}(x_2x_4))$$ $$+ \text{\rm tr}(s_3(x_1,x_2,x_4))(-\text{\rm tr}(x_1x_3)\text{\rm tr}(x_2x_5)
+\text{\rm tr}(x_1x_5)\text{\rm tr}(x_2x_3))$$ $$+ \text{\rm tr}(s_3(x_1,x_2,x_5))(\text{\rm tr}(x_1x_3)\text{\rm tr}(x_2x_4)
-\text{\rm tr}(x_1x_4)\text{\rm tr}(x_2x_3))$$ $$+ \text{\rm tr}(s_3(x_1,x_3,x_4))(\text{\rm tr}(x_1x_2)\text{\rm tr}(x_2x_5)-\text{\rm tr}(x_1x_5)\text{\rm tr}(x_2^2))$$ $$+ \text{\rm tr}(s_3(x_1,x_3,x_5))(-\text{\rm tr}(x_1x_2)\text{\rm tr}(x_2x_4)
+ \text{\rm tr}(x_1x_4)\text{\rm tr}(x_2^2))$$ $$+ \text{\rm tr}(s_3(x_1,x_4,x_5))(\text{\rm tr}(x_1x_2)\text{\rm tr}(x_2x_3)
-\text{\rm tr}(x_1x_3)\text{\rm tr}(x_2^2))$$ $$+ \text{\rm tr}(s_3(x_2,x_3,x_4))(\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_5)
-\text{\rm tr}(x_2x_5)\text{\rm tr}(x_1^2))$$ $$+ \text{\rm tr}(s_3(x_2,x_3,x_5))(-\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_4)
+ \text{\rm tr}(x_2x_4)\text{\rm tr}(x_1^2))$$ $$+ \text{\rm tr}(s_3(x_2,x_4,x_5))(\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3)
-\text{\rm tr}(x_2x_3)\text{\rm tr}(x_1^2))$$ $$+ \text{\rm tr}(s_3(x_3,x_4,x_5))(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)-\text{\rm tr}(x_1x_2)^2).$$
For $\lambda=(2,1^5)$: $$w_1=\sum_{\sigma\in S_6}\text{\rm sign}(\sigma)
\text{\rm tr}(s_5(x_{\sigma(1)},x_{\sigma(2)},x_{\sigma(3)},x_{\sigma(4)},x_{\sigma(5)}))
\text{\rm tr}(x_1x_{\sigma(6)}),$$ $$w_2=\sum_{\sigma\in S_6}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_{\sigma(1)},x_{\sigma(2)},x_{\sigma(3)})x_1))
\text{\rm tr}(s_3(x_{\sigma(4)},x_{\sigma(5)},x_{\sigma(6)}))$$ $$+\sum_{\sigma\in S_6}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_1,x_{\sigma(1)},x_{\sigma(2)})x_{\sigma(3)}))
\text{\rm tr}(s_3(x_{\sigma(4)},x_{\sigma(5)},x_{\sigma(6)})).$$ For $\lambda=(4,1^3), (3,2^2), (3,2,1^2), (3,1^4), (2^3,1), (2^2,1^3),(2,1^5)$, every highest weight vector $w\in W(\lambda)\subset \omega^2(S)$ is equal to a linear combination of $w_i$.
The computations are similar to those in the proof of Proposition \[hwv of degree up to 6\]. For a fixed $\lambda\vdash 7$, the number of $w_i$ can be calculated from the equations (\[W51xW2\]) – (\[W32xW2xW2\]) and Proposition \[decomposition of low degree\]. The concrete form of $w_i$ is found using Lemmas \[criterion for hwv\] and \[basis of module\]. We shall demonstrate the process in one case only.
Let $\lambda=(3,2^2)$. Consider the tensor product $W(3)\otimes(W(2)\otimes_sW(2))\subset\omega^2(S)$. The equation (\[W31xW2xW2\]) gives that the module $W(3,2^2)$ participates with multiplicity 1. More precisely, applying (\[Young rule A\]) to (\[W2xW2\]), we see that $W(3,2^2)$ appears as a submodule of the component $W(3)\otimes W(2^2)$ of $W(3)\otimes(W(2)\otimes_sW(2))$. Since $W(3)\subset G_3$ and $W(2)=G_2$ are generated by $\text{\rm
tr}(x_1^3)$ and $\text{\rm tr}(x_1^2)$, respectively, Lemma \[basis of module\] gives that they have bases $$\{\text{\rm tr}(x_i^3),\text{\rm tr}(x_i^2x_j), i\not=j,\quad
\text{\rm tr}(x_i(x_jx_k+x_kx_j)), i<j<k\},$$ $$\{\text{\rm tr}(x_i^2), \text{\rm tr}(x_ix_j),\quad i\not=j\}.$$ The submodule $W(2^2)$ of $W(2)\otimes_sW(2)$ is generated by $$u(x_1,x_2)=\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)-\text{\rm tr}^2(x_1x_2).$$ (Direct verification shows that $u(x_1,x_2+x_1)=u(x_1,x_2)$ and we apply Lemma \[criterion for hwv\].) Its partial linearization in $x_2$ is $$v(x_1,x_2,x_3)=\text{\rm tr}(x_1^2)\text{\rm tr}(x_2x_3)
-\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3).$$ We are looking for an element $w=w(x_1,x_2,x_3)$ in $W(3)\otimes W(2^2)\subset W(3)\otimes(W(2)\otimes_sW(2))$ which is homogeneous of multidegree $(3,2^2)$ and satisfies the conditions $\Delta_{21}(w)=\Delta_{31}(w)=\Delta_{32}(w)=0$, where $\Delta_{ij}$ is the derivation from Lemma \[criterion for hwv\]. All such elements are of the form $$w=\zeta_1\text{\rm tr}(x_1^3)u(x_2,x_3)
+ \zeta_2\text{\rm tr}(x_1^2x_2)v(x_3,x_1,x_2)
+ \zeta_3\text{\rm tr}(x_1^2x_3)v(x_2,x_1,x_3)$$ $$+ \zeta_4\text{\rm tr}(x_1x_2^2)u(x_1,x_3)
+\zeta_5\text{\rm tr}(x_1(x_2x_3+x_3x_2))v(x_1,x_2,x_3)
+\zeta_6\text{\rm tr}(x_1x_3^2)u(x_1,x_2).$$ Direct verifications show that $$\Delta_{21}(w)=(2\zeta_1+\zeta_2)\text{\rm tr}(x_1^3)v(x_3,x_1,x_2)$$ $$+(\zeta_2+2\zeta_4)\text{\rm tr}(x_1^2x_2)u(x_1,x_3)
+(-\zeta_3+2\zeta_5)\text{\rm tr}(x_1^2x_3)v(x_1,x_2,x_3),$$ $$\Delta_{31}(w)=(2\zeta_1+\zeta_3)\text{\rm tr}(x_1^3)v(x_2,x_1,x_3)$$ $$+(-\zeta_2+2\zeta_5)\text{\rm tr}(x_1^2x_2)v(x_1,x_2,x_3)
+(\zeta_3+2\zeta_6)\text{\rm tr}(x_1^2x_3)u(x_1,x_2),$$ $$\Delta_{32}(w)=(-\zeta_2+\zeta_3)\text{\rm tr}(x_1^2x_2)v(x_2,x_1,x_3)$$ $$+2(\zeta_4+\zeta_5)\text{\rm tr}(x_1x_2^2)v(x_1,x_2,x_3)
+(\zeta_5+\zeta_6)\text{\rm tr}(x_1(x_2x_3+x_3x_2))u(x_1,x_2).$$ Hence we obtain the homogeneous linear system $$2\zeta_1+\zeta_2=\zeta_2+2\zeta_4=-\zeta_3+2\zeta_5=0,$$ $$2\zeta_1+\zeta_3=-\zeta_2+2\zeta_5=\zeta_3+2\zeta_6=0,$$ $$-\zeta_2+\zeta_3=+2(\zeta_4+\zeta_5)=\zeta_5+\zeta_6=0.$$ Up to a multiplicative constant, the only solution of the system is $$\zeta_1=\zeta_4=\zeta_6=1,\quad \zeta_2=\zeta_3=-2,\quad\zeta_5=-1,$$ which is equal to $w_4$. In practice, in most of the cases we have used a slightly different algorithm to determine $w_i$. We have considered $w_i$ with unknown coefficients. Then we have evaluated it in the trace algebra $C_{3d}$ instead of in the symmetric algebra $S$, in order to use the programs which we already had. Requiring that $$g_{kl}w(x_1,\dots,x_d)=w(x_1,\ldots,x_d), \quad 1\leq k<l\leq d,$$ we have obtained the possible candidates for $w_i$. Since the number of candidates has coincided with the number predicted in Proposition \[decomposition of low degree\], we have concluded that the $w_i$’s really are the needed highest weight vectors.
\[decomposition of degree 8\] For $d=3$, the homogeneous component $(\omega^2(S))^{(8)}$ of degree $8$ of the square $\omega^2(S)$ of the augmentation ideal of the symmetric algebra of $G_2\oplus\cdots\oplus G_6$ decomposes as $$(\omega^2(S))^{(8)}=2W_3(8)\oplus W_3(7,1)\oplus 4W_3(6,2)\oplus 3W_3(6,1^2)$$ $$\oplus 2W_3(5,3)\oplus 6W_3(5,2,1)\oplus 4W_3(4^2)\oplus 7W_3(4,3,1)\oplus 9W_3(4,2^2)\oplus 4W_3(3^2,2).$$
The considerations are similar to those in the proof of Proposition \[decomposition of low degree\] and involves the equtions $$W_3(3^2)\otimes W_3(2)\cong W_3(5,3)\oplus W_3(4,3,1)\oplus W_3(3^2,2),$$ $$W_3(3,1^2)\otimes W_3(3)\cong W_3(6,1^2)\oplus W_3(5,2,1)\oplus W_3(4,3,1),$$ $$W_3(3,1^2)\otimes W_3(1^3)\cong W_3(4,2^2),$$ $$W_3(2^2,1)\otimes W_3(3)\cong W_3(5,2,1)\oplus W_3(4,2^2),$$ $$W_3(2^2,1)\otimes W_3(1^3)\cong W_3(3^2,2),$$ $$W_3(2^2)\otimes_sW_3(2^2)\cong W_3(4^2)\oplus W_3(4,2^2),$$ $$W_3(2^2)\otimes W_3(2,1^2)\cong W_3(4,3,1)\oplus W_3(3^2,2),$$ $$W_3(2,1^2)\otimes_sW_3(2,1^2)\cong W_3(4,2^2),$$ $$W_3(2^2)\otimes(W_3(2)\otimes_sW_3(2))\cong W_3(6,2)$$ $$\oplus W_3(5,2,1)\oplus W_3(4^2)\oplus W_3(4,3,1)\oplus 2W_3(4,2^2),$$ $$W_3(2,1^2)\otimes(W_3(2)\otimes_sW_3(2))\cong W_3(6,1^2)\oplus W_3(5,2,1)\oplus W_3(4,3,1)\oplus W_3(3^2,2),$$ $$(W_3(3)\otimes_sW_3(3))\otimes W_3(2)\cong W_3(8)$$ $$\oplus W_3(7,1)\oplus 2W_3(6,2)\oplus W_3(5,3)\oplus W_3(5,2,1)$$ $$\oplus W_3(4^2)\oplus W_3(4,3,1)\oplus W_3(4,2^2),$$ $$(W_3(3)\otimes W_3(1^3))\otimes W_3(2)
\cong W_3(6,1^2)\oplus W_3(5,2,1) \oplus W_3(4,3,1),$$ $$(W_3(1^3)\otimes_sW_3(1^3))\otimes W_3(2)\cong W_3(4,2^2),$$ $$W_3(2)\otimes_sW_3(2)\otimes_sW_3(2)\otimes_sW_3(2)\cong
W_3(8)\oplus W_3(6,2)\oplus W_3(4^2)\oplus W_3(4,2^2).$$ The proof of these equations uses the Young rule (\[Young rule A\]) and (\[Young rule B\]), the formulas (\[first Thrall\]) – (\[third Thrall\]), (\[usage of LRrule\]), and (\[symmetric square of W22\]). For example, by (\[first Thrall\]), $W_3(2)\otimes_sW_3(2)\cong W_3(4)\oplus W_3(2^2)$. Hence $$W_3(2^2)\otimes(W_3(2)\otimes_sW_3(2))\cong W_3(2^2)\otimes(W_3(4)\oplus W_3(2^2)),$$ by (\[Young rule A\]) $$W_3(2^2)\otimes W_3(4)\cong W_3(6,2)
\oplus W_3(5,2,1)\oplus W_3(4,2^2),$$ by (\[usage of LRrule\]) $$W_3(2^2)\otimes W_3(2^2)\cong W_3(4^2)\oplus W_3(4,3,1)\oplus W_3(4,2^2),$$ and we obtain the decomposition for $W_3(2^2)\otimes(W_3(2)\otimes_sW_3(2))$. In few cases we use also that $$W_d(\lambda_1+1,\ldots,\lambda_d+1)
\cong W_d(1^d)\otimes W_d(\lambda_1,\ldots,\lambda_d).$$ For example, $$W_3(2,1^2)\otimes_sW_3(2,1^2)
\cong (W_3(1^3)\otimes W_3(1))\otimes_s(W_3(1^3)\otimes W_3(1))$$ $$=(W_3(1^3)\otimes W_3(1^3))\otimes(W_3(1)\otimes_sW_3(1))
\cong W_3(2^3)\otimes W_3(2)\cong W_3(4,2^2).$$
\[hwv of degree 8\] For $d=3$, the following elements of $S=K[G_2\oplus\cdots\oplus G_6]$ are highest weight vectors:
For $\lambda=(4,3,1)$: $$w_1=\sum_{\sigma\in S_3}\text{\rm sign}(\sigma)
\text{\rm tr}([x_1,x_2]^2x_{\sigma(1)}x_{\sigma(2)})\text{\rm tr}(x_1x_{\sigma(3)}),$$ $$w_2=-\text{\rm tr}(s_3(x_1,x_2,x_3)x_1^2)\text{\rm tr}(x_1x_2^2)$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_3)(x_1x_2+x_2x_1)\text{\rm tr}(x_1^2x_2)
-\text{\rm tr}(s_3(x_1,x_2,x_3)x_2^2)\text{\rm tr}(x_1^3),$$ $$w_3=\text{\rm tr}([x_1,x_2]^2)\text{\rm tr}(s_3(x_1,x_2,x_3)x_1),$$ $$w_4=\text{\rm tr}([x_1,x_2][x_1,x_3])(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)$$ $$-\text{\rm tr}^2(x_1x_2)) -\text{\rm tr}([x_1,x_2]^2)(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2x_3)
-\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3)),$$ $$w_5=\text{\rm tr}(s_3(x_1,x_2,x_3)x_1)(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)-\text{\rm tr}^2(x_1x_2)),$$ $$w_6=(\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_2^2x_3)
-\text{\rm tr}(x_1^2x_3)\text{\rm tr}(x_2^3))\text{\rm tr}(x_1^2) +(-\text{\rm tr}(x_1^3)\text{\rm tr}(x_2^2x_3)$$ $$-\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_1(x_2x_3+x_3x_2))
+3\text{\rm tr}(x_1^2x_3)\text{\rm tr}(x_1x_2^2))\text{\rm tr}(x_1x_2)$$ $$+ (\text{\rm tr}(x_1^3)\text{\rm tr}(x_2^3)-\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_1x_2^2))\text{\rm tr}(x_1x_3)+$$ $$(\text{\rm tr}(x_1^3)\text{\rm tr}(x_1(x_2x_3+x_3x_2))
-2\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_1^2x_3))\text{\rm tr}(x_2^2)$$ $$+(-2\text{\rm tr}(x_1^3)\text{\rm tr}(x_1x_2^2) +2\text{\rm tr}^2(x_1^2x_2))\text{\rm tr}(x_2x_3),$$ $$w_7=\text{\rm tr}(s_3(x_1,x_2,x_3))(\text{\rm tr}(x_1^3)\text{\rm tr}(x_2^2)
-2\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_1x_2)+\text{\rm tr}(x_1x_2^2)\text{\rm tr}(x_1^2)).$$
For $\lambda=(4,2^2)$: $$w_1=\text{\rm tr}(s_3(x_1,x_2,x_3)x_1^2)\text{\rm tr}(s_3(x_1,x_2,x_3)),$$ $$w_2=\sum_{\sigma\in S_3}\text{\rm sign}(\sigma)
\text{\rm tr}(s_3(x_1,x_2,x_3)x_{\sigma(1)}x_{\sigma(1)})\text{\rm tr}(x_1^2x_{\sigma(3)}),$$ $$w_3=\text{\rm tr}([x_1,x_2]^2)\text{\rm tr}([x_1,x_3]^2)-\text{\rm tr}^2([x_1,x_2][x_1,x_3]),$$ $$w_4=\text{\rm tr}^2(s_3(x_1,x_2,x_3)x_1),$$ $$w_5=\text{\rm tr}([x_2,x_3]^2)\text{\rm tr}^2(x_1^2)
-2\text{\rm tr}([x_1,x_3][x_2,x_3])\text{\rm tr}(x_1^2)\text{\rm tr}(x_1x_2)$$ $$+2\text{\rm tr}([x_1,x_2][x_2,x_3])\text{\rm tr}(x_1^2)\text{\rm tr}(x_1x_3)
+\text{\rm tr}([x_1,x_3]^2)\text{\rm tr}^2(x_1x_2)$$ $$-2\text{\rm tr}([x_1,x_2][x_1,x_3])\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3)
+\text{\rm tr}([x_1,x_2]^2)\text{\rm tr}^2(x_1x_3),$$ $$w_6=\text{\rm tr}([x_1,x_3]^2)(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)-\text{\rm tr}^2(x_1x_2))$$ $$-2\text{\rm tr}([x_1,x_2][x_1,x_3])*(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2x_3)
-\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3))$$ $$+\text{\rm tr}([x_1,x_2]^2)(\text{\rm tr}(x_1^2)\text{\rm tr}(x_3^2)-\text{\rm tr}^2(x_1x_3)),$$ $$w_7=(-4\text{\rm tr}(x_1x_2^2)\text{\rm tr}(x_1x_3^2)+\text{\rm tr}^2(x_1(x_2x_3+x_3x_2)))\text{\rm tr}(x_1^2)$$ $$+ 4(2\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_1x_3^2)
-\text{\rm tr}(x_1^2x_3)\text{\rm tr}(x_1(x_2x_3+x_3x_2)))\text{\rm tr}(x_1x_2)$$ $$+4(2\text{\rm tr}(x_1^2x_3)\text{\rm tr}(x_1x_2^2)
-\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_1(x_2x_3+x_3x_2)))\text{\rm tr}(x_1x_3)$$ $$+4(-\text{\rm tr}(x_1^3)\text{\rm tr}(x_1x_3^2)+\text{\rm tr}^2(x_1^2x_3))\text{\rm tr}(x_2^2)$$ $$+4(\text{\rm tr}(x_1^3)\text{\rm tr}(x_1(x_2x_3+x_3x_2))
-2\text{\rm tr}(x_1^2x_2)\text{\rm tr}(x_1^2x_3))\text{\rm tr}(x_2x_3)$$ $$+4(\text{\rm tr}(x_1^3)\text{\rm tr}(x_1x_2^2)+\text{\rm tr}^2(x_1^2x_2))\text{\rm tr}(x_3^2),$$ $$w_8=\text{\rm tr}^2(s_3(x_1,x_2,x_3))\text{\rm tr}(x_1^2),$$ $$w_9=(\text{\rm tr}(x_1^2)(\text{\rm tr}(x_2^2)\text{\rm tr}(x_3^2)-\text{\rm tr}^2(x_2x_3))$$ $$-\text{\rm tr}^2(x_1x_2)\text{\rm tr}(x_3^2)+2\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3)\text{\rm tr}(x_2x_3)
-\text{\rm tr}^2(x_1x_3)\text{\rm tr}(x_2^2))\text{\rm tr}(x_1^2).$$
For $\lambda=(3^2,2)$: $$w_1=-\text{\rm tr}([x_2,x_3]^2[x_1,x_2])\text{\rm tr}(x_1^2)$$ $$+\text{\rm tr}([x_1,x_2]([x_1,x_3][x_2,x_3]+[x_2,x_3][x_1,x_3]))\text{\rm tr}(x_1x_2)$$ $$-2\text{\rm tr}([x_1,x_2]^2[x_2,x_3])\text{\rm tr}(x_1x_3) -\text{\rm tr}([x_1,x_3]^2[x_1,x_2])\text{\rm tr}(x_2^2)$$ $$+2\text{\rm tr}([x_1,x_2]^2[x_1,x_3])\text{\rm tr}(x_2x_3)-\text{\rm tr}([x_1,x_2]^3)\text{\rm tr}(x_3^2),$$ $$w_2=\text{\rm tr}(s_3(x_1,x_2,x_3)[x_1,x_2])\text{\rm tr}(s_3(x_1,x_2,x_3)),$$ $$w_3=\text{\rm tr}(s_3(x_1,x_2,x_3)x_1)\text{\rm tr}([x_1,x_2][x_2,x_3])
-\text{\rm tr}(s_3(x_1,x_2,x_3)x_2)\text{\rm tr}([x_1,x_2][x_1,x_3])$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_3)x_3)\text{\rm tr}([x_1,x_2]^2),$$ $$w_4=\text{\rm tr}(s_3(x_1,x_2,x_3)x_1)(\text{\rm tr}(x_1x_2)\text{\rm tr}(x_2x_3)
-\text{\rm tr}(x_1x_3)\text{\rm tr}(x_2^2))$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_3)x_2)(-\text{\rm tr}(x_1^2)\text{\rm tr}(x_2x_3)
+\text{\rm tr}(x_1x_2)\text{\rm tr}(x_1x_3))$$ $$+\text{\rm tr}(s_3(x_1,x_2,x_3)x_3)(\text{\rm tr}(x_1^2)\text{\rm tr}(x_2^2)-\text{\rm tr}^2(x_1x_2)).$$ In each of the cases, every highest weight vector $w\in W_3(\lambda)\subset \omega^2(S)$ is equal to a linear combination of $w_i$.
The considerations use the decomposition of $(\omega^2(S))^{(8)}$ given in Proposition \[decomposition of degree 8\]. The highest weight vectors have been found as those in Propositions \[hwv of degree up to 6\] and \[hwv of degree 7\]. Of course, if we already know the explicit form of the (candidates for) highest weight vectors $w_i$, we can check that they are linearly independent in $\omega^2(S)$ and satisfy the requirements of Lemma \[criterion for hwv\]. Since for each $\lambda$ the number of the highest weight vectors $w_i$ coincides with the multiplicity of $W_3(\lambda)\subset \omega^2(S)$ from Proposition \[decomposition of degree 8\], we conclude that every highest weight vector $w\in W_3(\lambda)\subset \omega^2(S)$ is equal to a linear combination of $w_i$.
Finally, we shall calculate the Hilbert series of the kernel of the natural homomorphism $S\to C_0$ for $d=3$.
\[Hilbert series of kernel\] Let $d=3$ and let $J$ be the kernel of the natural homomorphism $S\to C_0$. Let the Hilbert series of $J$ be $$H(J,t_1,t_2,t_3)=\sum_{k\geq 0}h_k(t_1,t_2,t_3),$$ where $h_k$ is the homogeneous component of degree $k$ of $H(J,t_1,t_2,t_3)$. Then $h_k=0$ for $k\leq 6$, $$h_7=S_{(3,2^2)}(t_1,t_2,t_3),$$ $$h_8=S_{(4,3,1)}(t_1,t_2,t_3) + 2S_{(4,2^2)}(t_1,t_2,t_3) + S_{(3^2,2)}(t_1,t_2,t_3).$$
Clearly, the Hilbert series of the kernel $J$ is equal to the difference of the Hilbert series of $S$ and $C_0$. For $d=3$ we have that $$G_2\oplus\cdots\oplus G_6=W(2)\oplus W(3)\oplus W(1^3)$$ $$\oplus W(2^2)\oplus W(2,1^2)
\oplus W(3,1^2)\oplus W(2^2,1)\oplus W(3^2)$$ and its Hilbert series is $$H(G_2\oplus\cdots\oplus G_6,t_1,t_2,t_3)=
\sum a_{k_1k_2k_3}t_1^{k_1}t_2^{k_2}t_3^{k_3}$$ $$=S_{(2)}+S_{(3)}+S_{(1^3)}+S_{(2^2)}+S_{(2,1^2)}
+S_{(3,1^2)}+S_{(2^2,1)}+S_{(3^2)}.$$ The Hilbert series of the symmetric algebra is $$H(S,t_1,t_2,t_3)=\prod\frac{1}{(1-t_1^{k_1}t_2^{k_2}t_3^{k_3})^{a_{k_1k_2k_3}}}.$$ Now the result follows by evaluation of the coefficients $a_{k_1k_2k_3}$ and expanding the first several homogeneous components of the difference of the Hilbert series of $S$ and of the Hilbert series of $C_0$, which is given in (\[Hilbert series of C0\]).
\[module of relations of degree 7 and 8\] For $d=3$, the algebra $C_0$ has a minimal system of defining relations with the property that the relations of degree $7$ and $8$ form $GL_3$-modules isomorphic, respectively, to $W_3(3,2^2)$ and $W_3(4,3,1) + 2W_3(4,2^2) + W_3(3^2,2)$.
If $J$ is the kernel of the natural homomorphism $S\to C_0$, then a minimal homogeneous system of generators of $J$ is obtained as a factor space of $J$ modulo $J\omega(S)$. Since $\omega(S)$ contains no homogeneous elements of degree 1, we obtain that the multihomogeneous components of total degree 7 and 8 of $J$ and $J/J\omega(S)$ are of the same dimension. Hence $J^{(7)}$ and $J^{(8)}$ are isomorphic as $GL_3$-modules to $(J/J\omega(S))^{(7)}$ and $(J/J\omega(S))^{(8)}$, respectively, and the conclusion follows from the expressions of $h_7$ and $h_8$ given in Lemma \[Hilbert series of kernel\].
Main results
============
Now we present the explicit defining relations of degree 7 of the algebra $C_{3d}$ for any $d\geq 3$ and of degree 8 for the algebra $C_{33}$, with respect to the generators of Abeasis and Pittaluga [@AP]. As we already mentioned, by (\[replacing with traceless matrices\]) it is sufficient to give the defining relations of the algebra $C_0$ generated by traces $\text{tr}(x_{i_1}\cdots x_{i_k})$ of products of the traceless matrices $x_i$. As in the previous sections, we denote by $S$ the symmetric algebra of the $GL_d$-module $G_2\oplus\cdots\oplus G_6$ of generators of $C_0$ and call defining relations of $C_0$ the expressions $f=0$, where $f$ is an element of the kernel $J$ of the natural homomorphisms $S\to C_0$.
\[relations of degree 7\] Let $d\geq 3$. The algebra $C_0$ does not have any defining relations of degree $\leq 6$. The $GL_d$-module structure of the homogeneous defining relations of degree $7$ of $C_0$, i.e., of the component $J^{(7)}$ in $S$ is $$J^{(7)}=W_d(4,1^3)\oplus W_d(3,2^2)\oplus W_d(3,2,1^2)\oplus W_d(2^3,1)
\oplus W_d(2^2,1^3)\oplus W_d(2,1^5).$$ In the notation of Proposition \[hwv of degree 7\], the defining relations of $C_0$ which are highest weight vectors are:
For $\lambda=(4,1^3)$: $$\label{relation hwv 4111}
12w_1-15w_2-20w_3=0;$$
For $\lambda=(3,2^2)$: $$2w_1-w_2+2w_3=0.$$
For $\lambda=(3,2,1^2)$: $$-6w_1+10w_3-15w_4+40w_6=0.$$
For $\lambda=(2^3,1)$: $$12w_1+w_2=0.$$
For $\lambda=(2^2,1^3)$: $$w_2=0.$$
For $\lambda=(2,1^5)$: $$2w_1-5w_2=0.$$
In all cases the idea is the same. We already know that there are no relations of degree $\leq 11$ for $d=2$ and the only $GL_3$-module of relations is isomorphic to $W_3(3,2^2)$. Hence, we have to consider the cases in Proposition \[hwv of degree 7\] only.
We shall consider in detail the case $\lambda=(4,1^3)$. The possible relations $w=0$ are linear combinations of $w_1,w_2,w_3$. We assume that $$\label{relation 4111}
w=\xi_1w_1+\xi_2w_2+\xi_3w_3=0$$ and evaluate $w$ on the traceless matrices (\[first matrix\]) and (\[other matrices\]). The coefficients of the monomials $(x_{11}^{(1)})^4x_{12}^{(2)}x_{22}^{(3)}x_{21}^{(4)}$ and $(x_{11}^{(1)})^4x_{13}^{(2)}x_{32}^{(3)}x_{21}^{(4)}$ are, respectively, $20\xi_1-8\xi_2+18\xi_3$ and $20\xi_1+12\xi_3$. Hence the equation (\[relation 4111\]) implies that $$20\xi_1-8\xi_2+18\xi_3=20\xi_1+12\xi_3=0.$$ Up to a multiplicative constant, the only solution of this system is $$\xi_1=12,\quad \xi_2=-15,\quad \xi_3=-20.$$ Hence, there is only one possible candidate for a defining relation which is a highest weight vector of some $W_d(4,1^3)$. We evaluate once again (\[relation 4111\]) on (\[first matrix\]) and (\[other matrices\]) for these values of $\xi_1,\xi_2,\xi_3$ and obtain that $w(x_1,x_2,x_3,x_4)=0$. Hence the multiplicity of $W_d(4,1^3)$ in $J$ is equal to 1 and the corresponding relation is (\[relation hwv 4111\]). We want to mention that the case $\lambda=(3,1^4)$ does not participate in the statement of the theorem, because the multiplicity of $W_d(3,1^4)$ in $J$ is 0.
The dimension of the defining relations of degree $7$ of the algebra $C_{3d}$ is equal to $$r_7=r_7(d)=\frac{2}{7!}(d+1)d(d-1)(d-2)(41d^3-86d^2+114d-360).$$
Since $C_{3d}$ does not satisfy relations of degree $\leq 6$, and the constants are the only elements of degree 0 in $K[\text{tr}(X_1),\ldots,\text{tr}(X_d)]$, the dimension of the relations of degree 7 of $C_{3d}$ coincides with this dimension in $C_0$. Now the proof is complete using Theorem \[relations of degree 7\] and the dimension formula (\[dim of W\]) for $W_d(\lambda)$.
\[relations of degree 8\] Let $d=3$. The $GL_d$-module structure of the homogeneous component $J^{(8)}$ of degree $8$ in $S$ is $$J^{(8)}=W_3(4,3,1)\oplus 2W_3(4,2^2)\oplus W_3(3,2^2,2).$$ In the notation of Proposition \[hwv of degree 8\], the defining relations which are highest weight vectors are:
For $\lambda=(4,3,1)$: $$\label{relation hwv 431}
-6w_1-18w_2+3w_3+3w_5-8w_7=0;$$
For $\lambda=(4,2^2)$: All nontrivial linear combinations of $$w_1-15w_2+3w_3+\frac{21}{4}w_4-\frac{5}{2}w_5+\frac{5}{2}w_6-3w_7+2w_9=0,$$ $$-36w_2+6w_3+\frac{27}{2}w_4-6w_5+6w_6-9w_7+w_8+6w_9=0.$$
For $\lambda=(3^2,2)$: $$6w_1+2w_2-3w_3-3w_4=0.$$ The number $r_8$ of the defining relations of degree $8$ of any homogeneous minimal system of defining relations of the algebra $C_{33}$ is equal to $30$.
The decomposition of $J^{(8)}$ is given in Corollary \[module of relations of degree 7 and 8\]. The explicit form of the highest weight vectors is obtained as in the proof of Theorem \[relations of degree 7\]. The number of defining relations of degree 8 in any homogeneous minimal system of defining relations for $C_{33}$ is equal to the dimension of the relations $J^{(8)}$ of $C_0$. For the proof that $r_8=30$ it is sufficient to use the dimension formula (\[dim of W\]) for $W_d(\lambda)$.
Using Lemma \[basis of module\] we can find an explicit basis of the set of defining relations of degree 7 for $C_0$, $d\geq 3$, and of degree 8 for $C_0$, $d=3$.
Acknowledgements {#acknowledgements .unnumbered}
================
This project was started when the second author visited the University of Palermo. He is very grateful for the hospitality and the creative atmosphere during his stay there.
[99]{}
S. Abeasis, M. Pittaluga, [*On a minimal set of generators for the invariants of $3 \times 3$ matrices*]{}, Commun. Algebra [**17**]{} (1989), 487-499.
G. Almkvist, W. Dicks, E. Formanek, [*Hilbert series of fixed free algebras and noncommutative classical invariant theory*]{}, J. Algebra [**93**]{} (1985), 189-214.
H. Aslaksen, V. Drensky, L. Sadikova, [*Defining relations of invariants of two $3 \times 3$ matrices*]{}, J. Algebra [**298**]{} (2006), 41-57.
F. Benanti, V. Drensky, [*Defining relations of noncommutative trace algebra of two $3 \times 3$ matrices*]{}, Adv. Appl. Math. [**37**]{} (2006), No. 2, 162-182.
A. Berele, J.R. Stembridge, [*Denominators for the Poincaré series of invariants of small matrices*]{}, Israel J. Math. [**114**]{} (1999), 157-175.
C. De Concini, D. Eisenbud, C. Procesi, [*Young diagrams and determinantal varieties*]{}, Invent. Math. [**56**]{} (1980), 129-165.
D.Z. Djokovic, [Poincaré series of some pure and mixed trace algebras of two generic matrices]{}, J. Algebra [**309**]{} (2007), No. 1, 654-671.
V. Drensky, [*Free Algebras and PI-Algebras*]{}, Springer-Verlag, Singapore, 1999.
V. Drensky, [*Computing with matrix invariants*]{}, Math. Balk., New Ser. [**21**]{} (2007) (to appear).
V. Drensky, E. Formanek, [*Polynomial Identity Rings*]{}, Advanced Courses in Mathematics, CRM Barcelona, Birkhäuser, Basel-Boston, 2004.
V. Drensky, L. Sadikova, [*Generators of invariants of two $4\times 4$ matrices*]{}, C.R. Acad. Bulg. Sci. [**59**]{} (2006), No. 5, 477-484.
E. Formanek, [*The Polynomial Identities and Invariants of $n \times n$ Matrices*]{}, CBMS Regional Conf. Series in Math. [**78**]{}, Published for the Confer. Board of the Math. Sci. Washington DC, AMS, Providence RI, 1991.
P. Koshlukov, [*Polynomial identities for a family of simple Jordan algebras*]{}, Commun. Algebra [**16**]{} (1988), 1325-1371.
I.G. Macdonald, [*Symmetric Functions and Hall Polynomials*]{}, Oxford Univ. Press (Clarendon), Oxford, 1979. Second Edition, 1995.
K. Nakamoto, [*The structure of the invariant ring of two matrices of degree $3$*]{}, J. Pure Appl. Algebra [**166**]{} (2002), No. 1-2, 125-148.
C. Procesi, [*The invariant theory of $n\times n$ matrices*]{}, Adv. Math. [**19**]{} (1976), 306-381.
Yu.P. Razmyslov, [*Trace identities of full matrix algebras over a field of characteristic zero*]{} (Russian), Izv. Akad. Nauk SSSR, Ser. Mat. [**38**]{} (1974), 723-756. Translation: Math. USSR, Izv. [**8**]{} (1974), 727-760.
Y. Teranishi, [*The ring of invariants of matrices*]{}, Nagoya Math. J. [**104**]{} (1986), 149-161.
Y. Teranishi, [*Linear diophantine equations and invariant theory of matrices*]{}, “Commut. Algebra and Combinatorics (Kyoto, 1985)”, Adv. Stud. Pure Math. [**11**]{}, North-Holland, Amsterdam-New York, 1987, 259-275.
R.M. Thrall, [*On symmetrized Kronecker powers and the structure of the free Lie ring*]{}, Trans. Amer. Math. Soc. [**64**]{} (1942), 371-388.
[^1]: The research of the first author was partially supported by MIUR, Italy.
[^2]: The research of the second author was partially supported by Grant MI-1503/2005 of the Bulgarian National Science Fund.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We obtain the global well-posedness to the 3D incompressible magnetohydrodynamics (MHD) equations in Besov space with negative index of regularity. Particularly, we can get the global solutions for a new class of large initial data. As a byproduct, this result improves the corresponding result in [@HHW]. In addition, we also get the global result for this system in $\mathcal{\chi}^{-1}({\mathbb{R}}^3)$ originally developed in [@LL]. More precisely, we only assume that the norm of initial data is exactly smaller than the sum of viscosity and diffusivity parameters.'
address: '$^1$ School of Mathematics Sciences, Zhejiang University, Hanzhou 310027, China'
author:
- 'Renhui Wan$^{1}$'
title: 'Global well-posedness to the 3D incompressible MHD equations with a new class of large initial data'
---
.2in
.2in
Introduction {#s1}
============
We are concerned with the 3D incompressible MHD equations: $$\label{1.1}
\left\{
\begin{aligned}
& \partial_t u + u\cdot\nabla u -\mu_1 \Delta u +\nabla p = B\cdot\nabla B, \\
& \partial_t B + u\cdot\nabla B-B\cdot\nabla u -\mu_2 \Delta B=0, \\
& {\rm div} u={\rm div}B=0,\\
& u(0,x) =u_0(x),\ B(0,x)=B_0(x),
\end{aligned}
\right.$$ here $(t,x)\in\mathbb{R}^{+}\times\mathbb{R}^3$, $u,p,B$ stand for velocity vector, scalar pressure and magnetic vector, respectively, $\mu_1$ and $\mu_2$ are nonnegative viscosity and diffusivity parameters, respectively.
.1in For $\mu_1>0$ and $\mu_2>0$, the local well-posedness and global existence with small data for (\[1.1\]) were obtained by Duvaut and Lions [@DL] in $d$ dimensional Sobolev space $H^s({\mathbb{R}}^d)$, $s\ge d$. Then Sermange and Termam [@ST] studied the regularity of weak solutions $(u,B)\in L^\infty(0,T;H^1({\mathbb{R}}^3))$. And some regularity criteria were established in [@Wu1; @Wu2; @Wu3]. For $\mu_1>0$ and $\mu_2=0$ (so-called non-resistive MHD equations), by the new Kato-Ponce commutator estimate, $$\|\Lambda^s(u\cdot\nabla B)-u\cdot\nabla \Lambda^s B\|_{L^2({\mathbb{R}}^d)}\le C\|\nabla u\|_{H^s({\mathbb{R}}^d)}\|B\|_{H^s({\mathbb{R}}^d)},\ s>\frac{d}{2},\ d=2,3,$$ Fefferman et al. [@Fefferman] proved the low regularity local well-posedness of strong solutions, which was extended to general inhomogeneous Besov space with initial data $(u_0,B_0)\in B_{2,1}^{\frac{d}{2}-1}({\mathbb{R}}^d)\times B_{2,1}^\frac{d}{2}({\mathbb{R}}^d)$ in the recent works [@Chemin; @1] and [@Wan]. Furthermore, for the non-resistive version with smooth initial data near some nontrivial steady state, we refer [@PZhnag1; @LT; @ZZhang; @PZhang2] for the related works.
.1in Due to a new observation that the velocity field plays a more important role than magnetic field. The new regularity criteria only involving the velocity were proved, see [@CMZ; @HX; @YZ1; @YZ2] and references therein.
.1in One can easily get a new formation of (\[1.1\]) by the following: $$W^+:=u+B,\ W^-:=u-B,\ \nu_{+}=\frac{\mu_1+\mu_2}{2},\ \nu_{-}=\frac{\mu_1-\mu_2}{2}$$ with initial data $W^{\pm}_0(x):=u_0(x)\pm B_0(x),$ that is, $$\label{1.12}
\left\{
\begin{aligned}
& \partial_t W^+ + W^-\cdot\nabla W^+ -\nu_{+}\Delta W^+ +\nabla p = \nu_{-}\Delta W^-, \\
& \partial_t W^- + W^+\cdot\nabla W^- -\nu_{+}\Delta W^- +\nabla p = \nu_{-}\Delta W^+, \\
& {\rm div} W^+={\rm div}W^-=0,\\
& W^+(0,x) =W^+_0(x),\ W^-(0,x)=W^-_0(x).
\end{aligned}
\right.$$
Very recently, He et al. [@HHW] obtained the global well-posedness for (\[1.12\]) with initial data $(u_0,B_0)$ satisfying:\
($i$) $\nu_-=0$ and $$\nu_+^{-3}\|W_0^-\|_{L^3}^3\exp\{C\nu_+^{-3}\|W_0^+\|_{L^3}^3\}<\epsilon_0$$ or $$\nu_+^{-3}\|W_0^+\|_{L^3}^3\exp\{C\nu_+^{-3}\|W_0^-\|_{L^3}^3\}<\epsilon_0;$$ ($ii$) $\nu_-\neq0$ and $$\label{1.121}
\left(\nu_+^{-2}\|W^-_0\|_{\dot{H}^\frac{1}{2}}^2+\frac{\nu_-^2}{\nu_+^2}(\nu_+^{-2}\|W_0^+\|_{\dot{H}^\frac{1}{2}}^2+\frac{\nu_-^2}{\nu_+^2})\right)
\exp\left\{C\nu_+^{-4}(\|W_0^+\|_{\dot{H}^\frac{1}{2}}^4+\nu_-^4)\right\}<\epsilon_0$$ or $$\left(\nu_+^{-2}\|W^+_0\|_{\dot{H}^\frac{1}{2}}^2+\frac{\nu_-^2}{\nu_+^2}(\nu_+^{-2}\|W_0^-\|_{\dot{H}^\frac{1}{2}}^2+\frac{\nu_-^2}{\nu_+^2})\right)
\exp\left\{C\nu_+^{-4}(\|W_0^-\|_{\dot{H}^\frac{1}{2}}^4+\nu_-^4)\right\}<\epsilon_0.$$ Here $\epsilon_0$ ia a sufficiently small positive constant. .1in
In this paper, we will prove the global well-posedness of (\[1.1\]) $(\mu_1>0,\mu_2>0)$ in generalized space, $\dot{B}_{p,r}^{\frac{3}{p}-1}({\mathbb{R}}^3)$, by make full use of the harmonic analysis tools. The details can be given as follows:
\[t1\] Consider (\[1.1\]) with initial data $(u_0,B_0)\in \dot{B}_{p,r}^{\frac{3}{p}-1}({\mathbb{R}}^3),$ $(p,r)\in (1,\infty)\times [1,\infty),$ satisfying ${\rm div}u_0={\rm div}B_0=0$. There exists a constant $C$ and a small constant $\eta>0$ such that if $$\label{1.14}
\left(\|W^-_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\frac{\nu_-}{\nu_+}(\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\nu_-)\right)
\exp\left\{C\nu_+^{-\frac{2}{1-\epsilon}}(\nu_-+\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}})^\frac{2} {1-\epsilon} \right\}<\eta \nu_+$$ or $$\label{1.15}
\left(\|W^+_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\frac{\nu_-}{\nu_+}(\|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\nu_-)\right)
\exp\left\{C\nu_+^{-\frac{2}{1-\epsilon}}(\nu_-+\|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}})^\frac{2}{1-\epsilon} \right\}<\eta\nu_+,$$ where $(\epsilon,r)$ satisfies
$$\label{1.16}
\left\{
\begin{aligned}
0\le \epsilon<1&, \ {\rm if }\ r=1; \\
0< \epsilon<1&, \ {\rm if }\ 1<r\le 2; \\
1-\frac{2}{r}\le \epsilon<1&,\ {\rm if }\ 2<r<\infty.
\end{aligned}
\right.$$
Then (\[1.1\]) admits a unique global solution $(u,B)$ satisfying $$(u,B)\in \tilde{C}([0,\infty); \dot{B}_{p,r}^{\frac{3}{p}-1}({\mathbb{R}}^3))\cap \tilde{L}^1([0,\infty); \dot{B}_{p,r}^{\frac{3}{p}+1}({\mathbb{R}}^3)).$$
.1in If $\nu_-=0$, i.e., $\mu_1=\mu_2=\nu_+$, we have a corollary immediately.
\[c1\] Consider (\[1.1\]) with initial data $(u_0,B_0)\in \dot{B}_{p,r}^{\frac{3}{p}-1}({\mathbb{R}}^3),$ $(p,r)\in (1,\infty)\times [1,\infty),$ satisfying ${\rm div}u_0={\rm div}B_0=0$. There exists a constant $C$ and a small constant $\eta>0$ such that if $$\label{1.141}
\|W^-_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}
\exp\left\{C\nu_+^{-\frac{2}{1-\epsilon}}\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}^\frac{2}{1-\epsilon} \right\}<\eta \nu_+$$ or $$\label{1.151}
\|W^-_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}
\exp\left\{C\nu_+^{-\frac{2}{1-\epsilon}}\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}^\frac{2}{1-\epsilon} \right\}<\eta \nu_+,$$ where $(\epsilon,r)$ satisfies (\[1.16\]). Then (\[1.1\]) admits a unique global solution $(u,B)$ satisfying $$(u,B)\in \tilde{C}([0,\infty); \dot{B}_{p,r}^{\frac{3}{p}-1}({\mathbb{R}}^3))\cap \tilde{L}^1([0,\infty); \dot{B}_{p,r}^{\frac{3}{p}+1}({\mathbb{R}}^3)).$$
\[r1\] ($i$) We will construct the global solution with a new class of large initial data. More precisely, assume that $\phi$ satisfies the condition in Proposition \[l2\], let $$u_0=(\partial_2\phi,-\partial_1\phi,0),\ \ B_0=2\sin^2\frac{x_3}{2\epsilon}(\partial_2\phi,-\partial_1\phi,0),$$ then ${\rm div}u_0={\rm div}B_0=0$ and $\|(u_0,B_0)\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\le \mathfrak{M}$ $(p>3)$, which is independent of $\epsilon$. Moreover, thanks to Proposition \[l2\], there exists a positive constant $C_1$ and $C_2$, $$\|u_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\ge C_1,\ \ \|B_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\ge \frac{C_1}{2},$$ $$\|u_0-B_0\|_{\dot{B_{p,r}^{\frac{3}{p}-1}}}\le C_2\epsilon^{1-\frac{3}{p}},$$ which ensures the conditions (\[1.14\])($\nu_+\gg \nu_-$) and (\[1.141\]) hold. Additionally, the assumption $\nu_+\gg\nu_-$ is reasonable in astrophysical magnetic phenomena, see Remark 2.3 in [@HHW]. Combining with the above explanations, this class of large data can lead the global well-posedness to (\[1.1\]).\
($ii$) One can easily check that condition (\[1.14\]) is equal to (\[1.121\]) when $p=r=2$ and choosing $\epsilon=\frac{1}{2}$. By Bernstein’s inequality, we have the following embedding relationship: $$\dot{H}^\frac{1}{2}\hookrightarrow \dot{B}_{p,r}^{\frac{3}{p}-1},\ \ p> 2, r\ge2.$$ So our result improves the corresponding work under (\[1.121\]) in [@HHW]. By the same way, similar improvements can also be obtained under (\[1.15\]) , (\[1.141\]) and (\[1.151\]).
We shall point out that the above result can not be extended to $p=\infty.$ As a matter of fact, by these works [@ill2] and [@ill1] concerning the well-known Navier-Stokes equations, (\[1.1\]) may ill-posedness in this endpoint Besov space. .1in Next, we consider the space $\chi^{-1}({\mathbb{R}}^3)$, which is smaller than $\dot{B}_{\infty,r}^{-1}$ due to Proposition \[l33\]. It was originally developed in [@LL] and applied to get the global well-posedness for the Navier-Stokes equations under $$\|u_0\|_{\chi^{-1}}<\mu.$$ For MHD equations (\[1.1\]), similar result holds under $$\label{1.2}
\|u_0\|_{\chi^{-1}}+\|B_0\|_{\chi^{-1}}<\min\{\mu,\eta\},$$ see [@KWang] for details. .1in We have some new result in $\chi^{-1}({\mathbb{R}}^3).$
\[t2\] Consider (\[1.1\]) with initial data $(u_0,B_0)\in \chi^{-1}({\mathbb{R}}^3)$ satisfying ${\rm div}u_0={\rm div}B_0=0$. There exists a constant $C$ such that if $$\label{1.21}
\left(\|W_0^-\|_{\chi^{-1}}+\frac{C\nu_-}{\nu_+}(\nu_-+\|W_0^+\|_{\chi^{-1}})\right)\exp\left\{\frac{C}{\nu_+^2}(\nu_-+\|W_0^+\|_{\chi^{-1}})^2\right\}<2\nu_+$$ or $$\label{1.22}
\left(\|W_0^+\|_{\chi^{-1}}+\frac{C\nu_-}{\nu_+}(\nu_-+\|W_0^-\|_{\chi^{-1}})\right)\exp\left\{\frac{C}{\nu_+^2}(\nu_-+\|W_0^-\|_{\chi^{-1}})^2\right\} <2\nu_+.$$ Then (\[1.1\]) admits a unique global solution $(u,B)$ satisfying $$(u,B)\in C([0,\infty); \chi^{-1}({\mathbb{R}}^3))\cap L^1([0,\infty);\chi^1({\mathbb{R}}^3)).$$
Similarly, we also have a corollary immediately when $\nu_-=0$.
\[c2\] Consider (\[1.1\]) with initial data $(u_0,B_0)\in \chi^{-1}({\mathbb{R}}^3)$ satisfying ${\rm div}u_0={\rm div}B_0=0$. There exists a constant $C$ such that if $$\|W_0^-\|_{\chi^{-1}}\exp\left\{\frac{C}{\nu_+^2}\|W_0^+\|_{\chi^{-1}}^2\right\} <2\nu_+$$ or $$\|W_0^+\|_{\chi^{-1}}\exp\left\{\frac{C}{\nu_+^2}\|W_0^-\|_{\chi^{-1}}^2\right\}<2\nu_+.$$ Then (\[1.1\]) admits a unique global solution $(u,B)$ satisfying $$(u,B)\in C([0,\infty); \chi^{-1}({\mathbb{R}}^3))\cap L^1([0,\infty);\chi^1({\mathbb{R}}^3)).$$
\[r2\] The authors in [@LL] proved the global well-posedness for Navier-Stokes equations by using $$\|u\cdot\nabla u\|_{\chi^{-1}}\le \|u\|_{\chi^{-1}}\|u\|_{\chi^1},$$ while we shall use the new estimate below in our proof, i.e., $$\|u\cdot\nabla v\|_{\chi^{-1}}\le \|u\|_{\chi^0}\|v\|_{\chi^0}.$$
\[r3\] Due to the symmetric structure of (\[1.12\]), we only give the proof of Theorem \[t1\] and Theorem \[t2\] under (\[1.14\]) and (\[1.21\]), respectively.
The present paper is structured as follows:\
In section \[s2\], we provide some definitions of spaces, establish several lemmas. The third section proves Theorem \[t1\], while the last section gives the proof of Theorem \[t2\].
.1in Let us complete this section by describing the notations we shall use in this paper.\
[**Notations**]{} The uniform constant $C$ is different on different lines. We also use $L^p$, $\dot{B}_{p,r}^s$ and $\chi^s$ to stand for $L^p(\mathbb{R}^d)$, $\dot{B}_{p,r}^s(\mathbb{R}^d)$ and $\chi^s({\mathbb{R}}^d)$ in somewhere, respectively. We use $A:=B$ to stands for $A$ is defined by $B$, and ${\bf 1}$ is the characteristic function.
.4in
Preliminaries {#s2}
==============
In this section, we give some necessary definitions, propositions and lemmas. .1in The Fourier transform is given by $$\widehat{f}(\xi)=\int_{{\mathbb{R}}^d}e^{-ix\cdot\xi}f(x)dx.$$
Let $\mathfrak{B}=\{\xi\in\mathbb{R}^d,\ |\xi|\le\frac{4}{3}\}$ and $\mathfrak{C}=\{\xi\in\mathbb{R}^d,\ \frac{3}{4}\le|\xi|\le\frac{8}{3}\}$. Choose two nonnegative smooth radial function $\chi,\ \varphi$ supported, respectively, in $\mathfrak{B}$ and $\mathfrak{C}$ such that $$\sum_{j\in\mathbb{Z}}\varphi(2^{-j}\xi)=1,\ \ \xi\in\mathbb{R}^d\setminus\{0\}.$$ We denote $\varphi_{j}=\varphi(2^{-j}\xi),$ $h=\mathfrak{F}^{-1}\varphi$ and $\tilde{h}=\mathfrak{F}^{-1}\chi,$ where $\mathfrak{F}^{-1}$ stands for the inverse Fourier transform. Then the dyadic blocks $\Delta_{j}$ and $S_{j}$ can be defined as follows $$\Delta_{j}f=\varphi(2^{-j}D)f=2^{jd}\int_{\mathbb{R}^d}h(2^jy)f(x-y)dy,\ \ S_{j}f=\sum_{k\le j-1}\Delta_{k}f$$
Formally, $\Delta_{j}$ is a frequency projection to annulus $\{\xi:\ C_{1}2^j\le|\xi|\le C_{2}2^j\}$, and $S_{j}$ is a frequency projection to the ball $\{\xi:\ |\xi|\le C2^j\}$. One easily verifies that with our choice of $\varphi$ $$\Delta_{j}\Delta_{k}f=0\ {\rm if} \ |j-k|\ge2\ \ {\rm and}\ \ \Delta_{j}(S_{k-1}f\Delta_{k}f)=0\ {\rm if}\ |j-k|\ge5.$$ Let us recall the definition of the Besov space.
\[HB\] Let $s\in \mathbb{R}$, $(p,q)\in[1,\infty]^2,$ the homogeneous Besov space $\dot{B}_{p,q}^s({\mathbb{R}}^d)$ is defined by $$\dot{B}_{p,q}^{s}({\mathbb{R}}^d)=\{f\in \mathfrak{S}'({\mathbb{R}}^d);\ \|f\|_{\dot{B}_{p,q}^{s}({\mathbb{R}}^d)}<\infty\},$$ where $$\|f\|_{\dot{B}_{p,q}^s({\mathbb{R}}^d)}=\left\{\begin{aligned}
&\displaystyle (\sum_{j\in \mathbb{Z}}2^{sqj}\|\Delta_{j}f\|_{L^p({\mathbb{R}}^d)}^{q})^\frac{1}{q},\ \ \ \ {\rm for} \ \ 1\le q<\infty,\\
&\displaystyle \sup_{j\in\mathbb{Z}}2^{sj}\|\Delta_{j}f\|_{L^p({\mathbb{R}}^d)},\ \ \ \ \ \ \ \ {\rm for}\ \ q=\infty,\\
\end{aligned}
\right.$$ and $\mathfrak{S}'({\mathbb{R}}^d)$ denotes the dual space of $\mathfrak{S}({\mathbb{R}}^d)=\{f\in\mathcal{S}(\mathbb{R}^d);\ \partial^{\alpha}\hat{f}(0)=0;\ \forall\ \alpha\in \ \mathbb{N}^d $ [multi-index]{}} and can be identified by the quotient space of $\mathcal{S'}/\mathcal{P}$ with the polynomials space $\mathcal{P}$.
The norm of the space $\tilde{L}^{r_1}_t(\dot{B}_{p,r}^s)$ and $\tilde{L}^{r_1}_{t,\omega}(\dot{B}_{p,r}^s)$ is defined by $$\|f\|_{\tilde{L}^{r_1}_t(\dot{B}_{p,r}^s)}:=\|2^{js}\|\Delta_j f\|_{L^{r_1}_tL^p}\|_{l^r({\mathbb{Z}})}$$ and $$\|f\|_{\tilde{L}^{r_1}_{t,\omega}(\dot{B}_{p,r}^s)}:=\|2^{js}\left(\int_0^t \omega(\tau)^{r_1}\|\Delta_j f(\tau)\|_{L^p}^{r_1} d\tau\right)^\frac{1}{r_1}\|_{l^r({\mathbb{Z}})}.$$ $f\in \tilde{C}(0,t;\dot{B}_{p,r}^s)$ means $f\in \tilde{L}^{\infty}_t(\dot{B}_{p,r}^s)$ and $\|f(t)\|_{\dot{B}_{p,r}^s}$ is continuous in time.
.1in The following proposition provide Bernstein type inequalities.
Let $1\le p\le q\le \infty$. Then for any $\beta,\gamma\in (\mathbb{N}\cup \{0\})^3$, there exists a constant $C$ independent of $f,j$ such that
1. If $f$ satisfies $$\mbox{supp}\, \widehat{f} \subset \{\xi\in \mathbb{R}^d: \,\, |\xi|
\le \mathcal{K} 2^j \},$$ then $$\|\partial^\gamma f\|_{L^q(\mathbb{R}^d)} \le C 2^{j|\gamma| +
j d(\frac{1}{p}-\frac{1}{q})} \|f\|_{L^p(\mathbb{R}^d)}.$$
2. If $f$ satisfies $$\label{spp}
\mbox{supp}\, \widehat{f} \subset \{\xi\in \mathbb{R}^d: \,\, \mathcal{K}_12^j
\le |\xi| \le \mathcal{K}_2 2^j \}$$ then $$\|f\|_{L^p(\mathbb{R}^d)} \le C2^{-j|\gamma|}\sup_{|\beta|=|\gamma|} \|\partial^\beta f\|_{L^p(\mathbb{R}^d)}.$$
For more details about Besov space such as some useful embedding relations, see [@BCD; @Grafakos; @Stein].
\[l1\][@Danchin] Let $1<p<\infty,$ $supp \widehat{u}\subset C(0,R_{1},R_{2})$ (with $0<R_{1}<R_{2}$). There exists a constant $c$ depending on $\frac{R_{2}}{R_{1}}$ and such that $$\label{Ber}
c\frac{R_{1}^2}{p^2}\int_{\mathbb{R}^3}|u|^p dx\le -\frac{1}{p-1}\int_{\mathbb{R}^3} \Delta u |u|^{p-2}u dx.$$
\[l2\] Let $\phi\in \mathcal{S}({\mathbb{R}}^3)$, whose Fourier transform supported in annulus contained in ${\mathbb{R}}^3\setminus\{0\}$, and $p>3$. If $u_0=(\partial_2 \phi, -\partial_1\phi,0)$ and $B_0=2\sin^2\frac{x_3}{2\epsilon}(\partial_2 \phi, -\partial_1\phi,0),$ then there exists a constant $C_1,C_2>0$ such that $$\|u_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\ge C_1,\ \ \|B_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\ge \frac{C_1}{2}$$ and $$\|u_0-B_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\le C_2\epsilon^{1-\frac{3}{p}},$$ here $\epsilon$ is sufficiently small.
The last estimate can be obtained by following the proof of Lemma 3.1 in [@Chemin; @2]. So we suffice to show both $\|u_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}$ and $\|B_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}$ has positive lower bound. With this $\phi$, there exists a finite $j_0\in{\mathbb{Z}}$, such that $\Delta_{j_0}\partial_2\phi\neq0$, which implies $$\|\Delta_{j_0} \partial_2\phi\|_{L^\infty}\ge \epsilon_0$$ for some positive constant $\epsilon_0$. Thanks to this, by Bernstein’s inequality, we have $$\|u_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\ge \|u_0\|_{\dot{B}_{\infty,\infty}^{-1}}\ge 2^{-j_0}\|\Delta_{j_0}\partial_2\phi\|_{L^\infty}\ge 2^{-j_0}\epsilon_0,$$ and by triangle inequality $$\|B_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\ge \|u_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}-\|u_0-B_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\ge
2^{-j_0}\epsilon_0-C_2\epsilon^{1-\frac{3}{p}}\ge 2^{-j_0-1}\epsilon_0$$ due to the sufficient small $\epsilon$. Choosing $C_1=2^{-j_{0}}\epsilon_0$ yields the desired result.
For some convenience, we provide the following definition of $\chi^{s}({\mathbb{R}}^d)$, $$\|f\|_{\chi^s}:=\int_{{\mathbb{R}}^d} |\xi|^s |\hat{f}(\xi)| d\xi,$$ and we refer [@LL] for some details.
\[l33\] Let $f\in \chi^{-1}$, then we have $$\|f\|_{\dot{B}_{\infty,r}^{-1}}\le \|f\|_{\dot{B}_{\infty,1}^{-1}}\le \|f\|_{\mathbb{{B}}_{1,1}^{-1}}\thickapprox \|f\|_{\chi^{-1}},$$ where $$\|f\|_{\mathbb{B}_{1,1}^{-1}}:=\sum_{j\in{\mathbb{Z}}}2^{-j}\|\widehat{\Delta_j f}\|_{L^1}.$$
The first inequality is obvious, while the second inequality can be proved by using $\|f\|_{L^\infty}\le \|\hat{f}\|_{L^1}.$ Now, we prove $\|f\|_{\mathbb{{B}}_{1,1}^{-1}}\thickapprox \|f\|_{\chi^{-1}}$. By the definition of $\Delta_j$, and using Monotone Convergence Theorem, $$\begin{aligned}
\|f\|_{\mathbb{{B}}_{1,1}^{-1}}=&\sum_{j\in {\mathbb{Z}}}2^{-j}\| \varphi(2^{-j}\xi)\hat{f}(\xi)\|_{L^1}\\
\thickapprox&
\sum_{j\in {\mathbb{Z}}}\| |\xi|^{-1}\varphi(2^{-j}\xi)\hat{f}(\xi)\|_{L^1}\\
=&\||\xi|^{-1}|\hat{f}(\xi)|\|_{L^1}\\
=&\|f\|_{\chi^{-1}},
\end{aligned}$$ where we have used $\sum_{j\in Z}\varphi(2^{-j}\xi)=1$ and $\varphi\ge 0$.
\[l3\] ($i$) Let $(p,r)\in [1,\infty)\times[1,\infty],$ ${\rm div}u=0$, then $$\label{SKP1}
\|u\cdot\nabla v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}\le C\left(\|u\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}\|v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}+
\|v\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}\|u\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\right);$$ ($ii$) Let $(p,r)\in [1,\infty)\times[1,\infty],$ ${\rm div}u=0$, then $$\label{SKP2}
\|u\cdot\nabla v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}\le
C\|v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}^\frac{1+\epsilon}{2}\|v\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}^\frac{1-\epsilon}{2},$$ where $0<\epsilon<1$ and $f=\|u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-1}}^\frac{2}{1-\epsilon}.$ In particular, (\[SKP2\]) also holds when $(\epsilon,r)=(0,1)$ and $f=\|u\|_{\dot{B}_{p,1}^\frac{3}{p}}^2.$
For the proof, we shall use homogeneous Bony’s decomposition: $$uv=T_uv+T_vu+R(u,v),$$ where $$T_uv=\sum_{j\in \mathbb{Z}}S_{j-1}u\Delta_{j}v,\ \ T_vu=\sum_{j\in \mathbb{Z}}\Delta_{j}uS_{j-1}v,\ \
R(u,v)=\sum_{j\in\mathbb{Z}}\Delta_{j}u\tilde{\Delta}_{j}v,$$ here $\tilde{\Delta}_{j}=\Delta_{j-1}+\Delta_{j}+\Delta_{j+1}.$
The estimate of (\[SKP1\]) can be established by using $$\|u\cdot\nabla v\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\le C\{\|u\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\|v\|_{\dot{B}_{p,r}^{\frac{3}{p}+1}}+
\|v\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}\|u\|_{\dot{B}_{p,r}^{\frac{3}{p}+1}}\},$$ whose proof is standard. Thus the goal is the estimate of (\[SKP2\]). By homogeneous Bony’s decomposition, $$\label{2.4}
\begin{aligned}
\|u\cdot\nabla v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}\le& \|T_{u_i}\partial_i v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}
+\|T_{\partial_i v}u_i\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+\|R(u,\nabla v)\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}\\
:=&I_1+I_2+I_3.
\end{aligned}$$ Let $\theta=\frac{1-\epsilon}{2}$, $0<\epsilon<1$. For $I_1$, using Hölder’s inequality and Bernstein’s inequality, $$\begin{aligned}
I_1\le& \left\|2^{j(\frac{3}{p}-1)}\sum_{|k-j|\le 4}\|\Delta_j(S_{k-1}u\cdot\nabla \Delta_k v)\|_{L^1_tL^p}\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}-1)}\|S_{j-1}u\cdot\nabla \Delta_j v\|_{L^1_tL^p}\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}-1)}\int_0^t \|S_{j-1}u\|_{L^\infty}\|\nabla \Delta_j v\|_{L^p}d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}+\epsilon)}\int_0^t \|u\|_{\dot{B}_{\infty,\infty}^{-\epsilon}}\| \Delta_j v\|_{L^p}d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}+\epsilon)}\int_0^t \|u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}\| \Delta_j v\|_{L^p}d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}+1)(1-\theta)}\|\Delta_j v\|_{L^1_tL^p}^{1-\theta}
(\int_0^t 2^{j(\frac{3}{p}-1)} \|u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}^\frac{1}{\theta}\| \Delta_j v\|_{L^p}d\tau)^\theta\right\|_{l^r({\mathbb{Z}})}\\
\le& C\|v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}^{1-\theta}\|v\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}^\theta
= C\|v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}^\frac{1+\epsilon}{2}\|v\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}^\frac{1-\epsilon}{2},
\end{aligned}$$ here $f=\|u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-1}}^\frac{2}{1-\epsilon}$ and we have used $$\left\|2^{js}\|S_j u\|_{L^p}\right\|_{l^r({\mathbb{Z}})}\approx \|u\|_{\dot{B}_{p,r}^{s}},\ \forall\ s<0.$$ Similarly, for $I_2$, by Hölder’s inequality and Bernstein’s inequality, $$\begin{aligned}
I_2\le& \left\|2^{j(\frac{3}{p}-1)}\sum_{|k-j|\le 4}\|\Delta_j(\Delta_{k}u\cdot\nabla S_{k-1} v)\|_{L^1_tL^p}\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}-1)}\|\Delta_{j}u\cdot\nabla S_{j-1} v\|_{L^1_tL^p}\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}-1)}\int_0^t \|\Delta_{j}u\|_{L^p}\|\nabla S_{j-1} v\|_{L^\infty} d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\epsilon-1)}\int_0^t \|u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}\sum_{j'\le j-2}2^{j'(\frac{3}{p}+1)}\| \Delta_{j'} v\|_{L^p} d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|\sum_{j'\le j-2}2^{(j-j')(\epsilon-1)}\int_0^t \|u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}2^{j'(\frac{3}{p}+\epsilon)}\| \Delta_{j'} v\|_{L^p} d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C \left\|2^{j(\frac{3}{p}+\epsilon)}\int_0^t \|u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}\| \Delta_{j} v\|_{L^p} d\tau\right\|_{l^r({\mathbb{Z}})}\\
\end{aligned}$$ where we have used Young’s inequality for series for the last inequality, i.e., $$\left\|\sum_{j'\le j-2}2^{(j-j')(\epsilon-1)}c_{j'}\right\|_{l^r({\mathbb{Z}})}\le C\|2^{j(\epsilon-1)}{\bf 1}_{j\ge 2}\|_{l^1({\mathbb{Z}})}\|c_j\|_{l^r({\mathbb{Z}})}\le C\|c_j\|_{l^r({\mathbb{Z}})}.$$ Following the same argument as $I_1$, one gets $$I_2\le C\|v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}^\frac{1+\epsilon}{2}\|v\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}^\frac{1-\epsilon}{2}.$$ Finally, we bound $I_3$. By Bernstein’s inequality, Young’s inequality for series and Hölder’s inequality, we have $$\begin{aligned}
I_3\le& \left\|2^{j(\frac{3}{p}-1)}\sum_{k\ge j-3}\|\Delta_j(\Delta_{k}u\cdot\nabla \tilde{\Delta}_k v)\|_{L^1_tL^p}\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j\frac{3}{p}}\sum_{k\ge j-3}\|\Delta_j(\Delta_{k}u\otimes \tilde{\Delta}_k v)\|_{L^1_tL^p}\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|\sum_{k\ge j-3}2^{(j-k)\frac{3}{p}}2^{k\frac{3}{p}}\|\Delta_j(\Delta_{k}u\otimes \tilde{\Delta}_k v)\|_{L^1_tL^p}\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{k\frac{3}{p}}\int_0^t \|\Delta_k u\|_{L^p}\|\tilde{\Delta}_k v\|_{L^\infty} d\tau \right\|_{l^r({\mathbb{Z}})}\ (p<\infty)\\
\le& C\left\|2^{k(\frac{3}{p}+\epsilon)}\int_0^t \| u\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}\|\tilde{\Delta}_k v\|_{L^p} d\tau \right\|_{l^r({\mathbb{Z}})},
\end{aligned}$$ and using the same way as the estimate of $I_1$ derives $$I_3\le C\|v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}^\frac{1+\epsilon}{2}\|v\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}^\frac{1-\epsilon}{2}.$$ Plugging the above estimates into (\[2.4\]) leads the desired result (\[SKP2\]).\
In addition, if $r=1$, the estimate of $I_1$ can be replaced as follows: $$\begin{aligned}
I_1\le& C\left\|2^{j(\frac{3}{p}-1)}\int_0^t \|S_{j-1}u\|_{L^\infty}\|\nabla \Delta_j v\|_{L^p}d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j\frac{3}{p}}\int_0^t \|u\|_{\dot{B}_{p,1}^\frac{3}{p}}\|\Delta_j v\|_{L^p}d\tau\right\|_{l^r({\mathbb{Z}})}\\
\le& C\left\|2^{j(\frac{3}{p}+1)\frac{1}{2}}\|\Delta_j v\|_{L^1_tL^p}^\frac{1}{2}(2^{j(\frac{3}{p}-1)}\int_0^t \|u\|_{\dot{B}_{p,1}^\frac{3}{p}}^2\|\Delta_j v\|_{L^p}d\tau)^\frac{1}{2}\right\|_{l^r}\\
\le& C\|v\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}^\frac{1}{2}\|v\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}^\frac{1}{2},
\end{aligned}$$ here $f=\|v\|_{\dot{B}_{p,1}^\frac{3}{p}}^2$. At the same time, one can get the new estimates of $I_2$ and $I_3$ with the similar procedure. Thus we complete the proof of this lemma.
.3in
Proof of Theorem \[t1\] {#s3}
=======================
As the Remark \[r3\], it suffices to prove the Theorem \[t1\] under (\[1.14\]). One can get the local existence and uniqueness for (\[1.1\]) by using the standard argument on the Navier-Stokes equations, namely, there exists a $T^\star>0$, such that $$(u,B)\in \tilde{C}([0,T^\star); \dot{B}_{p,r}^{\frac{3}{p}-1})\cap \tilde{L}^1([0,T^\star); \dot{B}_{p,r}^{\frac{3}{p}+1}).$$ Since the equivalence between (\[1.1\]) and (\[1.12\]), we will consider (\[1.12\]) and suffice to prove $T^\star=\infty$.\
Now, we begin the proof. Let us consider $0<\epsilon<1$ and $r\le \frac{2}{1-\epsilon}$, containing all cases in (\[1.16\]) except $(\epsilon,r)=(0,1).$ Define $$\label{3.1}
\bar{T}:=\sup\left\{t\in (0,T^\star):\ \|W^-\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+\nu_+\|W^-\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\le \epsilon_o \nu_+ \right\},$$ where $\epsilon_0$ is small positive constant and will be determined later on.\
[**Step 1. The estimate of $W^+$.**]{} Consider the first equation in (\[1.12\]), using (\[Ber\]), we get $$\frac{d}{dt}\|\Delta_j W^+\|_{L^p}+c\nu_+ 2^{2j}\|\Delta_j W^+\|_{L^p}\le C\|\Delta_j (W^-\cdot\nabla W^+)\|_{L^p}+C\nu_- 2^{2j}\|\Delta_j W^-\|_{L^p},$$ which yields by a standard procedure $$\begin{aligned}
\|W^+\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}&+c\nu_+\|W^+\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\\
\le& 2\|W^+_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\|W^-\cdot\nabla W^+\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}
+C\nu_-\|W^-\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}.
\end{aligned}$$ By (\[SKP1\]) and (\[3.1\]), we have for all $t\in (0,\bar{T}],$ $$\begin{aligned}
\|W^+\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+&c\nu_+\|W^+\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
\le 2\|W^+_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\nu_-\|W^-\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\\
&+C(\|W^-\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}
\|W^+\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}+\|W^+\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}
\|W^-\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})})\\
\le& 2\|W^+_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\epsilon_0 ( \nu_+\|W^+\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}+
\|W^+\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})})+C\epsilon_0\nu_-,\\
\end{aligned}$$ with the selection of $\epsilon_0<\min\{\frac{c}{2C},\frac{1}{2C}\}$ leads $$\label{3.2}
\|W^+\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+c\nu_+\|W^+\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
\le 4\|W^+_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+2c\nu_-.$$
[**Step 2. The estimate of $W^-$.**]{} Denote $$f(t):=\|W^+(t)\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}^\frac{2}{1-\epsilon},\ W_\lambda^\pm:=W^\pm\exp\{-\lambda\int_0^t f(\tau) d\tau\},
\ p_\lambda:=p\exp\{-\lambda\int_0^t f(\tau) d\tau\},$$ where $\lambda$ is large enough constant and will be determined later on. So we can rewrite the second equation in (\[1.12\]) as $$\partial_t W^-_\lambda+\lambda f(t)W^-_\lambda+W^+\cdot\nabla W^-_\lambda+\nabla p_\lambda-\nu_+\Delta W^-_\lambda=\nu_-\Delta W^+_\lambda.$$ By a similar procedure, we have $$\begin{aligned}
\|\Delta_j W^-_\lambda\|_{L^\infty_t L^p}+&\lambda \int_0^t f(\tau)\|\Delta_j W^-_\lambda\|_{L^p} d\tau+c\nu_+2^{2j}\|\Delta_j W^-_\lambda\|_{L^1_tL^p}\\
&\le \|\Delta_j W_0^-\|_{L^p}+C\|\Delta_j(W^+\cdot\nabla W^-_\lambda)\|_{L^1_tL^p}+C\nu_-2^{2j}\|\Delta_j W^+_\lambda\|_{L^1_tL^p}.
\end{aligned}$$ Then we obtain $$\begin{aligned}
\|W^-_\lambda\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+&c\nu_+\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
+\lambda \|W^-_\lambda\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}\\
&\le \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\nu_-\|W^+_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
+C\|W^+\cdot\nabla W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}.
\end{aligned}$$ Thanks to (\[SKP2\]), and by Young’s inequality, we obtain $$\label{3.3}
\begin{aligned}
&\ \ \ \|W^-_\lambda\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+c\nu_+\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
+\lambda \|W^-_\lambda\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}\\
\le& \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\nu_-\|W^+_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
+C\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}^\frac{1+\epsilon}{2}
\|W^-_\lambda\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}^\frac{1-\epsilon}{2}\\
\le& \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\nu_-\|W^+_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
+\frac{c\nu_+}{8}\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\\
&+C\nu_+^{-\frac{1+\epsilon}{1-\epsilon}}\|W^-_\lambda\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}.
\end{aligned}$$ Choosing $\lambda>2C\nu_+^{-\frac{1+\epsilon}{1-\epsilon}},$ absorbing the third and fourth term on the right hand side of last inequality by the left hand side in (\[3.3\]) follows $$\begin{aligned}
\|W^-_\lambda\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}&+\frac{7c}{8}\nu_+\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
+C\nu_+^{-\frac{1+\epsilon}{1-\epsilon}}\|W^-_\lambda\|_{\tilde{L}^1_{t,f}(\dot{B}_{p,r}^{\frac{3}{p}-1})}\\
\le& \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\nu_-\|W^+_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}.
\end{aligned}$$ Obviously, using (\[3.2\]), we have $$\begin{aligned}
\|W^-_\lambda\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+c\nu_+\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}
\le&2\|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+C\nu_-\|W^+_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\\
\le& C\left( \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\frac{\nu_-}{\nu_+}(\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\nu_-)\right).
\end{aligned}$$ This yields, after using (\[3.2\]) again, for all $t\in(0,\bar{T}),$ $$\begin{aligned}
&\ \ \ \|W^-_\lambda\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+c\nu_+\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\\
\le& C\left( \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\frac{\nu_-}{\nu_+}(\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\nu_-)\right)
\exp\left\{C\nu_+^{-\frac{1+\epsilon}{1-\epsilon}}\int_0^t \|W^+(\tau)\|_{\dot{B}_{p,\infty}^{\frac{3}{p}-\epsilon}}^\frac{2}{1-\epsilon}d\tau
\right\}\\
\le& C\left( \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\frac{\nu_-}{\nu_+}(\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\nu_-)\right)
\exp\left\{C\nu_+^{-\frac{1+\epsilon}{1-\epsilon}}\|W^+\|_{\tilde{L}^\frac{2}{1-\epsilon}_t(\dot{B}_{p,r}^{\frac{3}{p}-\epsilon})}^\frac{2}{1-\epsilon}\right\}\\
\le& C\left( \|W_0^-\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\frac{\nu_-}{\nu_+}(\|W_0^+\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}}+\nu_-)\right)
\exp\left\{C\nu_+^{-\frac{2}{1-\epsilon}}(\nu_-+\|W^+_0\|_{\dot{B}_{p,r}^{\frac{3}{p}-1}})^\frac{2}{1-\epsilon}\right\},
\end{aligned}$$ which implies that if we take $\eta$ small enough in (\[1.14\]), there holds for all $t\le \bar{T}$, $$\|W^-_\lambda\|_{\tilde{L}^\infty_t(\dot{B}_{p,r}^{\frac{3}{p}-1})}+\nu_+\|W^-_\lambda\|_{\tilde{L}^1_t(\dot{B}_{p,r}^{\frac{3}{p}+1})}\le C\eta \nu_+<\frac{\epsilon_0}{2}\nu_+.$$ Then by a standard continuous method, we get $\bar{T}=T^\star=\infty.$ .1in The remainder is $r=1, \epsilon=0$, by a similar arguments, using (\[SKP2\]) for this case and let $(\epsilon,r)=(0,1)$, $f=\|W^+\|_{\dot{B}_{p,1}^\frac{3}{p}}^2$ in (\[3.3\]), the desired result can be otained. Hence, we complete the proof of Theorem \[t1\].
.3in
Proof of Theorem \[t2\] {#s4}
=======================
One can easily get the local well-posedness of (\[1.1\]), that is, there exists a $T^\star>0$ such that $$(u,B)\in C([0,T^\star); \chi^{-1}({\mathbb{R}}^3))\cap L^1([0,T^\star); \chi^1({\mathbb{R}}^3)).$$ So we suffices to show $T^\star=\infty$. .1in Now, we begin the proof. (\[1.21\]) is indeed equal to $$\label{4.1}
\left(\|W_0^-\|_{\chi^{-1}}+\frac{C\nu_-}{\nu_+}(\nu_-+\|W_0^+\|_{\chi^{-1}})\right)\exp\left\{\frac{C}{\nu_+^2}(\nu_-+\|W_0^+\|_{\chi^{-1}})^2\right\}\le(2-\epsilon_0)\nu_+$$ for some $\epsilon_0>0.$ And next we suffices to prove the desired result under (\[4.1\]). Let $C_1,C_2\in (0,2)$ satisfying $2(2-\epsilon_0)^2<C_2(2-C_1)^2$ and $$a=C_2 \nu_+,\ \ a_1=C_1 \nu_+,\ b\in (\frac{2(2-\epsilon_0)}{2-C_1}\nu_+,\sqrt{2C_2}\nu_+).$$ Then consider the first equation in (\[1.12\]), by the procedure as [@LL], and using interpolation inequality, we have $$\begin{aligned}
\frac{d}{dt}\|W^+\|_{\chi^{-1}}+&\nu_+\|W^+\|_{\chi^1}\le \|W^+\cdot\nabla W^-\|_{\chi^{-1}}+\nu_-\|W^-\|_{\chi^1}\\
\le& \|W^+\|_{\chi^0}\|W^-\|_{\chi^0}+\nu_-\|W^-\|_{\chi^1}\\
\le& \|W^-\|_{\chi^0}^2\|W^+\|_{\chi^{-1}}^\frac{1}{2}\|W^+\|_{\chi^1}^\frac{1}{2}+\nu_-\|W^-\|_{\chi^1}\\
\le & \frac{1}{2a}\|W^-\|_{\chi^0}^2\|W^+\|_{\chi^{-1}}+\frac{a}{2}\|W^+\|_{\chi^1}+\nu_- \|W^-\|_{\chi^1},
\end{aligned}$$ which derives by integrating in time, $$\label{4.12}
\begin{aligned}
\|W^+\|_{L^\infty_t(\chi^{-1})}+&(\nu_+-\frac{a}{2})\|W^+\|_{L^1_t(\chi^1)}\\
\le& \frac{1}{2a}\|W^-\|_{L^2_t(\chi^0)}^2\|W^+\|_{L^\infty_t(\chi^{-1})}
+\nu_-\|W^-\|_{L^1_t(\chi^1)}+\|W_0^+\|_{\chi^{-1}}.
\end{aligned}$$ Define $$\label{4.2}
\bar{T}:=\sup\left\{t\in (0,T^\star):\ \|W^-\|_{L^\infty_t(\chi^{-1})}+\nu_+\|W^-\|_{L^1_t(\chi^1)}\le b \right\}.$$ Then we will prove $T^\star=\bar{T}=\infty$ under (\[4.1\]). Using (\[4.2\]), combining with (\[4.12\]), we have $$\label{4.3}
(1-\frac{b^2}{2a\nu_+})\|W^+\|_{L^\infty_t(\chi^{-1})}+(\nu_+-\frac{a}{2})\|W^+\|_{L^1_t(\chi^1)}\le \|W_0^+\|_{\chi^{-1}}+\frac{b\nu_-}{\nu_+}.$$ Following the similar way as (\[4.12\]), one gets $$\frac{d}{dt}\|W^-\|_{\chi^{-1}}+\nu_+\|W^-\|_{\chi^1}
\le \frac{1}{2a_1}\|W^+\|_{\chi^0}^2\|W^-\|_{\chi^{-1}}+\frac{a_1}{2}\|W^-\|_{\chi^1}+\nu_- \|W^+\|_{\chi^1}$$ and thanks to (\[4.3\]), $$\begin{aligned}
\|W^-(t)\|_{\chi^{-1}}+&(\nu_+-\frac{a_1}{2})\|W^-\|_{L^1_t(\chi^1)}\\
\le& \frac{1}{2a_1}\int_0^t \|W^+\|_{\chi^0}^2\|W^-\|_{\chi^{-1}} d\tau
+\frac{\nu_-}{\nu_+-\frac{a}{2}}(\frac{b\nu_-}{\nu_+}+\|W_0^+\|_{\chi^{-1}})+\|W^{-}_0\|_{\chi^{-1}},
\end{aligned}$$ with the application of Gronwall’s lemma, by interpolation’s inequality and (\[4.3\]) leads $$\begin{aligned}
&\ \ \ \ \ \ \|W^-\|_{L^\infty_t(\chi^{-1})}+(\nu_+-\frac{a_1}{2})\|W^-\|_{L^1_t(\chi^1)}\\
\le& (\|W_0^-\|_{\chi^{-1}}+\frac{\nu_-}{\nu_+-\frac{a}{2}}(\frac{b\nu_-}{\nu_+}+\|W_0^+\|_{\chi^{-1}})\exp\left\{\frac{1}{2a_1}\int_0^t \|W^+\|_{\chi^0}^2 d\tau \right\}\\
\le& (\|W_0^-\|_{\chi^{-1}}+\frac{\nu_-}{\nu_+-\frac{a}{2}}(\frac{b\nu_-}{\nu_+}+\|W_0^+\|_{\chi^{-1}})\exp\left\{\frac{1}{2a_1}\|W^+\|_{L^\infty_t(\chi^{-1})}
\|W^+\|_{L^1_t(\chi^1)}\right\}\\
\le& (\|W_0^-\|_{\chi^{-1}}+\frac{\nu_-}{\nu_+-\frac{a}{2}}(\frac{b\nu_-}{\nu_+}+\|W_0^+\|_{\chi^{-1}})\exp\left\{\frac{2a\nu_+}{a_1(2a\nu_+-b^2)(2\nu_+-a_1)}
(\frac{b\nu_-}{\nu_+}+\|W^+_0\|_{\chi^{-1}})^2
\right\}.
\end{aligned}$$ which indicates that there exists constant $C$ such that $$\begin{aligned}
\|W^-\|_{L^\infty_t(\chi^{-1})}+&(1-\frac{C_1}{2})\nu_+\|W^-\|_{L^1_t(\chi^1)}\\
\le& \left(\|W_0^-\|_{\chi^{-1}}+\frac{C\nu_-}{\nu_+}(\nu_-+\|W_0^+\|_{\chi^{-1}})\right)\exp\left\{\frac{C}{\nu_+^2}(\nu_-+\|W_0^+\|_{\chi^{-1}})^2\right\}\\
\le& (2-\epsilon_0)\nu_+.
\end{aligned}$$ This implies that $$\|W^-\|_{L^\infty_t(\chi^{-1})}+\nu_+\|W^-\|_{L^1_t(\chi^1)}<\frac{2(2-\epsilon_0)}{2-C_1}<b.$$ Therefore, by standard continuous method, we get $T^\star=\bar{T}=\infty.$ This concludes the proof of Theorem \[t2\].
.4in
[99]{}
H. Bahouri, J.-Y. Chemin, R. Danchin, *Fourier Analysis and Nonlinear Partial Differential Equations*, Grundlehren der mathematischen Wissenschaften, Springer, Heidelberg, 2011.
J. Bourgain and N. Pavlovic, Ill-posedness for the Navier-Stokes equations in a critical Besov space in 3D, [*J. Funct. Anal. **255***]{}, (2008), 2233-2247.
J.-Y. Chemin, I. gallagher, Wellposedness and stability results for the Navier-Stokes equations in ${\mathbb{R}}^3$, [*Ann. Inst. H. Poincaré Anal. Non Linéaire **26***]{}, (2009), 599-624
J.-Y. Chemin, D.S. McCormick, J.C. Robinson, J.L. Rodrigo, Local existence for the non-resistive MHD equations in Besov space, arXiv:1503.01651v1 \[math.AP\] 5 Mar 2015.
Q. Chen, C. Miao and Z. Zhnag, On the regularity criterion of weak solution for the 3D viscous magneto-hydrodynamics equations. [*Comm. Math. Phys. **284***]{}, (2008), 919-930.
R. Danchin, Local theory in critical spaces for compressible viscous and heat-conductive gases, [*Comm. Partial Differential Equations **26***]{}, (2001), 1183-1233.
G. Duraut and J.L. Lions, Inéqualition en thermoélasticité et magnéto-hydrodynamic equations, [*Arch. Ration. Mech. Anal. **46***]{} (1972), 241-247.
C.L. Fefferman, D.S. McCormick, J.C. Robinson, J.L. Rodrigo, Higher order commutator estimates and local existence for the non-resistive MHD equations and related models, [*J. Funct. Anal. **267***]{} (2014), 1035-1056.
L. Grafakos, *Modern Fourier Analysis.* 2nd Edition., Grad. Text in Math., **250**, Springer-Verlag, 2008.
C. He, X. Huang and Y. Wang, On some new global existence results for 3D magnetohydrodynamic equations, [*Nonlinearity **27***]{}, (2014), 343-352.
C. He and Z.P. Xin, On the regularity of solutions to magnetohydrodynamic equations, [*J. Diff. Eqns. **213***]{}, (2005), 235-254.
Z. Lei and F.H. Lin, Global mild solutions of Navier-Stokes equations, [*Comm. Pure Appl. Math. **64***]{}, (2011), 1297-1304.
F.H. Lin and P. Zhang, Global small solutions to an MHD-type system: the three-dimensional case, [*Comm. Pure Appl. Math. **67***]{}, (2014), 531-580.
F.H. Lin and T. Zhang, Global small solutions to a complex fluid model in three dimensional, [*Arch. Ration. Mech. Anal. **216***]{}, (2015), 905-920.
X. Ren, J. Wu, Z. Xiang and Z. Zhang, Global existence and decay of smooth solution for the 2-D MHD equations without magnetic diffusion, [*J. Funct. Anal. **267***]{}, (2014), 503-541.
M. Sermenge and R. Termam, Some mathemetical questions related to the MHD equations, [*Comm. Pure Appl. Math. **46***]{} (1983), 635-664.
E.M. Stein, *Singular Integrals and Differentiability Properties of Functions*, Princeton University Press, Princeton, 1970.
R. Wan, On the uniqueness for the 2D MHD equations without magnetic diffusion, arXiv:1503.03589v1 \[math.AP\] 12 Mar 2015.
B. Wang, Ill-posedness for the Navier-Stokes equations in critical Besov spaces $\dot{B}_{\infty,q}^{-1}$, [*Adv. Math. **268***]{}, (2015), 350-372.
Y. Wang and K. Wang, Global well-posedness of the three dimensional magnetohydrodynamics equations, [*Nonlinear Anal. Real World Appl. **17***]{}, (2014), 245-251.
J. Wu, Analytic results related to magneto-hydrodynamic turbulence, [*Physica D **136***]{}, (2000), 353-372. J. Wu, Bounds and new approaches for the 3D MHD equations, [*J. Nonliear Sci **12***]{}, (2002), 395-413.
J. Wu, Regularity results for weak solutions of the 3D MHD equations, [*Discrete Contin. Dyn. Syst. **10***]{}, (2004), 543-556.
L. Xu and P. Zhnag, Global small solutions to three-dimensional incompressible magnetohydrodynamical system, [*SIAM J. Math. Anal. **47***]{}, (2015), 26-65.
Y. Zhou, Remarks on the regularities for the 3D MHD equations, [*Discrete Contin. Dyn. Syst. **12***]{}, (2005), 881-886.
Y. Zhou, Regularity criteria for the generalized viscous MHD equations. [*Ann. Inst. H. Poincaré Anal. Non Linéaire **24***]{}, (2007), 491-505.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The results of MHD plane shock waves with infinite electrical conductivity are generalized for a plasma with a finite conductivity. We derive the adiabatic curves that describe the evolution of the shocked gas as well as the change in the entropy density. For a parallel shock (i.e., in which the magnetic field is parallel to the normal to the shock front) we find an expression for the shock thickness which is a function of the ambient magnetic field and of the finite electrical conductivity of the plasma. We give numerical estimates of the physical parameters for which the shock thickness is of the order of, or greater than, the mean free path of the plasma particles in a strongly magnetized plasma.'
address: |
Departamento de Astronomía, IAG-USP, Rua do Matão 1226,\
Cidade Universitaria, CEP: 05508-900, São Paulo, SP, Brazil.
author:
- Alejandra Kandus and Reuven Opher
title: Dependence of the MHD shock thickness on the finite electrical conductivity
---
Introduction
============
Previously, shock waves in plasmas, both relativistic and non-relativistic, were studied assuming ideal MHD, [@ecm-land; @priest; @anile]. Although this theory is suitable for studying most astrophysical shock waves, such as those in hot rarified astrophysical plasmas, where the electrical conductivity is extremely high and the magnetic field is weak, it is interesting to study the effect of the simplest dissipative process in non-ideal MHD, that due to a finite value of the electrical conductivity $\sigma $. We generalize here the results for planar shock waves to non-ideal, non-relativistic MHD. The junction conditions that must be satisfied across a shock wave in a non-ideal plasma are given in Sec. II. We then derive the adiabatic curve with corrections due to $\sigma $ and find the corresponding change in the entropy density across the shock. We perform this analysis for an oblique shock (i.e., in which $\vec{B}$ is neither parallel nor perpendicular to the normal to the shock surface) as well as for a parallel one (i.e., in which $\vec{B}$ is parallel to the normal to the shock surface). In Sec. III, we find a closed expression for the shock thickness for a parallel shock in a strongly magnetized plasma, and estimate its value for some physical situations. Our conclusions are discussed in Sec. IV.
Junction conditions and adiabatic curve
=======================================
For non ideal MHD, the fluxes of mass, energy, and momentum are given by Eqs. (\[OO\]-\[c\]), respectively, [@ecm-land; @flm-land],
$$\vec{M}=\rho \vec{v}, \label{OO}$$
$$\vec{q}=\rho \vec{v}\left( \frac{1}{2}v^{2}+{\sf w}\right) +\frac{1}{4\pi }
\vec{B}\times \left( \vec{v}\times \vec{B}\right) -\frac{c^{2}}{16\pi
^{2}\sigma }\vec{B}\times \left( \vec{\nabla}\times \vec{B}\right) ,
\label{a}$$
and
$$\Pi _{ik}=\rho v_{i}v_{k}+p\delta _{ik}-\frac{1}{4\pi }\left( B_{i}B_{k}-
\frac{1}{2}B^{2}\delta _{ik}\right) , \label{b}$$
where $\rho $ is the fluid density, $\vec{v}$ the velocity, ${\sf w}$ the enthalpy per unit mass, $\vec{B}$ the ambient magnetic field, and $p$ is the fluid pressure. The electric field is $$\vec{E}=\frac{c}{4\pi \sigma }\left( \vec{\nabla}\times \vec{B}\right) -
\frac{\vec{v}}{c}\times \vec{B}, \label{c}$$ where Ohm’s law in its simplest form [@priest; @spitzer] was used.
Junction conditions
-------------------
We assume a two dimensional, planar, shock wave in the $y-z$ plane. The normal to the transition surface is in the $-x$ direction. The velocity field can be decomposed into perpendicular and tangential components to the surface of transition, $\vec{v}=\left( v_{x},\vec{v}_{t}\right) $. It is assumed that all quantities vary as a function of $x$ . Let $n^{x}$ be a unit vector normal to the transition surface. We then have the hydrodynamical junction conditions, [@flm-land],
$$\left[ \rho v_{x}n^{x}\right] =0 \label{e}$$
$$\left[ q_{x}n^{x}\right] =0 \label{f}$$
$$\left[ \Pi _{ix}n^{x}\right] =0 \label{g}$$
where $i$ is $x$ ($t$) \[normal (tangential) to the shock surface\] and $\left[ {}\right] $ means the difference between the value of the corresponding quantities far upstream (which we denote by subscript “1”) and the value at some point in the shock (no subscript). It is also assumed that both far upstream and far downstream, all gradients vanish (i.e., the fields and flows are uniform).
In the MHD case that we are considering, Eqs. (\[e\]), (\[f\]), and (\[g\]) must be suplemented with the electromagnetic junction conditions, i.e., that the normal component of the magnetic field and the tangential component of the electric field must be constant across the shock surface:
$$\left[ B_{n}\right] =0, \label{p}$$
$$\left[ \vec{E}_{t}\right] =\left[ \frac{c}{4\pi \sigma }\left( \partial
_{x}\times \vec{B}\right) _{t}-\frac{v_{x}}{c}\vec{B}_{t}+\frac{B_{n}}{c}
\vec{v}_{t}\right] =0. \label{q}$$
Eq. (\[e\]) states that the mass flux along $x$ is conserved, i.e., $\rho v_{x}=j=const$. We write $\rho =1/V$, where $V$ is the specific volume, and replace $v_{x}=jV$ in the other junction conditions, obtaining
$$j\left[ \frac{1}{2}j^{2}V^{2}+\frac{1}{2}v_{t}^{2}+{\sf w}\right] +
\frac{1}{4\pi }j\left[ VB_{t}^{2}\right] -\frac{1}{4\pi }B_{x}\left[
\vec{v}_{t}.\vec{B}_{t}\right] -\frac{c^{2}}{16\pi ^{2}\sigma }\left[
\partial _{x}B_{t}^{2}
\right] =0, \label{t}$$
$$j^{2}\left[ V\right] +\left[ p\right] +\frac{1}{8\pi }\left[ B_{t}^{2}\right]
=0, \label{u}$$
$$j\left[ \vec{v}_{t}\right] -\frac{1}{4\pi }B_{n}\left[ \vec{B}_{t}\right] =0,
\label{v}$$
$$\frac{c}{4\pi \sigma }\left[ \left( \partial _{x}\times \vec{B}_{t}\right)
\right] -\frac{j}{c}\left[ V\vec{B}_{t}\right] +\frac{B_{n}}{c}\left[
\vec{v}_{t}\right] =0. \label{w}$$
Adiabatic curve and entropy density change
------------------------------------------
We derive the expression for the adiabatic curve and the corresponding entropy density change. From Eqs. (\[v\]) and (\[w\]), we obtain
$$\frac{1}{4\pi }B_{n}^{2}\left[ \vec{B}_{t}\right] =j^{2}\left[ V\vec{B}_{t}
\right] -\frac{jc^{2}}{4\pi \sigma }\left[ \left( \partial _{x}\times
\vec{B}_{t}\right) _{t}\right] . \label{z}$$
From Eq. (\[v\]), we have $\left[ \vec{v}_{t}\right] =B_{n}\left[ \vec{B}
_{t}\right] /4\pi j$. We can therefore complete the squares in Eq. (\[t\] ), obtaining
$$\begin{aligned}
&&\left[ {\sf w}\right] +\frac{1}{2}j^{2}\left[ V^{2}\right] +\frac{1}{2}
\left[ \left( v_{t}-\frac{1}{4\pi }\frac{B_{n}}{j}\vec{B}_{t}\right) ^{2}
\right] -\frac{1}{32\pi ^{2}}\frac{B_{n}^{2}}{j^{2}}\left[ B_{t}^{2}\right]
+
\frac{1}{4\pi j}B_{n}\left[ \vec{v}_{t}.\vec{B}_{t}\right] \nonumber \\
&&+\frac{1}{4\pi }\left[ VB_{t}^{2}\right] -\frac{1}{4\pi j}B_{n}\left[
\vec{v}_{t}.\vec{B}_{t}\right] -\frac{c^{2}}{16\pi ^{2}j\sigma }\left[
\partial_{x}\left( B_{t}^{2}\right) \right] \left. =\right. 0. \label{ac}\end{aligned}$$
From Eq. (\[v\]), the third term of Eq. (\[ac\]) is zero and we are left with
$$\left[ {\sf w}\right] +\frac{1}{2}j^{2}\left[ V^{2}\right] -\frac{1}{32\pi
^{2}}\frac{B_{n}^{2}}{j^{2}}\left[ B_{t}^{2}\right] +\frac{1}{4\pi }\left[
VB_{t}^{2}\right] -\frac{c^{2}}{16\pi ^{2}j\sigma }\left[ \partial
_{x}\left( B_{t}^{2}\right) \right] =0. \label{ad}$$
Using Eq. (\[z\]), we can write the third term of Eq. (\[ad\]) as
$$\frac{1}{32\pi ^{2}}\frac{B_{n}^{2}}{j^{2}}\left[ B_{t}^{2}\right] =
\frac{1}{8\pi }\left[ VB_{t}^{2}\right] +\frac{1}{8\pi }\left( V-V_{1}\right)
\vec{B}.\vec{B}_{1}-\frac{c^{2}}{32\pi ^{2}j\sigma }\left( \vec{\nabla}\times
\vec{B}
\right) .\left( \vec{B}_{t}+\vec{B}_{1t}\right) . \label{am}$$
From momentum conservation, we obtain
$$j^{2}=-\frac{\left( p-p_{1}\right) }{\left( V-V_{1}\right) }-\frac{1}{8\pi }
\frac{\left( B_{t}^{2}-B_{1t}^{2}\right) }{\left( V-V_{1}\right) }.
\label{an}$$
Using Eqs. (\[am\]) and (\[an\]) in Eq. (\[ad\]) and ${\sf w}
=\varepsilon +pV$, we have
$$\begin{aligned}
&&\varepsilon -\varepsilon _{1}+\frac{1}{2}\left( p+p_{1}\right) \left(
V-V_{1}\right) +\frac{1}{16\pi }\left( \vec{B}_{t}-\vec{B}_{1t}\right)
^{2}\left( V-V_{1}\right) \label{aq} \\
&&+\frac{c^{2}}{32\pi ^{2}j\sigma }\left[ \left( \vec{\nabla}\times \vec{B}
\right) .\left( \vec{B}_{t}+\vec{B}_{1t}\right) -2\partial _{x}\left(
B_{t}^{2}\right) \right] \left. =\right. 0. \nonumber\end{aligned}$$
The first four terms are found in the equation for ideal MHD (e.g. Ref. [@ecm-land]), while the last two are due to the finite electrical conductivity. The first term in the square brackets is due to the fact that $B_{n}\neq 0$; the second is the contribution from the tangential component of the magnetic field.
To obtain the entropy density change, we follow the procedure found in the standard literature, [@flm-land] and develop $V-V_{1}$ in powers of $\left( p-p_{1}\right) $. We also expand $\left( {\sf w}-{\sf w}_{1}\right) $ in powers of $\left( p-p_{1}\right) $ and to first order in powers of $\left( {\sf s}-{\sf s}_{1}\right) $. The resulting expression is
$$\begin{aligned}
T\left( {\sf s}-{\sf s}_{1}\right) &=&\frac{1}{12}\left( \frac{\partial
^{2}V}{\partial p^{2}}\right) _{s}\left( p-p_{1}\right) ^{3}-\frac{1}{16\pi }
\left( \frac{\partial V}{\partial p}\right) _{s}\left( B_{t}-B_{1t}\right)
^{2}\left( p-p_{1}\right) \nonumber \\
&&-\frac{c^{2}}{32\pi ^{2}j\sigma }\left[ \left( \vec{\nabla}\times \vec{B}
\right) .\left( \vec{B}_{t}+\vec{B}_{1t}\right) -2\partial _{x}\left(
B_{t}^{2}\right) \right] . \label{au}\end{aligned}$$
The first two terms are found in the equation for ideal MHD and the last two are the corrections due to a finite $\sigma $. In the following section, we repeat the calculations for a perpendicular shock and find that in the expressions for the adiabatic curve and the entropy density change, the only term present which depends on the conductivity is the last one.
Perpendicular shock
-------------------
Shock waves in a plasma permeated with a magnetic field show several features, of which the most well known is related to the orientation of the magnetic field with respect to the shock plane. Although perpendicular shocks can be considered to be a special case of oblique shocks, it is interesting to write the simplified expressions for the junction conditions explicitly, and re-derive the adiabatic curve and the entropy density change for this case.
### Hydrodynamical and electromagnetic junction conditions
For perpendicular shocks $B_{n}=0$, so that the junction conditions now read
$$\left[ \rho v_{x}\right] =0, \label{av}$$
$$\left[ \rho v_{x}\left( \frac{1}{2}v^{2}+{\sf w}\right) +\frac{1}{4\pi }
v_{x}B_{t}^{2}-\frac{c^{2}}{16\pi ^{2}\sigma }\partial _{x}\left(
B^{2}\right) \right] =0, \label{aw}$$
$$\left[ \rho v_{x}^{2}+p+\frac{1}{8\pi }B_{t}^{2}\right] =0, \label{ax}$$
$$\left[ \rho \vec{v}_{t}v_{x}\right] =0\Rightarrow \vec{v}_{t}=\vec{v}_{t1},
\label{ay}$$
$$\frac{c}{4\pi \sigma }\left[ \left( \partial _{x}\times \vec{B}_{t}\right)
\right] -\frac{1}{c}\left[ v_{x}\vec{B}_{t}\right] =0. \label{ba}$$
### Adiabatic curve and entropy density change
Proceeding as for an oblique shock and defining
$$\varepsilon ^{\ast }=\varepsilon +\frac{B_{t}^{2}V}{8\pi },\quad p^{\ast }=
p+\frac{1}{8\pi }B_{t}^{2}, \label{bi}$$
we obtain
$${\sf w}-{\sf w}_{1}-\frac{1}{2}\left[ p^{\ast }\right] \left( V+V_{1}\right)
+\frac{1}{4\pi }\left[ VB_{t}^{2}\right] -\frac{c^{2}}{16j\pi ^{2}\sigma }
\left[ \partial _{x}\left( B_{t}^{2}\right) \right] \left. =\right. 0
\label{be}$$
and
$$\varepsilon ^{\ast }-\varepsilon _{1}^{\ast }+\frac{1}{2}\left( p^{\ast
}+p_{1}^{\ast }\right) \left( V-V_{1}\right) -\frac{c^{2}}{16j\pi ^{2}\sigma
}\partial _{x}\left( B_{t}^{2}\right) =0. \label{bj}$$
There is now only one term that depends on the dissipative properties of the plasma, while for the oblique case, we had two such terms. The missing term is related to the normal component of the magnetic field.
Thickness of the shock wave
===========================
The calculation of a general expression for the shock thickness is very difficult, if not impossible. However for a perpendicular shock it is possible to calculate the shock thickness exactly. We then have $B_{n}=0$, $\vec{v}_{t}=0$ and consider a coordinate system in which the only non-zero component of the magnetic field is $B_{y}=B$, [@ecm-land]. In this case, the equation $\vec{\nabla}.\vec{B}=0$ is satisfied identically. The unidimensional ideal MHD equations are
$$\frac{\partial B}{\partial t}=\frac{\partial }{\partial x}\left(
v_{x}B\right) , \label{bs}$$
$$\frac{\partial \rho }{\partial t}+\frac{\partial }{\partial x}\left(
v_{x}\rho \right) =0, \label{bt}$$
$$\frac{\partial v_{x}}{\partial t}+v_{x}\frac{\partial v_{x}}{\partial x}+
\frac{1}{8\pi \rho }\frac{\partial B^{2}}{\partial x}=-\frac{1}{\rho }
\frac{
\partial p}{\partial x}. \label{bu}$$
From the first two equations, it is easy to see that the ratio $B/\rho \equiv \beta $ satisfies the equation $\partial \beta /\partial
t+v_{x}\partial \beta /\partial x=0$ or $d\beta /dt=0$ [@ecm-land]. Hence, if the fluid is homogeneous at some initial instant, so that $\beta
=const.$, then it will remain so at all subsequent times. Substituting $B=\rho \beta $ in the third equation, we obtain
$$\frac{\partial v_{x}}{\partial t}+v_{x}\frac{\partial v_{x}}{\partial x}=-
\frac{1}{\rho }\frac{\partial }{\partial x}\left[ p+
\frac{\beta ^{2}\rho ^{2}}{8\pi }\right] \label{bv}$$
Thus, the magnetic field has been eliminated from the equations. The equation for the velocity field, Eq. (\[bv\]), is formally identical to that for the ideal fluid case, provided we define the ‘true pressure’ as $p^{\ast }=p+\beta ^{2}\rho ^{2}/8\pi $. We can now proceed to evaluate the thickness, following Ref. [@flm-land]. We write
$$\frac{\partial }{\partial t}\delta p^{\ast }-v_{s}^{\ast }
\frac{\partial }{\partial x}\delta p^{\ast }-\alpha _{p}^{\ast }
\delta p^{\ast }\frac{\partial }{\partial x}\delta p^{\ast }=
cL^{\ast }\frac{\partial ^{2}}{\partial x^{2}}\delta p^{\ast }, \label{bw}$$
where $v_{s}^{\ast 2}\equiv \left( \partial p^{\ast }/\partial \rho \right)
_{s}$ and $L^{\ast }$ is the damping length, given by Eq. (\[a45\]) of the Appendix,
$$L^{\ast }=\frac{c}{8\pi \sigma }\frac{B_{0y}^{2}}{\left( B_{0y}^{2}+4\pi
\rho _{0}v_{s}^{2}\right) }. \label{bx}$$
From Ref. [@flm-land] we have
$$\alpha _{p}^{\ast }=\frac{1}{2}v_{s}^{\ast 3}\rho ^{2}\left[ \frac{\partial
^{2}}{\partial p^{\ast 2}}\left( \frac{1}{\rho }\right) \right] _{s}.
\label{by}$$
Equation (\[bw\]) can be solved using the procedure in Ref. [@flm-land] , obtaining the thickness of the shock wave as
$$\delta ^{\ast }=\frac{4cL^{\ast }}{\alpha _{p}^{\ast }\left( p_{2}^{\ast
}-p_{1}^{\ast }\right) }, \label{bz}$$
where $p_{2}^{\ast }$ and $p_{1}^{\ast }$ are the ‘true pressures’ far downstream and far upstream respectively. Equation (\[bz\]) is quantitatively valid for sufficiently small differences $\left( p_{2}^{\ast
}-p_{1}^{\ast }\right) $. However we can use it qualitatively to estimate the order of magnitude of the thickness in cases where the difference $\left( p_{2}^{\ast }-p_{1}^{\ast }\right) $ is of the same order of magnitude as $p_{2}^{\ast }$ and $p_{1}^{\ast }$ themselves. The velocity of sound in the gas, $v_{s}$ (not $v_{s}^{\ast }$) is of the same order of magnitude as the thermal velocity $v$. Let $\lambda $ be the mean free path of the atoms in the plasma. Then from dimensional analysis, the electric conductivity can be estimated as, $\sigma \sim \gamma v/\lambda \sim \gamma
v_{s}/\lambda $, where $\gamma $ takes into account anomalous effects and can have a value $10^{-6}\leq \gamma \leq 1$ [^1].
In Eq. (\[bx\]), we take $B_{0y}^{2}/\left( B_{0y}^{2}+4\pi \rho
_{0}v_{s}^{2}\right) \sim B_{0y}^{2}/p^{\ast }$ and in Eq. (\[by\]) $\alpha _{p}^{\ast }\left( p_{2}^{\ast }-p_{1}^{\ast }\right) \sim
v_{s}^{\ast 2}\sim p^{\ast }/\rho $. Using these relations in Eq. (\[bz\]) we obtain
$$\delta \sim c^{2}\frac{\rho ^{2}B_{0y}^{2}}{\gamma p^{\ast 2}p}\lambda .
\label{ca}$$
The shock thickness is larger than the mean free path when $c^{2}\rho
^{2}B_{0y}^{2}\geq \gamma p^{\ast 2}p$. Let us first assume that $B_{0y}^{2}\geq p$. We then have $p^{\ast 2}\sim B_{0y}^{4}$ and $p\leq B_{0y}^{2}\leq c^{2}\rho ^{2}/\gamma p$. As a specific numerical example, consider $\rho \sim 10^{2}$ gr/cm$^{3}$ and $T\sim 10^{8}$ K (characteristic parameters at the center of a massive star before collapse). Assuming hydrogen gas, we have $n\sim 10^{26}$ cm$^{-3}$ and $p\sim nT\sim 10^{18}$erg/cm$^{3}$ (for iron nuclei, the pressure would be two orders of magnitude smaller). For the above densities and pressures, the shock thickness is larger than the mean free path of the particles if the magnetic field is in the interval $10^{9}$ G $\leq B_{0y}\leq \gamma
^{-1/2}10^{14}$ G. If we now assume that $B_{0y}^{2}\leq p$, the shock thickness is larger than the mean free path if $\gamma p^{3}/c^{2}\rho
^{2}\leq B_{0y}^{2}$. For the above parameters we have $\gamma
^{1/2}10^{5}$ G $\leq B_{0y}\leq 10^{10}$ G [^2]. If neither of the two conditions above are fulfilled, the shock thickness is smaller than the mean free path. This means that the MHD approach breaks down and kinetic theory is needed to study the structure of the shock.
Conclusions
===========
In this article, we extended the results of shock waves treated in ideal MHD to the non-ideal case, in which the electrical conductivity is finite. We considered Ohm’s law in its simplest form, [@spitzer], but took into account phenomenologically (through the parameter $\gamma $) plasma effects that can modify the classical Spitzer electrical conductivity (e.g., turbulence). The expressions for the adiabatic curve and the entropy density change across the shock were generalized. Finally, we derived the expression for the shock thickness for a finite conductivity in the case of a parallel shock in strongly magnetized plasmas. The conditions that the ambient magnetic field must satisfy for the thickness to be of the order of the particle mean free path were estimated. We found that these conditions can be fulfilled for the plasma expected in the origin of a supernova explosion, [@star-1; @star-2]. Extensions of the results presented in this paper, using a more general Ohm’s law, as well as to relativistic shocks, are presently under investigation.
This work was partially supported by the Brazilian financing agency FAPESP (00/06770-2). A.K. acknowledges the FAPESP fellowship (01/07748-3). R. O. acknowledges partial support from the Brazilian financing agency CNPq (300414/82-0).
Appendix
========
In this appendix, we sketch the derivation of the expression for $L^{\ast }$ , the damping length used to calculate the shock thickness. Neglecting the displacement current (which is a good approximation in non-relativistic electrodynamics), the evolution equation for the magnetic field in a medium with electrical conductivity $\sigma $ moving with a velocity $\vec{v}$ is
$$\frac{\partial \vec{B}}{\partial t}-\vec{\nabla}\times \left( \vec{v}\times
\vec{B}\right) =\frac{c^{2}}{4\pi \sigma }\nabla ^{2}\vec{B}. \label{a5}$$
Adding the equations for the fluid, which we assume has neither viscosity nor thermal conduction, we have
$$\frac{\partial \rho }{\partial t}+\vec{\nabla}.\left( \rho \vec{v}\right) =0,
\label{a6}$$
$$\frac{\partial \vec{v}}{\partial t}+\left( \vec{v}.\vec{\nabla}\right)
\vec{v}=-\frac{1}{\rho }\vec{\nabla}p-\frac{1}{4\pi \rho }\vec{B}\times
\left( \vec{\nabla}\times \vec{B}\right) . \label{a7}$$
Hydromagnetic waves
-------------------
Let us assume that $B=\vec{B}_{0}+\vec{b}$, $\rho =\rho _{0}+\delta \rho $, $p=p_{0}+\delta p$, and $\vec{v}=\delta \vec{v}$. Replacing these terms in the above equations, keeping only terms to first order in the perturbations, expanding the density in powers of the perturbation in the pressure (i.e., $\delta \rho =\delta p/v_{s}^{2}+\left( \partial ^{2}\rho /\partial
p^{2}\right) _{s}\delta p^{2}$, where $v_{s}$ is the sound velocity of the medium) and taking the Fourier transform of the equations, we obtain
$$-\omega \delta p_{0}+v_{s}^{2}\rho _{0}\vec{k}.\delta \vec{v}_{0}=0,
\label{a23}$$
$$-\left( \omega +i\frac{c^{2}}{4\pi \sigma }k^{2}\right) \vec{b}_{0}=
\vec{k}\times \left[ \delta \vec{v}_{0}\times \vec{B}_{0}\right] ,
\label{a24}$$
$$-\omega \delta \vec{v}_{0}=-\frac{\vec{k}}{\rho _{0}}\delta p_{0}-
\frac{1}{4\pi \rho _{0}}\vec{B}_{0}\times \left( \vec{k}\times \vec{b}_{0}
\right) .
\label{a25}$$
From Eq. (\[a23\]), we find $\delta p_{0}=v_{s}^{2}\rho _{0}\left(
\vec{k}.\delta \vec{v}_{0}\right) /\omega $ and using this in Eq. (\[a25\] ), we obtain
$$-\delta \vec{v}_{0}=-\frac{\vec{k}}{\omega }v_{s}^{2}\left(
\frac{\vec{k}}{\omega }.\delta \vec{v}_{0}\right) -
\frac{1}{4\pi \rho _{0}}\vec{B}_{0}\times \left(
\frac{\vec{k}}{\omega }\times \vec{b}_{0}\right) .
\label{a26}$$
We define the scalar phase velocity as $u=\omega /k$, assuming that $\vec{k}$ is along the $x$-axis, (i.e., $\vec{k}=k\check{x}$) and that $\vec{B}_{0}$ is in the $x-y$ plane. Writing the previous equations in its components, we have
$$\delta p_{0}=\frac{v_{s}^{2}\rho _{0}}{u_{A}}\delta v_{0x}, \label{a27}$$
$$\left( u-\frac{v_{s}^{2}}{u}\right) \delta v_{0x}=\frac{1}{4\pi \rho _{0}}
b_{0y}B_{0y}, \label{a28}$$
$$u\delta v_{0y}=-\frac{1}{4\pi \rho _{0}}b_{0y}B_{0x}, \label{a29}$$
$$\left( u+i\frac{c^{2}}{4\pi \sigma }k\right) b_{0y}=\delta
v_{0x}B_{0y}-\delta v_{0y}B_{0x}, \label{a30}$$
$$u\delta v_{0z}=-\frac{1}{4\pi \rho _{0}}b_{0z}B_{0x}, \label{a31}$$
$$\left( u+i\frac{c^{2}}{4\pi \sigma }k\right) b_{0z}=-\delta v_{0z}B_{0x}.
\label{a32}$$
Generalized Alfven waves
------------------------
Using Eqs. (\[a31\]) and (\[a32\]) we obtain the compatibility relationship
$$u^{2}+i\frac{c^{2}k}{4\pi \sigma }u-\frac{B_{0x}^{2}}{4\pi \rho _{0}}=0,
\label{a33}$$
from which we obtain
$$u=\pm \frac{1}{2}\sqrt{\frac{B_{0x}^{2}}{\pi \rho _{0}}-
\frac{c^{4}k^{2}}{16\pi ^{2}\sigma ^{2}}}-i\frac{c^{2}k}{8\pi \sigma }.
\label{a34}$$
From Eq. (\[a34\]), the phase velocity is a complex number if $\sigma $ is finite. Rewriting $u_{A}$ in terms of $\omega $, we obtain the dispersion relationship:
$$\omega =\frac{1}{2}\frac{B_{0x}k}{\sqrt{\pi \rho _{0}}}\sqrt{1-\frac{\rho
_{0}}{B_{0x}^{2}}\frac{c^{4}k^{2}}{16\pi \sigma ^{2}}}-i
\frac{c^{2}k^{2}}{8\pi \sigma }. \label{a35}$$
We take the plus sign in eq. (\[a34\]) since the frequency is a positive quantity. The fact that the imaginary part is non-linear in $k$ means that the Alfven waves are damped and dissipated as a function of $k$. For $\sigma \rightarrow \infty $ we recover the known dispersion relationship for ideal MHD.
Assuming that the second term in the square root in Eq. (\[a35\]) is much smaller than unity, the group velocity is
$$v_{A}=\frac{\partial \omega }{\partial k}\simeq \frac{B_{0x}}{\sqrt{4\pi
\rho _{0}}}-i\frac{c^{2}k}{4\pi \sigma }. \label{a37}$$
When the electrical conductivity is infinite, we recover the known ideal MHD result, $v_{AI}=B_{0x}/2\sqrt{\pi \rho _{0}}$.
Generalized magnetosonic waves
------------------------------
From Eqs. (\[a28\]), (\[a29\]) and (\[a30\]), we obtain the generalized dispersion relationship for magnetosonic waves,
$$\omega ^{4}-\left( \frac{B_{0}^{2}}{4\pi \rho _{0}}+v_{s}^{2}\right)
k^{2}\omega ^{2}+\frac{B_{0x}^{2}}{4\pi \rho _{0}}v_{s}^{2}k^{4}+i
\frac{c^{2}k^{2}}{4\pi \sigma }\omega ^{3}-i
\frac{c^{2}v_{s}^{2}k^{4}}{4\pi \sigma
}\omega =0. \label{a38}$$
This relationship can be inverted to obtain $\omega =\omega \left( k\right) $ . However, for our purposes, it suffices to consider the dissipative terms as a correction to the ideal dispersion relationship,
$$\omega _{I}^{2}=\frac{1}{2}\left[ \left( \frac{B_{0}^{2}}{4\pi \rho _{0}}
+v_{s}^{2}\right) \pm \sqrt{\left( \frac{B_{0}^{2}}{4\pi \rho _{0}}
+v_{s}^{2}\right) ^{2}-\frac{B_{0x}^{2}}{\pi \rho _{0}}v_{s}^{2}}\right]
k^{2}\equiv v_{g0}^{2}k^{2}. \label{a40}$$
The plus sign corresponds to [*fast magnetosonic*]{} waves, while the minus sign to [*slow magnetosonic*]{} waves. Replacing $\omega _{I}$ in the last two terms of Eq. (\[a38\]), we obtain
$$\omega ^{4}-\left( \frac{B_{0}^{2}}{4\pi \rho _{0}}+v_{s}^{2}\right)
k^{2}\omega ^{2}+\frac{B_{0x}^{2}}{4\pi \rho _{0}}v_{s}^{2}k^{4}+i
\frac{c^{2}v_{g0}}{4\pi \sigma }\left( v_{g0}^{2}-v_{s}^{2}\right)
k^{5}\simeq 0.
\label{a41}$$
If the electrical conductivity is large (but not infinite), we have
$$\omega ^{2}\simeq \omega _{I}^{2}\mp \frac{i}{4}\frac{c^{2}v_{g0}}{\pi
\sigma }\frac{\left( v_{g0}^{2}-v_{s}^{2}\right) }{\left(
v_{B0}^{4}-v_{B0x}^{2}v_{s}^{2}\right) ^{1/2}}k^{3}, \label{a42}$$
where $v_{B0}^{2}=\left( B_{0}^{2}/4\pi \rho _{0}+v_{s}^{2}\right) $ and $v_{B0x}^{2}=B_{0x}^{2}/\pi \rho _{0}$. We thus have
$$\omega \simeq v_{g0}k-icL^{\ast }k^{2}, \label{a43}$$
with
$$L^{\ast }=\frac{c}{8\pi \sigma }\frac{\left( v_{g0}^{2}-
v_{s}^{2}\right) }{\left( v_{B0}^{4}-v_{B0x}^{2}v_{s}^{2}\right) ^{1/2}},
\label{a44}$$
where $L^{\ast }$ is the damping length. For a perpendicular shock, (i.e., $B_{0x}=0$), we have
$$L^{\ast }=\frac{c}{8\pi \sigma }\frac{B_{0t}^{2}}{\left( B_{0t}^{2}+4\pi
\rho _{0}v_{s}^{2}\right) }. \label{a45}$$
1999 [*Electrodynamics of Continuous Media*]{}. Second edition, Butterworth-Heinemann.
1982 [*Solar Magnetohydrodynamics*]{}. Reidel.
1989 [*Relativistic Hydrodynamics and Magnetohydrodynamics*]{}. Cambridge University Press.
1997 [*Fluid Mechanics*]{}. Second edition, Butterworth-Heinemann.
1962 [*Physics of Fully Ionized Gases*]{}. Second edition, John Wiley & Sons.
2001 [*Astrophys. J.*]{} [**560**]{}, 339.
2002 [ *astro-ph/0211458.*]{}
[^1]: We know that in accretion disks, protostars, galactic nuclei and neutron X-ray sources, for example, the plasma cannot have ideal Spitzer values for the conductivity and viscosity in order to obtain the observed accretion rates. Therefore it is generally assumed that these quantities are highly anomalous (due to turbulence, for example). Another example where the assumption of anomalous resistivity is used is in the treatment of solar flares, which are generally assumed to be due to magnetic reconnection. If ideal Spitzer values are used for the plasma in solar flares, reconnection times are $\sim 10^{6}$ times longer than the observed time scales for the flares. In general, plasmas near shocks are expected to be highly anomalous (i.e., $\gamma \ll 1$) due to turbulence.
[^2]: For $\gamma \sim 10^{-6}$, the magnetic field range is $10^{2}$ G $\leq B_{0y}\leq 10^{10}$ G. A magnetic field $B\geq 10^{2}$ G can be easily be present at the center of a massive star. The magnetic field increases in the collapse of the core of a massive star to $B\geq 10^{6}$ (for the collapse to a white dwarf) and to $B\geq 10^{9}$ (for the collapse to a neutron star). Thus, at the center of a massive star, we may expect that the magnetic field varie from $10^{2}$ G to $10^{10}$ G during the collapse of its core and the start of a supernova explosion.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Chris H.J. Evers, Jurriaan A. Luiken, Peter G. Bolhuis, Willem K. Kegel'
title: 'Self-assembly of Virus-like Shells via Colloidal Bond Hybridization and Anisotropy'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a comprehensive theory and an easy to follow method for the design and construction of a wideband homodyne detector for time-domain quantum measurements. We show how one can evaluate the performance of a detector in a specific time-domain experiment based on electronic spectral characteristic of that detector. We then present and characterize a high-performance detector constructed using inexpensive, commercially available components such as low-noise high-speed operational amplifiers and high-bandwidth photodiodes. Our detector shows linear behavior up to a level of over 13 dB clearance between shot noise and electronic noise, in the range from DC to 100 MHz. The detector can be used for measuring quantum optical field quadratures both in the continuous-wave and pulsed regimes with pulse repetition rates up to about 250 MHz.'
author:
- 'R. Kumar'
- 'E. Barrios'
- 'A. MacRae'
- 'E. Cairns'
- 'E. H. Huntington'
- 'A. I. Lvovsky'
title: Versatile Wideband Balanced Detector for Quantum Optical Homodyne Tomography
---
Introduction
============
The balanced homodyne detector (HD) is a useful tool in quantum optics and quantum information processing with continuous variables [@leo:97; @lvov:09] since it can be used to measure field quadratures of an electromagnetic mode. These measurements provide information for complete reconstruction of quantum states in the optical domain (optical homodyne tomography).
With developing tools of continuous-wave quantum-optical state engineering [@bim:10] as well as state and process tomography [@lob:08], the performance requirements for homodyne detectors continue to increase. The design of HDs for time-domain quantum tomography [@han:01; @vogel:89; @zav:02] is based on four main performance criteria: a) high bandwidth and a flat amplification profile within that bandwidth; b) high ratio of the measured quantum noise over the electronic noise; c) high common mode rejection ratio (CMRR); d) quantum efficiency of the photodiodes.
The *high bandwidth* requirement comes from the fact that an HD must be able to measure field quadratures with sufficient time resolution. In the case of pulsed lasers, this corresponds to the inverse of the repetition rate of the pulses; in the case of a continuous signal, the required resolution is determined by the duration of the optical mode in which the signal states are produced [@sas:06; @nie:09; @yue:83]. This is technically challenging because most amplifiers have a limited gain-bandwidth product. Increasing the bandwidth implies reducing the gain, which, in turn, increases the effect of the electronic noise. Also the high frequency circuit layout poses a challenge to designers.
Within its bandwidth range, the HD must feature a *flat amplification profile*. If this is not the case, the response of the HD to each individual pulse will exhibit ringing, which degrade the detector’s time resolution and distort the measurement. This requirement also presents a major design challenge.
Any non-desirable ambient noises, dark current noises from the photodiodes and the intrinsic noise of the amplifiers fall under the umbrella of *electronic noise*. The effect of this noise is to add a random quantity $Q_e$ to the measurement of the field quadrature $Q_{\rm meas}$. This effect is equivalent to an additional optical loss channel with transmission [@appel:07] $$\label{etae}
\eta_e=1-\langle\hat Q_e^2\rangle/\langle \hat Q_{\rm meas}^2\rangle.$$ As we show below, the value of $\eta_e$ depends not only on the characteristics of the detector, but also on the conditions of the measurement in which the detector is used.
A HD must have a *high subtraction capability* between the two photocurrents produced by the photodiodes. This can be expressed as a generalized common mode rejection ratio (CMRR) of the balanced detection [@vogel:89; @chi:11]. The CMRR measures the ability of the device to reject the classical noise of the local oscillator [@bac:04; @sas:06]. This is particularly important in the pulsed case because a low CMRR (which implies a poor subtraction) will result in contamination of the signal with the repetition rate of the pulse and harmonics. Additionally, this lack of subtraction capability will make the HD more susceptible to saturation by the amplified signal from the photodiodes. High CMRR is difficult to achieve because the response functions of the photodiodes are not exactly the same. Therefore a pair of photodiodes with response functions as similar as possible must be chosen.
Experimentally, these performance benchmarks can be measured using an electronic spectrum analyzer. The spectrum of the homodyne output photocurrent gives information about the detector’s bandwidth and amplification profile. Observing the output current in the absence of the local oscillator provides information about the magnitude and spectrum of the electronic noise. The lower bound on CMRR is determined by comparing the HD spectra when both photodiodes are illuminated and when only one is illuminated while the other is blocked.
In the present work, we quantitatively relate the measured electronic spectra to added noise in quadrature measurements. We show that the limited bandwidth and electronic noise can be translated into equivalent optical losses such as in [Eq. (\[etae\])]{}. We show how to estimate and reduce these losses for a specific time-domain experiment. In fact, in many cases (particularly, in the continuous-wave regime) electronic spectral measurements on the HD photocurrent in the presence and absence of the local oscillator are sufficient to precisely calculate the equivalent loss associated with the electronics.
The theoretical discussion in this paper is limited to the effects of the bandwidth and the electronic noise in two practically relevant regimes. The effect of the non-unitary quantum efficiency on quantum state reconstruction is well known [@leo:97]. A discussion of CMRR has been presented in detail in Ref. [@chi:11].
We then demonstrate an easy to follow method for the design and construction of a wideband homodyne detector using commercial available components such as low-noise high-speed operational amplifiers and high-bandwidth photodiodes. Aside from high performance benchmarks, a special feature of our detector is its versatility: it is designed and tested to operate in both the continuous-wave or pulsed regimes. Therefore the unit presented here may be useful for a wide range of quantum optics experiments.
Theoretical analysis
====================
Balanced homodyne detection consists of overlapping the signal mode carrying the quantum state in question and a strong reference field in a matching mode (the local oscillator, or LO) on a symmetric beam splitter. The two output signals of this beam splitter are directed to the two photodiodes of the HD, where these fields are detected and subtracted. Neglecting experimental imperfections, the subtraction photocurrent is then $$\label{idealHD}
\hat i(t)=A\alpha(t)\hat q_\theta(t),$$ where $\hat{q}_\theta(t)$ is the instantaneous field quadrature value in the signal mode, $\alpha(t)$ and $\theta$ are the local oscillator amplitude and phase, respectively, and $A$ is a proportionality coefficient related to the HD amplifier gain. It is assumed that the local oscillator phase is constant.
The instantaneous quadrature observable can be written as $$\label{qtheta}
\hat{q}_\theta(t)=\hat a(t)e^{i\theta}+\hat a^\dag(t)e^{-i\theta},$$ where $\hat a(t)$ is the time dependent photon annihilation operator[@fed:05]. In a practical HD, the relationship between the quadrature measurement and the output current is more complex. It can be approximated by $$\label{cpracHD}
\hat i(t)=\hat i_e(t)+A{\int\limits_{-\infty}^{+\infty}}\alpha(t')\hat q(t')r(t-t'){\textrm{d}}t',$$ where $i_e(t)$ is the detector’s electronic noise and $r(\cdot)$ is its response function. An ideal detector would have $i_e(t)=0$ and $r(\tau)=\delta(\tau)$. In practice these conditions are not met.
As evident from [Eq. (\[cpracHD\])]{}, the impact of the electronic noise is minimized by raising the power of the local oscillator and the amplifier gain. However, practical possibilities of increasing the gain without proportionally increasing the electronic noise are limited. The local oscillator power must also be restricted to avoid saturation of the photodiodes and eliminating the classical noise [@bac:04]. Therefore in the analysis below we assume $A$ and $\alpha$ to equal their optimal values for the given experimental setting.
Continuous regime
-----------------
In the continuous regime, the amplitude of the local oscillator is a constant: $\alpha(t)\equiv\alpha$. We are interested in measuring the quadrature of the signal field associated with a particular (normalized) temporal mode function $\phi(t)$, which we assume real: $$\label{QCV}
\hat Q={\int\limits_{-\infty}^{+\infty}}\hat q(t)\phi(t){\textrm{d}}t$$ To that end, we integrate the homodyne photocurrent with a certain weight function $\psi(t)$, obtaining a *measured* quadrature value, $$\begin{aligned}
\label{QmeasCV}
\hat Q_{\rm meas}&=&{\int\limits_{-\infty}^{+\infty}}\hat i(t)\psi(t){\textrm{d}}t \\ {\nonumber}&=&A\alpha{\int\limits_{-\infty}^{+\infty}}{\int\limits_{-\infty}^{+\infty}}\hat q(t')\psi(t)r(t-t'){\textrm{d}}t {\textrm{d}}t'+\hat Q_e,\end{aligned}$$ in which the last term, $$\label{}
\hat Q_e={\int\limits_{-\infty}^{+\infty}}\hat i_e(t)\psi(t){\textrm{d}}t,$$ corresponds to the electronic noise contribution, which we will discuss later. First, we discuss the effect of finite detector response function (bandwidth) on the quadrature measurement.
Equation can be rewritten as $$\label{QmeasCV1}
\hat Q_{\rm meas}=A\alpha{\int\limits_{-\infty}^{+\infty}}\hat q(t')\psi'(t'){\textrm{d}}t'+\hat Q_e,$$ where $$\label{}
\psi'(t')={\int\limits_{-\infty}^{+\infty}}\psi(t)r(t-t'){\textrm{d}}t.$$ By comparing Eqs. and we find that, by choosing $\psi(t)$ such that $\psi'(t)=\phi(t)$, we have $\hat Q_{\rm meas}=A\alpha\hat Q+\hat Q_e$, i.e., the distortions associated with the detector’s finite bandwidth are completely eliminated. This may however be difficult in practice, because the required weight function is a *deconvolution* of the temporal mode of interest and the detector’s response. Lack of precise knowledge of either of the above may lead to significant errors in deconvolving.
If $\psi'(t)\ne\phi(t)$, the detection efficiency is degraded by the mode matching factor [@aic:02] $$\label{MMCV}
\eta_b=\left.\left|{\int\limits_{-\infty}^{+\infty}}\psi'(t)\phi(t){\textrm{d}}t\right|^2\middle/{\int\limits_{-\infty}^{+\infty}}|\psi'(t)|^2{\textrm{d}}t\right.,$$ where the denominator normalizes $\psi'(t)$. A practically important particular case is when the temporal mode of the signal is known and the finite bandwidth of the detector is neglected, so $\psi(t)$ is set to equal $\phi(t)$. In Fig. \[MMCVfig\], the efficiency obtained in this setting is plotted for Gaussian $\phi(t)$ and $r(t)$ as a function of the detector bandwidth, which, as we show below, is obtained from the Fourier transform of the response function. As we see, the detector bandwidth has no significant degrading effect on the measurement ($\eta_b>0.99$) as long as it is comparable to or larger than the inverse temporal width of the signal temporal mode.
![Effective efficiency of the HD associated with its finite bandwidth. The 3-dB bandwidth is plotted along the horizontal axis in units of the inverse full-width-at-half-maximum of the signal temporal mode $\phi(t)$. Both functions are assumed Gaussian. \[MMCVfig\]](MMCV.eps){width="0.8\columnwidth"}
Let us now calculate the contribution of the electronic noise to the measured quadrature, so the equivalent efficiency can be estimated. We start with the electronic spectra of the HD output photocurrent in the absence and in the presence of the local oscillator, with the signal in the vacuum state. According to the Wiener-Khintchine theorem, these spectra are given, respectively, by $$\label{}
S_e(\nu)={\int\limits_{-\infty}^{+\infty}}\langle \hat i_e(t)\hat i_e(t+\tau)\rangle e^{2\pi i\nu\tau}{\textrm{d}}\tau$$ and $$\label{Snu}
S(\nu)={\int\limits_{-\infty}^{+\infty}}\langle \hat i(t)\hat i(t+\tau)\rangle e^{2\pi i\nu\tau}{\textrm{d}}\tau,$$ where $\nu$ is the frequency and the averaging is performed over both time $t$ and the quantum ensemble of the vacuum signal state. The autocorrelation function in the latter equation can be further simplified as $$\begin{aligned}
\langle \hat i(t)\hat i(t+\tau)\rangle&& \\{\nonumber}&&\hspace{-1.8cm}
= A^2\alpha^2 \left\langle{\int\limits_{-\infty}^{+\infty}}{\int\limits_{-\infty}^{+\infty}}\hat q(t')r(t-t')\hat q(t'')r(t+\tau-t''){\textrm{d}}t'{\textrm{d}}t''\right\rangle \\{\nonumber}&&\hspace{-1cm}+\langle \hat i_e(t)\hat i_e(t+\tau)\rangle\\{\nonumber}&&\hspace{-1.8cm}=A^2\alpha^2{\int\limits_{-\infty}^{+\infty}}\langle r(t-t')r(t+\tau-t')\rangle{\textrm{d}}t'+\langle \hat i_e(t)\hat i_e(t+\tau)\rangle\\{\nonumber}&&\hspace{-1.8cm}=A^2\alpha^2{\int\limits_{-\infty}^{+\infty}}r(t)r(t+\tau){\textrm{d}}t+\langle \hat i_e(t)\hat i_e(t+\tau)\rangle\end{aligned}$$ because, in the vacuum state, $\langle \hat q(t')\hat q(t'')\rangle=\delta(t'-t'')$. From the above, we find $$\label{SnuCV}
S(\nu)=A^2\alpha^2|\tilde r(\nu)|^2+S_e(\nu),$$ with $$\label{}
\tilde r(\nu)={\int\limits_{-\infty}^{+\infty}}r(t) e^{2\pi i \nu\tau}{\textrm{d}}t$$ being the Fourier image of $r(t)$. In other words, neglecting the electronic noise, the spectrum of the HD output current in the continuous regime is simply the squared amplitude of the Fourier transform of the detector’s response function. Note, however, that the response function cannot be obtained from this spectrum because inverse Fourier transformation also requires data on the phase of $\tilde r(\nu)$. The response function can be measured directly in the time domain with a pulsed LO as discussed below.
Let us now discuss the contribution of electronic noise to quadrature measurements. From [Eq. (\[QmeasCV1\])]{}, and because in the vacuum state $\langle \hat q(t)\hat q(t')\rangle=\delta(t-t')$, we can write $$\begin{aligned}
\label{QmeasCVvar}
\langle \hat Q_{\rm meas}^2\rangle &=& A^2\alpha^2{\int\limits_{-\infty}^{+\infty}}|\psi'(t)|^2{\textrm{d}}t + \langle \hat Q_e^2\rangle\\ {\nonumber}&=& A^2\alpha^2{\int\limits_{-\infty}^{+\infty}}|\tilde\psi'(\nu)|^2{\textrm{d}}\nu+ \langle \hat Q_e^2\rangle\\ {\nonumber}&=& A^2\alpha^2{\int\limits_{-\infty}^{+\infty}}|\tilde\psi(\nu)|^2|\tilde r(\nu)|^2{\textrm{d}}\nu+ \langle \hat Q_e^2\rangle,\end{aligned}$$ where the variance of the electronic noise component is given by $$\begin{aligned}
\langle \hat Q_e^2\rangle
&=& {\int\limits_{-\infty}^{+\infty}}{\int\limits_{-\infty}^{+\infty}}\langle \hat i_e(t)\hat i_e(t+\tau)\rangle\psi(t)\psi(t+\tau){\textrm{d}}t{\textrm{d}}\tau{\nonumber}\\
&=& {\int\limits_{-\infty}^{+\infty}}S_e(\nu)|\tilde\psi(\nu)|^2{\textrm{d}}\nu.\label{QeCV}\end{aligned}$$ Combining the above two results with Eq. , we find $$\label{QmeasCVvar1}
\langle \hat Q_{\rm meas}^2\rangle = {\int\limits_{-\infty}^{+\infty}}S(\nu) |\tilde\psi(\nu)|^2 {\textrm{d}}\nu.$$ Equations and lead us to an important conclusion: by knowing the homodyne output spectra $S(\nu)$ and $S_e(\nu)$, as well as the weight function $\psi(t)$, one can predict the fraction of the electronic noise in the measured quadrature variance in an arbitrary temporal mode. This, as discussed above, directly translates into an equivalent optical loss.
In the case of a high-bandwidth detector, when $S(\nu)$ and $S_e(\nu)$ can be assumed constant over the support of $\tilde\psi(\nu)$, we have $$\label{}
1-\eta_e\approx \left[\frac {S_e(\nu)}{S(\nu)}\right]_{\textrm{signal bandwidth}}.$$ This quantity, which we call *clearance* of the detector’s shot noise over the electronic noise, is one of the primary characteristics of any HD circuit.
Pulsed regime
-------------
Now let us suppose that the LO is pulsed, with the pulse width much shorter than the time resolution of the electronics. In this case, [Eq. (\[cpracHD\])]{} takes the form $$\label{pracHD}
\hat i(t)= A\alpha_p\hat Q r(t)+\hat i_e(t),$$ where the pulse is assumed to occur at $t=0$, $\hat Q={\int\limits_{-\infty}^{+\infty}}\alpha(t) \hat q(t){\textrm{d}}t/\alpha_p$ is the normalized quadrature operator corresponding to the signal mode defined by the shape of the LO pulse, with $\alpha_p=\sqrt{{\int\limits_{-\infty}^{+\infty}}|\alpha(t)|^2{\textrm{d}}t}$ being the effective amplitude of the local oscillator pulse. In other words, neglecting the electronic noise, the shape of the HD response to a single short pulse is given by the detector’s response function. The quadrature measurement is obtained by integrating the homodyne photocurrent over a certain time interval: $$\label{Qmeas}
\hat Q_{\rm meas}=\int\limits_{t_1}^{t_2}\hat i(t){\textrm{d}}t=A\alpha_p \hat Q\int\limits_{t_1}^{t_2}r(t){\textrm{d}}t+\hat Q_e,$$ with $$\hat Q_e=\int\limits_{t_1}^{t_2}\hat i_e(t){\textrm{d}}t.$$ The optimal choice of the integration limits is determined by the bandwidths of the detector’s electronic noise and the temporal width of its response function. When the local oscillator is a train of pulses with repetition period $T$, the HD output current is given by $$\label{pracHDrep}
\hat i(t)= A\alpha_p\sum\limits_{j=-\infty}^{\infty}\hat Q_j r(t-jT)+\hat i_e(t),$$ where $\hat Q_j$ is the quadrature operator of the $j$th pulsed signal mode, with the pulse of interest having index $j=0$. If the response function is nonzero over an interval longer than $T$, the quadrature measurement is contaminated by that of the neighboring pulses: $$\label{QmeasSum}
\hat Q_{{\rm meas},0}=A\alpha_p \sum\limits_{j=-\infty}^{\infty}R_j\hat Q_j +\hat Q_e,$$ where $R_j=\int\limits_{t_1}^{t_2}r(t-jT){\textrm{d}}t$. The sum in [Eq. (\[QmeasSum\])]{} defines a new measured mode whose state is not necessarily identical to that in the $j=0$th pulsed mode. The corresponding mode matching efficiency (neglecting the electronic noise) is given by $$\label{}
\eta_b=\frac{R_0^2}{\sum\limits_{j=-\infty}^{\infty}R_j^2}.$$ This efficiency is plotted in Fig. \[pulsedefftheory\] for the response function of Gaussian shape as a function of the 3-dB bandwidth of the detector response function spectrum. As we see, a detector bandwidth of at least $~0.4/T$ is required for $\eta_b$ to exceed 99%.
![Effective efficiency of the HD associated with temporal overlap of responses to different pulsed modes, as a function of the 3-dB bandwidth of the electronics. The response function is assumed Gaussian. The solid line corresponds to a short integration interval $[-\epsilon,+\epsilon]$; the dashed line to the integration interval of length $T$. Both integration intervals are centered at the peak of the response function. \[pulsedefftheory\]](pulsedefftheory.eps){width="0.8\columnwidth"}
If $r(t)$ is known, so are all $R_j$ and the effect of finite bandwidth can be reversed by means of discrete deconvolution, akin to the continuous case. However, partial reversal can be implemented even if the response function is not known, provided that $\eta_b$ is sufficiently high, i.e. $|R_j|\ll|R_0|$ for $j\ne 0$, as follows. In a typical detector, $R_j$ are negligibly small for $j>0$: most of the ringings occur *after* the optical pulse that generates the response. This is the situation, for example, with the detector assembled in this work.Then we have, according to [Eq. (\[QmeasSum\])]{}, and because $\langle\hat Q_j\hat Q_k\rangle=\delta_{jk}$ in the vacuum state, $$\label{}
\langle \hat Q_{\rm meas,0}\hat Q_{{\rm meas},i}\rangle=A^2\alpha^2_p\sum\limits_j R_jR_{j-i}\approx A^2\alpha^2_p R_0R_{-i},$$ where $\hat Q_{{\rm meas},i}$ denotes the quadrature measurement for the $i$th pulse. The above correlation can be easily obtained experimentally, from which one can determine $$\label{}
\frac {R_{-i}}{R_0}\approx \frac{\langle \hat Q_{\rm meas,0}\hat Q_{{\rm meas},i}\rangle}{\langle \hat Q_{\rm meas,0}\hat Q_{{\rm meas},0}\rangle}.$$ One then calculates $$\label{}
Q'_{{\rm meas},0}=Q_{{\rm meas},0}- \sum\limits_{i=1}^{\infty}\frac{\langle \hat Q_{\rm meas,0}\hat Q_{{\rm meas},i}\rangle}{\langle \hat Q_{\rm meas,0}\hat Q_{{\rm meas},0}\rangle} Q_{{\rm meas},i}$$ for each experimentally measured quadrature, thereby eliminating contamination from neighboring pulses. The resulting quadrature values are then renormalized and used in quantum state reconstruction. By means of this technique, the reconstruction efficiency has been improved in Ref. [@hui:09].
We now use [Eq. (\[Snu\])]{} to determine the spectral power of the HD output in the pulsed regime. The ensemble average of $i(t)i(t+\tau)$ is a periodic function of time $t$, hence we can write $$\begin{aligned}
\label{}
\langle \hat i(t)\hat i(t+\tau)\rangle&=&\frac 1T (A\alpha_p)^2\\ {\nonumber}&&\hspace{-2cm}\times\sum\limits_{j,k=-\infty}^{\infty}\langle\hat Q_j\hat Q_k\rangle\int\limits_{-T/2}^{T/2}r(t-jT)r(t+\tau-kT){\textrm{d}}t\\ {\nonumber}&&\hspace{-2cm}+\langle \hat i_e(t) \hat i_e(t+\tau)\rangle_t.\end{aligned}$$ In the vacuum state, $$\label{}
\langle \hat i(t)\hat i(t+\tau)\rangle=\frac 1T (A\alpha_p)^2{\int\limits_{-\infty}^{+\infty}}r(t)r(t+\tau){\textrm{d}}t+\langle \hat i_e(t) \hat i_e(t+\tau)\rangle_t.$$ Accordingly, $$\label{Sm}
S(\nu)=\frac 1T (A\alpha_p)^2 |\tilde r(\nu)|^2+S_e(\nu).$$ We see that in spite of the pulsed character of the local oscillator, the HD spectrum is determined by the Fourier transform of its response function akin to the continuous case. An important difference is the multiplication by the pulse repetition rate: when the separation between the pulses is increased, the spectral power reduces proportionally.
In contrast to the continuous regime, in the pulsed case the evaluation of the equivalent efficiency requires knowledge of $r(t)$; information on the spectra $S(\nu)$ and $S_e(\nu)$ is not sufficient. Let us, however, consider a practically important particular case when the bandwidth of both the detector’s response and the electronic noise greatly exceed the laser repetition rate. Suppose the integration in [Eq. (\[Qmeas\])]{} is done over the time interval $[-T_0/2,T_0/2]$ with $T_0<T$. Then we have $$\label{Qmeas2}
\langle \hat Q_{\rm meas}^2 \rangle=\left\langle \left(A\alpha_p\hat Q\int\limits_{-T_0/2}^{T_0/2}r(t){\textrm{d}}t\right)^2\right\rangle+\langle \hat Q_e^2\rangle.$$ We assume that the temporal width of the function $r(t)$ is much less than $T_0$, so the integration limits can be replaced by $\pm\infty$. We then have $$\label{Qmeas2hb}
\langle \hat Q_{\rm meas}^2 \rangle=(A\alpha_p)^2|\tilde r(0)|^2+\langle \hat Q_e^2\rangle,$$ where $$\begin{aligned}
\label{Qe2hb}
Q_e &=& \left\langle \left(\int\limits_{-T_0/2}^{T_0/2}\hat i_e(t){\textrm{d}}t\right)^2\right\rangle \\ {\nonumber}&=& \left\langle \int\limits_{-T_0/2}^{T_0/2}\int\limits_{-T_0/2}^{T_0/2}\hat i_e(t_1)\hat i_e(t_2){\textrm{d}}t_1{\textrm{d}}t_2\right\rangle\\ {\nonumber}&=& \int\limits_{-T_0/2}^{T_0/2}\int\limits_{-T_0/2-t_1}^{T_0/2-t_1}\langle \hat i_e(t_1)\hat i_e(t_1+\tau)\rangle{\textrm{d}}\tau{\textrm{d}}t_1\\ {\nonumber}&\approx& \int\limits_{-T_0/2}^{T_0/2}\int\limits_{-\infty}^{+\infty}\langle \hat i_e(t_1)\hat i_e(t_1+\tau)\rangle{\textrm{d}}\tau{\textrm{d}}t_1\\ {\nonumber}&=& T_0 S_e(0).\end{aligned}$$ Here we again took advantage of the high bandwidth of the electronic noise to modify the integration limits. We also used the fact that the electronic noise $i_e(t)$ is a stationary stochastic process to eliminate the dependence on $t_1$. Substituting Eqs. – into [Eq. (\[etae\])]{}, we find $$\label{etaeS}
\eta_e=1-\frac {T_0} T \frac{\hat S_e(0)}{ S(0)}.$$ In other words, the shot-to-electronic-noise clearance measured in the pulsed regime yields a too conservative estimate for the equivalent loss associated with the electronic noise in the high-bandwidth limit. In order to minimize this loss, one needs to choose the integration interval that is as short as possible, but still accommodates the entire detector response function. However, if the detector bandwidth does not greatly exceed the laser repetition rate, we have $T_0\simeq T$ and the factor of $T_0/T$ in the above equation can be neglected. The exact value for the equivalent loss in this case cannot be determined from the spectra because it depends on the shape of the detector’s response function $r(t)$.
Design and challenges {#designSec}
=====================
Now we present an efficient wideband HD constructed with easily available components, as well as tips to solve the most common challenges found when building such a device. We show the results given by our HD in experiments with a pulsed laser source (76 MHz repetition rate) as well as with a continuous wave laser. These features give our HD a great versatility for applications in different kinds of experiments.
Our electronic cicruit is shown in Fig. \[fig:schm\]. The values and identification names are as noted on the schematic. Provision has been made for a second stage of voltage gain which, however, proved to be unnecessary. We use a homemade $\pm 15$ V DC power supply. To avoid possible leakage of any ambient noise to the HD through the voltage supply lines, each of them is filtered for both high and low frequencies (components L1, L2, C13, C14, C15, C16 that include a ferrite choke, a 100 $\mu$f electrolytic and a 0.1 $\mu$f ceramic capacitors). The use of long power supply lines in the circuit can create parasitic inductances making the circuit susceptible to instabilities. Because it is not possible to make all the tracks short, we perform the filtering in two stages, at the beginning and the end of each power supply track.
The filtered $\pm 15$ Volt supply rails are each further separated into 3 isolated lines by incorporating 3 networks consisting of a fast Schottky diode (for improving the stability of the power supplies), a tantalum 6.8 $\mu$f capacitor and a 0.1 $\mu$f ceramic capacitor (components C1–C12 and D3–D8). These isolated supplies in turn feed $\pm 12$ V regulators for the photodiodes (IC3 and IC8) and two pairs of $\pm 5$ V regulators for the first and second stage operational amplifiers (IC2, IC5, IC4 and IC7). Care has been taken to isolate and regulate the supply voltage for the detectors and amplifiers in order to provide as stable and oscillation free platform as possible. Bypass 0.1 $\mu$f capacitors are located at the regulator outputs very close to the amplifiers (C18–C23).
One of the primary challenges in achieving a high-bandwidth HD is associated with the photodiodes. Their P-N junction has a terminal capacitance which has to be significant for high quantum efficiency of detection. The feedback resistance of the amplifier will form a low pass filter with the terminal capacitance, thus limiting the high-frequency response of the circuit. Furthermore, this capacitance combined with other capacitances and inductances in the circuit board can give rise to instabilities and oscillations associated with a low pass filter configuration.
The terminal capacitance of the photodiodes can be reduced by increasing the reverse bias voltage across them. This, however, increases the dark current that contributes to the electronic noise [@zav:02]. Consequently, special care must be taken for choosing photodiodes with very low capacitance and dark current, as well as for a printed circuit board (PCB) design that avoids or compensates any capacitance that could produce oscillations (instabilities) in the frequency response of the HD.
We have chosen to use Hamamatsu S5972 photodiodes. This photodiode has a 91% quantum efficiency at a wavelength of 780 nm and a high cut-off frequency (500 MHz) when supplied in a 12 V reverse voltage configuration. Other reasons for this choice include a very low dark current (11 pA at 12V reverse voltage) and a low terminal capacitance (2.8 pf at 12V reverse voltage).
The amplification circuit starts at the junction of the two S5972 photodiodes, which is the differential sum point for the input to the first stage amplifier. The amplification of the difference signal is carried out by a single Texas Instruments OPA847 operational amplifier (op-amp) in trans-impedance configuration with a trans-impedance gain of 4 k$\Omega$ (IC1). For testing purposes we place a jumper (J1) in the inverting input of the op-amp. In order to have control of the DC offset of output signal produced by the the op-amp we place in its non-inverting input a capacitor of 0.1 $\mu$f in parallel with a variable resistor of 10 k$\Omega$ both connected to ground (C17 and R4). With the help of the variable resistor we can maintain the DC offset at the zero reference point to avoid saturation in the op-amp. The optional second stage is designed to use an OPA847 in inverting amplifier configuration. J2 and J3 are the positions of its input and feedback resistances.
An important component of the circuit is a custom made variable capacitor (C24) placed in parallel to the feedback resistance R2. This capacitor is constructed by twisting two wires with a variable number of twists. By twisting (or untwisting) these wires we can vary their capacitance; although it is very small, it is enough to change the response of the circuit. In our case it is used to flatten the spectral response of the HD.
Finally, resistors R3, R5 and R6 are used for impedance matching of the output signal of the amplifier with a 50$\Omega$ coaxial connector.
All components are soldered on a specially designed printed circuit board (Fig. \[fig:pcb\]). When designing the layout for the circuit we take the following factors into consideration.
- The tracks on the PCB are kept as short as possible to avoid parasitic inductances.
- A ground (GND) plane, essential for high frequency circuits, and a double-sided board design are used. The back side of the PCB is mainly reserved for the GND plane, with the fewest possible discontinuities. This separation between tracks and components in one side and GND on the other minimize possible capacitances between the tracks and the ground plane of the circuit.
- The use of surface mount components is preferred for the high frequency regime of electronics.
- In order to provide a mechanically stable platform for the HD, the circuit module is mounted on a 0.25-inch aluminum plate.
- To ensure a maximally balanced response, we designed our circuit to be as symmetrical as possible.
- The HD is shielded by a custom-made metal box to avoid any environmental noise impact on the circuit.
- To minimize the parasitic capacitance the two photodiodes were placed closely together in the upper left corner of the board.
- Due to its relatively large size, the offset adjustment potentiometer (R4) for the first stage is placed at the bottom of the printed circuit board to avoid the issue of having long tracks on the upper side.
{width="6in"}
![Balanced HD experimental set-up\[fig:setup\]](setup.eps){width="2.5in"}
Performance and characterization
================================
Pulsed regime {#pusedSec}
-------------
After having built the HD we proceed to test its performance. To that end, we place it in the experimental set-up shown in Fig. \[fig:setup\]. Our LO is obtained from a mode-locked Coherent MIRA 900 Ti:Sapphire laser producing 1.8 ps pulses with a repetition rate of 76 MHz, central wavelength of 791 nm. The HD beam splitter configuration is implemented by using a half wave plate and a polarizing beam splitter. The two beams coming out of the beam splitter are focused into the photodiodes.
The HD is balanced by first adjusting the waveplate in order to equalize the responses of the photodiodes, compensating for a possible difference in their quantum efficiencies, thereby minimizing the spurious signal at the local oscillator repetition rate. A further crucial step in balancing the HD is to equalize the path lengths of the two beams entering the photodiodes by using a XY translation stage on which the HD is mounted. This is necessary because the pulses, if arriving at the photodiodes at different times, can lead to subtraction pulse having a bipolar shape. Even after these alignment steps, the subtraction may not be perfect due to different capacitances of the two photodiodes.
Subsequently, the HD is tested for electronic instabilities and oscillations. This is done by observing the output signal of the HD (with the LO power of 6 mW) with a spectrum analyzer in a range from 100 kHz up to 3 GHz. If present, instabilities produce peaks in the spectral response at frequencies different from the repetition rate of the LO and its harmonics. They can be removed by adjusting the values of resistances R2, R3, R5 and R6 as well as the custom made variable capacitor C24. Additional instability sources can be associated with the ambient noise or leakage through the power supply lines; in this case, the elements of the power supply filtering sections must be changed.
In the final step of adjustments we flatten the spectral response of the HD. This is done by changing the value of the capacitance C24 while observing the spectrum of the HD output. Fig. \[fig:spectralresponsepulse\] (dashed and dotted lines) shows the effect of this adjustment in the spectral profile of the HD.
In order to characterize the detector, we first measure the shot noise clearance. To this end, we increase the LO power to 12 mW. The choice of this power level is determined by the need to achieve the highest shot-to-electronic noise clearance, while at the same time ensuring stable and saturation-free operation of the HD. With the given wavelength and repetition rate, this LO power corresponds to $6.2 \times 10^{8}$ photons per pulse, and will produce $5.6 \times 10^{8}$ photoelectrons per pulse, which corresponds to a shot noise of $24,000$ electrons per pulse. In Fig. \[fig:spectralresponsepulse\] (dotted vs. dash-dotted lines) we show the clearance between electronic noise and shot noise, which has a value of 13dB, correponding to $\eta_e=0.95$. The 3-dB bandwidth of the HD is 80 MHz, which implies that it can be applied, without significant efficiency loss, to pulsed lasers with repetition rates up to 250 MHz (Fig. \[pulsedefftheory\]).
We verify that the power of the observed HD output signal grows linearly with the LO power, which is the signature of the shot noise [@bac:04; @lvov:09]. In Fig. \[fig:linear\] we show measurements of the shot noise level for five different frequencies depending on the LO power. For frequencies up to 100 MHz, the HD behaves linearly, although the slope is reduced for higher frequencies.
[ ![Spectral response of the HD at 12 mW of LO power. The large peaks are the laser repetition rate and its second harmonic. *Inset*: The response at 100 $\mu$W of LO power, necessary to calculate the CMRR. Grey (top) trace: one photodiode blocked; Black (bottom) trace: both photodiodes illuminated\[fig:spectralresponsepulse\]](spectralresponsepulse.eps "fig:"){width="0.9\columnwidth"} ]{}
![Response of the HD respect to the LO power for different fixed frequencies. The linear fits are also shown. Red dash line (plus): 80MHz; Green dot line (diamond): 60MHz; Black continous line (circle): 40MHz; Blue dot dash line (cross): 100MHz.\[fig:linear\]](linear.eps){width="0.8\columnwidth"}
In order to determine the CMRR value, we measure the spectral response of the HD at the repetition frequency of the LO at a power of 100 $\mu$W in two cases: when both photodiodes are illuminated, and when only one photodiode is illuminated (inset of Fig. \[fig:spectralresponsepulse\]). The measurement is performed with a low LO power to avoid saturation of the HD in the absence of subtraction. We find that our HD has a CMRR of 52.4 dB, which corresponds to a reduction by a factor of $2.4\times10^{-3}$ in the photocurrent corresponding to the common mode, yielding $6.7\times 10^5$ residual photoelectrons per pulse due to imperfect subtraction [@chi:11], which is much greater than the average number of photoelectrons per pulse due to the shot noise. This is the reason why, even with the optimized alignment, the pulsed HD spectrum exhibits a peak at the pulse repetition rate.
Theoretically, this residual periodic signal leads to a constant displacement in the quadrature measurement, which is easy to average out. In practice, however, the signal tends to spontaneously change its magnitude every few seconds. Furthermore, the presence of that signal complicates computer acquisition of the HD output because of the limited digitizer resolution. Therefore we use two notch filters (MFC 6367) to eliminate the output components at 76 and 152 MHz.
The HD output signal is further processed by a 100 kHz high pass filter to remove the DC offset. To clean the signal and to avoid any noise at higher frequencies, two low pass filters with a 100 MHz cut-off are used. The resulting HD spectrum is shown in Fig. \[fig:spectralresponsepulse\] (solid line).
Continuous regime
-----------------
In order to check the versatility of our scheme, we analyze the performance of another HD, constructed similarly to the first one, in the continuous wave regime by using a TekhnoScan Ti:Sapphire laser with a center wavelength of 795 nm and a bandwidth of 5 kHz. We use the same experimental set-up as in Fig. \[fig:setup\]. We analyze the response of the HD in the time domain by acquiring 30 traces of 100-$\mu$s duration, each containing $2 \times 10^{5}$ points (sampling frequency of 2GHz). Thereafter, we compute the autocorrelation function for each trace, and then average it over the 30 traces \[Fig. \[fig:spectralresponsecont\](a)\].
[ ![characterization of the HD in the continuous regime. a) Autocorrelation function. b) Electronic spectrum of the shot noise obtained with a spectrum analyzer (blue trace — top); the black (bottom) trace is the electronic noise, and by Fourier transform of the autocorrelation function (red trace — middle). The vertical position on the logarithmic scale of the red trace with respect to the spectrum analyzer traces is chosen arbitrarily. c) Effective efficiency associated with the electronic noise as a function of the signal wavefunction bandwidth (in units of the inverse FWHM of its temporal mode). \[fig:spectralresponsecont\]](cvspectrum.eps "fig:"){width="0.9\columnwidth"} ]{}
We then acquire the HD output spectrum, both in the presence and absence of LO, using a spectrum analyzer. As evidenced by Fig. \[fig:spectralresponsecont\](b), the shot noise spectrum is similar, up to a constant factor, to that obtained by Fourier transform to the autocorrelation function in agreement with the Wiener-Khintchine theorem. We find this detector to exhibit a 3-dB bandwidth of 100 MHz and the shot-to-electronic noise clearance ranging between 10 and 18 dB.
In Fig. \[fig:spectralresponsecont\](c) we plot the effective loss associated with the electronic noise under the assumption that the detector is used to measure field quadratures in a Gaussian temporal mode. The loss is calculated according to Eqs. , and given the measured spectra \[Fig. \[fig:spectralresponsecont\](b)\]. We find that the effective loss strongly varies dependent on the shape and width of the temporal mode.
Table \[tab:table1\] shows the features of our device in comparison with other HDs reported in the literature. As we see, our detector compares favorably with its counterparts: it shows a unique combination of the bandwidth, CMRR and the electronic noise clearance.
Characteristics Ours [@han:01] [@chi:11] [@zav:02] [@nie:09] [@oku:08] [@had:09]
---------------------- ------ ----------- ----------- ----------- ----------- ----------- -----------
Wavelength (nm) 791 790 1550 786 860 1064 –
3 dB bandwidth (MHz) 100 1 100 $\sim$50 100 250 54
CMRR (dB) 52.4 85 46 42 – 45 61.8
Clearance (dB) 13 14 13 $\sim$5 10 $\sim$7.5 12
: \[tab:table1\]Comparison between various HD’s.
Summary
=======
We presented a comprehensive theory and a detailed recipe for designing, building and troubleshooting a wideband balanced detector for highly accurate time-domain quantum measurements. We showed that by following these recommendations a homodyne detector with a shot-noise clearance of 13 dB at 12 mW of local oscillator power, a CMRR value of 52.4dB, a flat response up to 100 MHz can be constructed in an easy way and by only using a single trans-impedance amplification stage. We also have shown that this HD can be applied to both pulsed and CW configurations.
We thank Frank Vewinger, Nitin Jain and Ryan Thomas for helpful discussions. This work has been supported by NSERC, CIFAR, Quantum*Works* and ARC (Grant DP1094650).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We numerically construct asymptotically anti-de Sitter boson star solutions using a minimally coupled $\frac{D-1}{2}$-tuplet complex scalar field in $D=5,7,9,11$ dimensions. The metric admits multiple Killing vector fields in general, however the scalar fields are only invariant under a particular combination, leading to such boson star solutions possessing just a single helical Killing symmetry. These boson stars form a one parameter family of solutions, which can be parametrized by the energy density at their center. As the central energy density tends to infinity, the angular velocity, mass, and angular momentum of the boson star exhibit damped harmonic oscillations about finite central values, while the Kretschmann invariant diverges, signaling the formation of a black hole in this limit.'
author:
- Sean Stotyn
- 'C. Danielle Leonard'
- Marius Oltean
- 'Laura J. Henderson'
- 'Robert B. Mann'
title: 'Numerical Boson Stars with a Single Killing Vector I: the $D\ge5$ Case'
---
Introduction
============
Boson stars are smooth, horizonless geometries composed of self-interacting and self-gravitating bosonic matter[@Kaup:1968zz; @Ruffini:1969qy]. They describe localized bundles of field energy that typically do not display a sharp edge, like those of ordinary stars or neutron stars. It is currently unknown whether boson stars are physically realized in our universe, but they have nevertheless maintained a steady level of interest both from a theoretical gravitational point of view and from an astrophysical point of view. In the latter context, boson stars in four dimensional asymptotically flat space-time have been put forth as candidates for dark matter halos that may help explain galaxy rotation curves – see Ref [@Jetzer:1991jr] for a review. While boson stars in an excited state typically produce a more physically realistic, flatter rotation curve than boson stars in the ground state, such excited states are known to decay to the ground state unless the boson stars are in rather particular mixed states [@Bernal:2009zy]. If not coupled to a Maxwell field, they also provide dark alternatives to astrophysical black hole candidates, which could potentially be discerned by gravitational wave astronomy [@Berti:2006qt; @Kesden:2004qx]. On the purely theoretical side, boson stars in asymptotically anti-de Sitter (AdS) space-times are believed to play important roles in holographic gauge theories through the AdS/CFT correspondence. There is much work yet to be done in this respect, but some progress has been made, at least in mapping out the possible boson star solutions. For a detailed review of boson stars and their various applications, we refer the reader to Ref [@Liebling:2012fv].
There are a wide range of boson star solutions, which can be composed of a complex scalar field with mass[@Astefanesei:2003qy; @Brihaye:2013hx] or without[@Stotyn:2011ns; @Dias:2011at], with self-interactions[@Colpi:1986ye; @Ho:2002vz], with gauge charges[@Brihaye:2004nd; @Dias:2011tj], with rotation[@Stotyn:2011ns; @Dias:2011at], with de Sitter[@Fodor:2010hg; @Cai:1997ij], flat[@Colpi:1986ye; @Hartmann:2012gw], or anti-de Sitter[@Astefanesei:2003qy; @Bjoraker:1999yd] boundary conditions, etc. Despite this diversity, boson stars all have one important feature in common: since they are horizonless, they are zero temperature objects that describe finite energy excitations above the vacuum state. The stability of these objects is of central interest if AdS boundary conditions are present. If they are perturbatively unstable to the formation of a black hole, then they are themselves of little interest gravitationally. If they are perturbatively stable then their non-linear stability determines whether the corresponding state in the holographic dual CFT thermalizes and on what time-scale. Due to the evidence of a gravitational turbulent instability in asymptotically AdS space-times [@Dias:2011ss; @Bizon:2011gg; @Jalmuzna:2011qw; @Maliborski:2013jca], one might expect all AdS boson stars to be non-linearly unstable to black hole formation. However, recently in [@Buchel:2013uba] a wide range of initial data were discovered such that boson stars are immune to the turbulent instability, leading on the gauge theory side to a family of strongly coupled CFT states that do not thermalize in finite time. Furthermore, the analysis of [@Dias:2012tq] suggests that the turbulent instability of global AdS is due to the high level of symmetry: the normal mode frequencies are all integer multiples of the AdS frequency, leading to a large number of resonances responsible for the nonlinear instability. Solutions possessing less symmetry, such as boson stars, are not plagued by this problem and tend to be nonlinearly stable as a result. Although a clear mapping between boson star properties and physically observed states in CFTs is currently unknown, it is hoped that a better understanding of the gravitational aspects of boson stars will lead to insights into physically realizable systems.
Most of the boson star solutions constructed to date have relied on a relatively high level of symmetry to yield equations that can feasibly be solved numerically. Indeed, space-time symmetries often play an indispensable role in constructing analytic solutions to the Einstein equations. Such symmetries readily manifest in the familiar form of Killing vector fields, which generate the space-time’s isometry group. Furthermore, various theorems show that Killing symmetries are ubiquitous under physically reasonable assumptions. For instance, the rigidity theorem states that if a space-time is stationary, it must also possess an axis of symmetry, leading to a minimum of 2 Killing symmetries[@Hawking:1971vc; @Hollands:2006rj; @Moncrief:2008mr]. Concordantly, it has become an almost universal feature of exact solutions to Einstein gravity to admit multiple Killing fields.
However, this need not necessarily be the case: for example, the gravitational field outside of a generic matter source may only have approximate Killing symmetries, or none at all. Even in more idealized settings, there is considerable interest in exact solutions which possess only a single Killing field. For instance, due to a judicious choice of metric and scalar field ansatz first put forth in [@Hartmann:2010pm], spinning boson star solutions exhibiting a single helical Killing vector have been found in 5 dimensions with anti-de Sitter (AdS) boundary conditions [@Dias:2011at]; the scalar field has a harmonic time dependence that breaks the continuous rotational symmetry. In [@Stotyn:2011ns] these techniques were used to construct analogous boson stars in odd dimensions $D\ge5$ in the perturbative regime where energy and angular momentum are low. However full numerical results away from the perturbative regime have yet only been presented in the 5 dimensional treatment in [@Dias:2011at]. In this paper, we complete this analysis and obtain numerical solutions for asymptotically AdS boson stars with a single Killing field in $D=7,9,11$ dimensions; this one parameter family of solutions is parameterized by the central energy density of the boson star. This central energy density is unbounded above and in the limit that it tends to infinity, the mass, angular momentum, and angular velocity all tend to finite values, while the Kretschmann scalar at the centre diverges. Analogous boson stars in $D=3$ have qualitatively distinct features and require separate considerations, even in the perturbative regime [@Stotyn:2012ap]; the numerical construction of these lower dimensional boson star solutions is discussed in a separate paper [@Stotyn:2013spa]. We focus on odd-dimensions since there is a useful ansatz for the scalar fields (given in section 2) that allows the full solutions to have only a single Killing vector, but yields a stress-energy with the same symmetries as the metric.
The remainder of this paper is structured as follows: in section \[Setup\] we present our ansatz and extract the ordinary differential equations (ODEs) and constraint equations we must solve. In section \[BosonConditions\], we discuss the boundary conditions and the physical properties of our boson stars. In section \[numconst\] we describe our numerical methods. In section \[Results\] we present and discuss our results, and in section \[Conclusion\] we provide some concluding remarks.
Setup {#Setup}
=====
We begin with $D=n+2$ dimensional Einstein gravity with negative cosmological constant minimally coupled to an $\frac{n+1}{2}$-tuplet complex scalar field $$S=\frac{1}{16\pi}\int{d^{D}x\sqrt{-g}\left(R+\frac{n(n+1)}{\ell^{2}}-2\big|\nabla\vec{\Pi}\big|^{2}\right)}\label{eq:action}$$ where we take the usual convention $\Lambda=-\frac{n(n+1)}{2\ell^{2}}$ with AdS length $\ell$ and restrict attention to $n=3,5,7,9$. The equations of motion resulting from this action are $G_{ab}-\frac{n(n+1)}{2\ell^{2}}g_{ab}=T_{ab}$ and $\nabla^{2}\vec{\Pi}=0$, where the stress tensor of the scalar field is given by $$T_{ab}=\left(\partial_{a}\vec{\Pi}^{*}\partial_{b}\vec{\Pi}+\partial_{a}\vec{\Pi}\partial_{b}\vec{\Pi}^{*}\right)-g_{ab}\left(\partial_{c}\vec{\Pi}\partial^{c}\vec{\Pi}^{*}\right).\label{eq:Tab}$$
We take the following ansätze for the metric and scalar field $$ds^{2}=-f(r)g(r)dt^{2}+\frac{dr^{2}}{f(r)}+r^{2}\bigg(h(r)\big(d\chi+A_{i}dx^{i}-\Omega(r)dt\big)^{2}+g_{ij}dx^{i}dx^{j}\bigg)\label{eq:metric}$$ $$\Pi_{i}=\Pi(r)e^{-i\omega t}z_{i},\quad\quad\quad\quad i=1...\frac{n+1}{2}\label{eq:ScalarField}$$ where $z_{i}$ are complex coordinates such that ${\displaystyle \sum_{i}dz_{i}d\bar{z}_{i}}$ is the metric of a unit $n-$sphere. An explicit and convenient choice for the $z_{i}$ is $$z_{i}=\left\{ {\genfrac{.}{.}{0pt}{0}{e^{i(\chi+\phi_{i})}\cos\theta_{i}{\displaystyle \prod_{j<i}\sin\theta_{j},\quad\quad\quad i=1...\frac{n-1}{2}}}{{e^{i\chi}{\displaystyle \prod_{j=1}^{\frac{n-1}{2}}\sin\theta_{j}}},\quad\quad\quad\quad\quad i=\frac{n+1}{2}}}\right.\label{eq:zi}$$ in which case ${\displaystyle \sum_{i}dz_{i}d\bar{z}_{i}=(d\chi+A_{i}dx^{i})^{2}+g_{ij}dx^{i}dx^{j}}$ is the Hopf fibration of the unit $n-$sphere where $$A_{i}dx^{i}=\sum_{i=1}^{\frac{n-1}{2}}{\cos^{2}\theta_{i}\left[\prod_{j<i}\sin^{2}\theta_{j}\right]d\phi_{i}}$$ and $g_{ij}$ is the metric on a unit complex projective space $\mathbb{CP}^{\frac{n-1}{2}}$. In these coordinates, $\chi$ and the $\phi_{i}$ all have period $2\pi$ while the $\theta_{i}$ take value in the range $[0,\frac{\pi}{2}]$.
The form of the scalar fields is crucial to this construction and was first considered in [@Hartmann:2010pm]: it is clear from Eq. (\[eq:ScalarField\]) that the scalar fields can be viewed as coordinates on ${\mathbb{C}}^{\frac{n+1}{2}}$. For each value of $r$, $\vec{\Pi}$ traces out a round $n$-sphere with a time-varying but otherwise constant phase. On the other hand, constant $r$ surfaces in the metric (\[eq:metric\]) correspond to squashed rotating $n$-spheres. The stress tensor has the same symmetries as the metric (\[eq:metric\]) since the first term is the pull-back of the round metric of the $n$-sphere and the second term is proportional to $g_{ab}$.
Although the matter stress tensor has the same symmetries as the metric, the scalar fields themselves do not. Indeed, the metric (\[eq:metric\]) is invariant under $\partial_{t}$, $\partial_{\chi}$ as well as the rotations of $\mathbb{CP}^{\frac{n-1}{2}}$ while the scalar field (\[eq:ScalarField\]) is only invariant under the combination $$K=\partial_{t}+\omega\partial_{\chi}.\label{eq:KV}$$ Therefore, any solution with non-trivial scalar field will only be invariant under the single Killing vector field given by (\[eq:KV\]).
The equations of motion yield the following system of coupled second order ODEs: $$\begin{split}f''-\frac{6f'}{fr}\left(\frac{rf'}{6}-\frac{f}{6}+\Xi\right)+\frac{4h'}{r}+\frac{n^{2}-1}{r^{2}}+\frac{n^{2}-1}{\ell^{2}}+\frac{8\Pi'\Pi}{r}+\frac{4\Pi^{2}(\omega-\Omega)^{2}}{fg}\\
-\frac{8\Xi^{2}}{fr^{2}}-\frac{4\Pi^{2}\left(1+\frac{(n-1)}{2}h\right)}{hr^{2}}-\frac{2(n-3)\Xi}{r^{2}}=0, & \label{eq:fEq}
\end{split}$$ $$\begin{split}g''-g'\left(\frac{4\Xi}{fr}+\frac{g'}{g}-\frac{1}{r}\right)-4g\left(\frac{\big(\Xi r^{\frac{n-1}{2}}\sqrt{h}\big)'}{fr^{\frac{n+1}{2}}\sqrt{h}}+\frac{\frac{(n-1)}{2}h^{2}-\Pi^{2}}{fhr^{2}}+\frac{3(n+1)}{2f\ell^{2}}-\frac{(n-3)\Xi}{2fr^{2}}\right)\\
-\frac{8\Pi^{2}(\omega-\Omega)^{2}}{f^{2}}-\frac{hr^{2}\Omega'^{2}}{f}=0, & \label{eq:gEq}
\end{split}$$ $$h''+\frac{h'}{r}-\frac{2h'}{fr}\left(\Xi+\frac{frh'}{2h}\right)+\frac{h^{2}r^{2}\Omega'^{2}}{fg}+\frac{4(1-h)}{fr^{2}}\left(\Pi^{2}+\frac{(n+1)}{2}h\right)=0,\label{eq:hEq}$$ $$\Omega''+\frac{4\Pi^{2}}{fhr^{2}}(\omega-\Omega)+\Omega'\left(\frac{f'}{f}+\frac{2h'}{h}+\frac{2\Xi}{fr}+\frac{2n+1}{r}\right)=0,\label{eq:OmegaEq}$$ $$\Pi''-\frac{2\Pi'}{fr}\left(\Xi-\frac{f}{2}\right)+\frac{\Pi(\omega-\Omega)^{2}}{f^{2}g}-\frac{\big(1+(n-1)h\big)\Pi}{fhr^{2}}=0,\label{eq:PiEq}$$ where $\Xi=h+\Pi^{2}-\frac{n+1}{2}-\frac{(n+1)r^{2}}{2\ell^{2}}$ and a $'$ denotes differentiation with respect to $r$.
In addition to these second order ODEs, the Einstein equations further impose two first order ODEs in the form of constraint equations, $C_{1}=0$ and $C_{2}=0$. Explicitly, these are $$C_{1}=\frac{(f^{2}ghr^{2(n-1)})'}{fr^{2n-3}}+4gh\Xi, \label{eq:C1Eq}$$ $$C_{2}=\frac{\Pi^{2}(\omega-\Omega)^{2}}{f^{2}g}+\Pi'^{2}-\frac{r^{2}h\Omega'^{2}}{4fg}+\frac{\Xi(r^{n+1}h)'}{fhr^{n+2}}+\frac{(hf)'h'}{4fh^{2}}+\frac{n(fhr^{n-1})'}{2fhr^{n}}+\frac{\frac{(n-1)}{2}h^{2}-\Pi^{2}}{fhr^{2}}+\frac{n+1}{2f\ell^{2}}. \label{eq:C2Eq}$$ Furthermore, under the flow of the equations of motion above, these constraint equations obey $$\begin{aligned}
C_1'={}&{\mathcal F}_1(r) C_1, \label{eq:C1deriv} \\
C_2'={}&{\mathcal F}_2(r)C_1+{\mathcal F}_3(r)C_2, \label{eq:C2deriv}\end{aligned}$$ for some functions ${\mathcal F}_1,~{\mathcal F}_2,~{\mathcal F}_3$ that depend on the dimensionality of the space-time. The explicit form of these functions is neither important nor illuminating, but what is important is that the derivatives of the constraints are proportional to the constraints themselves.
Boundary Conditions and Physical Charges {#BosonConditions}
========================================
In order to solve the equations of motion, we need to impose boundary conditions at the boson star origin, as well as asymptotically. In this section we present these boundary conditions and we write the asymptotic charges in terms of them.
Boundary Conditions at the Origin
---------------------------------
The geometry must be smooth and horizonless, which means that all metric functions must be regular at the origin. Furthermore, due to the slow physical rotation of points as $r\rightarrow0$, surfaces of constant $t$ in the vicinity of the origin ought to be described by round $n$-spheres with $r$ being the proper radial distance. To find the boundary condition on $\Pi$, we multiply (\[eq:PiEq\]) by $r^{2}$ and note that $\Pi$ must vanish at the origin in order to yield consistent equations of motion. Thus, the boundary conditions at the origin take the form $$\begin{split}
\left.f\right|_{r\rightarrow0}=1+\mathcal{O}(r^{2}),\quad\left.g\right|_{r\rightarrow0}=g(0)+\mathcal{O}(r),\quad\left.h\right|_{r\rightarrow0}=1+\mathcal{O}(r^{2}),\label{eq:OriginBC}\\
\left.\Omega\right|_{r\rightarrow0}=\Omega(0)+\mathcal{O}(r),\quad\left.\Pi\right|_{r\rightarrow0}=q_{0}\frac{r}{\ell}+{\cal O}(r^{2}),\quad\quad\quad\quad
\end{split}$$ for all $n$, where $q_{0}$ is a dimensionless parameter such that the energy density of the scalar field, $T^{00}$, at the origin is proportional to $q_{0}^{2}$. In fact, $q_{0}$ uniquely parameterizes the one-parameter family of boson star solutions in each dimension. Formally, it is defined by $q_0\equiv\ell\Pi'(0)$.
Asymptotic Boundary Conditions
------------------------------
In order to simplify the asymptotic boundary conditions, we first make note of a residual gauge freedom. It is straightforward to show that the transformation $$\chi=\tilde{\chi}+\lambda t,\quad\quad\Omega(r)=\tilde{\Omega}(r)+\lambda,\quad\quad\omega=\tilde{\omega}+\lambda\label{eq:PsiGaugeFreedom}$$ for some arbitrary constant $\lambda$, leaves both the metric (\[eq:metric\]) and scalar field (\[eq:ScalarField\]) unchanged. We find it convenient in our numerical analysis to set $\lambda=\omega$ so that we can set $\tilde{\omega}=0$: in this frame, the coordinates are rigidly rotating asymptotically so that $\tilde{\Omega}(r)\rightarrow-\omega$ as $r\rightarrow\infty$. In what follows, we use $\tilde{\chi}$, $\tilde{\Omega}(r )$, and $\omega$ but we drop the tildes for notational convenience.
In the $r\rightarrow\infty$ limit the boundary conditions for the boson star will asymptote to AdS with corrections for mass and angular momentum, which determine the metric functions up to constants $C_{f},~C_{h},~\mathrm{and}~C_{\Omega}$. The boundary condition on the scalar field is set by requiring $\Pi$ to be normalizable, which means it must decay like $r^{-(n+1)}$. Explicitly, the asymptotic boundary conditions are given by $$\begin{split}
\left.f\right|_{r\rightarrow\infty}={} & \frac{r^{2}}{\ell^{2}}+1+\frac{C_{f}\ell^{n-1}}{r^{n-1}}+\mathcal{O}(r^{-n}),\quad\>\>\left.g\right|_{r\rightarrow\infty}=1-\frac{C_{h}\ell^{n+1}}{r^{n+1}}+\mathcal{O}(r^{-(n+2)}), \\
\left.h\right|_{r\rightarrow\infty}={} & 1+\frac{C_{h}\ell^{n+1}}{r^{n+1}}+\mathcal{O}(r^{-(n+2)}),\quad\quad\left.\Omega\right|_{r\rightarrow\infty}=-\omega+\frac{C_{\Omega}\ell^{n}}{r^{n+1}}+\mathcal{O}(r^{-(n+2)}),\label{eq:AsymptoticBC}\\
\left.\Pi\right|_{r\rightarrow\infty}={} & \frac{\epsilon\ell^{n+1}}{r^{n+1}}+\mathcal{O}(r^{-(n+2)}),
\end{split}$$ where $\epsilon$ is a dimensionless measure of the amplitude of the scalar field at infinity. A perturbative analysis of this problem [@Stotyn:2011ns] indicates that $\epsilon$ uniquely parameterizes the boson star solutions when the energy and angular momentum are low; however in the non-perturbative regime it does not, as will be explicitly demonstrated in section \[Results\].
Physical Charges
----------------
As we have seen, these boson stars are invariant under the single Killing field (\[eq:KV\]), that is, a linear combination of $\partial_t$ and $\partial_\chi$. However, the scalar field vanishes at infinity with sufficient fall-off to imply that $\partial_{t}$ and $\partial_{\chi}$ are each asymptotic Killing fields (the metric alone being invariant under them). Thus they define conserved charges, which can be computed via the Astekhar-Das formalism [@Ashtekar:1999jx; @Das:2000cu]; in particular, $\partial_{t}$ and $\partial_{\chi}$ are readily associated with a conserved energy and angular momentum, respectively. For details on how these charges are defined and computed, we refer the reader to Ref [@Stotyn:2011ns] and here simply state the result: $$\begin{aligned}
\label{EJ}
M & =\frac{(n+1)\pi^{\frac{n-1}{2}}\ell^{n-1}}{16\left(\frac{n+1}{2}\right)!}\big((n+1)C_{h}-nC_{f}\big),\\
J & =\frac{(n+1)^{2}\pi^{\frac{n-1}{2}}\ell^{n}C_{\Omega}}{16\left(\frac{n+1}{2}\right)!}.\end{aligned}$$ Here, $C_{f},~C_{h},~\mathrm{and}~C_{\Omega}$ are the constants appearing in the asymptotic boundary conditions (\[eq:AsymptoticBC\]).
The existence of these asymptotic Killing symmetries also guarantees that the boson stars satisfy the first law of thermodynamics. They have vanishing temperature and entropy, so the first law takes the form $$dM=\omega dJ.$$ The fact that these boson stars satisfy the first law is an important numerical tool. Indeed one of the primary cross-checks on the validity of our numerical methods discussed below is an explicit verification that the first law holds to at least one part in $10^{6}$.
Numerical Construction {#numconst}
======================
To numerically construct boson star solutions, we used a relaxation method on a Chebyshev grid with the metric and scalar field functions approximated by Chebyshev polynomials. A detailed review of approximating analytic functions by Chebyshev polynomials as well as the Chebyshev relaxation procedure can be found in [@Pfeiffer:2005zm].
In the case under consideration here there are 5 functions $\{f(r),g(r),h(r),\Omega(r),\Pi(r)\}$ defined over the domain $r\in[0,\infty)$. To employ the Chebyshev relaxation procedure, we need to compactify the domain, which we do by introducing the new coordinate, $y$, defined by $$y=\frac{r^{2}}{r^{2}+\ell^{2}}.$$ Note that since $r$ takes value in the range $[0,\infty)$, $y$ takes value in the range $[0,1]$. In terms of the Chebyshev grid points $x_k=\cos\left(\frac{k\pi}{N}\right)$, $k=0,...,N$, defined as the extrema of the $N^{\mathrm{th}}$ Chebyshev polynomial $T_N(x)$ over the range $[-1,1]$, we define our grid by the simple transformation $y_{k}\equiv\frac{x_{k}+1}{2}$.
To obtain a set of analytic functions, we first extract the singular behavior from each and introduce auxiliary functions with the boundary conditions (\[eq:OriginBC\]) and (\[eq:AsymptoticBC\]) in mind. In terms of the coordinate $y$, this leads to $$\begin{aligned}
& f(y)=\frac{y}{1-y}+1+(1-y)^{\frac{n-1}{2}}q_{f}(y), \label{fq}\\
& g(y)=1+(1-y)^{\frac{n+1}{2}}q_{g}(y), \label{gq}\\
& h(y)=1+(1-y)^{\frac{n+1}{2}}q_{h}(y), \label{hq}\\
& \Omega(y)=\frac{q_{\Omega}(y)}{\ell}, \label{Oq}\\
& \Pi(y)=\sqrt{y}(1-y)^{\frac{n+1}{2}}q_{\Pi}(y). \label{Pq}\end{aligned}$$ The set of functions $\{q_{f},q_{g},q_{h},q_{\Omega},q_{\Pi}\}$ are analytic over the range $y\in[0,1]$; in the remaining discussion the coordinate $y$ will be used, unless otherwise stated. In particular, a prime $'$ will denote a derivative with respect to $y$.
To find the boundary conditions on the $q$ functions, we Taylor expand the equations of motion (\[eq:fEq\])–(\[eq:OmegaEq\]) and the constraint equations (\[eq:C1Eq\]) and (\[eq:C2Eq\]) around the two boundary points, $y_0=0,1$ and require them to vanish order by order in $(y-y_0)$. This leads to the following nontrivial relationships between the various functions and their first derivatives: at $y_0=0$ $$\begin{aligned}
\label{eq:BC1}
& q_{f}(0)=0,\qquad\qquad q_{g}'(0)-\frac{n+1}{2}q_{g}(0)-\frac{3}{n}q_{h}'(0)\big(1+q_{g}(0)\big)-\frac{2}{n}\big(1+q_{g}(0)\big)q_{0}^{2}=0, \\
& q_{h}(0)=0,\qquad\qquad q_{\Omega}'(0)-\frac{2}{n+3}q_{0}^{2}q_{\Omega}(0)=0,\qquad\qquad q_{\Pi}(0)-q_{0}=0,\end{aligned}$$ and at $y_0=1$ $$\begin{aligned}
&q_{f}'(1)+\frac{n-1}{2}q_{f}(1)+q_{h}(1)=0,\qquad\qquad q_{g}(1)+q_{h}(1)=0, \qquad\qquad q_{\Omega}'(1)=0,\\
&q_{h}'(1)+\frac{n+1}{2}q_{h}(1)=0, \qquad\qquad q_{\Pi}'(1)+\frac{1}{2(n+3)}\big((n+2)^{2}-q_{\Omega}(1)^{2}\big)q_{\Pi}(1)=0. \label{eq:BC2}\end{aligned}$$
Note that since the constraint equations (\[eq:C1Eq\]) and (\[eq:C2Eq\]) are explicitly satisfied at the boundary points by these boundary conditions, the derivatives of the constraints, (\[eq:C1deriv\]) and (\[eq:C2deriv\]), are also satisfied, meaning that the constraints are guaranteed to be satisfied over the whole domain.
To solve the equations of motion, we approximated each of the analytic $q$-functions by an order $N$ expansion in terms of Chebyshev polynomials, as described in [@Pfeiffer:2005zm]. Next, this expansion was inserted into the equations of motion at each of the $N-1$ interior Chebyshev grid points (excluding the boundary points $y=0,1$). At the two boundary points, the Chebyshev expansions were inserted into the boundary conditions (\[eq:BC1\])–(\[eq:BC2\]). The equations of motion and boundary conditions then became a system of non-linear algebraic equations for the spectral coefficients of the Chebyshev approximations. A Newton-Raphson method was then employed by linearizing the equations of motion with respect to the spectral coefficients.
The initial seed solution at $q_0=0.01$ was obtained using the perturbative results in [@Stotyn:2011ns], while subsequent solutions used the previously generated solution as a seed with a step size of $\Delta q_0=0.01$. In all cases, convergence to a solution occurred after 5 iterations, where convergence was determined by the changes to the spectral coefficients being the same order of magnitude for 2 subsequent iterations; typically at convergence these changes are on the order of $10^{-30}$ or smaller.
The fact that Chebyshev approximations have exponential convergence[@Pfeiffer:2005zm] leads to a number of checks on the validity of the solution generated by this method. First of all, since the Chebyshev polynomials are all of order 1, and the $j^{\mathrm{th}}$ spectral coefficient is approximately $C_j\approx C_0e^{-kj}$ for some constant $k$ that depends on the approximated function, the expansion at any point obeys the inequality $$q\le\sum_{j=0}^{N}C_0 e^{-kj}\sim \int_{0}^{N}{C_0 e^{-kj}dj}.$$ A rough estimate of the error of the approximation is then given by $$\mathrm{error}\approx \int_{N}^{\infty}{C_0 e^{-kj}dj} = \frac{C_0 e^{-kN}}{k}$$ where the term in the numerator is the $N^{\mathrm{th}}$ spectral coefficient, $C_N$. Therefore, an order of magnitude estimate of the percent-error in the Chebyshev expansion is simply given by $$\mathrm{\%~error}\approx \frac{C_N}{C_0},$$ which is easily computable by inspection of the output coefficients of an iteration. Indeed, ensuring this quantity maintained a value below $10^{-8}$ was one of the criteria used to determine whether the Chebyshev grid used was sufficiently dense.
Similarly, because of the exponential convergence of the spectral coefficients, a plot of $\log|C_j|$ vs $j$ should be approximately a straight line with a negative slope $-k$. This property was used extensively to determine whether there was a sufficient number of grid points or, equally as important, a sufficient amount of digit precision in the calculations. Since the spectral coefficients decrease exponentially, after a certain point, $j=n$, the digit precision of the calculations, denoted Prec, cannot resolve the difference between spectral coefficients: the remaining coefficients $C_j$ for $j=n,...,N$ are then numerical noise. To overcome this obstacle, we always worked with Prec$~\gtrsim N$ and found that plots of $\log|C_j|$ vs $j$ behaved as expected.
Unfortunately, as $q_0$ increases, this becomes computationally quite expensive and time intensive. When $q_0$ is sufficiently small, the curvature scale is small everywhere and a sparse Chebyshev grid of $N\sim 20$ suffices to solve the equations of motion. As $q_0$ begins to rise, both $N$ and Prec need to correspondingly increase to resolve the curvature, with the time for each iteration being at least quadratic in both. We found that we typically reached the point of diminishing returns at approximately $N=100$, which was the upper limit used to generate the results of the following section.
Finally, we note that the above considerations are not always sufficient to determine whether the generated solutions are viable. To determine this, we imposed that the first law of thermodynamics must be satisfied to one part in $10^6$. We found that our code would continue to generate solutions well beyond this limit on the first law if we blindly pushed it forward. However, if we evaluated the equations of motion for such spurious solutions at points in-between the Chebyshev grid points, we found that they were indeed not satisfied within the required tolerance, whereas they were satisfied for solutions obeying the first law bound. This was primarily how we decided when to increase the grid density and digit precision from one solution to the next.
Representative plots of the metric functions and scalar field function constructed by the above numerical methods are plotted in figure \[fig:functions\] in appendix A: these plots are for $D=5$ and $q_0=5.0$, although the generic features of the plots are found to change very little with varying dimension or varying $q_0$. It is clear that all of these functions have their expected behaviours and are free from singularities, except for $f(y)$ which diverges as $y\rightarrow 1$ as one expects due to the AdS boundary conditions.
Results and Discussion {#Results}
======================
With confidence in the validity of our generated solutions assured by the methods detailed above, we can now present the results. Of course, the physically interesting data for these boson stars amount to their physical properties, all of which can be written explicitly in terms of the boundary values of the analytic $q$ functions. For instance, the Kretschmann invariant, $K=R_{abcd}R^{abcd}$, at the centre of the boson star is given by $$K_{n}=\frac{2\left(n+1\right)}{\ell^4}\left[\left(n+2\right)+\frac{3\left(n-1\right)\left(n+3\right)}{n}q_{h}'(0)^{2}-4q_{0}^{2}+\frac{4}{n}q_{0}^{4}\right],$$ while the thermodynamic quantities take the form [@Ashtekar:1999jx; @Das:2000cu] $$\begin{aligned}
M={}&\frac{\pi^{\frac{n-1}{2}}\ell^{n-1}}{8\left(\frac{n-1}{2}\right)!}\big((n+1)q_h(1)-nq_f(1)\big),\\
J={}&\frac{(-1)^{\frac{n+1}{2}}\pi^{\frac{n-1}{2}}\ell^n}{4\left(\left(\frac{n-1}{2}\right)!\right)^2}q_\Omega^{(\frac{n+1}{2})}(1),\\
\omega={}&-q_\Omega(1).\end{aligned}$$ as can be shown using equations (\[EJ\]) and (\[fq\] – \[Pq\]). Similarly, the perturbative parameter $\epsilon$ used in [@Stotyn:2011ns] is given by $$\epsilon=q_\Pi(1).$$ Given these quantities, we present here the results for $n=3,5,7,9$, where we include $n=3$ both for completeness, and to demonstrate that we reproduce the existing results of [@Dias:2011at]. 0.5cm
\
In all dimensions we find behaviour that is qualitatively similar to that which was found in [@Dias:2011at] for $n=3$. In figure \[fig:subfigures\], we plot mass, $M$, angular momentum, $J$, angular velocity, $\omega$, and asymptotic scalar field amplitude, $\epsilon$, as functions of $q_{0}$, which can be seen to uniquely parameterize the family of boson stars in each dimension.
We find numerically that each of these quantities displays damped harmonic oscillations as $q_0$ increases, with the amplitude of the oscillation decreasing as the dimension increases. Furthermore, as $q_0$ tends toward infinity, these quantities all asymptote to finite central values that depend on the dimension.
The damped oscillations have interesting consequences. For instance, the maxima of the $M$ vs $q_0$ graph represent instabilities in the boson stars: adding a small but finite perturbation with energy $\delta M$ will necessarily force the solution off the family of boson stars, likely forming a (potentially hairy) black hole as a result. It is tantalizing to conjecture that this effect may be intimately tied to the turbulent instability in AdS spaces. Recall that AdS is non-perturbatively unstable to forming a black hole because of the large number of resonant frequencies coming from the normal modes being integer multiples of the AdS frequency[@Dias:2012tq]. Perhaps similar normal mode resonances are occurring at the boson star mass maxima, leading to non-perturbative instabilities toward black hole formation at these particular values of $q_0$. From the holographic dual point of view, this would imply that generally the strongly coupled dual CFT states would not thermalize, except at particular points in parameter space corresponding to an infinite tower of resonances.
The interpretation of the mass minima is less straightforward, although these points are perturbatively unstable in parameter space in the sense that a small perturbation could push the solution either to the left or the right along the family of boson stars. The holographic dual interpretation is similarly obscured due to a lack of a map between the gravitational degrees of freedom and those in the CFT for these objects. Nevertheless, these minimum mass states still represent CFT states that never thermalize since small but finite perturbations keep the solution within the family of boson stars. That being said, from the perturbative results of [@Stotyn:2011ns] we know that in the perturbative regime (*i.e.* in the vicinity of the first minimum located at $q_0=0$) the boson stars are perturbatively joined to small hairy black holes. This is presumably because the geometry of small boson stars is still sufficiently close to AdS that approximate resonances still exist in the normal modes, causing the instability. It is far from obvious whether the subsequent mass minima have similar behaviour, like the mass maxima do. We leave this, and other issues of stability, as an open question for future consideration.
In the perturbative regime of Ref. [@Stotyn:2011ns], the parameter $\epsilon$ uniquely determined the boson star solutions. In the full numerical treatment this is no longer the case, as can be seen in figure \[fig:subfigures2\]. The damped oscillations with respect to $q_0$ seen in figure \[fig:subfigures\] show up as damped spirals with respect to $\epsilon$ in figure \[fig:subfigures2\], with the centre of the spirals corresponding to the asymptotic values attained as $q_0\rightarrow\infty$. Furthermore, plots of the Kretschmann scalar, $K$, against $\epsilon$ shows that the curvature at the centre of the boson star is quickly diverging as $q_0$ increases. In the limit $q_0\rightarrow\infty$, the boson star develops a curvature singularity and hence collapses to form a black hole. We note, however, that as $q_0$ increases, the family of solutions is spiralling tighter and tighter around the asymptotic black hole values and the boson star will become non-perturbatively unstable to forming a black hole before the limit $q_0\rightarrow\infty$ is reached. This can also be seen in figure \[fig:subfigures\] by the fact that the $M$ vs $q_0$ graphs asymptote to horizontal lines in all dimensions, meaning that adding a small but finite perturbation will necessarily force the solution off the boson star branch, leading to the formation of a black hole.
\
Finally, we note in figure \[fig:2fourth\] that plots of $M$ vs $J$ form jagged zig-zags in each dimension. This is a direct consequence of the first law of thermodynamics, combined with the boson stars being a one parameter family of solutions. Rewriting the first law as $$\frac{dM}{dq_0}=\omega\frac{dJ}{dq_0} \label{eq:FirstLaw2}$$ demonstrates that the extrema of the $M$ vs $q_0$ and the $J$ vs $q_0$ graphs of figure \[fig:subfigures\] must occur at the same value of $q_0$ in a given dimension. Such points represent the cusps of figure \[fig:2fourth\], while the slopes of the zigs and zags represent the value of $\omega$ by virtue of (\[eq:FirstLaw2\]). As the space-time dimension increases, the amplitude of oscillation for $\omega$ decreases, meaning that the zig-zags of figure \[fig:2fourth\] become more and more collinear as can be seen by comparing $n=3,5,7,9$. Lastly, since the amplitudes of oscillation for $M$ and $J$ decrease with increasing $q_0$, the amplitude of the $M$ vs $J$ zig-zags decrease, as can be clearly seen for $n=3,5$ in figure \[fig:2fourth\].
Conclusions {#Conclusion}
===========
Using a relaxation procedure on a Chebyshev grid to solve the equations of motion, we have numerically obtained asymptotically AdS boson star solutions with only one Killing field in all odd space-time dimensions of interest in string theory, *i.e.* up to $D=11$.[^1] Previously, these solutions had only been produced perturbatively in $D=3,5,7,9,11$, and numerically in $D=5$. Our results reveal that the behaviour of boson stars in 7, 9 and 11 dimensions is qualitatively very similar to their 5-dimensional counterparts originally found in [@Dias:2011at]. In particular, their various properties ($M$, $J$, $\omega$, and $\epsilon$) undergo damped harmonic oscillations about finite critical values as the central energy density is increased. Accordingly, we see that the amplitude of these oscilations is lower in higher space-time dimension.
Since the publication of Ref. [@Dias:2011at] a number of developments in the non-linear (in)stability of asymptotically AdS space-times have been made. For instance, it was argued in [@Dias:2011ss] that global AdS is non-perturbatively unstable to the formation of a black hole; the heuristic idea is that the reflecting boundary conditions of AdS means that eventually the finite energy perturbations will come together in such a way that the energy is sufficiently concentrated to form a black hole. In light of this heuristic picture, it seems plausible that *any* asymptotically AdS space-time will be non-perturbatively unstable to the formation of a black hole. However, it was then discovered in [@Dias:2012tq] that this instability is really due to the high level of symmetry present in global AdS, since the normal mode frequencies are all resonant with the AdS frequency. This was vindicated in Ref. [@Buchel:2013uba], which found a large class of boson star initial data that were non-perturbatively stable.
The current work mirrors the 5 dimensional solution found in [@Dias:2011at] and given the above developments, we are able to further elucidate the rather rich physics of these boson star solutions. For instance, the damped oscillations in the mass implies that the boson stars corresponding to the mass maxima are all non-perturbatively unstable to black hole formation. This could be related to the resonances of global AdS, although this remains conjecture at this point. From the holographic dual perspective, this implies that the boson stars correspond to strongly coupled CFT states that never thermalize, except at the infinite tower of resonances corresponding to the boson star mass maxima. Little else is known about the holographic dual description, other than that the CFT state is a localized scalar operator defining a current that breaks rotational symmetry. Whether there are any physically realizable quantum systems that share some of these properties remains to be seen, although such systems would be very intriguing indeed, on account of the fact that they never thermalize in finite time (modulo the aforementioned exceptions).
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada. D.L. was funded in part by the Rhodes Trust. M.O. acknowledges additional support from the University of Waterloo Department of Applied Mathematics. We would like to thank M.Chanona, H. Pfeiffer, O. Dias, and J. Santos for helpful discussions at various stages of this project.
Representative Plots of the Metric and Scalar Field Functions
=============================================================
\
\
[99]{}
D. J. Kaup, “Klein-Gordon Geon,” Phys. Rev. [**172**]{}, 1331 (1968). R. Ruffini and S. Bonazzola, “Systems of selfgravitating particles in general relativity and the concept of an equation of state,” Phys. Rev. [**187**]{}, 1767 (1969). P. Jetzer, “Boson stars,” Phys. Rept. [**220**]{}, 163 (1992). A. Bernal, J. Barranco, D. Alic and C. Palenzuela, “Multi-state Boson Stars,” Phys. Rev. D [**81**]{}, 044031 (2010) \[arXiv:0908.2435 \[gr-qc\]\]. E. Berti and V. Cardoso, “Supermassive black holes or boson stars? Hair counting with gravitational wave detectors,” Int. J. Mod. Phys. D [**15**]{}, 2209 (2006) \[gr-qc/0605101\]. M. Kesden, J. Gair and M. Kamionkowski, “Gravitational-wave signature of an inspiral into a supermassive horizonless object,” Phys. Rev. D [**71**]{}, 044015 (2005) \[astro-ph/0411478\]. S. L. Liebling and C. Palenzuela, “Dynamical Boson Stars,” Living Rev. Rel. [**15**]{}, 6 (2012) \[arXiv:1202.5809 \[gr-qc\]\]. Y. Brihaye, B. Hartmann and S. Tojiev, “Stability of charged solitons and formation of boson stars in 5-dimensional Anti-de Sitter space-time,” Class. Quant. Grav. [**30**]{}, 115009 (2013) \[arXiv:1301.2452 \[hep-th\]\]. D. Astefanesei and E. Radu, “Boson stars with negative cosmological constant,” Nucl. Phys. B [**665**]{}, 594 (2003) \[gr-qc/0309131\]. O. J. C. Dias, G. T. Horowitz and J. E. Santos, “Black holes with only one Killing field,” JHEP [**1107**]{}, 115 (2011) S. Stotyn, M. Park, P. McGrath and R. B. Mann, “Black Holes and Boson Stars with One Killing Field in Arbitrary Odd Dimensions,” Phys. Rev. D [**85**]{}, 044036 (2012) \[arXiv:1110.2223 \[hep-th\]\]. M. Colpi, S. L. Shapiro and I. Wasserman, “Boson Stars: Gravitational Equilibria of Selfinteracting Scalar Fields,” Phys. Rev. Lett. [**57**]{}, 2485 (1986). J. -w. Ho, F. C. Khanna and C. H. Lee, “Boson stars with selfinteracting quantum scalar fields,” gr-qc/0207073. O. J. C. Dias, P. Figueras, S. Minwalla, P. Mitra, R. Monteiro and J. E. Santos, “Hairy black holes and solitons in global $AdS_5$,” JHEP [**1208**]{}, 117 (2012) \[arXiv:1112.4447 \[hep-th\]\]. Y. Brihaye, B. Hartmann and E. Radu, “Boson stars in SU(2) Yang-Mills-scalar field theories,” Phys. Lett. B [**607**]{}, 17 (2005) \[hep-th/0411207\]. R. -G. Cai, J. -Y. Ji and K. -S. Soh, “Hairs on the cosmological horizons,” Phys. Rev. D [**58**]{}, 024002 (1998) \[gr-qc/9708064\]. G. Fodor, P. Forgacs and M. Mezei, “Boson stars and oscillatons in an inflationary universe,” Phys. Rev. D [**82**]{}, 044043 (2010) \[arXiv:1007.0388 \[gr-qc\]\].
B. Hartmann and J. Riedel, “Supersymmetric Q-balls and boson stars in (d+1) dimensions,” Phys. Rev. D [**87**]{}, 044003 (2013) \[arXiv:1210.0096 \[hep-th\]\]. J. Bjoraker and Y. Hosotani, “Stable monopole and dyon solutions in the Einstein-Yang-Mills theory in asymptotically Anti-de Sitter space,” Phys. Rev. Lett. [**84**]{}, 1853 (2000) \[gr-qc/9906091\]. O. J. C. Dias, G. T. Horowitz and J. E. Santos, “Gravitational Turbulent Instability of Anti-de Sitter Space,” Class. Quant. Grav. [**29**]{}, 194002 (2012) \[arXiv:1109.1825 \[hep-th\]\]. P. Bizon and A. Rostworowski, “On weakly turbulent instability of anti-de Sitter space,” Phys. Rev. Lett. [**107**]{}, 031102 (2011) \[arXiv:1104.3702 \[gr-qc\]\]. J. Jalmuzna, A. Rostworowski and P. Bizon, “A Comment on AdS collapse of a scalar field in higher dimensions,” Phys. Rev. D [**84**]{}, 085021 (2011) \[arXiv:1108.4539 \[gr-qc\]\]. M. Maliborski and A. Rostworowski, Phys. Rev. Lett. [**111**]{}, 051102 (2013) \[arXiv:1303.3186 \[gr-qc\]\]. A. Buchel, S. L. Liebling and L. Lehner, “Boson Stars in AdS,” arXiv:1304.4166 \[gr-qc\]. O. J. C. Dias, G. T. Horowitz, D. Marolf and J. E. Santos, “On the Nonlinear Stability of Asymptotically Anti-de Sitter Solutions,” Class. Quant. Grav. [**29**]{}, 235019 (2012) \[arXiv:1208.5772 \[gr-qc\]\]. V. Moncrief and J. Isenberg, “Symmetries of Higher Dimensional Black Holes,” Class. Quant. Grav. [**25**]{}, 195015 (2008) \[arXiv:0805.1451 \[gr-qc\]\]. S. Hollands, A. Ishibashi and R. M. Wald, “A Higher dimensional stationary rotating black hole must be axisymmetric,” Commun. Math. Phys. [**271**]{}, 699 (2007) \[gr-qc/0605106\]. S. W. Hawking, “Black holes in general relativity,” Commun. Math. Phys. [**25**]{}, 152 (1972). B. Hartmann, B. Kleihaus, J. Kunz and M. List, “Rotating Boson Stars in 5 Dimensions,” Phys. Rev. D [**82**]{}, 084022 (2010) S. Stotyn and R. B. Mann, “Another Mass Gap in the BTZ Geometry?,” J. Phys. A [**45**]{}, 374025 (2012) \[arXiv:1203.0214 \[gr-qc\]\]. S. Stotyn, M. Chanona and R. B. Mann, “Numerical Boson Stars with a Single Killing Vector II: the D=3 Case,” arXiv:1309.2911 \[hep-th\]. A. Ashtekar and S. Das, “Asymptotically Anti-de Sitter space-times: Conserved quantities,” Class. Quant. Grav. [**17**]{}, L17 (2000) \[hep-th/9911230\]. S. Das and R. B. Mann, “Conserved quantities in Kerr-anti-de Sitter space-times in various dimensions,” JHEP [**0008**]{}, 033 (2000) \[hep-th/0008028\]. H. P. Pfeiffer, “Initial data for black hole evolutions,” gr-qc/0510016.
[^1]: The full numerical boson star solution in $D=3$ can be found separately in Ref. [@Stotyn:2013spa] due to its significantly different behaviour than its higher dimensional counterparts presented here.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present cluster-DMFT (CTQMC) calculations based on a downfolded tight-binding model in order to study the electronic structure of vanadium dioxide (VO$_2$) both in the low-temperature ($M_1$) and high-temperature (rutile) phases. Motivated by the recent efforts directed towards tuning the physical properties of VO$_2$ by depositing films on different supporting surfaces of different orientations we performed calculations for different geometries for both phases. In order to investigate the effects of the different growing geometries we applied both contraction and expansion for the lattice parameter along the rutile $c$-axis in the 3-dimensional translationally invariant systems miming the real situation. Our main focus is to identify the mechanisms governing the formation of the gap characterizing the $M_1$ phase and its dependence on strain. We found that the increase of the band-width with compression along the axis corresponding to the rutile $c$-axis is more important than the Peierls bonding-antibonding splitting.'
author:
- 'Bence Lazarovits$^{1,2}$, Kyoo Kim$^1$, Kristjan Haule$^1$, and Gabriel Kotliar$^1$'
title: 'Effects of strain on the electronic structure of VO$_2$'
---
Introduction
============
Vanadium oxides compounds exhibiting exotic transport phenomena are subjects of extensive interest. In particular vanadium dioxide, VO$_2$, undergoes a first-order transition from a high-temperature metallic phase to a low-temperature insulating phase at almost the room temperature ($T=340\,$K) [@morin]. There are intensive efforts around the world to make devices such as switches, transistors, detectors, varistors, phase change memory, exploiting the unique properties of VO$_2$.
At low-temperature VO$_2$ has a simple monoclinic ($M_1$) structure with space group $P2_1/c$ ($M_1$ phase) while at high temperature it has a simple tetragonal lattice with space group $P4_2/mnm$ rutile ([*R*]{}-phase) as displayed in Fig. \[fig:structure\]. The lattice structures of the two phases are closely related as emphasized in Fig. \[fig:structure\] by showing similar wedges of the different lattices: the $M_1$ unit cell is similar to the [*R*]{} unit cell, when the latter is doubled along the rutile $c$-axis. For the sake of simplicity, we will use the notation $c$-axis for both the $M_1$ and [*R*]{} phase when referring to the axis equivalent to the rutile $c$-axis (in $M_1$ phase this axis is sometimes called $a$-axis). The $M_1$ phase is characterized by a dimerization of the vanadium atoms into pairs, as well as a tilting of these pairs with respect to the $c$-axis [@KL70; @andresson] as indicated by showing bonds between the paired V atoms in Fig. \[fig:structure\]b.
The electronic and transport properties are dramatically different for the two phases. The resistivity jumps by several orders of magnitude through the phase transition, and the crystal structure changes from [*R*]{}-phase at high-temperature to monoclinic $M_1$-phase at low-temperature [@morin; @allen]. While the rutile phase is a conductor, the $M_1$ phase is an insulator with a gap of $\sim0.6$ eV [@KHH+06] at the Fermi energy. VO$_2$ has also attracted a great deal of attention for its ultrafast optical response, switching between the [*R*]{} and the $M_1$ phases [@lysenko; @cavalleri; @Baum]. Despite the large number of experimental [@Baum; @MH02; @Maekawa; @Qazilbash1; @Qazilbash2; @Qazilbash3; @QBW+08; @Eguchi; @Ruzmetov; @Braichovich; @Dmitry; @KHH+06] and theoretical [@BPL+05; @Tomczak1; @Tomczak2; @Sakuma; @Gatti; @E02; @RSA94; @LCM06; @LIB05] studies focusing on this material the physics driving this phase transition and the resulting optical properties are still not identified undoubtedly. In the theoretical works using LDA [@E02; @RSA94; @allen] the formation of the gap was not found. The single site DMFT approach [@LCM06; @LIB05] is known to correctly describe the metallic phase but can not take into account the formation of the bonding states of the dimers in the insulating phase which requires the cluster DMFT method [@BPL+05]. As shown in Ref. [@Tomczak2] the fully interacting one-electron spectrum of the $M_1$ phase can be reproduced by an effective band structure description. The optical conductivities calculated by using cluster DMFT shows good agreement with the experiment for both phases [@TB09].
The metal-insulator transition (MIT) of VO$_2$ is usually attributed to two different physical mechanisms. One of them is the Peierls physics, i.e. the dimerization of the V atoms along the rutile $c$-axis [@g60], and consequently opening of the gap in the Brillouin-zone reduced in the direction along the $c$-axis. In this case, the gap opening can be explained in the framework of effective one-electron theories. The other one is the Mott mechanism where the gap opens due to the strong Coulomb repulsion between the localized V-$3d$ orbitals [@ZM76] and the related dynamical effects. Understanding in detail the interplay and relative importance of both the Peierls and the Mott mechanisms for the electronic structure is crucial for controlling this material with an eye towards applications. For example, whether the driving force of this transition is electronic (i.e. occurring on femtosecond timescales) or structural (occurring on the picosecond timescale) is important to understand the speed of the switching from the $M_1$ to the rutile phase. From the perspective of applications, in order to control the properties of vanadium dioxide, it is essential to identify the effects of compressive and tensile stress resulting from the various substrates, on which the films are deposited [@MH02; @Maekawa].
The experimental results of Ref. [@MH02] showed that compressive uniaxial strain (along the $c$-axis) stabilizes the metallic phase. This result cannot be explained by applying a simple Peierls picture exclusively. The Peierls mechanism predicts that a compression along the $c$-axis increases the splitting between the bonding and antibonding state, formed by the combinations of particular $3d$ orbitals residing on the different V atoms in a dimer, and hence would promote insulating state. The Peierls mechanism thus increases the tendency to open the gap in the $M_1$ phase, which would stabilize the insulating phase at the expense of the metallic phase. In this picture, the transition temperature would increase under uniaxial compressive strain, opposite to experiment [@MH02].
The LDA calculations fail to reproduce the gap in the $M_1$ phase [@E02]. This suggests that the correlations are important in this material. Motivated by this and the above mentioned experiment we have examined the influence of strain on the electronic structure of VO$_2$ microscopically applying the dynamical mean-field theory [@GKK+96] combined with the local density approximation of density functional theory (LDA+DMFT) [@KSH+06]. We carried out LDA+DMFT calculations of VO$_2$ under strain. We established that in addition to the Peierls distortion, other factors like the position of the $e_g^\pi$ band and the band-width of the bonding state in the $M_1$ phase showing definite sensitivity to the strain, play a defining role in driving the MIT in VO$_2$. Theoretical predictions for the strain dependence of many spectroscopic quantities are made, including the photoemission, the optical conductivity, the inverse photoemission and XAS in the $R$ and $M_1$ phases. The insights achieved in this study together with the computational machinery developed, will serve as a basis for rational material design of VO$_2$ based applications.
Method
======
The unit cell of the rutile structure contains two vanadium and four oxygen atoms while the $M_1$ structure contains four vanadium (2 dimers) and eight oxygen atoms. The calculations in the rutile phase were performed with the doubled unit cells (four V and eight O atoms) to allow the formation of a bonding state between the V atoms separated along the $c$-axis, a mechanism which has dramatic effect in the $M_1$ phase. The lattice structure parameters for the rutile (referred as $a_0$, $b_0$ and $c_0$ in the following) and $M_1$ cases ($\tilde{a}_0$, $\tilde{b}_0$, $\tilde{c}_0$ and $\beta_0$) were published by McWhan [*et al.*]{} in Ref. [@MMR+74] and Kierkegaard and Longo in Ref. [@KL70], respectively. Note that our notation of the axis differs from the published data [@E02], in order to emphasize the similarity of the [*R*]{} and $M_1$ structures.
The strain was applied by changing the lattice constants but not the internal parameters of the atomic positions. The strain is characterized by the ratio between the experimental lattice constant along the $c$-axis of the unstrained crystal, and the one used in the calculation, $r=c/c_0$ in the rutile and $r=\tilde{c}/\tilde{c}_0$ in the $M_1$ case. The applied strains correspond to $r=0.98$, 1.00 and 1.02 ratios both for the [*R*]{} and $M_1$ phases in order to make calculations which are compatible with the experiment in Ref. [@MH02]. The lattice parameters were changed with the constraint to preserve the volume of the original unit cell and to form a structure to closely resemble the original structure. This is illustrated in Fig. \[fig:structure\] for both phases. The constant volume constraint in the rutile phase requires for the other two lattice constants $a=a_0/\sqrt{r}$ and $b=b_0/\sqrt{r}$. For the $M_1$ structure, two other constraints were introduced: the two ratios $(\tilde{b}\sin\beta)/\tilde{a}=0.9988$ and $-2(\tilde{b}\sin\beta)/\tilde{c}=1.0096$ were kept constant. Hence we used the following lattice constants and angles $\tilde{c}=r\tilde{c}_0$, $\tilde{a}= \tilde{a}_0 /\sqrt{r}$, $\tan(\beta)= \tan(\beta_0) /r^{3/2} $ and $\tilde{b}=\tilde{b}_0\sqrt{r^2\cos^2(\beta_0)+\sin^2(\beta_0)/r}$. For the [*R*]{}-phase, we used a local coordinate system introduced by Eyert in Ref , as shown in Fig. \[fig:orbitals\]. In the rutile geometry the $x$-axis is parallel to the rutile $c$-axis while the $z$-axis points along the $[110]$ direction, pointing to the adjacent oxygen atom in the vanadium plane. In the $M_1$ phase, our local coordinate system is somewhat tilted to align to monoclinic geometry. The symmetry classification of the electronic orbitals is also adopted, namely $d_{3z^2-r^2}$ and $d_{xy}$ stand for the $e_g$ states and $d_{xz}$, $d_{yz}$ and $d_{x^2-y^2}$ for the $t_{2g}$ orbitals. The $t_{2g}$ manifold is divided further: the notation $a_{1g}$ is used for $d_{x^2-y^2}$ state and $e^{\pi}_g$ for the $d_{xz}$ and $d_{yz}$ states. In this choice of the local coordinate system the dimerization strongly affects the $d_{x^2-y^2}$ states, which have large electron density along the V-V bond.
$
\begin{array}{ccc}
\includegraphics[height=0.3\linewidth,clip=true, viewport=0pt 40pt 500pt 500pt]{fig2a.png}&
\includegraphics[height=0.3\linewidth,clip=true, viewport=0pt 40pt 500pt 500pt]{fig2b.png}&
\includegraphics[height=0.3\linewidth,clip=true, viewport=0pt 40pt 500pt 500pt]{fig2c.png}\\
d_{x^2-y^2}&d_{xz}&d_{yz}
\end{array}$
First, a self-consistent LDA calculation was performed using the linear muffin-tin orbitals method combined with the atomic sphere approximation (LMTO-ASA) [@lmto]. For a satisfactory description of the interstitial region 48 empty spheres were added in the LDA-ASA calculations. In the next step we determined a [*downfolded parameters*]{} of the low energy, effective Hamiltonian, $H_{eff}$, including only the $3d$ $t_{2g}$ subset of electronic states lying in the vicinity of the Fermi energy. The effective Hamiltonian is constructed as: $$H_{\mathrm{eff}}=
\sum\limits_{\sigma}
\sum\limits_{\langle i,i'\rangle}
\sum\limits_{ \alpha,\alpha'}
\left(
\epsilon_{i,\alpha,\sigma}
\delta_{\alpha,\alpha'}
\delta_{i,i'}
+
t_{i,\alpha,\sigma;i',\alpha',\sigma}
\right)
c_{i,\alpha,\sigma}^{\dagger}
c_{i',\alpha',\sigma}$$ with off-diagonal hopping ($t$) and diagonal one-electron energy ($\epsilon$) parameters belonging to different sites $(i)$ and states ($\alpha =d_{xz}$, $d_{yz}$, $d_{x^2-y^2}$) with a spin character $\sigma$. The applicability and accuracy of the above downfolding method is determined by the mutual positions and characters of the electronic bands. In the studied case of VO$_2$ both in the $M_1$ and [*R*]{} phases the bands in the proximity of the Fermi level have mainly $t_{2g}$ character and are well separated from both the low-lying $2p$-bands of the oxygen, and from the $e_g$ bands of vanadium, due to the strong crystal-field splitting ($\sim 3.0-3.5$ eV [@E02]). The crystal field splitting is due to oxygen octahedron surrounding the individual V atoms. This mapping of the LDA band structure onto a tight-binding type Hamiltonian provides a reasonably good description of the electronic structure as compared to the LDA serving as a good starting point for the following DMFT calculations. The evolution of the intra dimer V-V hopping matrix elements corresponding to the three different $t_{2g}$ orbitals induced by the strain are shown in Fig. \[fig:hoppings\]. Although, the relative changes in the parameters are small, the manifestations of these changes in the low energy spectral function, and therefore in the transport properties, are significant.
The effects of the electron-electron interaction beyond LDA are treated by means of the dynamical mean-field theory [@GKK+96; @KSH+06]. The coulomb interaction was taken into account for the V-V dimers by using the cluster extension of the DMFT [@KSH+06]. In this approach the quantities describing the electronic states of a cluster (Green’s function, self-energy, [ *etc.*]{}) are matrices of both the site and electronic state indices. Having two sites (two V atoms) and three states ($t_{2g}$ states) the resulted matrices have dimensions of $6\times6$ for non-magnetic calculations. For solving the impurity problem, we used the continuous time quantum Monte Carlo method (CTQMC) [@H07] at the temperatures $T=232$ K and $T=390$ K, below and above the critical temperature, for the $M_1$ and [*R*]{} cases, respectively. In the present calculations we assumed on-site Coulomb interaction on the V atoms written as: $$H_U=
U
\sum\limits_{i=1,2}
\sum\limits_{ \alpha \alpha'}
\sum\limits_{ \sigma \sigma'}
c_{i,\alpha\sigma}^{\dagger}
c_{i,\alpha\sigma}
c_{i,\alpha'\sigma'}^{\dagger}
c_{i,\alpha'\sigma'}
(1-\delta_{\alpha\alpha'}
\delta_{\sigma\sigma'})$$ excluding the terms with similar orbital and spin character at same time. We fixed the parameter $U$ at $U=2.2$ eV, to reproduce the experimentally measured gap in the $M_1$ phase, which shows a weak temperature dependence between 100 and 340 K, varying between $\sim0.75$ and $\sim0.6$ eV [@OFO01; @KHH+06]. The strong dependence of the gap size on the $U$ parameter is shown in Fig. \[fig:var\_U\].
The linear $U$ dependence of the renormalization factor $Z$ calculated from the real part of the self-energy of the different orbitals $(\alpha)$ as $Z_{\alpha}=(1-\partial \mathrm{Re}\Sigma_{\alpha}
(\omega)/\partial \omega)^{-1}$ and the electronic specific heat, $\gamma=\sum\limits_{\alpha}\rho_{\alpha}(0)/Z_{\alpha}$, of the rutile phase are shown in Fig. \[fig:Z\_gamma\]. The $Z$ value obtained at $U=2.2$ eV is in good agreement with the one published in Ref. [@BPL+05]. An alternative method to determine the value of $Z$ factor is to take the ratio of the experimentally measured [*plasma frequency*]{} in the rutile phase ($\omega_p^{exp}=2.75$ eV)[@QBW+08] and devide it by the band theoretical LDA-LAPW calculation [@wien2k] ($\omega_p^{LDA}\approx4.1$ eV), i.e., $\omega_p^{exp}/\omega_p^{LDA}\approx0.67$, which agrees well with $Z\approx 0.62$ in the case of $U=2.2$ eV. Although the $U$ was chosen to reproduce the gap in the $M_1$-phase, it is satisfying that this value is also compatible with the alternative estimation, based exclusively on the optical and thermodynamic properties of the rutile phase.
In order to achieve a structured (almost diagonal) self-energy matrix for the cluster, we used a basis of symmetric $(s)$ and anti-symmetric $(as)$ combination of the states localized on the individual V atoms of the dimers defined as $$(c_{\alpha,\sigma }^{s(as)})^\dagger
=
\frac{1}{\sqrt{2}}(c_{1,\alpha,\sigma}^{\dagger} \pm c_{2,\alpha,\sigma}^{\dagger})
,\quad
(\alpha \in t_{2g})
\quad.
\label{eq:b-ab}$$ To obtain the physical properties at real energies we performed analytic continuation to the real axis using a recently developed method of expansion in terms of modified Gaussians and a polynomial fit at low frequencies, as described in details in Ref. [@HYK09].
Results and Discussion
======================
Fig. \[fig:lda\_dos\] displays the LDA $t_{2g}$ densities of states (DOS) of the V atoms in both the rutile and $M_1$ phases for different $r$ ratios ($r=0.98$, 1.00, 1.02). The total band-width ($\sim2.6$ eV for both phases) and also the fine details of the DOS obtained by the LDA-LMTO agree well with previous studies [@E02; @LCM+06]. The minor discrepancies are probably a consequence of the different electronic structure method or the slightly different geometry.
The effect of the stress applied along the rutile [*c*]{}-axis on the electronic DOS can be clearly seen. The position of the $a_{1g}$ peak is changed considerably, while the total band-width is only slightly changed ($<0.1$ eV) under application of stress. Only the $d_{x^2-y^2}$ orbital changes significantly, while the two $e_{g}^\pi$ orbitals remain mostly unchanged. This is due to the sensitivity of the overlap integral of $d_{x^2-y^2}$ states along the $c$-axis. It is clear from Fig. \[fig:lda\_dos\] that the splitting of the $d_{x^2-y^2}$ peaks strongly increases with decreasing lattice parameter $c$ roughly following the linear dependence of the intra-dimer hopping parameter $t_{x^2-y^2,x^2-y^2}$ on ratio $r$, shown in Fig. \[fig:hoppings\]. However, even under compressive stress of $r=0.98$, the splitting is not large enough to open a gap in the $M_1$ phase. The value of the intra-dimer hopping parameters corresponding to the $d_{x^2-y^2}$ states, which plays a significant role in the formation of the electronic structure of VO$_2$, are $-0.30$ eV and $-0.61$ eV for the [*R*]{} and $M_1$ cases, respectively. These values are in a good agreement with published values of Ref. . One can observe that the splitting of the $d_{x^2-y^2}$ peaks are larger for the $M_1$ geometry ($\sim1.4$ eV) than for the rutile structure ($\sim0.76$ eV). This can be attributed to the reduced V-V distance along the $c$-axis due to the dimerization. The splitting of the bands can be roughly approximated by $2t_{x^2-y^2,x^2-y^2}$ for the different $r$ ratios. This behavior resembles the bonding and anti-bonding splitting of a dimer molecule, suggesting that the splitting of these states is determined mainly by the intra dimer hopping, especially in the $M_1$ case. For the $e_g^\pi$ states, this correspondence is less clear, showing the importance of the inter-dimer hoppings. It is worth to note that the LDA calculations can capture only the Peierls physics, which alone is not sufficient to explain the formation of the gap even if a reduced V-V distance is considered.
![ (Color online) LDA partial DOS of the $t_{2g}$ states of a V atom in the rutile (left) and $M_1$ (right) phases in the proximity of the Fermi energy with different $r$ ratios: $r=0.98$ (dotted, red), 1.00 (full, black) and 1.02 (dashed, green). Note the large splitting of the $d_{x^2-y^2}$ states and the barley changed DOS at the Fermi level. []{data-label="fig:lda_dos"}](fig6a.png "fig:"){width="0.49\linewidth"} ![ (Color online) LDA partial DOS of the $t_{2g}$ states of a V atom in the rutile (left) and $M_1$ (right) phases in the proximity of the Fermi energy with different $r$ ratios: $r=0.98$ (dotted, red), 1.00 (full, black) and 1.02 (dashed, green). Note the large splitting of the $d_{x^2-y^2}$ states and the barley changed DOS at the Fermi level. []{data-label="fig:lda_dos"}](fig6b.png "fig:"){width="0.49\linewidth"}
To trace the effect of the electron-electron interactions [*e.g.*]{}, appearance of the gap in the $M_1$ phase and the reduction of the band-width, we carried out LDA+DMFT calculations. Our theoretically calculated [*orbitally resolved* ]{} spectral functions can be directly compared to the measured angle integrated photoemission (PES) and x-ray absorption spectroscopy spectra (XAS). In the last few years a large number of experimental studies of these kind were carried out probing also the many body character of the occupied and the unoccupied states in VO$_2$ and serving as a stringent test of the theoretical approach. This validation of our LDA+DMFT results is crucial before proceeding to make reliable predictions for the strained materials for which these spectroscopic information is not yet available.
The upfolded density of states of the LDA+DMFT, which includes besides the V-$3d-t_{2g}$ states the O-$2p$ and the V-$3d-e_g$ states, are shown in Fig. \[fig:total\]. The dynamical correlation effects, not included in LDA, decrease the width of the quasiparticle $t_{2g}$ states for $\sim 0.6$ times, in agreement with our calculated quasiparticle renormalization amplitude $Z\approx 0.62$. In the rutile structure two Hubbard bands appear: a very weak lower Hubbard band around -1.0 eV (was previously reported in theoretical [@BPL+05] and experimental [@KHH+06] studies) and a stronger upper hubbard band at 2.0 eV which overlaps with the $e_g$ band (can be seen better in Fig. \[fig:resolved\] where the $e_g$ band is not shown). In the $M_1$ phase, the most prominent effect of the DMFT theory is the appearance of a gap of $\sim 0.7$ eV at the chemical potential. For both the rutile and $M_1$ phase the positions of the O-$2p$ and the V-$3d-e_g$ states show good agreement with the experiments. Our calculated $Z$ factors for the $M_1$ phase are $\sim 0.9$ for the $e_{g}^\pi$ state and $\sim 0.7$ for the $a_{1g}$ state confirming that the $M_1$ phase is “less correlated” than the rutile phase, as emphasized in Ref. .
Fig. \[Akw\] shows the momentum resolved spectra in the [*R*]{} and $M_1$-phase. Notice the downshift of the two bonding bands in the $M_1$ phase, primarily of $x^2-y^2$ character (notice that there are four V atoms per unit cell, and hence two bonding $x^2-y^2$ bands). The states above the Fermi level move slightly up and shrink due to many-body renormalization similarly to the $t_{2g}$ states in the [*R*]{} phase. The bands around $1\,$eV also acquire substantial lifetime in both phases. The oxygen bands below $-1.5\,$eV and the $eg$ bands above $2\,$eV are almost unchanged compared to LDA.
The results for the orbitally resolved $3d-t_{2g}$ spectral functions are shown in Fig. \[fig:resolved\]. In the rutile phase one can see that the three $t_{2g}$ states are approximately equally occupied predicting isotropic transport properties. For the $a_{1g}$ state the bonding anti-bonding structure can be also recognized similarly to the $M_1$ phase with a splitting of $\sim0.4$ eV which is by a factor of 2 smaller then in the LDA results. The calculated positions of the $t_{2g}$ states agree well with the XAS results. In the $M_1$ phase, the weight redistribution is very different: only a single state, namely the bonding $a_{1g}$ orbital is occupied. One can observe that the bonding $a_{1g}$ state is shifted for $\sim0.8$ eV lower, which is in good agreement with the previous theoretical results [@BPL+05]. The spectral function at the upper edge of the gap has predominantly the $e_g^\pi$ character, but there is also some weight of the anti-bonding $a_{1g}$ character, which is in agreement with experiment showing that the spectral density has not purely $e_{g}^\pi$ character above the gap [@KHH+06]. The first two peaks at $\sim0.6$ eV above the chemical potential are attributed mainly to the $e_g^\pi$ states, and the third one is due to the $a_{1g}$ state. This is consistent with the recent results of polarization dependent O K XAS experiments from Koethe [*at al.*]{} [@KHH+06] where the orbital character of the states can be deduced by changing the polarization of the x-ray from parallel (O K XAS $\parallel$) to the $c$-axis to perpendicular polarization (O K XAS $\perp$). The anti-bonding $a_{1g}$ state lies at 1.3 eV above the Fermi level. This peak did not appear in previous theoretical studies of Ref , but agrees well with current XAS results, which show that there is a prominent $a_{1g}$ $\sim1.0$ eV above the $e_g^\pi$ peak [@KHH+06; @AGF+91; @SST+90]. It is interesting to note that the position of the anti-bonding $a_{1g}$ state is roughly the same in both the DMFT results and the pure LDA results. The calculated separation between the bonding and anti-bonding peaks of the $a_{1g}$ state is $\sim 2.1$ eV, which agrees reasonably with the experimentally found value ($2.5-2.8$ eV)[@KHH+06]. Finally, the peak around $3\,$eV of the XAS spectra may be assigned to the contribution from the $e_g$ states, included in Fig. \[fig:total\], but excluded in Fig. \[fig:resolved\]. This is consistent with the experimental finding of Ref. [@KHH+06] where negligible change of the peak weight was observed across the MIT, but strong sensitivity to the polarization was noticed.
![ (Color online) Comparison of the orbitally resolved $t_{2g}$ DOS of LDA+DMFT calculation corresponding to a V atom in rutile (left) and $M_1$ (right) phases ($r=1.00$) with angle integrated photoemission (PES) and x-ray absorption spectroscopy spectra (XAS) measurement. The experimental results are reproduced from Ref. [@KHH+06]. The XAS result is shifted to obtain the best agreement with our theoretical results. []{data-label="fig:resolved"}](fig9a.png "fig:"){width="0.4\linewidth"} ![ (Color online) Comparison of the orbitally resolved $t_{2g}$ DOS of LDA+DMFT calculation corresponding to a V atom in rutile (left) and $M_1$ (right) phases ($r=1.00$) with angle integrated photoemission (PES) and x-ray absorption spectroscopy spectra (XAS) measurement. The experimental results are reproduced from Ref. [@KHH+06]. The XAS result is shifted to obtain the best agreement with our theoretical results. []{data-label="fig:resolved"}](fig9b.png "fig:"){width="0.4\linewidth"}
To elucidate the effect of the strain along the $c$-axis the spectral functions of the $t_{2g}$ orbitals for different $r$ ratios ($r=0.98$, 1.00 and 1.02) are shown in Fig. \[fig:dos\_strain\] for both phases. In the rutile phase the $e_{g}^\pi$ states are hardly affected by the strain. In contrast, the bonding anti-bonding splitting of the $a_{1g}$ states shows a strong sensitivity to strain. The width of the upper $a_{1g}$ peak at $\sim0.5$ eV decreases with increasing $r$, which is due to the weaker hybridization between the V atoms, as indicated by the decreasing hopping integrals. Surprisingly, the spectral weight at the chemical potential is practically unaffected by the changes of the lattice constant $c$. Similar calculation using cluster DMFT was carried out for the rutile phase, allowing for the formation of split bonding-anti-bonding pairs along the $c$-axis. We did not find an appreciable sign for the development of the bonding-antibonding splitting even for the $r=0.98$ case, confirming that the single site DMFT is quite accurate in the rutile phase.
Fig. \[fig:dos\_strain\] clearly shows that the width of the bonding $a_{1g}$ peak is increased with decreasing $r$ ratio, which can be attributed to the increase in inter dimer hoppings. For all ratios $r$, the gap in orbital space is indirect, i.e., the valence band is of $a_{1g}$ character and the conduction band of $e_{g}^\pi$ character. This is in agreement with the experimental findings of Ref. demonstrating that the Peierls physics is playing a secondary role in the gap opening in $M_1$ phase. Due to the decreasing length of the $c$-axis, the $e_{g}^\pi$ states are *shifted to* slightly *lower energy*, which together with the broadening of the bonding $a_{1g}$ peak results in the *contraction* of the gap, despite the increase in the bonding anti-bonding splitting of the $a_{1g}$ peaks. The decrease of the gap size due to decreasing $c$-axis length is more apparent in the inset of Fig. \[fig:optics\], which shows the total $t_{2g}$ DOS. This result is supported by the experimental result of Muraoka and Hiroi [@MH02] demonstrating that the decrease in lattice parameter $c$ leads to decrease in the metal-insulator transition temperature. This is a clear indication that smaller $c$-axis length leads to a weakened stability of the insulator in the $M_1$ phase, and consequently a smaller gap in $M_1$ phase. This behavior is not expected for a Peierls type gap, which increase as the lattice parameter decreases along the dimerized chains.
![ (Color online) Orbitally resolved V$-3d-t_{2g}$ DOS of a V atom in rutile (left) and $M_1$ (right) phases in case of $r=0.98$ (dotted, red line), 1.00 (full, black line) and 1.02 (dashed, green line). []{data-label="fig:dos_strain"}](fig10a.png "fig:"){width="0.4\linewidth"} ![ (Color online) Orbitally resolved V$-3d-t_{2g}$ DOS of a V atom in rutile (left) and $M_1$ (right) phases in case of $r=0.98$ (dotted, red line), 1.00 (full, black line) and 1.02 (dashed, green line). []{data-label="fig:dos_strain"}](fig10b.png "fig:"){width="0.4\linewidth"}
The evolution of the gap is reflected also in the gap of the optical conductivity [@millis04]. In Fig. \[fig:optics\] the real part of the average optical conductivity $\sigma_{av}=\frac{1}{3}(\sigma_{\parallel}+ 2\sigma_{\perp})$ is shown, where $\sigma_{\parallel(\perp)}$ is the optical conductivity in the case where the polarization of incident light is parallel (perpendicular) to the $c$-axis. The calculated optical gap and the intensity of the first peak corresponding to the $t_{2g}-t_{2g}$ excitations compare well with the experimental results for polycrystalline VO$_2$ films [@QBW+08]. The shoulder around $\sim
2.5$ eV in the experimental optical conductivity, which is primarily due to the inter-band transitions, is shifted slightly upwards ($\sim
0.5$ eV) in the theoretical result. This is an indication that the applied downfolding method describes well the low-energy properties, but not so well the higher energy interband excitations.
The fact that the gap in the density of states arises between orbitals of different symmetry, indicates that the anisotropy of the transport properties in this phase will be very sensitive to disorder and grain boundaries which can drastically alter the orientation of these orbitals changing the matrix elements for hopping and conductivity.
Fig. \[fig:optics\_anis\] shows the calculated and experimentally measured optical conductivity for differently polarized light. It can be seen that the trends of the dependence of the optical conductivity on the polarization are in a good agreement with experimental results, although the values are slightly different. When the polarization is perpendicular to the $c$-axis the optical response is practically unaffected by the strain. In case of parallel polarization the optical conductivity is strongly modified by changing the lattice parameter $\tilde{c}$, especially in the frequencies between $1.5-2.5$ eV. This region can be attributed to the $d_{x^2-y^2}-d_{x^2-y^2}$ excitations as can be concluded from the positions of the bonding and antibonding peaks in Fig. \[fig:dos\_strain\]. This results strongly indicate that the anisotropy of the transport properties is due to directed V-V bonds along the $c$-axis.
In Fig. \[fig:imsig\_optics\], the orientation average optical conductivity of the [*R*]{} phase and the imaginary part of the self-energy at the Matsubara frequencies close to zero are shown. While the area below the calculated and experimentally measured [@QBW+08] optical conductivity (plasma frequency) agree fairly well, the width of the two Drude peaks are different. In order to improve the agreement with experimental results [@QBW+08] an imaginary part of 0.55 eV (scattering rate) was added to the self-energy for the low frequency part of the optical conductivity, to simulate the experimentally measured broadening of the Drude peak. Inspecting the inset of Fig. \[fig:imsig\_optics\], one can see that even by employing larger $U$ values, the scattering rate ($\mathrm{Im}\Sigma(\omega \rightarrow 0 )$) is not large enough to reproduce the experimental results, and the calculated optical conductivity will not show [*bad metal*]{} behaviour at this temperature. From this result one can draw the conclusion that the experimentally measured large scattering rate is a consequence of an inhomogeneity of the system as indicated in recent experiments [@Qazilbash3; @CES+09].
Phase diagram and limits of downfolding
=======================================
In Fig. \[fig:phase\], a sketch of the phase diagram of the rutile structure based on calculated points within the cluster DMFT is displayed as a function of the temperature and the Coulomb repulsion parameter $U$. Three different phases are distinguished: the metal (M), the insulator (I) and the coexistent (C) regions. To identify them, the imaginary part of the local Matsubara Green’s function, $\mathrm{Im}\mathcal{G}(\omega_n)$, of the cluster was investigated. We considered that the solution is insulating when $\mathrm{Im}\mathcal{G}(\omega_n)$ converges to 0 at low imaginary frequencies for all orbitals, as shown in an example in the right inset of Fig. \[fig:phase\]. The metallic solutions are the ones where $\mathrm{Im}\mathcal{G}(\omega_n)$ tends to a finite value for at least one of the orbitals (see the left inset in Fig. \[fig:phase\]). In the metallic region only the metallic solution is stable. In the insulator region only the insulating solution is stable, and in the coexistence region, both are stable. In order to decide whether a solution is stable or not, the DMFT calculations were started from an ansatz of a specified type (metal or insulator) and if it remained of the same type after the self-consistent solution is reached, one can regard it as a stable mean-field solution. In Fig. \[fig:phase\] it can be observed that below $U\approx2.9$ eV only the metallic solutions are stable while above $\sim3.8$ eV only the insulator is stable. The nature of the insulating state in rutile phase is very different from the insulating state in $M_1$ phase. The lower Hubbard band in the rutile insulator is an almost equal mixture of all three $t_{2g}$ orbitals. However, the interaction strength needed to open the true Mott gap without the help of the Peierls mechanism, is considerably larger. In the coexistence region, one can expect a crossover between the two phases governed by the free energy of the system [@PHK08]. There is a strong experimental indication that the rutile phase resides in the vicinity (but on the metallic side) of this crossover. Pouget and Launois showed that the metallic feature of the rutile phase is very sensitive by substitutional alloying of VO$_2$ with Nb (V$_{1-x}$Nb$_{x}$O$_2$) which increases the $c/a$ ratio and results in the appearance of a gap keeping the rutile structure at $x=0.2$ [@JL76]. Recently, Holman [*et al.*]{} [@HMW+09] reported insulator to metal transition in V$_{1-x}$Mo$_x$O$_2$ system at $x\approx0.2$. All those experiments suggest that VO$_2$ is in the crossover region near the coexistence of two solutions in the cluster DMFT phase diagram. On the basis of our calculation we conjecture that the rutile phase might be able to support either metallic or insulating solution (with a very small gap), and hence either of the two phases can be stabilized depending on small external stimuli. On the other hand, the $M_1$ phase at the same interaction strength supports only insulating solution, and it is unlikely that small external perturbation can turn it to metallic state. One possibility to improve the agreement between the experimental and theoretical results is that one should use different parameters in the downfolded model. The second possibility is that the material is strongly inhomogeneous which was not taken into account so far in any theoretical calculations. Finally, it is most likely that calculation without invoking the downfolding approximation will result in a more accurate description of this material. This would place the VO$_2$ closer to the Mott charge transfer insulator boundary in the Zaanen-Sawatzky-Allen phase diagram. Work in this direction is in progress.
Conclusions
===========
Our exploratory theoretical research set up the machinery for describing the subtle interplay of Coulomb correlations, orbital degeneracy and strain in determining the mechanism of the MIT in VO$_2$. Our theory, coupled with existing strain experiments, clearly shows that the Peierls distortion is only one element affecting the MIT and the switching mechanism of this material. The LDA+DMFT calculations in the unstrained material are in good agreement with experiments. We performed the first LDA+DMFT studies of the electronic structure of VO$_2$ under strain. Besides the Peierls increase in $a_{1g}$ bonding-antibonding splitting, the lowering in energy of the $e^\pi_g$ orbital, and the rapid change in bandwidth of the $a_{1g}$ orbital due to the varying overlaps, play an equally important role in controlling the position of the MIT. These theoretical insights can be used for understanding and improving material properties by means of chemical substitutions. For a more accurate description it is mandatory to take into account the oxygen degrees of freedom, and calculations of the total energy in the regions suggested by this exploratory work.
Acknowledgments
===============
This research has been supported by Grants No. DARPA W911NF-08-1-0203, NSF-DMR 0806937, and OTKA F68726. We are grateful to Jan Tomczak for stimulating discussions.
[99]{} , Phys. Rev. Lett. [**[3]{}**]{}, [34]{} ([1959]{}). , [Acta Chem. Scand.]{} [**[24]{}**]{}, [420]{} ([1970]{}). G. Andresson, Acta Chem. Scand. **10**, 623-628 (1956). P. B. Allen, R. M. Wentzcovitch, W. W. Schulz, and P. C. Canfield, Phys. Rev. B **48**, 4359 (1993). , [Phys. Rev. Lett.]{} [**97**]{}, [116402]{} ([2006]{}). S. Lysenko, V. Vikhnin, F. Fernandez, A. Rua, and H. Liu, Phys. Rev. B **75**, 75109 (2007). A. Cavalleri, Cs. Tóth, C. W. Siders, J. A. Squier, F. Ráksi, P. Forget and J. C. Kieffer, Phys. Rev. Lett. **87**, 237401 (2001). P. Baum, D.-S. Yang, and A. H. Zewail, Science **318**, 788 (2007) , [Z. Hiroi]{}, [App. Phys. Lett.]{} [**80**]{}, [583]{} ([2002]{}). K. Maekawa, M. Takizawa, H. Wadati, T. Yoshida, A. Fujimori, H. Kumigashira, M. Oshima, Y. Muraoka, Y. Nagao, and Z. Hiroi, Phys. Rev. B **76**, 115121 (2007) M. M. Qazilbash, M. Brehm, G. O. Andreev, A. Frenzel, P.-C. Ho, Byung-Gyu Chae, Bong-Jun Kim, Sun Jin Yun, Hyun-Tak Kim, A. V. Balatsky, O. G. Shpyrko, M. B. Maple, F. Keilmann, and D. N. Basov, Phys. Rev. B **79**, 075107 (2009). M. M. Qazilbash, A. A. Schafgans, K. S. Burch, S. J. Yun, B. G. Chae, B. J. Kim, H. T. Kim, and D. N. Basov, Phys. Rev. B [**77**]{}, 115121 (2008). M. M. Qazilbash, M. Brehm, Byung-Gyu Chae, P.-C. Ho, G. O. Andreev, Bong-Jun Kim, Sun Jin Yun, A. V. Balatsky, M. B. Maple, F. Keilmann, Hyun-Tak Kim, and D. N. Basov, Science **318**, 1750 (2007) M. M. Qazilbash, K. S. Burch, D. Whisler, D. Shrekenhamer, B. G. Chae, H. T. Kim, and D. N. Basov, [Phys. Rev. B]{} [**[74]{}**]{}, [205118]{} ([2006]{}). R. Eguchi, M. Taguchi, M. Matsunami, K. Horiba, K. Yamamoto, Y. Ishida, A. Chainani, Y. Takata, M. Yabashi, D. Miwa, Y. Nishino, K. Tamasaku, T. Ishikawa, Y. Senba, H. Ohashi, Y. Muraoka, Z. Hiroi, and S. Shin, Phys. Rev. B **78**, 075115 (2008). D. Ruzmetov, S. D. Senanayake, V. Narayanamurti, and S. Ramanathan, Phys. Rev. B **77**, 195442 (2008). L. Braicovich, G. Ghiringhelli, L. H. Tjeng, V. Bisogni, C. Dallera, A. Piazzalunga, W. Reichelt, and N. B. Brookes, Phys. Rev. B **76**, 125105 (2007). D. Ruzmetov S. D. Senanayake, and S. Ramanathan, Phys. Rev. B **75**, 195102 (2007) , [A. Poteryaev]{}, [A. I. Lichtenstein]{}, [A. Georges]{}, [Phys. Rev. Lett.]{} **94**, 026404 (2005) J. M. Tomczak, F. Aryasetiawan, and S. Biermann, Phys. Rev. B **78**, 115103 (2008). J. M. Tomczak, and S. Biermann, J. Phys.: Condens. Matter **19**, 365206 (2007) R. Sakuma, T. Miyake, and F. Aryasetiawan, Phys. Rev. B **78**, 075106 (2008). M. Gatti, F. Bruneval, V. Olevano, and L. Reining, Phys. Rev. Lett. **99**, 266402 (2007). , [Ann. Phys. (Leipzig)]{} [**[11]{}**]{}, [9]{} ([2002]{}). , [Phys. Rev. Lett.]{} [**[72]{}**]{}, [3389]{} ([1994]{}). M. S. Laad, L. Craco, and E. Müller-Hartmann, Phys. Rev. B **73**, 195120 (2006). A. Liebsch, H. Ishida, and G. Bihlmayer, Phys. Rev. B [**71**]{}, 085109 (2005). J. B. Goodenough, Phys. Rev. [**117**]{}, 1442 - 1451 (1960) , [N. F Mott]{} , [Phys. Rev. B]{} [**[11]{}**]{}, [4383]{} ([1976]{}). , [G. Kotliar]{}, [W. Krauth]{}, [M. J. Rozenberg]{}, [Rev. Mod. Phys.]{} [**[68]{}**]{}, [13]{} ([1996]{}). , [S. Y. Savrasov]{}, [K. Haule]{}, [V. S. Oudovenko]{}, [O. Parcollet]{}, [C. A. Marianetti]{}, [Rev. Mod. Phys.]{} [**[78]{}**]{}, [865]{} ([2006]{}) O. K. Andersen, Phys. Rev. B [**12**]{}, 3060 (1975). , [Phys. Rev. B]{} [**75**]{}, [155113]{} ([2007]{}). , [J. Phys. Soc. Jpn]{} [**71**]{}, [822]{} ([2002]{}). K. Haule, C.-H. Yee, K. Kim, arXiv:0907.0195. , [M. Marezio]{}, [J. P. Remeika]{}, [P. D. Dernier ]{}, [Phys. Rev. B]{} [**[10]{}**]{}, [490]{} ([1974]{}) M. S. Laad, L. Craco, and E. Müller-Hartmann, Phys. Rev. B [**73**]{}, 195120 (2006) and references therein. M. Abbate, F. M. F. de Groot, J. C. Fuggle, Y. J. Ma, C. T. Chen, F. Sette, A. Fujimori, Y. Ueda, and K. Kosuge, Phys. Rev. B [**43**]{}, 7263 (1991). S. Shin, S. Suga, M. Taniguchi, M. Fujisawa, H. Kanzaki, A. Fujimori, H. Daimon, Y. Ueda, K. Kosuge, and S. Kachi, Phys. Rev. B [**41**]{}, 4993 (1990). A. J. Millis, in Strong Interactions in Low Dimensions, edited by D. Baeriswyl and L. DeGiorgi Springer Verlag, Berlin, 2004 H. W. Verleur, A. S. Barker, and C. N. Berglund, Phys. Rev. [**172**]{}, 788 (1968). J. M Tomczak and S. Biermann, Eur. Phys. Lett., [**86**]{}, 37004 (2009). J. Cao, E. Ertekin, V. Srinivasan, W. Fan, S. Huang, H. Zheng, J. W. L. Yim, D. R. Khanal, D. F. Ogletree, J. C. Grossman, and J. Wu, arXiv:0907.4769. H. Park, K. Haule, and G. Kotliar Phys. Rev. Lett. [**101**]{}, 186403 (2008). J.P. Pouget and H. Launois, Journal de Physique [**37**]{}, C4-49 (1976). K. L. Holman, T. M. McQueen, A. J. Williams, T. Klimczuk, P. W. Stephens, H. W. Zandbergen, Q. Xu, F. Ronning, and R. J. Cava, Phys. Rev. B [**79**]{}, 245114 (2009)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article, we define general normal forms for any logic that has propositional part and whose non-propositional connectives distribute over the finite disjunctions. We do not require the non-propositional connectives to be closed on the set of formulas, so our normal forms cover logics with partial connectives too. We also show that most of the known normal forms in the literature are in fact particular cases of our general forms. These general normal forms are natural improvement of the distributive normal forms of J. Hintikka [@hintikka] and their modal analogues, e.g. [@anderson] and [@fine].'
author:
- 'Mohamed Khaled[^1]'
title: General normal forms for any additive logic
---
Introduction
============
It was shown that every propositional formula can be rewritten equivalently in the disjunctive normal form. Such form is a disjunction of one or more conjunctive clauses, each of these clauses consists of statement letters or negations of statement letters. Similar forms were introduced for more complex logics, e.g. distributive normal forms for first order logic [@hintikka]. Normal forms were also extended to some modal logics and then they were used to give elegant and constructive proofs for many standard results, see [@anderson] and [@fine]. Recently, the method of normal forms was used in new directions to solve problems where all the standard techniques fail, e.g. [@van], [@myrsl], [@myphd] and [@myigpl].
Here, we generalize the normal forms in a novel way. We prove a general theorem that can be applied for arbitrary logics that extend the propositional one. But, we also require the non-propositional connectives to be distributive over the finite disjunctions. For each $k\in\omega$, we define a set of formulas that we call . Each one of these forms is a conjunction of (constituents consist of) any one of the following or their negations: propositional letters, and normal forms of the first smaller degree joined together by a non-propositional connective. That is, we build a hierarchy of normal forms by considering their connections, via the non-propositional connectives, with the normal forms of the first smaller degree. Then we prove that any formula of the logic in question can be rewritten in an equivalent way as a disjunction of finitely many formulas that live in the same level of the hierarchy of the normal forms.
Having these degrees, one can view each normal form $\varphi$ of degree $k$ as a hierarchy in itself. This hierarchy consists of $k$-levels, its first level consists of a unique element that represents $\varphi$. Then, to go up one level, one needs to consider brand new elements to represent the normal forms that appeared together joined by a non-propositional connective and without negation symbol in the conjuncts of $\psi$, where $\psi$ is a normal form represented by an element in the current level of the hierarchy. Thus, this hierarchy will end with a level that contains elements represent forms of degree $0$ only. Such hierarchy (or its modified versions) was shown, in most of the cases, to be a fair witness for the validity of the normal form $\varphi$.
Thus, since such hierarchies are finite, one may obtain some decidability results through the finite model property [@fine; @myigpl]. Also these hierarchies can provide easy and transparent proofs for some completeness results [@fine]. Moreover, these finite hierarchies associated with the normal forms can be extended by adding more levels. With a careful choice of such extensions one might be able to investigate the so-called Gödel’s incompleteness properties, e.g. [@myigpl]. These properties talk about the possibilities of refining the definable concepts by adding more conjuncts to the formulas defining them. These properties were defined by I. Németi in 1985 and named by him in appreciation of the pioneering work of K. Gödel. For more details concerning these incompleteness properties, see [@nem86; @zalan; @myphd; @myrsl; @myigpl].
Normal forms are essentially used in establishing results in both algebra and logic, e.g. [@sagi99], [@sagi01] and [@sagi02]. In the application section, we apply our general theorem to produce normal forms for Boolean algebras with operators. These ‘algebraic’ normal forms were used recently to give an answer for a problem that was raised by I. Németi in 1985. Indeed, the non-atomicity of the free algebras of several classes of algebras of logics was shown (c.f. [@myphd], [@myrsl] and [@myigpl]). Moreover, normal forms were also used to give a complete description for some of these free algebras from the point of view of atoms [@myphd].
One more feature of our result is, according to our definition of arbitrary additive logics, that the non-propositional connectives can be partial operators on the set of formulas, i.e., they do not need to be closed. That allows us to apply our normal forms to several important logics, e.g. guarded fragment of first order logic [@andvannem], the loosely guarded fragment [@vB97] and the packed fragment [@mar01]. Those are versions of first order logic in which the quantifiers are bounded and can not be applied to any formula. These logics were investigated by many logicians and it was shown that they have a number of desirable properties, e.g. decidability and finite model property. They have applications in linguistics (dynamic semantics of natural language) and computer science.
Additive logics
===============
In the present paper, an arbitrary logic is a tuple $\mathcal{L}=\langle Fm,\Vdash\rangle$, where $Fm$ is the set of formulas of $\mathcal{L}$ and $\Vdash$ is a unary relation on the set of formulas. Instead of writing $\varphi\in\Vdash$, we rather write $\Vdash\varphi$, for any formula $\varphi\in Fm$. This unary relation could be the set of provable formulas, the set of valid formulas, or may be something else. Let $P$ be a non-empty set of propositions and let $Cn$ be a set of connectives each of which is associated with a non-zero finite rank. We say that $\mathcal{L}$ is generated by $P$ and $Cn$ if $Fm$ is built up recursively from the set $P$ using the connectives in $Cn$, i.e.,
1. A string $\varphi$ of propositions and connectives is a formula of $\mathcal{L}$ only if $\varphi\in P$ or $\varphi$ is of the form $\varphi=\Box(\psi_0,\ldots,\psi_{k-1})$ for some $\Box\in Cn$ whose rank is $k$ and some $\psi_0,\ldots,\psi_{k-1}\in Fm$.
Suppose that $\mathcal{L}$ is generated by $P$ and $Cn$. Let $\Box\in Cn$ be a connective, say of rank $k$, and let $\psi_0,\ldots,\psi_{k-1}\in Fm$. Note that $\Box(\psi_0,\ldots,\psi_{k-1})$ is not necessarily a formula of $\mathcal{L}$. Define $$D(\Box){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{(\psi_0,\ldots,\psi_{k-1})\in{^kFm}:\Box(\psi_0,\ldots,\psi_{k-1})\text{ is a formula in }Fm\}.$$ We say that $\mathcal{L}$ has a propositional part if, in addition to above, it satisfies the following:
1. The usual propositional connectives $\lor$, $\land$ and $\lnot$ are members of $Cn$. Moreover, for any formulas $\varphi,\psi\in Fm$,
- $\lnot\varphi{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\lnot(\varphi)$ is a formula in $Fm$.
- $\varphi\lor\psi{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\lor(\varphi,\psi)$ is a formula in $Fm$.
- $\varphi\land\psi{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\land(\varphi,\psi)$ is a formula in $Fm$.
In other words, $D(\lnot)=Fm$, $D(\lor)={^2Fm}$ and $D(\land)={^2Fm}$. As usual, we define the derived connectives $\varphi\rightarrow\psi{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\lnot\varphi\lor\psi$ and $\varphi\leftrightarrow\psi{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}(\varphi\rightarrow\psi)\land(\psi\rightarrow\varphi)$.
2. For each $\Box\in Cn$ whose rank is $h$ and $\varphi_0,\psi_0,\ldots,\varphi_{h-1},\psi_{h-1}\in Fm$, $$\begin{aligned}
&&\text{if }(\varphi_0,\ldots,\varphi_{h-1})\in D(\Box)\ \text{ and }\ (\psi_0,\ldots,\psi_{h-1})\in D(\Box)\text{ then }\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\
&&\ \ \ \ \ \ \ \ \ \ \Big[\bigwedge\{\Vdash \varphi_i\longleftrightarrow\psi_i: i\in h\} \implies \Vdash \Box(\varphi_0,\ldots,\varphi_{h-1})\longleftrightarrow \Box(\psi_0,\ldots,\psi_{h-1})\Big].\end{aligned}$$
3. For each non-propositional connective $\Box\in Cn'{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}Cn\setminus\{\lnot,\lor,\land\}$, say of rank $h$, and formulas $\varphi_0,\ldots,\varphi_{h-1}\in Fm$ with $(\varphi_0,\ldots,\varphi_{h-1})\in D(\Box)$, $$\bigvee\{\Vdash \lnot\varphi_i: i\in h\} \implies\Vdash\lnot\Box(\varphi_0,\ldots,\varphi_{h-1}).\footnote{This is a normality condition, ituitively it says that applying non-propostional connectives to a contradiction will give a contradiciton again.}$$
4. The propositional tautologies are members of the unary relation $\Vdash$.
The above conditions are indeed relevant to what is meant in the literature for a logic to have a propositional part. Now, we define what we mean by an additive logic.
\[additive\] An additive logic $\mathcal{L}=\langle Fm, \Vdash\rangle$ is an arbitrary logic that has propositional part and satisfies the following condition:
1. For any non-propositional connective $\Box\in Cn'$, say of rank $k$, any $i\in k$ and any formulas $\varphi_0,\ldots,\varphi_{i-1},\varphi,\psi,\varphi_{i+1},\ldots,\varphi_{k-1}$, if $(\varphi_0,\ldots,\varphi_{i-1},\varphi\lor\psi,\varphi_{i+1},\ldots,\varphi_{k-1})$, $(\varphi_0,\ldots,\varphi_{i-1},\varphi,\varphi_{i+1},\ldots,\varphi_{k-1})$ and $(\varphi_0,\ldots,\varphi_{i-1},\psi,\varphi_{i+1},\ldots,\varphi_{k-1})$ are all in the domain $D(\Box)$ then $$\begin{aligned}
\Vdash\Box(\varphi_0,\ldots,\varphi_i,\varphi\lor\psi,\varphi_{i+1},\ldots,\varphi_{k-1})\leftrightarrow && \Box(\varphi_0,\ldots,\varphi_{i-1},\varphi,\varphi_{i+1},\ldots, \varphi_{k-1})\lor\\
&&\Box(\varphi_0,\ldots,\varphi_{i-1},\psi,\varphi_{i+1},\ldots,\varphi_{k-1})
\end{aligned}$$
Most of the important logics, together with their set of provable formulas or their set of valid formulas, are in fact additive, e.g. propositional logic, first order logic (with or without equality), guarded fragment of first order logic [@andvannem] and its liberal versions [@vB97; @mar01], non-permutable first order logic [@richard; @nem86], first order logic with general assignment models [@nem86], modal logics, etc. However, there are few examples of logics that are not additive: logics that do not have propositional part, e.g. intuitionistic logic, and logics that have propositional part but they are not additive, e.g. instantial neighbourhood logic [@van].
\[system\] Let $\mathcal{L}=\langle Fm, \Vdash\rangle$ be an additive logic. We say that $\mathcal{L}$ has a domain-representation system if there is a quadrable $(V,\iota,\jmath_1,\jmath_2)$, where $V$ is an arbitrary non-empty set, and $\iota:Fm\rightarrow\mathcal{P}(V)$ and $\jmath_1,\jmath_2:Cn'\rightarrow\mathcal{P}(V)$ are arbitrary maps such that the following conditions are true for any connective $\Box \in Cn'$ of rank $h\in\omega$ and any formulas $\varphi,\psi,\varphi_0,\ldots,\varphi_{h-1}\in Fm$:
(a) $\iota(\lnot\varphi)=\iota(\varphi)$, $\iota(\varphi\land\psi)=\iota(\varphi)\cup\iota(\psi)$ and $\iota(\varphi\lor\psi)=\iota(\varphi)\cup\iota(\psi)$.
(b) $\iota(\Box(\varphi_0,\ldots, \varphi_{h-1}))=\jmath_2(\Box)\setminus\jmath_1(\Box)$.
(c) \[c\] $(\varphi_0,\ldots,\varphi_{h-1})\in D(\Box)\iff \bigwedge\{\iota(\varphi_k)\subseteq\jmath_2(\Box):k\in h\}$.
Suppose that every non-propositional connective is unary. The domain-representation system is acting as a “chart of domain areas” that tells us which formula is in the domain of which connective of $Cn'$. The set $V$ is a broadsheet paper and $\jmath_2$ draws the borders of each non-propositional connective on that paper. The formulas are also represented via $\iota$ as bounded figures on the chart. According to condition (c) above, the bounded figure represents a formula $\varphi$ has to be inside the borders of a non-propositional connective $\Box$ if and only if $\varphi\in D(\Box)$ (See Figure \[fig\]). Condition (b) indicates that $\jmath_1$ is there only to represent any restrictions appear on the formula $\Box(\varphi)$ (after applying the connective $\Box$).
(-3,-2) – (-3,2) – (3,2) – (3,-2) – (-3,-2); at (2.5,-1.5) [$V$]{}; (0,0) ellipse (2.5cm and 1.5cm); at (1.6,-0.7) [$\jmath_2(\Box)$]{}; (-1.3,0) circle (0.5cm); at (-1.3,0) [$\iota(\varphi)$]{}; at (0,0) [$\ldots$]{};
We will construct normal forms for any additive logic associated with a fixed domain-representation system. We note that there might be more than one domain-representation system for the same logic. Definition \[system\] is liberal enough to guarantee the existence of domain-representation systems for all the additive logics mentioned above. In fact, we do not know any example of an additive logic that does not have such system[^2]. For example, if all the non-propositional connectives in an additive logic $\mathcal{L}$ are full operators, then we can define a domain-representation system as follows. Let $V\not=\emptyset$ be an arbitrary set. For each formula $\varphi$ of $\mathcal{L}$, define $\iota(\varphi)=V$. For each non-propositional connective $\Box$, define $\jmath_1(\Box)=\emptyset$ and $\jmath_2(\Box)=V$. It is easy to see that $(V,\iota,\jmath_1,\jmath_2)$ satisfies all the above conditions.
Main result
===========
Fix an additive logic $\mathcal{L}=\langle Fm,\Vdash\rangle$ and a domain-representation system $(V,\iota,\jmath_1,\jmath_2)$. We note that the normal forms we construct here depend on the choice of this system. Some domain-representation systems may be suitable for a certain task, some may not. Let $\Sigma$ be a non-empty finite set of formulas and fix any enumeration of $\Sigma=\{\varphi_0,\ldots,\varphi_{k-1}\}$, where $k$ is the cardinality of the set $\Sigma$. Define $\bigwedge\Sigma$ and $\bigvee\Sigma$ as follows: $$\bigwedge\Sigma{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\varphi_0\land\cdots\land\varphi_{k-1} \ \text{ and } \ \bigvee\Sigma{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\varphi_0\lor\cdots\lor\varphi_{k-1}.$$ This might seem ambiguous because the definitions of $\bigwedge\Sigma$ and $\bigvee\Sigma$ depend on the chosen enumeration for $\Sigma$. But, by condition (P4), one can see that considering any other enumeration will yield to an equivalent formula, up to the relation $\Vdash$. A function $\alpha\in{^{\Sigma}\{-1,1\}}$ allows us to choose, for each formula in $\Sigma$, either the formula itself or its negation, i.e. we define $\varphi^{\alpha}{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\varphi$ if $\alpha(\varphi)=1$ and $\varphi^{\alpha}{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\lnot\varphi$ otherwise. Then, we can define the formula $$\Sigma^{\alpha}{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\bigwedge\{\varphi^{\alpha}:\varphi\in\Sigma\}.$$
We need more conventions. Recall that $Cn'=Cn\setminus\{\lnot,\lor,\land\}$. Let $X\subseteq P$ be a finite set and let $A\subseteq V$. Define $\widetilde{X}{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{p\in X:\iota(p)\subseteq A\}$. We say that $A$ is large enough for $X$ if $\widetilde{X}\not=\emptyset$. Let $\Box\in Cn'$, we say that $\Box$ is compatible with $(X,A)$ if $\jmath_2(\Box)\setminus\jmath_1(\Box)\subseteq A$ and $\jmath_2(\Box)$ is large enough for $X$. Let $Y\subseteq Cn'$ be finite. Now, we are ready to define our normal forms. For each $k\in\omega$, we will build a set $N_k(X,Y;A)$ of normal forms of degree $k$. Each $\varphi\in N_k(X,Y;A)$ is built up from the propositions in $X$, the Boolean connectives and the connectives in $Y$. Throughout the construction, we restrict ourself to the condition that $\iota(\varphi)\subseteq A$, for each $\varphi\in N_k(X,Y;A)$. Then we will prove that any formula of $\mathcal{L}$ can be rewritten in an equivalent form, up to the relation $\Vdash$, as a disjunction of normal forms of the same degree (actually in the same $N_k(X,Y;A)$).
Let $X\subseteq P$ and $Y\subseteq Cn'$ be finite sets, and let $A\subseteq V$ be large enough for $X$. Let $k\in \omega$, we define the set $N_k(X,Y;A)$ (of normal forms of degree $k$) inductively as follows.
- Normal forms of degree $0$: $$N_0(X,Y;A){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{(\widetilde{X})^{\alpha}: \alpha\in{^{\widetilde{X}}\{-1,1\}}\}.$$
- Normal forms of degree $k+1$: $$N_{k+1}(X,Y;A){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{(\widetilde{X})^{\alpha}\land (\overline{N_k(X,Y;A)})^{\beta}:\alpha\in{^{\widetilde{X}}\{-1,1\}}, \beta\in{^{\overline{N_k(X,Y;A)}}\{-1,1\}}\},$$ where $\overline{N_k(X,Y;A)}$ is defined as follows. If there is at least one connective $\Box\in Y$ which is compatible with $(X,A)$, then $\overline{N_k(X,Y;A)}$ consists of all formulas of the form $\diamondsuit(\varphi_0,\ldots,\varphi_{h-1})$, where $\diamondsuit\in Y$ is a connective compatible with $(X,A)$, $h\in\omega$ is the rank of $\diamondsuit$ and $\varphi_0,\ldots,\varphi_{h-1}$ are normal forms in $N_k(X,Y;\jmath_2(\diamondsuit))$ [^3]. Otherwise, $\overline{N_k(X,Y;A)}{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}N_k(X,Y;A)$
- Finally, define $N(X,Y;A){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\bigcup\{N_k(X,Y;A):k\in\omega\}$.
By induction on $k\in\omega$, one can prove that $\overline{N_k(X,Y;A)}\not=\emptyset$, hence the above sets of normal forms are well defined.
A normal form $\varphi\in N_k(X,Y;A)$ is a formula that is built up from the set $X$ using the connectives in $Y$ together with the propositional connectives such that $k$ is the depth of the non-propositional connectives nesting in $\varphi$. Each normal form carries some information indicating its connections (via the non-propositional connectives) with the normal forms of the first smaller degree. We introduce the following notions that allow us to handle this information easily when it is needed.
Let $X\subseteq P$ and $Y\subseteq Cn'$ be finite sets, and let $A\subseteq V$ be large enough for $X$.
- Let $\varphi\in N_0(X,Y;A)$ be a normal form of degree $0$. Then there is $\alpha\in{^{\widetilde{X}}\{-1,1\}}$ such that $\varphi=(\widetilde{X})^{\alpha}$. Define $color(\varphi){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{p\in \widetilde{X}: \alpha(p)=1\}$.
- Let $k\in\omega$ and let $\varphi\in N_{k+1}(X,Y;A)$. Then there are choices $\alpha\in{^{\widetilde{X}}\{-1,1\}}$ and $\beta\in{^{\overline{N_k(X,Y;A)}}\{-1,1\}}$ such that $\varphi=(\widetilde{X})^{\alpha}\land (\overline{N_k(X,Y;A)})^{\beta}$. Define $$color(\varphi){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{p\in \widetilde{X}: \alpha(p)=1\}.$$ If $\Box\in Y$ is compatible with $(X,A)$ and its rank is $h\in\omega$, then define $$sub_{\Box}(\varphi){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{(\varphi_0,\ldots,\varphi_{h-1})\in {^hN_k(X,Y,\jmath_2(\Box))}:\beta(\Box(\varphi_0,\ldots,\varphi_{h-1}))=1\}.$$
The lemma below follows immediately from the construction of $N_k(X,Y;A)$ and (P4).
\[lem\] Let $X\subseteq P$ and $Y\subseteq Cn'$ be finite sets, and let $A\subseteq V$ be large enough for $X$. Let $k\in \omega$, then we have the following: $\Vdash\bigvee N_k(X,Y;A)$, and for every normal forms $\varphi,\psi\in N_k(X,Y;A)$ if $\varphi\not=\psi$ then $\Vdash\lnot(\varphi\land\psi)$.
Let $\psi\in Fm$ be an arbitrary formula of $\mathcal{L}$ and let $d$ be the maximum depth of the non-propositional connectives nesting in $\psi$. We define $P(\psi)$ and $Cn(\psi)$ to be the set of all propositions and the set of all connectives from $Cn'$, respectively, that appear in $\psi$. Let $k\in\omega$, let $X\subseteq P$ and $Y\subseteq Cn'$ be finite sets, and let $E\subseteq V$. We say that $(k,X,Y,E)$ is a generator suitable for $\psi$ if $k\geq d$, $X\supseteq P(\psi)$, $Y\supseteq Cn(\psi)$ and $E\supseteq\iota(\psi)$, $A$ is large enough for $X$ and, for each $\Box\in Cn(\psi)$, $\jmath_2(\Box)$ is large enough for $X$.
\[thm\] Let $\varphi$ be any formula of $\mathcal{L}$. Then, for any generator $(k, X, Y, E)$ suitable for $\varphi$, there is a finite $\Sigma\subseteq N_k(X,Y;E)$ such that $\Vdash\varphi\longleftrightarrow\bigvee\Sigma$[^4]. Moreover, there is a finite algorithm (that does not depend on the relation $\Vdash$) which generates such $\Sigma\subseteq N_k(X,Y;E)$ for any $\varphi$ and $(k,X,Y,E)$ as above.
Let $\varphi$ be any formula and let $(k,X,Y,E)$ be a generator suitable for $\varphi$. We use induction on the complexity of $\varphi$.
- Suppose that $\varphi$ is a proposition. Then we have $k\geq 0$ and $\varphi\in X\subseteq P$. Set $$\Sigma=\{\psi\in N_k(X,Y;E): \varphi\in color(\psi)\}.$$ Then, by (P4), it is easy to see that $\Vdash\varphi\longleftrightarrow\bigvee\Sigma$.
- Suppose that $\varphi=\varphi_1\land\varphi_2$ for some formulas $\varphi_1,\varphi_2\in Fm$. Let $j\in\{1,2\}$. Note that $(k,X,Y,E)$ is also a generator suitable for $\varphi_j$. Thus, by induction hypothesis, there is $\Sigma_j\subseteq N_k(X,Y;E)$ such that $\Vdash\varphi_j\longleftrightarrow\bigvee\Sigma_j$. By Lemma \[lem\], for each $\psi_1\in \Sigma_1$ and $\psi_2\in\Sigma_2$, if $\psi_1\not=\psi_2$ then $\Vdash \lnot(\psi_1\land\psi_2)$. Hence, by (P4), there is a finite set $\Sigma(=\Sigma_1\cap\Sigma_2)\subseteq N_k(X,Y;E)$ such that $$\Vdash\bigvee\Sigma\longleftrightarrow\bigvee\{\psi_1\land \psi_2: \psi_1\in \Sigma_1\text{ and }\psi_2\in\Sigma_2\}.$$ Now, (P4) implies that $$\Vdash\varphi\longleftrightarrow\bigvee\{\psi_1\land \psi_2: \psi_1\in \Sigma_1\text{ and }\psi_2\in\Sigma_2\}.$$ Therefore, $\Vdash \varphi\longleftrightarrow\bigvee\Sigma$ as desired. The induction step goes in a similar way for the disjunction $\varphi_1\lor\varphi_2$.
- Suppose that $\varphi=\lnot\psi$ for some formula $\psi\in Fm$. Again $(k,X,Y,E)$ is a generator suitable for $\psi$. Then, by the induction hypothesis, there is a finite set of normal forms $\Sigma'\subseteq N_k(X,Y;E)$ such that $\Vdash\psi\longleftrightarrow\bigvee\Sigma'$. Let $\Sigma=N_k(X,Y;E)\setminus\Sigma'\subseteq N_k(X,Y;E)$. Therefore, by Lemma \[lem\] and (P4), we have $\Vdash\varphi\longleftrightarrow\bigvee\Sigma$.
- Suppose that $\varphi=\diamondsuit(\varphi_0,\ldots,\varphi_{h-1})$ for some $\diamondsuit\in Cn'$ whose rank is $h\in\omega$ and some formulas $\varphi_0,\ldots,\varphi_{h-1}$ such that $(\varphi_0,\ldots,\varphi_{h-1})\in D(\diamondsuit)$. Let $j\in h$. By condition (c) in Definition \[system\], we have $\iota(\varphi_j)\subseteq\jmath_2(\diamondsuit)$. Hence, $(k-1,X,Y,\jmath_2(\diamondsuit))$ is a generator suitable for $\varphi_j$. Thus, by induction hypothesis, there is $\Sigma_j\subseteq N_{k-1}(X,Y;\jmath_2(\diamondsuit))$ such that $\Vdash \varphi_j\longleftrightarrow\bigvee\Sigma_j$. Hence, by (P2), $\Vdash \varphi\longleftrightarrow \diamondsuit(\bigvee\Sigma_0,\ldots,\bigvee\Sigma_{h-1})$. Then, the additivity of $\mathcal{L}$ (Add) together with (P3) imply $$\label{eq}\Vdash \varphi\longleftrightarrow \bigvee\{\diamondsuit(\varphi_0,\ldots,\varphi_{h-1}):(\forall j\in h) \varphi_j\in \Sigma_j\}.\tag{*}$$ Note that $\jmath_2(\diamondsuit)\setminus\jmath_1(\diamondsuit)=\iota(\varphi)\subseteq A$ and by assumptions $\jmath_2(\diamondsuit)$ is compatible with $(A,X)$. So, by the construction, each disjunct in is a member of $\overline{N_{k-1}(X,Y;A)}$. Therefore, (P4) implies that $\Vdash\varphi\longleftrightarrow\bigvee\Sigma$, where $$\Sigma=\{\psi\in N_k(X,Y;E):sub_{\diamondsuit}(\psi)\cap (\Sigma_0\times\cdots\times\Sigma_{h-1})\not=\emptyset\}.$$
The above steps also write a finite algorithm that generates the set $\Sigma\subseteq N_k(X,Y;E)$ for any $\varphi$ and $(k,X,Y;E)$ as required.
The above theorem can be explained intuitively as follows. Recall the chart of the domain areas and let $E$ be any part of the paper sheet. Then there are normal forms that are represented on the part $E$ of the chart, and these forms generate any other formula (whose elementary components are those components of the normal forms) represented inside $E$.
Applications
============
In this section, we apply Theorem \[thm\] to several logics and also to Boolean algebras with operators. We start with the first known example of disjunctive normal forms.
Propositional logic
-------------------
Suppose that $\mathcal{L}$ is the propositional logic with set of propositions $P$. Let $\emptyset\not=X\subseteq P$ be a finite set of propositions, define the following set: $$F(X){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{X^{\alpha}:\alpha\in{^X\{-1,1\}}\}.$$
There is a finite algorithm that generates, for every formula $\varphi$ of the propositional logic $\mathcal{L}$ and every finite $X\subseteq P$ that contains all the propositions appearing in $\varphi$, a finite $\Sigma\subseteq F(X)$ such that $\vdash\varphi\longleftrightarrow\bigvee\Sigma$ and $\models\varphi\longleftrightarrow\bigvee\Sigma$.
Let $V\not=\emptyset$ be any set and define $\iota:Fm\rightarrow\mathcal{P}(V)$ such that $\iota(\varphi)=V$ for every $\varphi\in Fm$. Note that $Cn'$ in this case is empty, so we can choose $\jmath_1=\jmath_2=\emptyset$. Thus, $(V,\iota,\jmath_1,\jmath_2)$ is a domain-representation system for $\mathcal{L}$. Moreover, $(0,X,\emptyset,V)$ is suitable generator for $\varphi$ and $F(X)=N_0(X,\emptyset;V)$. Therefore, we are done by Theorem \[thm\].
The distinction between the provability relation and the validity relation of propositional logic above is superfluous, indeed it is well known that propositional logic is (strongly) sound and complete. Now, we will give more interesting examples. We will construct normal forms for several version of first order logic that keeps the propositional axioms and the distributivity of the existential quantifiers over finite disjunctions[^5]. For convenience, we consider first order language $L$ that contains only variables $VAR$ (we also assume that $\mid VAR\mid\geq 2$) and relation symbols $REL$ (may or may not contain the equality symbol $=$). The connectives are the propositional ones $\lnot,\lor,\land$ together with the existential quantifiers $\exists v$. We will also suppose a relation $\Vdash$ on the set $Fm$ of formulas of the first order formulas (on the language $L$) that makes $\mathcal{L}=\langle Fm,\Vdash\rangle$ additive in the sense of Definition \[additive\][^6].
First order-like logics
-----------------------
For any $X\subseteq VAR$ and any $Y\subseteq REL$, let $P(X,Y)$ be the set of all atomic formulas that use variables from $X$ and relation symbols form $Y$. To define the normal forms, we need to restrict ourselves to some finite subset of the language. If $X\subseteq VAR$ and $Y\subseteq REL$ are both finite, then $P(X,Y)$ is also finite.
Let $X\subseteq VAR$ and $Y\subseteq REL$ be finite subsets such that $P(X,Y)\not=\emptyset$, and let $k\in\omega$. Define the following inductively:
1. $F_0(X,Y){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{P(X,Y)^{\alpha}:\alpha\in{^{P(X,Y)}\{-1,1\}}\}$.
2. $F_{k+1}(X,Y){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{P(X,Y)^{\alpha}\land(\overline{F_k(X,Y)})^{\beta}:\alpha\in{^{P(X,Y)}\{-1,1\}},\beta\in{^{\overline{F_k(X,Y)}}\{-1,1\}}\}$,
where if $X=\emptyset$ then $\overline{F_k(X,Y)}=F_k(X,Y)$, otherwise $$\overline{F_k(X,Y)}=\{\exists v \ \varphi: v\in X\text{ and }\varphi\in F_k(X,Y)\}.$$
\[FOL\] There is a finite algorithm that generates, for every first order formula $\varphi$, every $k$ bigger than or equal to the maximum depth of quantifiers nesting in $\varphi$, every finite $X\subseteq VAR$ that contains all the variables appearing in $\varphi$ and every finite $Y\subseteq REL$ that contains all relation symbols appearing in $\varphi$, a finite $\Sigma\subseteq F_k(X,Y)$ such that $\Vdash\varphi\longleftrightarrow\bigvee\Sigma$.
Let $V\not=\emptyset$ be any set. For each formula $\psi$, define $\iota(\psi)=V$. For each $v\in VAR$, define $\jmath_1(\exists v)=\emptyset$ and $\jmath_2(\exists v)=V$. Then, $(V,\iota,\jmath_1,\jmath_2)$ is a domain-representation system. Let $\mathbb{X}$ be the set of all atomic formulas built up from $X$ and $Y$, and let $\mathbb{Y}=\{\exists v:v\in X\}$. Thus, $(k,\mathbb{X},\mathbb{Y},V)$ is a generator suitable for $\varphi$ and $N_k(\mathbb{X},\mathbb{Y};V)=F_k(X,Y)$. Therefore, the statement follows immediately by Theorem \[thm\].
Suppose that $L$ is finite, that is suppose $VAR\cup REL$ is finite. Then the following is an immediate corollary of Theorem \[FOL\].
There is a finite algorithm that generates, for every first order formula $\varphi$ on the finite language $L$ and every $k$ bigger than or equal to the maximum depth of quantifiers nesting in $\varphi$, a finite $\Sigma\subseteq F_k(VAR,REL)$ such that $\Vdash\varphi\longleftrightarrow\bigvee\Sigma$.
Another application of Theorem \[thm\] can be the normal forms for modal logics appeared in [@anderson] and [@fine]. This can be shown easily in a similar way to the one used above. Now, we will give an example of a version of first order logic whose quantifiers are not full operators, i.e. the domains of the quantifiers are not the whole set of formulas. Namely, we will construct normal forms for the so-called guarded fragment (GF) of first order logic. The guarded fragment was introduced by H. Andréka, J. van Benthem and I. Németi in [@andvannem]. What follows can be mimicked to obtain normal forms for the more liberal guarded logics, e.g. the loosely guarded fragment and the packed fragment.
Guarded fragment of first order logic
-------------------------------------
Keep the first order language $L$ that consists of $VAR$ and $REL$. Let $\bar{v}\subseteq VAR$ be finite and suppose that $\bar{v}=\{v_0,\ldots,v_{k-1}\}$, then by $\exists\bar{v}$ we mean $\exists v_0\cdots\exists v_{k-1}$. For any first order formula $\varphi$, we write $free(\varphi)$ to mean the set of all free variables in $\varphi$. The set of *GF-formulas* on $L$ is defined recursively to be the smallest subset of the set of first order formulas on $L$ that satisfies the following.
(a) Any first order atomic formula is a GF-formula (GF-atom).
(b) If $\varphi$ and $\psi$ are GF-formulas, then $\varphi\land\psi$, $\varphi\lor\psi$ and $\lnot\varphi$ are GF-formulas.
(c) \[guards\] Let $\varphi$ be a GF-formula and $G$ be an atom such that $free(\varphi)\subseteq free(G)$. Then, for any finite tuple $\bar{v}\subseteq free(G)$, $\exists\bar{v}\ (G\land\varphi)$ is a GF-formula and such $G$ is called a *suitable GF-guard for $\varphi$*.
The semantics of the guarded logic GF is the standard semantics of the classical first order logic and the set of valid formulas $\models$ is defined in the usual way. We define the normal forms for GF as follows.
\[2.1.8\]Let $X'\subseteq X\subseteq VAR$ and $Y\subseteq REL$ be finite subsets. Let $At$ be the set consists of all atoms built up from $X'$ and $Y$. Assume that $At\not=\emptyset$. For any $k\in\omega$, we define the following recursively.
1. $F_0(X,Y;X'){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{At^{\alpha}:\alpha\in{^{At}\{-1,1\}}\}$.
2. $F_{k+1}(X,Y;X'){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{At^{\alpha}\wedge(\overline{F_k(X,Y;X')})^{\beta}:\alpha\in{^{At}\{-1,1\}}\text{ and }\beta\in{^{\overline{F_k(X,Y;X')}}\{-1,1\}}\}$,
where $\overline{F_k(X,Y;X')}$ consists of all GF-formulas of the form $\exists \bar{u}(\gamma\land\varphi)$ such that $\gamma$ is built up from $X$ and $Y$, $\varphi\in F_k(X,Y;free(\gamma))$, $\bar{u}\subseteq free(\gamma)\subseteq X$ and $free(\gamma)\setminus\bar{u}\subseteq X'$.
3. $F(X,Y;X'){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\bigcup\{F_{\lambda}(X,Y;X'):\lambda\in\omega\}$.
Note that every form in $F(X,Y;X')$ is a GF-formula built up from the variables in $X$ and the relation symbols in $Y$, and whose free variables belong to the set $X'$. Before we proceed, we need to define a domain-representation system. This time we do this in a different way. Let $V=VAR$ and define $\iota(\varphi)= free(\varphi)$ for every GF-formula $\varphi$. Now, the non-propositional connectives are of the form $\exists\bar{u}(\gamma\land - )$, for some finite set of variables $\bar{u}$ and an atom $\gamma$ such that $\bar{u}\subseteq free(\gamma)$. By definition, the domain of such bounded quantifier is given by $D(\exists\bar{u}(\gamma\land - ))=\{\varphi:\varphi\text{ is GF-formula and }free(\varphi)\subseteq free(\gamma)\}$. Define $\jmath_1(\exists\bar{u}(\gamma\land - ))=\bar{u}$ and $\jmath_2(\exists\bar{u}(\gamma\land - ))= free(\gamma)$. One can easily verify that $(V,\iota,\jmath_1,\jmath_2)$ is a domain-representation system for the guarded logic $GF$.
\[GF\]There is a finite algorithm that generates, for every guarded formula $\varphi$, every $k$ bigger than or equal to the maximum depth of quantifiers nesting in $\varphi$, every finite $X\subseteq VAR$ that contains all the variables appearing in $\varphi$, every $\emptyset\not=X'\subseteq X$ that contains $free(\varphi)$ and every finite $Y\subseteq REL$ that contains all relation symbols appearing in $\varphi$, a finite $\Sigma\subseteq F_k(X,Y;X')$ such that $\models\varphi\longleftrightarrow\bigvee\Sigma$.
In a similar way to the one used in Theorem \[FOL\]. Recall the domain-representation system defined above. Let $\mathbb{X}$ be the set of all atomic formulas that are constructed using the variables in $X$ and the relation symbols in $Y$. Let $\mathbb{Y}=\{\exists\bar{u}(\gamma\land - ):\bar{u}\subseteq free(\gamma)\text{ and }\gamma\in\mathbb{X}\}$. We need to show that $(k,\mathbb{X},\mathbb{Y}, X')$ is a generator suitable for $\varphi$. One can see that there is at least one relation symbol appeared in $\varphi$, this relation symbol together with the variables in $X'$ can form an atomic formula that guarantees the largeness of $X'$ for $\mathbb{X}$. Now, suppose that $\exists\bar{u}(\gamma\land - )\in Cn(\varphi)$, for some GF-guard $\gamma$ and $\bar{u}\subseteq free(\gamma)$. The guard $\gamma$ itself shows that $free(\gamma)$ is large enough for $\mathbb{X}$. It remains to note that $N_k(\mathbb{X},\mathbb{Y};X')=F_k(X,Y;X')$. Therefore, we are done by Theorem \[thm\].
Again, the following is a corollary of Theorem \[GF\] for the case when the language is finite, i.e. when $VAR\cup REL$ is finite.
There is a finite algorithm that generates, for every GF-formula $\varphi$ on the finite language $L$, every non-empty set of variables $X\supseteq free(\varphi)$ and every $k$ bigger than or equal to the maximum depth of quantifiers nesting in $\varphi$, a finite $\Sigma\subseteq F_k(VAR,REL;X)$ such that $\models\varphi\longleftrightarrow\bigvee\Sigma$.
A slightly different version of GF, namely the solo-GF, was introduced and discussed in [@myphd]. The solo GF-formulas are defined analogously to the GF-formulas except that the solo GF-guarded existential quantification $\exists\bar{u}(\gamma\land - )$ is now allowed only if the block of quantifiers $\bar{u}$ is of length $\leq 1$. In [@myphd], similar normal forms to the ones in Definition \[2.1.8\] were introduced for the solo guarded fragment on a finite language. These normal forms then were used to prove that the solo GF on finite languages enjoys weak Gödel incompleteness property but lacks Gödel’s incompleteness property (recall the fourth paragraph of the introduction). That gives more importance to the guarded logics indeed this is the first known example that differentiate the two properties for finite languages. In the following section we will show that the general normal forms for Boolean algebras with operators that was given in [@myphd appendix A] is also a special case of our general theorem.
Boolean algebras with operators
-------------------------------
Let $I$ and $J$ be any two index sets and let $t=\{+,\cdot,-,0,1,f_i,d_j:i\in I\text{ and }j\in J\}$ be a similarity type such that $\{+,\cdot,-,0,1\}$ is the type of Boolean algebras, $f_i$ is an operator symbol of positive rank $n_i{\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}rank(f_i)\geq 1$ (for any $i\in I$), and $d_j$ is a constant symbol (for any $j\in J$). Let $K$ be the class of all Boolean algebras with operators of type $t$, that is the class of all algebras of type $t$, $\a{A}=\langle A,+,\cdot,-,0,1,f_i,d_j:i\in I, j\in J\rangle$, that satisfy the following:
(1) The Boolean part $\langle A,+,\cdot,-,0,1\rangle$ is a Boolean algebra.
(2) The operators of positive ranks are additive, i.e. for any $i\in I$, any $k\in n_i$ and any $a_0,\ldots,a_{k-1},a,b,a_{k+1},\ldots,a_{n_i-1}\in A$, $$\begin{aligned}
f_i(a_0,\ldots,a_{k-1},a+b,a_{k+1},\ldots,a_{n_i-1})&=&f_i(a_0,\ldots,a_{k-1},a,a_{k+1},\ldots,a_{n_i-1})\\
&+&f_i(a_0,\ldots,a_{k-1},b,a_{k+1},\ldots,a_{n_i-1}).
\end{aligned}$$
(3) The operators of positive ranks are normal, i.e. for any $i\in I$, any $k\in n_i$ and any $a_0,\ldots,a_{k-1},a,a_{k+1},\ldots,a_{n_i-1}\in A$, $$a=0\implies f_i(a_0,\ldots,a_{k-1},a,a_{k+1},\ldots,a_{n_i-1})=0.$$
Fix a set of free variables $X$ and let $T$ be the set of all terms of type $t$ generated by the variables in $X$. The set $T$ can be viewed as the set of formulas (here we call them terms) built up recursively, using the Boolean operators and the operators $(f_i:i\in I)$, from the free variables in $X$ and the constant symbols $(d_j:j\in J)$. Let $\tau,\sigma\in T$ be arbitrary terms. We write $K\models\tau=\sigma$ if, for any algebra $\a{A}\in K$ and any evaluation $\nu:X\rightarrow A$, the interpretations $[\tau]^{\a{A}}_{\nu}\in A$ and $[\sigma]^{\a{A}}_{\nu}\in A$ are equal.
Let $\Vdash$ be the set of all terms that are equal to $1$, i.e., $\Vdash=\{\tau\in T:K\models\tau=1\}$. Thus, $\langle T,\models\rangle$ is an additive logic. We note that the Boolean operators $+,\cdot$ and $-$ correspond to the propositional connectives $\lor,\land$ and $\lnot$, respectively. We also note that the equality $=$ between terms corresponds to the derived connective $\longleftrightarrow$. Let $V$ be any non-empty set. For every term $\tau\in T$, define $\iota(\tau)=V$. For every $i\in I$, define $\jmath_1(f_i)=\emptyset$ and $\jmath_2(f_i)=V$. Clearly, $(V,\iota,\jmath_1,\jmath_2)$ satisfies the conditions of Definition \[system\].
Let $\emptyset\not=Y\subseteq T$ be finite. Let $\prod$ and $\sum$ be the grouped versions of $\cdot$ and $+$, respectively. That is, we fix any enumeration of $Y$ and then we define $\prod Y$ and $\sum Y$ inductively according to this enumeration. The algebras in the class $K$ are Boolean algebras with operators, hence $\cdot$ and $+$ are both commutative on the elements of these algebras. Thus, it doesn’t matter which enumeration we use to define $\prod$ and $\sum$, all are equal terms in the class $K$. Let $\alpha\in{{^Y}\{-1,1\}}$, define $Y^{\alpha}=\prod\{\tau^{\alpha}:\tau\in Y\}$, where for every $\tau\in Y$, $\tau^{\alpha}=\tau$ if $\alpha(\tau)=1$ and $\tau^{\alpha}=-\tau$ otherwise. We define our normal forms as follows.
Let $I'\subseteq I$, $J'\subseteq J$ and $X'\subseteq X$ be finite sets. Let $D=X'\cup\{d_j:j\in J'\}$ and suppose that $D\not=\emptyset$. For every $n\in\omega$, we define the following inductively.
1. The normal forms of degree $0$, $F_{0}(I',J',X'){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{D^{\beta}:\beta\in{^{D}\{-1,1\}}\}$.
2. The set of normal forms of degree $n+1$, $$F_{n+1}(I',J',X'){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\{D^{\beta}\cdot (\overline{F_n(I',J',X')})^{\alpha}:\beta\in{^{D}\{-1,1\}}\text{ and }\alpha\in{^{\overline{F_n(I',J',X')}}\{-1,1\}}\},$$
where $\overline{F_n(I',J',X')}=\{f_i(\tau_0,\ldots,\tau_{n_i-1}):i\in I'\text{ and }\tau_0,\ldots,\tau_{n_i-1}\in F_n(I',J',X')\}$, the one-step closure of $F_n(I',J',X')$ by the operations $\langle f_i: i\in I'\rangle$.
3. The set of all forms, $F(I',J',X'){\mathrel{\stackrel{\makebox[0pt]{\mbox{\normalfont\tiny def}}}{=}}}\bigcup\{F_k(I',J',X'):k\in\omega\}$.
The normal forms in $F_n(I',J',X')$ are built up from the free variables in $X'$ and the constant symbols $(d_j:j\in J')$ using the operators $(f_i:i\in I')$. Moreover, every normal form is constructed based on the normal forms of the first smaller degree. Again, one can show that the following theorems are special instances of Lemma \[lem\] and Theorem \[thm\].
Let $I'\subseteq I$, $J'\subseteq J$ and $X'\subseteq X$ be finite sets. Let $D=X'\cup\{d_j:j\in J'\}$ and suppose that $D\not=\emptyset$. Let $n\in\omega$, then the following are true.
(i) \[and1\] $K\models\sum F_n(I',J',X')=1$.
(ii) \[and2\] For every $\tau,\sigma\in F_n(I',J',X')$, if $\tau\not=\sigma$ then $K\models\tau\cdot \sigma=0$.
There is an effective method (finite algorithm) to find, for each $\tau\in T$, each $k\in\omega$ bigger than or equal the maximum depth of non-Boolean operators nesting in $\tau$, and each finite sets $I'\subseteq I$, $J'\subseteq J$ and $X'\subseteq X$ such that $\tau$ is built up using $X'$, $(d_j:j\in J')$, the Boolean operators and $(f_i:i\in I')$, a finite $S\subseteq F_k(I',J',X')$ such that $K\models\tau=\sum S$.
Suppose that $I$, $J$ and $X$ are all finite. Then there is a finite algorithm that generates, for each $\tau\in T$, an $k\in\omega$ and a finite $S\subseteq F_k(I,J,X)$ such that $K\models\tau=\sum S$.
[100]{} A. R. Anderson (1954). Improved decision procedures for [L]{}ewis’s calculus $S4$ and von [W]{}right’s calculus $M$. The Journal of Symbolic Logic, 19 (03), pp. 201-214. J. Hintikka (1965). Distributive Normal Forms in First-Order Logic. In: J.N. Crossley and M.A.E. Dummett, Editor(s), Studies in Logic and the Foundations of Mathematics, Elsevier, Volume 40, pp. 48-91. K. Fine (1975). Normal forms in modal logic. Notre Dame journal of formal logic, 16 (12), pp. 229-237. I. N[é]{}meti (1986). Free algebras and decidability in algebraic logic. Academic Doctoral Dissertation (in Hungarian), Hungarian Academy of Sciences, Budapest. <http://www.renyi.hu/~nemeti/NDis/NDis86.pdf>. English version is [@nem95]. R. Thompson (1988). Noncommutative cylindric algebras and relativization of cylindric algebras. Bulletin of the Section of Logic, 17 (2), pp. 75-81. I. N[é]{}meti (1995). Decidable versions of first order logic and cylindric-relativized set algebras. In: Logic Colloquium’92 (Proc. Veszprem, Hungary 1992), eds: L. Csirmaz and D. M. Gabbay and M. de Rijke, Studies in Logic, Language and Computation, CSLI Publications, pp. 177-241. J. van Benthem (1996). Exploring Logical Dynamics. Studies in Logic, Language and Information. CSLI Publications & Cambridge University Press. J. van Benthem (1997). Dynamic bits and pieces. Technical report, Institute for Logic, Language and Computation (ILLC), University of Amsterdam. The firsrt annual update of the book [@vB96]. H. Andr[é]{}ka, J. van Benthem and I. N[é]{}meti (1998). Modal languages and bounded fragments of predicate logic. Journal of Philosophical Logic, 27 (3), pp. 217-274. G. S[á]{}gi (1999). Defining relations for non-permutational finite transformations. Semigroup Forum, 58 (1), pp. 94-105. M. Marx (2001). Tolerance logic. Journal of Logic, Language and Information, 10 (3), pp. 353-374. G. S[á]{}gi (2001). Non-computability of the equational theory of polyadic algebras. Bulletin of the Section of Logic, 30 (3), pp. 155-164. G. S[á]{}gi (2002). A note on algebras of substitutions. Studia Logica, 72 (2), pp. 265-284. Z. Gyenis (2011). On atomicity of free algebras in certain cylindric-like varieties. Logic Journal of the IGPL, 19 (1), pp. 44-52. M. Khaled (2015). Weak Gödel’s incompleteness property for some decidable version of the calculus of relations. Preprint, arXiv:1511.01383 \[math.LO\]. N. Bezhanishvili, J. van Benthem, S. Enqvist and J. Yu (2016). Instantial Neighbourhood logic. The Review of Symbolic Logic, 10 (1), pp. 116-144. M. Khaled (2016). Gödel’s incompleteness properties and the guarded fragment: An algebraic approach. PhD thesis, Central European University, Budapest, Hungary. <https://mathematics.ceu.edu/sites/mathematics.ceu.hu/files/attachment/basicpage/27/phdthesis.pdf>. M. Khaled (2016). The free non-commutative cylindric algebras are not atomic. Logic Journal of the IGPL, jzw058.
[^1]: [email protected]
[^2]: We are aware that such an example can be constructed. What we meant here is that additive logics with no domain-representation systems have not been studied yet in mathematical logic.
[^3]: This guarantees that $(\varphi_0,\ldots,\varphi_{h-1})\in D(\diamondsuit)$ because of condition in Definition \[system\].
[^4]: Here, if $\Sigma=\emptyset$ then we define $\bigvee\Sigma$ to be a contradiction.
[^5]: These normal forms are in fact the distributive normal forms of J. Hintikka [@hintikka]
[^6]: For instance, $\Vdash$ can be the set of provable formulas of the standard or the non permutable first order logic, also $\Vdash$ can be the set of valid formulas of first order logic over standard models or general assignments models.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The bremsstrahlung energy loss distribution of electrons propagating in matter is highly non Gaussian. Because the Kalman filter relies solely on Gaussian probability density functions, it might not be an optimal reconstruction algorithm for electron tracks. A Gaussian-sum filter (GSF) algorithm for electron track reconstruction in the CMS tracker has therefore been developed. The basic idea is to model the bremsstrahlung energy loss distribution by a Gaussian mixture rather than a single Gaussian. It is shown that the GSF is able to improve the momentum resolution of electrons compared to the standard Kalman filter. The momentum resolution and the quality of the estimated error are studied with various types of mixture models of the energy loss distribution.'
author:
- 'W. Adam, R. Fr[ü]{}hwirth'
- 'A. Strandlie'
- 'T. Todorov'
title: 'Reconstruction of electrons with the Gaussian-sum filter in the CMS tracker at LHC'
---
Introduction
============
Modern track detectors based on semiconductor technologies contain larger amounts of material than gaseous detector types, partially due to the detector elements themselves and partially due to additional material required for on-sensor electronics, power, cooling, and mechanical support. A precise modelling of material effects in track reconstruction is therefore necessary to obtain the best estimates of the track parameters. Such material effects are particularly relevant for the reconstruction of electrons which, in addition to ionization energy loss and multiple Coulomb scattering, suffer from large energy losses due to bremsstrahlung.
A well-known model of the bremsstrahlung energy loss is due to Bethe and Heitler [@BetheHeitler34]. In this model, the probability density function (PDF), $f(z)$, of the energy loss of an electron is $$f(z) = \frac{\left[- \ln z \right]^{c-1}}{\Gamma (c)},$$ where $c = t / \ln 2$, $t$ is the thickness of material traversed by the electron (in units of radiation length), and $z$ is the fraction of energy remaining after the material layer is traversed. The probability of a given fractional energy loss is assumed to be independent of the energy of the incoming particle. This PDF is shown in Fig. \[fig:BHpdf\] for different thickness values.
\[fig:BHpdf\]
The baseline for track reconstruction in the CMS tracker is the Kalman filter [@Rudi87]. Throughout the filter tracks are described by a five-dimensional state vector, containing the information about the momentum, the direction and the position at some reference surface. The material effects are currently assumed to be concentrated in the active elements of the detector layers. In this context the optimal treatment of radiative energy loss is to correct the momentum with the mean value of energy loss and to increase the variance of the momentum by adding the variance of the energy loss distribution. This procedure should ensure unbiased estimates of the track parameters and of the associated uncertainties [@Stampfer94]. The Kalman filter is a linear least-squares estimator, and is proved to be optimal only when all probability densities encountered during the track reconstruction procedure are Gaussian. The implicit assumption of approximating the Bethe-Heitler distribution with a single Gaussian is quite crude. It is therefore plausible that a non-linear estimator which takes the actual shape of the distribution into account can do better.
A non-linear generalization of the Kalman filter (KF), the [*Gaussian-sum filter (GSF)*]{} [@Rudi97; @Rudi98], has therefore been implemented in the reconstruction software of the CMS tracker [@ORCA]. In the GSF the distributions of all state vectors are Gaussian mixtures, i.e. weighted sums of Gaussians instead of single Gaussians. The algorithm is therefore appropriate if the probability densities involved in track reconstruction can be adequately described by Gaussian mixtures. The basic idea of the present work is to approximate the Bethe-Heitler distribution as a Gaussian mixture rather than a single Gaussian, in which the different components of the mixture model different degrees of hardness of the bremsstrahlung in the layer under consideration. The resulting estimator resembles a set of Kalman filters running in parallel, where each Kalman filter corresponds to one of the components of the mixture describing the distribution of the state vector.
Approximating the fractional energy loss distribution
=====================================================
An important issue with the GSF reconstruction of electrons is to obtain a good Gaussian-mixture approximation of the Bethe-Heitler distribution. The parameters to be obtained are the weights, the mean values and the variances of each of the components in the approximating mixture. The parameters are determined by minimizing the following two distances: $$\begin{aligned}
D_{\mbox{{\scriptsize CDF}}} & = & \int_{- \infty}^{\infty} \left| F(z) - G(z)
\right| dz, \\
\label{equation:KLdist}
D_{\mbox{{\scriptsize KL}}} & = & \int_{- \infty}^{\infty} \ln \left[ f(z)/g(z)
\right] f(z) dz, \end{aligned}$$ where $f(z)$ and $F(z)$ are the PDF and cumulative distribution function (CDF) of the model distribution and $g(z)$ and $G(z)$ are the PDF and CDF of the Gaussian mixture, respectively. The distance $D_{\mbox{{\scriptsize KL}}}$ is the so-called Kullback-Leibler distance between the model distribution and the mixture. Hereafter, the mixtures obtained by minimizing $D_{\mbox{{\scriptsize CDF}}}$ are called CDF-mixtures, whereas the mixtures obtained by minimizing $D_{\mbox{{\scriptsize KL}}}$ are called KL-mixtures. The minimizations have been done independently on a set of discrete values of $t$, ranging from 0.02 to 0.20. Figures \[fig:KLdist\] and \[fig:CDFdist\] show the resulting distances as a function of thickness for a varying number of components in the approximating mixture.
\[fig:KLdist\]
\[fig:CDFdist\]
In order to obtain mixtures for arbitrary values of the thickness, fifth-degree polynomials have been fitted to the parameters as a function of $t$. Due to the fast access to the parameters from the polynomials, the calculation of the mixture is done on the fly during reconstruction, using the effective thickness of a detector layer from the knowledge of the incident angle of inclination.
Reducing the number of components
=================================
The approximation of energy loss by a Gaussian mixture amounts to a convolution of this mixture with the current state, which in general is also composed of several Gaussian components. The strict application of the GSF algorithm therefore quickly leads to a prohibitively large number of components due to the combinatorics involved each time a layer of material is traversed.
In a realistic implementation of the GSF the number of components must repeatedly be reduced to a predefined maximum. As little information as possible should be lost in this procedure. Two strategies have been tested:
1. Only the [*N*]{} components with the largest weights are kept;
2. Components are merged into clusters, according to a given metric.
The first option has the advantage of being computationally light, but it turns out to be inferior. Even the first two moments of the estimated parameters are not described correctly.
In the second approach, the component with the largest weight is merged with the one closest to it, and this procedure is repeated until the required number of components is reached. The results below have been obtained by using the Kullback-Leibler distance – defined in Equation (\[equation:KLdist\]) – as a measure of distance.
Results from simulated tracks in the CMS tracker
================================================
First, results from the reconstruction of data originating from a simplified simulation are shown. In this simulation multiple scattering and ionization energy loss are turned off, all the material is concentrated on the detector units, and the exact amount of material used in the simulation is known by the reconstruction program. Single electron tracks with $p_T = 10 $ GeV/$c$ have been simulated for absolute values of $\eta$ less than 1.0 . Reconstructed hits have been collected using the knowledge of the associated simulated hits, so no pattern recognition has been involved. The following results all refer to the quantity $q/p$ (charge over absolute value of the momentum) recorded at the point of closest approach to the vertex in the transverse plane – the transverse impact point (TIP) – after a fit going from the outside towards the inside of the tracker. Figure \[fig:qpSingleTrack\] shows an example of the estimated $q/p$ for one single track, both for the KF and for the GSF.
\[fig:qpSingleTrack\]
Figures \[fig:pullProbsMixture\] and \[fig:pullProbs\] show probability distributions for the estimated $q/p$ of the KF and the GSF with a varying maximum number of components kept during the reconstruction. Given the estimated PDF (a single Gaussian for the KF, a Gaussian mixture for the GSF), each entry in the histogram amounts to the integral from $-\infty$ to the true value of $q/p$. If the estimated PDF is a correct description of the real distribution of the parameter, the corresponding histogram should be flat.
{width="6.75cm"}
\[fig:pullProbsMixture\]
![Probability distribution for the estimated $q/p$ for the KF (solid) and the GSF with a maximum of six (dashed-dotted), twelve (dashed), $18$ (solid) and $36$ (dotted) components kept during reconstruction. The same six-component mixture as the one described in the caption of Fig. \[fig:pullProbsMixture\] has been used in reconstruction, but the simulation of the disturbance of the momentum in a detector unit has been done by sampling from the Bethe-Heitler distribution. The distributions for the GSF are seen to be less flat than those shown in Fig. \[fig:pullProbsMixture\].](pullprobs1.eps){width="6.75cm"}
\[fig:pullProbs\]
The deviation from flatness can be quantified by the $\chi^2$ of the difference between the probability distributions of $q/p$ and the flat distribution. This $\chi^2$ per bin is shown in Fig. \[fig:probChi2\] for a set of different mixtures as a function of the maximum number of components kept. The CDF-mixtures are superior to the KL-mixtures concerning the quality of the estimated $q/p$. The main trend seems to be related to the maximum number of components kept rather than the number of components in the mixture describing the energy loss, even though the mixtures with five and six components are best in the limit of keeping a large number of components.
\[fig:probChi2\]
Figure \[fig:qpPredTipBH\] shows the residuals of the estimated $q/p$ of the GSF and the KF with respect to the true value of the parameter. The estimated $q/p$ for the GSF is the mean value of the state vector mixture, and the mixture used for this specific plot is a CDF-mixture with six components. In order to quantify the difference between the GSF and the KF residuals, the full-width at half-maximum (FWHM) and the half-width of intervals covering 50% and 90% of the distribution have been considered. The covering intervals have been chosen to be symmetric about zero. The FWHM and the half-widths of the covering intervals are shown in Figs. \[fig:resFwhm\], \[fig:resQ5\] and \[fig:resQ9\]. The different flavours of the GSF in these figures are the same as those described in the caption of Fig. \[fig:probChi2\].
\[fig:qpPredTipBH\]
\[fig:resFwhm\]
\[fig:resQ5\]
\[fig:resQ9\]
The GSF and the KF have also been run on tracks from a full simulation using the official CMS simulation program [@CMSIM]. The $p_T$ and the $\eta$ range are the same as in the simplified simulation, but the amount and spatial distribution of the material are different. Probability distributions of the estimated $q/p$ for the GSF and the KF are shown in Fig. \[fig:pullProbsFull\].
\[fig:pullProbsFull\]
The probability distribution of the GSF exhibits no large deviation from flatness, indicating that the estimated PDF of $q/p$ describes reasonably well the actual PDF of $q/p$. This observation is all the more remarkable since, with the full simulation, the energy loss is not generated by the simple Bethe-Heitler model, and neither the exact amount nor the exact location of the material are known to the GSF.
The corresponding residuals of the estimated $q/p$ with respect to the true value are shown in Figs. \[fig:qpPredTipFull\] and \[fig:qpUpdTipFull\]. The residuals shown in Fig. \[fig:qpUpdTipFull\] have been obtained by including a vertex constraint in the fit. Such a constraint allows the momentum to be measured in the innermost part of the track and thus gives a handle on possible radiation in the first two layers. The result of including this constraint is a less skew distribution with the mode being moved closer towards zero, and the amount of tracks in the tails is also reduced. Even though the results from the full simulation qualitatively seem to confirm those from the simplified simulation, more studies are needed to understand the differences in detail.
\[fig:qpPredTipFull\]
\[fig:qpUpdTipFull\]
Conclusion
==========
The Gaussian-sum filter has been implemented in the CMS reconstruction program. It has been validated with electron tracks with a simplified simulation in which the energy loss distribution (Bethe-Heitler model), the exact amount of material and its exact location are known to the reconstruction program. It has been shown that the quality of the momentum estimate depends mainly on the number of mixture components kept during reconstruction, and to some extent also on the number of components in the mixture approximation to the energy loss distribution. A comparison with the best linear unbiased estimator, the Kalman filter, shows a clear improvement of the momentum resolution. Remarkably, a similar improvement can be seen with electron tracks from the full simulation, although in this case neither the exact energy loss distribution nor the precise amount and location of material are known to the reconstruction program. More systematic studies with electrons from the full simulation are clearly needed, but it seems safe to conclude that in electron reconstruction the Gaussian-sum filter yields a substantial gain in precision as compared to the Kalman filter.
[9]{} H. Bethe and W. Heitler, Proc. R. Soc. London [**A 146**]{} (1934) 83. R. Fr[ü]{}hwirth, Nucl. Instrum. and Methods [**A 262**]{} (1987) 444. D. Stampfer, M. Regler and R. Fr[ü]{}hwirth, Comp. Phys. Comm. [**79**]{} (1994) 157. R. Fr[ü]{}hwirth, Comp. Phys. Comm. [**100**]{} (1997) 1. R. Fr[ü]{}hwirth and S. Fr[ü]{}hwirth-Schnatter, Comp. Phys. Comm. [**110**]{} (1998) 80.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Sparse code multiple access (SCMA) is a promising multiplexing approach to achieve high system capacity. In this paper, we develop a novel iterative detection and decoding scheme for SCMA systems combined with Low-density Parity-check (LDPC) decoding. In particular, we decompose the output of the message passing algorithm (MPA) based SCMA multiuser detection into intrinsic part and prior part. Then we design a joint detection and decoding scheme which iteratively exchanges the intrinsic information between the detector and the decoder, yielding a satisfied performance gain. Moreover, the proposed scheme has almost the same complexity compared to the traditional receiver for LDPC-coded SCMA systems. As numerical results demonstrate, the proposed scheme has a substantial gain over the traditional SCMA receiver on AWGN channels and Rayleigh fading channels.'
author:
- |
Baicen Xiao, Kexin Xiao, Shutian Zhang, Zhiyong Chen, Bin Xia and Hui Liu\
Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, P. R. China\
Email: [{xinzhiniepan, kexin.xiao, zhangshutian, zhiyongchen, bxia, huiliu}@sjtu.edu.cn]{}
bibliography:
- 'SCMA\_iterative.bib'
title: Iterative detection and decoding for SCMA systems with LDPC codes
---
Introduction
============
Sparse code multiple access (SCMA) has attracted much attention since it is capable of supporting massive connections simultaneously, yielding a competitive candidate for Fifth Generation (5G) communications[@nikopour2013sparse]. Commonly, SCMA can be viewed as a generalization of sparse spread CDMA[@guo2008multiuser], with a few numbers of nonzero elements within a signature. For an uplink SCMA system, when each user is assigned a specific codebook, the multiplexing becomes a superposition scheme which will obtain the shaping gain. However, on the receiver side serious multiple address interference (MAI) is the main obstacle to implement multiuser detection. The optimum maximum a posterior (MAP) algorithm obviously shows the best performance with considerable complexity. In order to tackle the high complexity of MAP algorithm, some low complexity algorithm are proposed to handle this NP-complete problem[@verdu1998multiuser] within tolerable performance loss. Especially, thanks to the sparse structure of SCMA, the complex MAP formula can be solved iteratively with sum-product algorithm or message passing algorithm (MPA)[@kschischang2001factor].
Lately, in order to improve the bit error rate (BER) performance of SCMA, [@huawei] has introduced the Turbo-principle, which is widely used in detection and decoding problems such as joint source-channel coding [@hagenauer1997turbo; @hagenauer2003turbo] and multiuser detection [@wang1999iterative; @sanderovich2005ldpc], to exchange information between the SCMA detector and the channel decoder. However, the proposed Turbo-like scheme in [@huawei] doesn’t take full advantage of the iterative structure of SCMA detection and suffers from high complexity proportional to the number of outer iterations. For the sake of taking full advantage of the iterative characteristic of SCMA detection, we need a kind of channel coding which applies iterative decoding. Low-Density Parity-Check (LDPC) codes which are excellent error correcting codes providing a large coding gain [@richardson2003renaissance] and adopt iterative decoding, are especially suitable for our requirments.
The goal of this paper is to apply a novel Turbo-like combination of SCMA multiuser detection and LDPC decoding. The difference from [@huawei] should be noted that this paper do a novel Turbo-like combination of iterative detection and iterative decoding, i.e, during each outer loop only partial inner iterations in detector and decoder are implemented, to obtain a satisfied performance gain with almost the same complexity compared to the receiver without Turbo-like scheme. Firstly, we investigate in detail the MPA-based SCMA multiuser detector from the perspective of solving marginal function and then deduce the SCMA multiuser detection algorithm in logarithmic form. Furthermore, the intrinsic information is decomposed from the output of both SCMA multiuser detector and LDPC decoder and the way intrinsic information interacts between detector and decoder is presented. As numerical results show, this scheme achieves a 0.9 dB performance gain in terms of BER with almost the same complexity for both AWGN channels and Rayleigh fading channels compared to traditional receiver for LDPC-coded SCMA systems.
For the sake of clarity, throughout this paper, the sets of binary and complex numbers are denoted by $\mathbb{B}$ and $\mathbb{C}$, respectively. Upper-case calligraphic symbols $\mathcal{X}$ denote constellation sets and log$(\cdot)$ denotes natural logarithm. To represent a scalar, a vector and a matrix, we use $x$, $\mathbf{x}$ and $\mathbf{X}$, respectively.
System model
============
We consider an uplink LDPC-coded SCMA system with $J$ users and $K$ resources, and signaling through fading channels with additive white Gaussian noise (AWGN), as shown in Fig. [\[system\_diagram\]]{}.
For each user $j$, $j$ = 1, $\cdots$ , $J$, data bits {$d_{n_0}^{j}\mid n_0=1,2,\cdots, n$} are first encoded into {$b_{m_0}^{j}\mid m_0=1,2,\cdots, m$} by an LDPC encoder with code rate $R_j=n/m$. In order to reduce error bursts and take advantage of diversity gain, the coded bits are permuted by an interleaver $\pi_j$. Every log$_2$$(M)$ interleaved coded bits $\{b_{m^{\pi}}^{~j} | m^{\pi}=1,2,\cdots,\text{log}_2(M)\}$ is grouped together and then mapped by SCMA mapper $f_j$ into a $K$-dimensional complex symbol as $f_j$: $\mathbb{B}^{\text{log}_2(M)}\rightarrow{\mathbf{x}}_{j}\in\mathcal{X}^j\subset\mathbb{C}^K$ with cardinality $|\mathcal{X}^j|$ = $M$. Because of the sparsity of SCMA, a $K$-dimensional symbol ${\mathbf{x}}_{j}$ consists of $N_j$ $<$ $K$ non-zero elements, each corresponding to an OFDMA tone or other resources. For the receiver end, the received siganls are the superposition of $J$ users’ signals and ambient noise, which can be written in a discrete-form, if memoryless channel considered, as $$\ \mathbf{y} = \sum_{j=1}^J{\text{diag}(\mathbf{h}_{j}){\mathbf{x}}_j} + \mathbf{n}\text{,}$$ where $\mathbf{y} = (y_{1},\cdots,y_{K})^\text{T}$ is the received signal vector, $\mathbf{h}_{j} = (h_{1j},\cdots,h_{Kj})^\text{T}$ is the channel vector for user $j$, $\mathbf{x}_{j} = (x_{1j},\cdots,x_{Kj})^\text{T}$ is the symbol transmitted by user $j$, and $\mathbf{n}$ is a white Gaussian noise vector subject to $\mathcal{CN}(0,N_0\mathbf{I})$. Apparently, from (1) the received signal at resource $k$ can be written as $$\ y_k = \sum_{j=1}^J{h_{k,j}x_{k,j}}+n_k, k=1,\cdots, K.$$ It is easy to recognize from (2) that each user sees interference from other $K-1$ users. However, since the signal vector from arbitrary user is sparse, i.e., not all users contribute to the k-th resource, the interference is hence reduced and (2) can be re-written as $$\ y_k = \sum_{j\in{\partial k}}{h_{k,j}x_{k,j}}+n_k, k=1,\cdots, K.$$ where $\partial k$ denotes the users contributing to the $k$-th resource, called the neighborhood of node $k$, and this relationship which is decided by the SCMA mapper can be presented by factor graph and indicator matrix $\mathbf{C}$. Let $d_j$ and $d_k$ be the number of resources occupied by user $j$ and the number of users resource $k$ is connected, respectively. For the sake of clarity, we give an example of SCMA factor graph. Assuming $d_j = 2$ for all $j$ and $d_k = 3$ for all $k$, the factor representation is shown as Fig. [\[factor-graph\]]{}, and the corresponding indicator matrix is
$$\mathbf{C}=
\left[
\begin{array}{cccccc}
1 & 1 & 1 & 0 & 0 & 0\\
1 & 0 & 0 & 1 & 1 & 0\\
0 & 1 & 0 & 1 & 0 & 1\\
0 & 0 & 1 & 0 & 1 & 1
\end{array}
\right]$$ therein, $c_{k,i}=1$ means resource $k$ is occupied by user $i$.
{width="3.2in"}
It’s observed that this factor graph is regular, i.e., all the user nodes have the same degree, so do the resource nodes. It should be emphasized that this regular structure may not be the best, in another word, it’s possible for a irregular structure to play a better performance if there are diversities between the quality of different resources. In this paper, however, our attention is only focused on regular SCMA structure, but some of the results can be easily extended to the irregular scenario and the irregular structure will be our future work.
Iterative detection and decoding
==================================
In this section, we analyse SCMA detection based on MPA iterative algorithm, and propose an effective combination method of SCMA multiuser detection and LDPC decoding.
SCMA detection
--------------
Firstly, we describe why the MPA algorithm can be applied to SCMA detection from the perspective of solving marginal function. Given received signal $\mathbf{y}$ and assuming ideal channel estimation, a SCMA detection based on MAP is to choose a matrix $\hat{\mathbf{X}}=(\mathbf{x}_1, \cdots, \mathbf{x}_J)$ to maximizing the joint $a~posterior$ pmf, which is expressed as $$\ \hat{\mathbf{X}} = \arg\max_{\mathbf{X}\in{\mathcal{X}^{KJ}}} p(\mathbf{X}|\mathbf{y})\text{,}$$ where $\mathcal{X}^{KJ}$ denotes the set of all possible symbols, i.e., the $j$-th column of $\mathcal{X}^{KJ}$ is the set $\mathcal{X}^j$ described in Section . In order to estimate the information of user $j$, we can choose a $\mathbf{\hat{x}}_j$ to maximize the marginal $a~posterior$ pmf with respect to $\mathbf{x}_j$ as $$\ \hat{\mathbf{x}}_j = \arg\max_{z\in\mathcal{X}^j} \sum_{\substack{\mathbf{X}\in{\mathcal{X}^{K\!J}}\\{\mathbf{x}_j=z}}}{p(\mathbf{X}|\mathbf{y})}\text{.}$$ Applying Bayes’s rule, we can get $$\label{SCMA_f(x|y)}
\ p(\mathbf{X}|\mathbf{y})=\frac{p(\mathbf{y}|\mathbf{X})p(\mathbf{X})}{p(\mathbf{y})}$$ $$\ \propto p(\mathbf{y}|\mathbf{X})p(\mathbf{X})=p(\mathbf{X})\prod_{k=1}^K{p(y_k|\mathbf{X})}\text{.}$$ The last equation follows the fact that the elements of noise vector is identically independent distributed (i.i.d.) and uncorrelated with transmitted symbols, hence once the transmitted symbols are given, different dimension of received signal $\mathbf{y}$ are independent.
Since $y_k$ is influenced by parts of users, i.e., users from $\partial k$, equation (5) can be reduced to $$\label{SCMA_x_j}
\ \hat{\mathbf{x}}_j = \arg\max_{z\in\mathcal{X}^j} \sum_{\substack{\mathbf{X}\in{\mathcal{X}^{K\!J}}\\{\mathbf{x}_j=z}}}{p(\mathbf{X})\prod_{k=1}^K{p(y_k|\mathbf{x}_p, p\in{\partial k})}}~~\forall j\text{.}$$ Furthermore transmitted symbols from different users are independent, $p(\mathbf{X})$ can be written in a product form as $p(\mathbf{X})=\prod_{q=1}^{J}{p(\mathbf{x}_q)}$, and it’s clear that the sum terms in equation (\[SCMA\_x\_j\]) can be given by $$\label{SCMA_f(x)}
\ f(\mathbf{x}_1,\cdots,\mathbf{x}_J) = \prod_{q=1}^{J}{p(\mathbf{x}_q)}\prod_{k=1}^K{p(y_k|\mathbf{x}_p, p\in{\partial k})}\text{.}$$
Traditionally, to solve equation (\[SCMA\_x\_j\]) needs $J$ operations where redundant computing exits. Because of the product form (\[SCMA\_f(x)\]), thanks to the method in [@kschischang2001factor], we can solve a marginal function problem of multivariable function like (\[SCMA\_x\_j\]) iteratively based on factor graph, which is usually called sum-product algorithm or MPA algorithm. Each node in the factor graph sends “belief message" to its neighbors during each iteration and the “belief message" shouldn’t be sent back during the next iteration, hence the inference can be made sufficiently in the graph after some iterations. It should be emphasized that when the factor graph of equation (\[SCMA\_f(x)\]) is cycle-free MPA algorithm is able to produce the accurate marginal function [@guo2008multiuser], but for the factor graph with cycles this algorithm is suboptimal.
The iterative SCMA multiuser detection algorithm based on MPA is presented in algorithm 1.
**Variable definition**\
$V_{j\rightarrow k}^{t}(\mathbf{x}_j)$: the message sent from $j$-th user node to $k$-th resource node during $t$-th iteration\
$U_{k\rightarrow j}^{t}(\mathbf{x}_j)$: the message sent from $k$-th resource node to $j$-th user node during $t$-th iteration\
**Initialization**\
$p(\mathbf{x}_j) \leftarrow \frac{1}{M}$, for all $\mathbf{x}_j\in\mathcal{X}^j$ and $j=1, \cdots, J$.\
**for all** $j=1, \cdots, J$ and $k=1, \cdots, K$ **do**\
$U_{k\rightarrow j}^{0}(\mathbf{x}_j) \leftarrow 1 $, for all $\mathbf{x}_j\in\mathcal{X}^j$\
**for all** $j,k$ with $c_{jk}=1$ and $\mathbf{x}_j \in \mathcal{X}^j$ $\textbf{do}$\
$$V_{j\rightarrow k}^{t}(\mathbf{x}_j) \leftarrow p(\mathbf{x}_j)\prod_{s\in{\partial j}\backslash k}{U_{s\rightarrow j}^{t-1}(\mathbf{x}_j)}$$ **for all** $j,k$ with $c_{jk}=1$ and $\mathbf{x}_j \in \mathcal{X}^j$ $\textbf{do}$\
$$\begin{aligned}
U_{k\rightarrow j}^{t}(\mathbf{x}_j)&\leftarrow \sum_{(\mathbf{x}_p)\partial k \backslash j}\frac{1}{\pi N_0}\text{exp}[\frac{1}{N_0}\parallel y_k-h_{k,j}x_{k,j}\nonumber\\
&-\sum_{p\in\partial k \backslash j}{h_{k,p}x_{k,p}}\parallel ^2]\prod_{p\in{\partial k}\backslash j}V_{p\rightarrow k}^{t}(\mathbf{x}_p)\end{aligned}$$ **return** $\textbf{for~all}$ $\mathbf{x}_j\in\mathcal{X}^j$ and $j=1,\cdots,J$\
$V_j(\mathbf{x}_j)\leftarrow p(\mathbf{x}_j)\prod_{s\in{\partial j}}{U_{s\rightarrow j}^{t-1}(\mathbf{x}_j)}$
Therein£¬ “$\partial k \backslash j$” denotes the neighborhood of node $k$ excluding node $j$ and the notation “$\sum_{(\mathbf{x}_p)\partial k \backslash j}$" denotes sum over all possible values of $\mathbf{x}_p \in \mathcal{X}^p$ for all $p \in \partial k \backslash j$. In this algorithm, we assume all possible symbols are equiprobable, that is to say there is no prior information. When prior information is provided by channel decoder as described later, $p(\mathbf{x}_j)$ should be updated. For the convenience of implementation, we then derive the SCMA detection algorithm in logarithmic form, given prior information in the form of the log-likelihood ratio. Note that one of our goal is to decompose the output of SCMA detector into intrinsic information and prior information in order to do a novel Turbo-like combination with LDPC decoder in subsection $C$.
Once the prior information $\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J}$ is given, here the superscript “s" denotes the LLR is with respect to symbol, the posterior probability of $\mathbf{X}$ can be written as $$\label{SCMA_f(x|y,l)}
\begin{split}
\ p(\mathbf{X}|\mathbf{y},&\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})=\\
&C\cdot p(\mathbf{y}|\mathbf{X})p(\mathbf{X}|\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})\text{,}
\end{split}$$ where $C=\frac{p(\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})}{p(\mathbf{y},\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})}$ is a constant during each inner loop of SCMA detection. Comparing equation (\[SCMA\_f(x|y)\]) with (\[SCMA\_f(x|y,l)\]), it’s easy to show that only the initialization part of algorithm 1 should be modified when prior information is available. Then “$p(\mathbf{x}_j) \leftarrow \frac{1}{M}$" should be replaced by “$p(\mathbf{x}_j) \leftarrow p(\mathbf{x}_j|L^{s,p}(\mathbf{x}_j))$". Let us fix a reference point $\mathbf{\tilde{x}}_j$ , which denotes all “1" transmitted to facilitate the following deduction, for each user $j$. Let $LV_{j\rightarrow k}^{t}(\mathbf{x}_j)$ and $LU_{k\rightarrow j}^{t}(\mathbf{x}_j)$ be the information in logarithmic form from user node $j$ to resource node $k$ and from resource node $k$ to user node $j$, respectively, then, $$\label{LV}
\begin{split}
LV_{j\rightarrow k}^{t}(\mathbf{x}_j) &= \text{log}\frac{V_{j\rightarrow k}^{t}(\mathbf{x}_j)}{V_{j\rightarrow k}^{t}(\mathbf{\tilde{x}}_j)}\\
&= \text{log}\frac{p(\mathbf{x}_j|L^{s,p}(\mathbf{x}_j))\prod_{s\in{\partial j}\backslash k}{U_{s\rightarrow j}^{t-1}(\mathbf{x}_j)}}{p(\mathbf{\tilde{x}}_j|L^{s,p}(\mathbf{\tilde{x}}_j)\prod_{s\in{\partial j}\backslash k}{U_{s\rightarrow j}^{t-1}(\mathbf{\tilde{x}}_j)}}\\
&=L^{s,p}(\mathbf{x}_j)+\sum_{s\in{\partial j}\backslash k}LU_{k\rightarrow j}^{t-1}(\mathbf{x}_j)\text{,}
\end{split}$$ $$\label{LU}
\begin{split}
LU_{k\rightarrow j}^{t}&(\mathbf{x}_j) = \text{log}\frac{\sum_{(\mathbf{x}_p)\partial k \backslash j}\text{exp}[f_k(\mathbf{x}_j)]\prod_{p\in{\partial k}\backslash j}V_{p\rightarrow k}^{t}(\mathbf{x}_p)}{\sum_{(\mathbf{x}_p)\partial k \backslash j}\text{exp}[f_k(\mathbf{\tilde{x}}_j)]\prod_{p\in{\partial k}\backslash j}V_{p\rightarrow k}^{t}(\mathbf{x}_p)}\\
&=\text{log}\frac{\sum_{(\mathbf{x}_p)\partial k \backslash j} \text{exp}[f_k(\mathbf{x}_j)+\sum_{p\in{\partial k}\backslash j}LV_{p\rightarrow k}^{t}(\mathbf{x}_p)]}{\sum_{(\mathbf{x}_p)\partial k \backslash j}\text{exp}[f_k(\mathbf{\tilde{x}}_j)+\sum_{p\in{\partial k}\backslash j}LV_{p\rightarrow k}^{t}(\mathbf{x}_p)]}\text{.}
\end{split}$$
In equation (\[LU\]), $f_k(\mathbf{x}_j)=\frac{1}{N_0}\parallel y_k-h_{k,j}x_{k,j}\nonumber-\sum_{p\in\partial k \backslash j}{h_{k,p}x_{k,p}}\parallel ^2$. And in the final round, the output of SCMA detector is $$\label{scma_output}
LV_j(\mathbf{x}_j)=L^{s,p}_{j}(\mathbf{x}_j)+\sum_{s\in{\partial j}}LU_{k\rightarrow j}^{T}(\mathbf{x}_j)\text{.}$$ Hence, the output can be decomposed into two parts: one is the prior information and another is the intrinsic information derived from the structure of SCMA factor graph. From another perspective, the LLR of $\mathbf{x}_j$ given prior information can be written as $$\begin{split}
L^s_j&(\mathbf{x}_j)=\text{log}\frac{p(\mathbf{x}_j|\mathbf{y},\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})}{p(\mathbf{\tilde{x}}_j|\mathbf{y},\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})}\\
&=\text{log}\frac{p(\mathbf{y}|\mathbf{x}_j,\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})}{p(\mathbf{y}|\mathbf{\tilde{x}}_j,\{L^{s,p}(\mathcal{X}^j)\}_{j=1,\cdots,J})}
+\text{log}\frac{p(\mathbf{x}_j|L^{s,p}(\mathbf{x}_j))}{p(\mathbf{\tilde{x}}_j|L^{s,p}(\mathbf{\tilde{x}}_j))}\\
&=L^{s,i}_{j}(\mathbf{x}_j)+L^{s,p}_{j}(\mathbf{x}_j)\text{.}
\end{split}$$ It is apparent that $L^s_j(\mathbf{x}_j)$ also consists of two parts, the intrinsic parts and prior parts. when using MPA SCMA detection, we apply the term $\sum_{s\in{\partial j}}LU_{k\rightarrow j}^{T}(\mathbf{x}_j)$ to approximate $L^{s,i}_{j}(\mathbf{x}_j)$.
LDPC decoding
-------------
In this section, we briefly introduce the decoding process of LDPC for the sake of clear description of the combination of SCMA detection with LDPC decoder in the next section.
For LDPC is a sparse liner block code, it can be effectively expressed as factor graph which is a bipartite graph with check nodes and variable nodes. Take into consideration the complexity of implementation, LDPC decoder usually adopts BP algorithm using LLRs to reduce multiplication and avoid normalization. During each iteration, belief information is exchanged between variable nodes and check nodes based on check matrix. Let $L_{1,j}^{b,p}(b_i)$ be the LLR of the $i$-th bit of user $j$ input to the decoder, i.e., $L_{1,j}^{b,p}(b_i)=\text{log}\frac{p(b_i=0|\mathbf{y})}{p(b_i=1|\mathbf{y})}$, then the output of LDPC decoder is $$L_{2,j}^{b}(b_i)=L_{1,j}^{b,p}(b_i)+\sum_{m\in\partial i}L_{j,m\rightarrow i}\text{,}$$ where the notation $\partial i$ denotes the set of parity check functions variable $i$ belongs to and $L_{j,m\rightarrow i}$ is the belief information passed from check node $m$ to variable node $i$ in the final iteration. And the summation $\sum_{m\in\partial i}L_{j,m\rightarrow i}$ can be seen LDPC intrinsic information $L_{2,j}^{b,i}(b_i)$ based on LDPC structure. Hence, the output of LDPC decoder can also be decomposed to intrinsic information parts and prior information parts. Note that to distinguish the information generated by SCMA detector and LDPC decoder, we use subscript “1” and “2” respectively.
In the following section, we’ll cope with some obstacle to implement the novel combination of SCMA detection and LDPC decoding and present the iterative scheme.
\[information\_flow\]
Combination of LDPC decoding and SCMA detection
-----------------------------------------------
From equation (\[scma\_output\]), the output of SCMA detector is at symbol level, but unfortunately the bit LLRs are required for input of LDPC decoder. In order to do a combination of LDPC decoding and SCMA detection, one problem we should deal with is the transformation between symbol LLRs and bit LLRs. The input to LDPC decoder should be intrinsic information parts of the multiuser detector output for a better performance like the decoding method for Turbo codes. Then another problem arises that the prior information part for a symbol comprise several bits’ prior information and for a specific bit of this symbol, the other bits’ prior information can be seen as intrinsic information, so we should decompose the output of SCMA detector into two parts, prior bit LLR and intrinsic bit LLR.
Here we assume that an SCMA symbol consists of $n=\text{log}_2(M)$ bits. Without loss of generality, the interleaved coded bits for a symbol can be considered independent. For a specific symbol $\mathbf{x}_j$, let $\mathcal{X}_j^{+}=\{i~|\text{if the i-th bit of $\mathbf{x}_j$ is} ``1"\}$ and $\mathcal{X}_j^{-}=\{i~|\text{if the i-th bit of $\mathbf{x}_j$ is} ``0"\}$. Then, the symbol LLR can be calculated based on bit LLRs as $$L^{s,p}_{2,j}(\mathbf{x}_j)=\sum_{i\in \mathcal{X}_j^{-}}L_{2,j}^{b,p}(b_i)\text{.}$$ Let $\mathcal{X}_i^{j,0}$ and $\mathcal{X}_i^{j,1}$ denote the set of the symbols of which the $i$-th bit corresponding to “0" and “1" for user $j$, respectively. The bit LLR can be written as $$\begin{split}
&L_{1,j}^{b}(b_i)=\text{log}\frac{\sum_{\mathbf{x}_j\in\mathcal{X}_i^{j,0}}\text{exp}[LV_j(\mathbf{x}_j)]}{\sum_{\mathbf{x}_j\in\mathcal{X}_i^{j,1}}\text{exp}[LV_j(\mathbf{x}_j)]}\\
&=\text{log}\frac{\sum_{\mathbf{x}_j\in\mathcal{X}_i^{j,0}}\text{exp}[L^{s,p}_{2,j}(\mathbf{x}_j)+\sum_{s\in{\partial j}}LU_{k\rightarrow j}^{T}(\mathbf{x}_j)]}{\sum_{\mathbf{x}_j\in\mathcal{X}_i^{j,1}}\text{exp}[L^{s,p}_{2,j}(\mathbf{x}_j)+\sum_{s\in{\partial j}}LU_{k\rightarrow j}^{T}(\mathbf{x}_j)]}\\
&=L_{1,j}^{b,i}(b_i)+L_{2,j}^{b,p}(b_i)\text{.}
\end{split}$$ The last equation follows the fact that $L^{s,p}_{2,j}(\mathbf{x}_j)$ is a combination of $\{{L_{2,j}^{b,p}(b_i)}\}_{i=1, \cdots, n}$ and we can extract a same product term $\text{exp}[L_{2,j}^{b,p}(b_i)]$ out of each sum term in the numerator.
Let $\text{I}_{T}$ and $\text{I}_{L}$ be the number of iterations of SCMA detector and LDPC decoder, respectively. The scheme of the novel iterative combination is described as follows. Besides outer iterations, this scheme consists of two inner stages: an SCMA detection stage with $\text{I}_{T}$ iterations, followed by LDPC decoding stage with $\text{I}_{L}$ iterations and the number of inner iterations of both stages can be much less than the number of iterations required by the SCMA receiver with or without traditional Turbo-like outer iterations. The two inner stages are connected by a connection module containing LLR converters, interleavers and deinterleavers. During each outer iteration, the symbol LLRs $\{L^{s,p}_{1,j}(\mathbf{x}_j)\}$ output by SCMA detector are converted into bit LLRs $\{L_{1,j}^{b}(b_i)\}$ by LLR converter.In order to obtain a better performance, the intrinsic bit LLRs $\{L_{1,j}^{b,i}(b_i)\}$ from SCMA detector should be extracted and then feed to the deinterleaver to get $\{L_{1,j}^{b,p}(b_i)\}$ which is fed into LDPC decoder as prior information. The process to feed information from LDPC decoder output to SCMA detector is similar. During each outer iteration, the detector and decoder only do a few number of iterations, less than the receiver without Turbo-like scheme need, although the total number of iterations may be the same. We call this characteristics partial inner iterations. For example, numerical results in the next section show that to obtain a satisfied BER performance 8 iterations are needed in traditional SCMA receiver, but 2 inner iterations are enough in our scheme with 4 outer iterations and the BER performance is even better. Hence this iterative combination can achieve a more satisfied BER performance gain with almost the same complexity as shown in the next section. The connection module and the information flow of a special case where $\text{I}_{T}=\text{I}_{L}=1$ is shown as Fig.\[information\_flow\]. We call this special case mode 1. In particular, in mode 1 the detector does 1 demapping iteration and then the LLR goes to LDPC decoder through connection modules. The LDPC decoder only does 1 iteration and fed back the LLR to the detector. Then we goes to the next iteration between detector and LDPC decoder. The best number of inner iterations under the constraint of total iteration number is unsolved and will be our future work. Intuitively, the more outer iterations the more intrinsic information exchanged, and hence the better performance. However, it will be shown in the next section that when take the extreme case where $\text{I}_{T}=\text{I}_{L}=1$ the BER performance is poor. One reason may lie in the fact that if $\text{I}_{T}=1$, it can be seen from Fig.\[information\_flow\] that the information interacts in SCMA detector is far from sufficient, i.e., there is no red arrow line connecting user node 1 with user node 6. Hence, in this scenario the information given out from SCMA detector is particularly poor so that the information cannot be corrected by LDPC decoder with only 1 iteration.
Complexity analysis
-------------------
In this section, considering regular SCMA systems, we evaluate the complexity of traditional scheme for LDPC-coded SCMA receiver and our proposed scheme, i.e., the average number of operations we need for each symbol. Since the main complexity is introduced by multiplication, division, logarithm and exponent operations, we mainly consider these four kinds of operations for real number. From equation (\[LV\]) and (\[LU\]), it is clear that each inner iteration of SCMA detection needs $(2d_kKM^{d_k}+4{d_k}^2KM)/J$ multiplication operations, $d_kKM^{d_k}/J$ division operations $d_kKM^{d_k}/J$ exponent operations and $KMd_k/J$ logarithm operations. For regular LDPC codes with 1/2 code rate, $(11P-9)\text{log}_2(M)$ multiplications and $(P+1)\text{log}_2(M)$ divisions are needed during each iteration of decoding, [@fossorier1999reduced], where P denotes the degree of variable nodes. For other kind of code rates, the order of complexity is the same [@mackay1999good]. $M\text{log}_2(M)$ exponent operations and $2\text{log}_2(M)$ logarithm operations are needed in connection module during each outer iterations and these operations are caused by the converting from symbol LLRs to bit LLRs. It is now obvious that the main complexity is introduced by SCMA detection. Let $\text{I}_{O}$ be the number of outer iterations. From the deduction above and ignore the lower oder terms, it is easy to get that the total number of multiplication operations is approximately equal to $2I_OI_Td_kKM^{d_k}/J$, division and exponent operations are both approximately equal to $I_OI_Td_kKM^{d_k}/J$.
Here, with a slight abuse of notation, let the superscript t and p denote traditional scheme and proposed scheme, respectively. For the traditional scheme $I^t_O=1$, and for the proposed scheme $I^p_OI^p_T$ can be approximately equal to $I^t_T$. Hence, the complexity of our proposed scheme is almost the same as traditional receiver for LDPC-coded SCMA systems.
Numerical results
=================
In this section, numerical simulations are conducted to evaluate the performance of our proposed scheme. This simulation is based on the LDPC codes with finite block size. The regular systematic LDPC code in China Mobile Multimedia Broadcasting (CMMB) systems is used in the simulations [@hu2013low]. The code length we adopted is 9216 bits and code rate is $R=1/2$. Bit error rates (BER) over AWGN channel and Rayleigh fading channel with $J=6,~K=4,~M=4$, and $N_j=2$ for all $j$ are presented. The user-specific codebooks are designed according to [@taherzadeh2014scma]. The channel coefficient $h$ is set to unit power for both AWGN channel and Rayleigh fading channel, i.e., $E(h^2)=1$. To keep numerically stable, when computing the term like $\text{log}(e^a+e^b)$, we adopt the equation as $\text{log}(e^a+e^b)=\text{max}(a,b)+\text{log}(1+e^{-|a-b|})$ [@hoshyar2008novel].
{width="3.2in"}
Here, we compare four different modes, of which the total number of iterations is almost the same excepting the mode 1 “$\text{I}_{T}=\text{I}_{L}=1, \text{I}_{O}=32$". The purpose to compare this mode with great difference in total iteration number is to show that the BER performance of this extreme case deteriorates seriously as shown in Fig. \[AWGN\] and Fig. \[Rayleigh\]. This poor BER performance of Mode 1 is consistent with the analysis in the last section. Fig. \[AWGN\] shows that the required ${E_s/N_0}$ of the Mode 2 to achieve a BER of $10^{-4}$ is 4.5 dB on AWGN channels and 5.4 dB on Rayleigh fading channels, obtaining 0.3 dB gain over Mode 3 and 0.9 dB gain over Mode 4 on both kinds of channels. Note that Mode 4 is the traditional SCMA receiver scheme, i.e., there is no information fed back from decoder to detector. It can also be recognized that the BER *waterfall* region of Mode 3 is more narrow than Mode 4 and that of Mode 2 is even more narrow than Mode 3. The phenomenon that a better BER performance and more narrow *waterfall* region of Mode 2 and Mode 3 than Mode 4 owes to the ability of iterative combination that more *diversity* gain can be achieved. Note that the complexity of the last three modes is almost the same if ignore the complexity caused by interleaver. If let Mode 2 and Mode 4 have the same BER performance, Mode 2 should reduce the number of iterations. Hence the proposed iterative scheme can be viewed as a method to reduce complexity from another perspective.
{width="3.2in"}
Conclusions
===========
In this paper, we have proposed a novel Turbo-like combination scheme of SCMA detector and LDPC decoding with almost the same complexity compared to traditional SCMA receiver (non-iterative structure). Firstly, we have investigated the SCMA detection algorithm based on MPA from the aspect of solving marginal functions and shown this MPA-based detection is an approximation of MAP criterion since the factor graph of SCMA is usually not cycle-free. Then the MPA-based SCMA detection in Logarithmic form as well as the intrinsic information of the output have also been deduced in order to apply a novel soft Turbo-like combination with LDPC decoder. The combination scheme is described in detail and the information flows of a special case are presented. Numerical results show that the proposed scheme can achieve an excellent performance gain and more narrow waterfall region in terms of BER compared with traditional SCMA receiver for both AWGN channels and Rayleigh fading channels.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Franz Achleitner[^1], Anton Arnold[^2], and Eric A. Carlen[^3]'
title: 'On multi-dimensional hypocoercive BGK models'
---
We study hypocoercivity for a class of linearized BGK models for continuous phase spaces. We develop methods for constructing entropy functionals that enable us to prove exponential relaxation to equilibrium with explicit and physically meaningful rates. In fact, we not only estimate the exponential rate, but also the second time scale governing the time one must wait before one begins to see the exponential relaxation in the $L^1$ distance. This waiting time phenomenon, with a long plateau before the exponential decay “kicks in” when starting from initial data that is well-concentrated in phase space, is familiar from work of Aldous and Diaconis on Markov chains, but is new in our continuous phase space setting. Our strategies are based on the entropy and spectral methods, and we introduce a new “index of hypocoercivity” that is relevant to models of our type involving jump processes and not only diffusion. At the heart of our method is a decomposition technique that allows us to adapt Lyapunov’s direct method to our continuous phase space setting in order to construct our entropy functionals. These are used to obtain precise information on linearized BGK models. Finally, we also prove local asymptotic stability of a nonlinear BGK model.
**keywords:** kinetic equations, BGK models, hypocoercivity, Lyapunov functionals, perturbation methods for matrix equations
Introduction
============
This paper is concerned with the large time behavior of nonlinear BGK models (named after the physicists Bhatnagar-Gross-Krook [@BGK54]) and their linearizations around their Maxwellian steady state. With respect to position, we consider here only models on $\tilde{\mathbf{T}}^d:=\big(\frac{L}{2\pi}{\mathbf{T}}\big)^d$, the $d$-dimensional torus of side length $L$ without confinement potential. Then, the usual BGK model for a phase space density $f(x,v,t)$; $x\in\tilde{\mathbf{T}}^d,\,v\in{{\mathbb{R}}^d}$ satisfies the kinetic evolution equation $$\label{bgk}
\partial_t f+v\cdot\nabla_x f= {\mathbf{Q}}f := M_f(x,v,t) -f(x,v,t)\ ,\quad t\ge0\ ,$$ where $M_f$ denotes the local Maxwellian corresponding to $f$; i.e., the local Maxwellian with the same hydrodynamic moments as $f$: $$M_f(x,v,t)=\frac{\rho(x,t)}{(2\pi T(x,t))^{\frac{d}2}}\,e^{-\frac{|v-u(x,t)|^2}{2T(x,t)}}
=\frac{\rho(x,t)^{1+\frac{d}{2}}}{(2\pi P(x,t))^\frac{d}2}\,e^{-\frac{|v-u(x,t)|^2\rho(x,t)}{2P(x,t)}}\ ,$$ with density $$\rho(x,t):={\int_{{\mathbb{R}}^d}}f(x,v,t)\,{\,\operatorname{d}\!v}\ ,$$ mean velocity $$u(x,t):=\frac1{\rho(x,t)}{\int_{{\mathbb{R}}^d}}vf(x,v,t)\,{\,\operatorname{d}\!v}\ ,$$ temperature $$T(x,t):=\frac1{d\rho(x,t)}{\int_{{\mathbb{R}}^d}}|v-u(x,t)|^2 f(x,v,t)\,{\,\operatorname{d}\!v}\ ,$$ and pressure (setting the gas constant $R=1$) $$P(x,t):=T(x,t)\rho(x,t)=\frac1{d}{\int_{{\mathbb{R}}^d}}|v-u(x,t)|^2 f(x,v,t)\,{\,\operatorname{d}\!v}\ .$$
Let ${\,\operatorname{d}\!\tilde x}:=L^{-d}{\,\operatorname{d}\!x}$ denote the normalized Lebesgue measure on $\tilde{\mathbf{T}}^d$, and consider normalized initial data $f^I(x,v)$ such that $$\label{f0-normal}
\int_{\tilde{\mathbf{T}}^d\times{\mathbb{R}}^d} f^I(x,v){\,\operatorname{d}\!\tilde x} {\,\operatorname{d}\!v} =1\ ,\quad
\int_{\tilde{\mathbf{T}}^d\times{\mathbb{R}}^d} vf^I(x,v){\,\operatorname{d}\!\tilde x} {\,\operatorname{d}\!v} =0\ ,\quad
\int_{\tilde{\mathbf{T}}^d\times{\mathbb{R}}^d} |v|^2f^I(x,v){\,\operatorname{d}\!\tilde x} {\,\operatorname{d}\!v} =d\ .$$ This means, our system has unit mass, zero mean momentum, and unit position-averaged pressure (w.l.o.g. this can be obtained by a Galilean transformation and choice of units). One easily checks that this normalization is conserved under the flow of . Hence the system is expected to have the unique, space-homogeneous steady state $$f^\infty(v) = M_1(v):=(2\pi)^{-\frac{d}2}e^{-\frac{|v|^2}2} \ ,$$ the centered Maxwellian at unit temperature, which clearly has the same normalization as . A standard argument involving the Boltzmann entropy confirms that this is indeed the case, but it gives no information on the rate of convergence to equilibrium, nor does it even prove convergence. We remark that involves two different time scales: the generic transport time is $O(L)$, while the relaxation time is $O(1)$. The goal of this paper is to prove the large time convergence to this $f^\infty$ for solutions of and its linearizations in 1, 2, and 3D with explicitly computable exponential rates.
This extends our previous work [@AAC16], which considered the 1D linear BGK model: $$\label{bgk:linear}
\partial_t f +v\cdot\nabla_x f = {\mathbf{Q_{lin}}}f := M_{T}(v)\,\int_{\mathbb{R}}f(x,v,t)\,{\,\operatorname{d}\!v} -f(x,v,t)\ ,\quad t\ge0\ ,$$ where $M_{T}$ denotes the normalized Maxwellian at some temperature $T>0$: $$M_T(v) = (2\pi T)^{-1/2}e^{-|v|^2/2T}\ .$$ In [@AAC16] we studied the rate at which normalized solutions of approach the steady state $f^\infty = M_T$ as $t\to\infty$. This problem is interesting since the collision mechanism drives the local velocity distribution towards $M_T$, but a more complicated mechanism involving the interaction of the streaming term $v \partial_x$ and the collision operator ${\mathbf{Q_{lin}}}$ is responsible for the emergence of spatial uniformity.
To elucidate this key point, let us define the operator ${\mathbf{L}}$ by $${\mathbf{L}}f(x,v) := -v\ \partial_x f (x,v) + {\mathbf{Q_{lin}}}f(x,v)\ .$$ The evolution equation can be written $\partial_t f = {\mathbf{L}}f$. Let $\mathcal{H}$ denote the weighted space $L^2(\tilde{\mathbf{T}}\times{\mathbb{R}}^d;M_T^{-1}(v){\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})$, where in the current discussion $d=1$. Then ${\mathbf{Q_{lin}}}$ is self-adjoint on $\mathcal{H}$, ${\mathbf{L}}f^\infty = 0$, and a simple computation shows that if $f(t)$ is a solution of , $$ \frac{{\rm d}}{{\rm d}t} \| f(t) - f^\infty\|_\mathcal{H}^2 = 2 \langle f(t), {\mathbf{L}}f(t)\rangle_\mathcal{H} = 2 \langle f(t), {\mathbf{Q_{lin}}}f(t)\rangle_\mathcal{H}
= -2\| f - M_T \rho\|_{\mathcal{H}}^2\ ,$$ where, as before, $\rho(x,t) := \int_{{\mathbb{R}}}f(x,v,t) {\,\operatorname{d}\!v}$. Thus, while the norm $ \| f(t)- f^\infty\|_\mathcal{H}$ is monotone decreasing, the derivative is zero whenever $f(t)$ has the form $f(t) = M_T \rho$ for [*any*]{} smooth density $\rho$. In particular, the inequality $$\label{coercive}
\langle f - f^\infty , {\mathbf{L}}(f - f^\infty) \rangle_\mathcal{H} \leq - \lambda \|f - f^\infty\|_\mathcal{H}^2$$ is valid in general for $\lambda = 0$, but for no positive value of $\lambda$. If (\[coercive\]) were valid for some $\lambda>0$, we would have had $\|f(t) - f^\infty\|_\mathcal{H}^2 \leq e^{-t\lambda}\|f^I - f^\infty\|_\mathcal{H}^2$ for all solutions of our equation, and we would say that the evolution equation is [*coercive*]{}. However, while this is not the case, it does turn out that one still has constants $1< c < \infty$ and $\lambda> 0$ such that $$\|f(t) - f^\infty\|_\mathcal{H}^2 \leq ce^{-t\lambda}\|f^I - f^\infty\|_\mathcal{H}^2\ .$$ (The fact that there exist initial data $f(0) \neq f^\infty $ for which the derivative of the norm is zero shows that necessarily $c>1$.) In Villani’s terminology (see §3.2 of [@ViH06]), this means that our evolution equation is [*hypocoercive*]{}.
Since $f(t)$ and $f^\infty$ are probability densities, a natural norm in which to measure the distance between them is the $L^1$ distance, or, what is the same up to a factor of $2$, the total variation distance between the corresponding probability measures. However, as is well known, the norm $\|\cdot \|_\mathcal{H}$ controls the $L^1$ norms. Specifically, by the Cauchy-Schwarz inequality, $$\begin{aligned}
\label{L1dom}
\|f(t) - f^\infty\|_\mathcal{H}^2 &=& \int_{\tilde{\mathbf{T}}\times{\mathbb{R}}^d} | f(x,v,t)M_T^{-1}(v) -1|^2 M_T(v){\,\operatorname{d}\!\tilde x} {\,\operatorname{d}\!v}\nonumber\\
&\geq& \left(\int_{\tilde{\mathbf{T}}\times{\mathbb{R}}^d} | f(x,v,t)M_T^{-1}(v) -1| M_T(v){\,\operatorname{d}\!\tilde x} {\,\operatorname{d}\!v}\right)^2\nonumber\\
&=& \| f(t) - f^\infty\|^2_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})}\ .\end{aligned}$$
Many hypocoercive equations have been studied in recent years [@ViH06; @He06; @DoMoScH09; @DoMoScH10; @ArEr14], including BGK models in §1.4 and §3.1 of [@DoMoScH10], but sharp decay rates were rarely an issue there. In our earlier work [@AAC16], we established hypocoercivity for such models in 1D by an approach that yields explicit – and quite reasonable – values for $c$ and $\lambda$. To this end, our main tools have been variants of the *entropy–entropy production method*.
The articles [@AAC16] and [@DoMoScH10] only consider BGK models with conserved mass, and partly with also conserved energy. But the tools presented there did not apply to BGK equations that also conserve momentum. This is in fact an important structural restriction that we shall formalize in §\[sec-hypo-index\] with the notion *hypocoercivity index*. The common feature of all models analyzed in [@AAC16] as well as in [@DoMoScH10] is that their hypocoercivity index is 1. The main goal of this paper is to extend the methods from [@AAC16] (i.e. constructing feasible Lyapunov functionals) to models with higher hypocoercivity index. Applied to BGK equations this then also includes models with conserved momentum.
The existence of global solutions for the Cauchy problem of has been proven in case of unbounded domains [@Pe89] and bounded domains [@Ri91; @PePu93], respectively. In case of bounded domains (such as $x\in\tilde{\mathbf{T}}^d$), these solutions are essentially bounded and unique [@PePu93]. For a space-inhomogeneous nonlinear BGK model with an external confinement potential, the global existence of solutions for its Cauchy problem and their strong convergence in $L^1$ to a Maxwellian equilibrium state has been proven recently [@BoCa09].
In the first part of this paper we shall study the linearization of the BGK equation around the centered Maxwellian with constant-in-$x$ temperature equal to one.
To this end we consider $f$ close to the global equilibrium $M_1(v)$, with $h$ defined by $f(x,v,t) = M_1(v)+h(x,v,t)$. Then $$\begin{aligned}
\rho(x,t) &= 1 + \sigma(x,t) \qquad\text{with }\: \sigma(x,t):=\int_{{{\mathbb{R}}^d}} h(x,v,t) {\,\operatorname{d}\!v}\ , \nonumber \\
(\rho u)(x,t) &= \int_{{{\mathbb{R}}^d}} vf(x,v,t) {\,\operatorname{d}\!v} =\mu(x,t) \qquad\text{with the vector function }\: \mu(x,t):=\int_{{{\mathbb{R}}^d}} vh(x,v,t) {\,\operatorname{d}\!v}\ ,\label{f:perturb} \\
P(x,t) &= \frac1d \int_{{{\mathbb{R}}^d}} (|v-u|^2) f(x,v,t) {\,\operatorname{d}\!v}
= 1+\frac1d\big[\tau(x,t)-\frac{|\mu(x,t)|^2}{1+\sigma(x,t)}\big] \quad\text{with }\: \tau(x,t):={\int_{{\mathbb{R}}^d}}|v|^2 h(x,v,t) {\,\operatorname{d}\!v}\ . \nonumber $$ The conservation of the normalizations implies $$\label{mloc2}
\int_{\tilde{\mathbf{T}}^d} \sigma(x,t){\,\operatorname{d}\!\tilde x} =0\ , \qquad \int_{\tilde{\mathbf{T}}^d} \mu(x,t){\,\operatorname{d}\!\tilde x} =0 \ , \qquad{\rm and}\qquad \int_{\tilde{\mathbf{T}}^d} \tau(x,t){\,\operatorname{d}\!\tilde x} =0\ .$$ The perturbation $h$ then satisfies $$\partial_t h (x,v,t) + v\cdot \nabla_x h(x,v,t) = [M_f(x,v,t) -M_1(v)] - h(x,v,t)\ , \quad t\geq 0 \ .$$ For $\sigma$, $\mu$, and $\tau$ small we have $$\begin{aligned}
\label{taylor}
& M_f(x,v) - M_1(v) \nonumber \\
&= \frac{(1+\sigma)^{1+\frac{d}2}(x)}{\left(2\pi \big\{1+\frac1d\big[\tau(x)-\frac{|\mu|^2(x)}{1+\sigma(x)}\big]\big\}\right)^\frac{d}2}
\,\exp\Big\{-\frac{|v(1+\sigma(x))-\mu(x)|^2}{2\big(1+\frac1d\big[\tau(x)-\frac{|\mu|^2(x)}{1+\sigma(x)}\big]\big) (1+\sigma(x))}\Big\}
- (2\pi)^{-\frac{d}2}e^{-\frac{|v|^2}2}\nonumber \\
&\approx M_1(v)\left[\big( 1+\frac{d}2 - \frac{|v|^2}{2}\big)\sigma(x)+v\cdot \mu(x)
+ \big( -\frac{1}{2}+ \frac{|v|^2}{2d}\big)\tau(x)\right] \ ,\end{aligned}$$ which yields the linearized BGK model that we shall analyze in dimensions 1, 2, and 3 in this paper: $$\begin{aligned}
\label{linBGK:torus}
& \partial_t h (x,v,t) + v\cdot \nabla_x h(x,v,t) \\
&\quad = M_1(v)\left[\big( 1+\frac{d}2 - \frac{|v|^2}{2}\big)\sigma(x,t)+v\cdot \mu(x,t)
+ \big( -\frac{1}{2}+ \frac{|v|^2}{2d}\big)\tau(x,t)\right] - h(x,v,t)\ , \quad t\geq 0 \ .\nonumber\end{aligned}$$ Here and in the sequel we only have $h(x,v,t)\approx f(x,v,t)-M_1(v)$, but for simplicity of notation we shall still denote the perturbation by $h$.
\[linBGK-decay\] For each side length $L>0$ and for dimensions $d=1,2,3$, there exists a (quadratic) entropy functional ${\mathcal{E}}^d(f)$ and a decay rate $\lambda^d(L)>0$ satisfying $$\label{entropy-equiv}
c_d(L) {\mathcal{E}}^d(f) \leq \| f - M_1\|^2_\mathcal{H} \le C_d(L) {\mathcal{E}}^d(f)\ ,$$ with some positive constants $c_d,\,C_d$ given explicitly in the proofs. Moreover, any solution $h(t)$ to with ${\mathcal{E}}^d(h^I+M_1)<\infty$, normalized according to , then satisfies $$\label{ineq:linBGK-decay}
{\mathcal{E}}^d(h(t)+M_1) \leq e^{-\lambda^d(L)\ t}\ {\mathcal{E}}^d(h^I+M_1)\ ,\qquad t\ge0\ .
$$
1. Combining and the bound on the right in , we obtain a [*Pinsker type inequality*]{} [@Pi64] for our entropy. Let $\tilde f := h+M_1$. Then $$\label{pinsker}
\| \tilde f - M_1\|_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})} \le \sqrt{C_d(L) {\mathcal{E}}^d(\tilde f)}\ .$$
2. For any solution $h(t)$ to with ${\mathcal{E}}^d(h^I+M_1)<\infty$, normalized according to , the function $\tilde f(t):=h(t)+M_1$ satisfies $$\label{twoscale1}
\| \tilde f(t) - M_1\|_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})} \le \sqrt{C_d(L) {\mathcal{E}}^d(\tilde f^I)} e^{-\lambda^d(L)\ t / 2}\,,$$ due to and with $\tilde f^I :=h^I +M_1$. However, since $\tilde f(t)$ and $M_1$ are both probability measures, we also have $$\label{twoscale2}
\| \tilde f(t) - M_1\|_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})} \le 2$$ for all $t$. Moreover, if most of the mass density is initially located in a small portion of $\tilde{\mathbf{T}}$; e.g., if the gas molecules are initially released from a small container into a vacuum in the rest of $\tilde{\mathbf{T}}$, then $\| \tilde f(t) - M_1\|_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})}$ will be close to $2$ until the streaming has had time to distribute the particles more uniformly over $\tilde {\mathbf{T}}$. Our estimates bound the time that it takes for this to happen.
Combining with yields $$\label{Hnorm-decay-est}
\| \tilde f(t) - M_1\|_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})} \le \min\left\{2,\,\sqrt{C_d(L)\,{\mathcal{E}}^d(\tilde f^I)}\,e^{-\lambda^d(L)\ t/2} \right\}\ ,\qquad t\ge0\ .$$ Our bound improves the trivial bound only for $t> t_{{\rm init}}$ where $$t_{{\rm init}} := \frac{ \log C_d(L) + \log {\mathcal{E}}^d(\tilde f^I)-2\log 2}{ \lambda^d(L) } \ .$$
For the one dimensional case, it is shown in Remark \[rem3.2\] that $\lambda^1(L) = O(1/L^2)$ in the limit $L\to\infty$. Moreover, the constant $C_1(L)$ approaches $1$ in the limit $L\to\infty$ by using the limiting behavior $\alpha_*(L)=O(1/L)$ in expression . For initial data $\tilde f^I$ with all of the gas molecules initially located in a small region of $\tilde{\mathbf{T}}$ with a volume fraction of order $\epsilon$, the initial entropy ${\mathcal{E}}^1(\tilde f^I)$ will satisfy ${\mathcal{E}}^1(\tilde f^I) = O(\epsilon^{-2})$. In this case, $t_{{\rm init}}$ is approximately given by $$O(-L^2 (C +\log \epsilon) )\quad \mbox{for } \epsilon\ll 1,\;L\gg1$$ and some positive constant $C$. Thus one time scale in our problems is given, or at least bounded, by $t_{{\rm init}}$. After this time, the solution satisfies $$\label{twoscale3b}
\| \tilde f(t) - M_1\|_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})} \leq 2 e^{-\lambda^d(L)\ (t- t_{{\rm init}})/2}\ ,$$ and the second time scale, is given by $2/\lambda^d(L)$, the waiting time after $t_{{\rm init}}$ for $\| \tilde f(t) - M_1\|_{L^1(\tilde{\mathbf{T}}\times{\mathbb{R}}^d,{\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})}$ to decrease by a factor of $1/e$; see Figure \[fig:H-decay\].
These two times scales are quite similar to what one observes in interacting particle systems or even in card shuffling; see [@AlDi86; @Di96]. In particular, [@AlDi86 Fig. 2] is quite similar to our Fig. \[fig:H-decay\] below.
3. The resemblance of to the results of Aldous and Diaconis for finite Markov chains in [@AlDi86; @Di96], and in particular for card shuffling, is not a coincidence. The equation can be interpreted as the Kolmogorov forward equation for a Markov process. Exponential rates for related Markov process with exponential rates have been obtained by probabilistic methods; see [@BL] for an early study of this type. However, the approach in [@BL] relies on compactness arguments and does not yield explicit values for $c$ or $\lambda$. One difference between our results and those for finite Markov chains is that in our case, the initial relative entropy can be infinite. In card shuffling, starting form a perfectly ordered deck of cards, one starts from a state of maximal—but finite—relative entropy, and the waiting time for uniformization from this state dominates that of any other starting point. For this reason, the initial waiting time for finite Markov chains is a universal “worst case”, while this is impossible in our setting; our result must refer to ${\mathcal{E}}^d(\tilde f^I)$.
4. Our bound on the decay rate is monotonically decreasing in $L$ and satisfies $\lambda^d(L=0)>0$ and $\lambda^d(L=\infty)=0$ (for $d=1$ see Fig. \[fig:muL\] below). Moreover $c_d(L=0)=C_d(L=0)=1$ (see , below).
![These two functions illustrate the time dependent decay estimate from . The values of $C_d,\,\lambda^d$ correspond to the 1D case with $L=2\pi$, and we chose ${\mathcal{E}}^d(\tilde f^I)=15$. We also show the two time scales of the BGK equation: $t_{{\rm init}}$ marks the intersection point of the two blue curves and it corresponds to the generic transport time. $t_2:=t_{{\rm init}}+\frac2\lambda$ marks the intersection point of the exponential curve with the value $2/e$, and $t_2-t_{{\rm init}}$ corresponds to the relaxation time scale. For larger values of $L$, $t_{{\rm init}}$ will be much larger. []{data-label="fig:H-decay"}](remark12.eps)
To prove local asymptotic stability for the nonlinear BGK equation in 3D, we make use of another set of norms: For $\gamma \geq 0$, let $H^\gamma(\tilde{\mathbf{T}}^3)$ be the Sobolev space consisting of the completion of smooth functions $\varphi$ on $\tilde{\mathbf{T}}^3$ in the Hilbertian norm $$\|\varphi\|_{H^\gamma}^2 := \sum_{k\in {\mathbb{Z}}^3} (1+|k|^2)^{\gamma}|\varphi_k|^2\ ,$$ where $\varphi_k$ ($k\in{\mathbb{Z}}^3$) is the $k$th Fourier coefficient of $\varphi$. Let $\mathcal{H}_\gamma$ denote the Hilbert space $H^\gamma(\tilde{\mathbf{T}}^3)\otimes L^2({\mathbb{R}}^3;M_1^{-1})$. Then the inner product in $\mathcal{H}_\gamma$ is given by $$\langle f,g\rangle_{\mathcal{H}_\gamma} = \int_{\tilde{\mathbf{T}}^3\times{\mathbb{R}}^3} \overline{f}(x,v) \left[\left(1 - \Delta_x\right)^{\gamma} g(x,v) \right]M_1^{-1}(v) {\,\operatorname{d}\!\tilde x} {\,\operatorname{d}\!v} \ .$$
\[BGK-decay\] Let $L=2\pi$ and let the initial data $f^I$ satisfy the normalization .
1. \[BGK-decay:a\] For all $\gamma \ge 0$ there is an entropy functional ${\mathcal{E}}_\gamma(f)$ satisfying $$ \frac34 {\mathcal{E}}_\gamma(f) \leq \|f-M_1\|_{\mathcal{H}_\gamma}^2 \leq \frac32 {\mathcal{E}}_\gamma(f)$$ such that, if $h$ is a solution of the linearized BGK equation in 3D with initial data $h^I$ and ${{\mathcal{E}}_\gamma(h^I+M_1)< \infty}$, then $$ {\mathcal{E}}_\gamma(h(t)+M_1) \leq e^{-t/2820}{\mathcal{E}}_\gamma(h^I+M_1)\ ,\qquad t\ge0\ .$$
2. \[BGK-decay:b\] Moreover, for all $\gamma>3/2$, there is an explicitly computable $\delta_\gamma>0$ such that, if $f$ is a solution of the nonlinear BGK equation with initial data $f^I$ and $\|f^I-M_1\|_{\mathcal{H}_\gamma} < \delta_\gamma$, then for the same entropy function ${\mathcal{E}}_\gamma$, the following decay estimate holds: $${\mathcal{E}}_\gamma(f(t)) \leq e^{-t/2820}{\mathcal{E}}_\gamma(f^I)\ ,\qquad t\ge0\ .$$
Note that part \[BGK-decay:a\] of this theorem generalizes Theorem \[linBGK-decay\] to the Sobolev-type entropies ${\mathcal{E}}_\gamma(f)$ in the case $d=3$, $L=2\pi$.
This paper is organized as follows: In §\[sec:P-method\] we review from [@AAC16] a Lyapunov-type method for hypocoercive ODEs that yields their sharp exponential decay rate. While this approach requires all eigenvectors of the system matrix, we also develop an approach using simplified Lyapunov functionals. This alternative strategy comes at the price of yielding only a suboptimal decay rate, but it can be extended to infinite dimensional systems and BGK equations. In §\[sec:linBGK:1D\] we apply the second strategy to the linearized BGK model in 1D, proving exponential decay of the solution towards the spatially uniform Maxwellian, as stated in Theorem \[linBGK-decay\]. This is based on decomposing into spatial Fourier modes and introducing a Hermite function basis in velocity direction. In the Sections \[sec:linBGK:2D\] and \[sec:linBGK:3D\] we extend our result to 2D and 3D, respectively. But this is not straightforward, as it is already not obvious how to choose a convenient Hermite function basis in multi dimensions. Finally, in §\[sec:6\] we prove local exponential stability of the nonlinear BGK equation in 3D as stated in Theorem \[BGK-decay\](b).
Decay of hypocoercive ODEs {#sec:P-method}
==========================
The local convergence result in Theorem \[BGK-decay\](b) is obtained from the global convergence result in Theorem \[linBGK-decay\] and a relatively straightforward control of the errors involved in linearization. Therefore, the essential content of the paper concerns the linearized BGK equations. To this end we shall rewrite them as ODEs – of infinite dimension – in fact. We therefore begin this section with a discussion of the hypocoercivity structure of ODEs and review (from [@AAC16]) a Lyapunov-type method that yields their sharp decay rate.
Lyapunov’s direct method
------------------------
To illustrate the method we start with linear, finite dimensional ODEs. Consider an ODE for a vector $f(t)=(f_1(t),\,f_2(t),\ldots,f_n(t))^\top\in {\mathbb{C}}^n$: $$\label{ODE:general:n}
\begin{cases}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}{f}= -{\mathbf{C}}f,\quad t\ge 0\,,\\
f(0) = f^I \in{\mathbb{C}}^n\,,
\end{cases}$$ for some (typically non-Hermitian) matrix ${\mathbf{C}}\in{\mathbb{C}}^{n\times n}$. The stability of the steady state $f^0\equiv 0$ is determined by the eigenvalues of matrix ${\mathbf{C}}$:
Let ${\mathbf{C}}\in{\mathbb{C}}^{n\times n}$ and let $\lambda_j$ ($j=1,\ldots,n)$ denote the eigenvalues of ${\mathbf{C}}$ (counted with their multiplicity).
1. The equilibrium $f^0$ of is stable if and only if (i) $\Re (\lambda_j) \geq 0$ for all $j=1,\ldots,n$; and (ii) all eigenvalues with $\Re(\lambda_j)=0$ are non-defective[^4].
2. The equilibrium $f^0$ of is asymptotically stable if and only if $\Re (\lambda_j)> 0$ for all $j=1,\ldots,n$.
3. The equilibrium $f^0$ of is unstable in all other cases.
For positive definite Hermitian matrices ${\mathbf{C}}$, using the Lyapunov functional $\|f\|^2$ in the energy method allows to obtain the sharp decay rate, which is the smallest eigenvalue $\mu$ of ${\mathbf{C}}$: The derivative of $\|f\|^2$ along solutions $f(t)$ of satisfies $$ {\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}\|f(t)\|^2 = -{\langle {f(t)}\ ,\, {({\mathbf{C}}^* + {\mathbf{C}})f(t)} \rangle} = -2 {\langle {f(t)}\ ,\, {{\mathbf{C}}f(t)} \rangle} \leq -2\mu \|f(t)\|^2 \,,$$ where ${\mathbf{C}}^*$ denotes the Hermitian transpose of ${\mathbf{C}}$. Note that the derivative of $\|f\|^2$ depends only on the Hermitian part $\tfrac12 ({\mathbf{C}}^* + {\mathbf{C}})$ of matrix ${\mathbf{C}}$, such that for a Hermitian matrix ${\mathbf{C}}$ there is no loss of information.
But for non-Hermitian matrices it is more natural to use a modified norm: $$\|f\|^2_{\mathbf{P}}:= \langle f,{\mathbf{P}}f\rangle\,,$$ for some positive definite Hermitian matrix ${\mathbf{P}}\in{\mathbb{C}}^{n\times n}$, to be derived from ${\mathbf{C}}$. The derivative of $\|f\|^2_{\mathbf{P}}$ along solutions $f(t)$ of satisfies $${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}\|f(t)\|^2_{\mathbf{P}}= -{\langle {f(t)}\ ,\, {({\mathbf{C}}^* {\mathbf{P}}+{\mathbf{P}}{\mathbf{C}})f(t)} \rangle} \,. $$ Then, $f^0\equiv 0$ is asymptotically stable, if there exists a positive definite Hermitian matrix ${\mathbf{P}}$ such that ${\mathbf{C}}^* {\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}$ is positive definite. To determine the decay rate to $f^0$, and to choose ${\mathbf{P}}$ conveniently we shall use the following algebraic result.
\[lemma:Pdefinition\] For any fixed matrix ${\mathbf{C}}\in{\mathbb{C}}^{n\times n}$, let $\mu:=\min\{\Re(\lambda)|\lambda$ is an eigenvalue of ${\mathbf{C}}\}$. Let $\{\lambda_{j}|1\leq j\leq j_0\}$ be all the eigenvalues of ${\mathbf{C}}$ with $\Re(\lambda_j)=\mu$, only counting their geometric multiplicity. If all $\lambda_j$ ($j=1,\dots,j_0$) are non-defective, then there exists a positive definite Hermitian matrix ${\mathbf{P}}\in{\mathbb{C}}^{n\times n}$ with $$\begin{aligned}
\label{matrixestimate1}
{\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}&\geq 2\mu {\mathbf{P}}\,.
\end{aligned}$$ But ${\mathbf{P}}$ is not uniquely determined. Moreover, if all eigenvalues of ${\mathbf{C}}$ are non-defective, such matrices ${\mathbf{P}}$ satisfying are given by $$\begin{aligned}
\label{simpleP1}
{\mathbf{P}}:= \sum\limits_{j=1}^n b_j \,\overline{w_j}\otimes w_j^\top \,,
\end{aligned}$$ where $w_j$ ($j=1,\dots,n$) denote the left eigenvectors of ${\mathbf{C}}$, and $b_j\in{\mathbb{R}}^+$ ($j=1,\dots,n$) are arbitrary weights.
1. This result was proven in [@ArEr14 Lemma 4.3] for real matrices ${\mathbf{C}}\in{\mathbb{R}}^{n\times n}$, and in [@AAC16 Lemma 2] for complex matrices ${\mathbf{C}}$. In particular, if ${\mathbf{C}}$ is a real matrix, then the inequality of Lemma \[lemma:Pdefinition\] holds true for some real positive definite symmetric matrices ${\mathbf{P}}\in{\mathbb{R}}^{n\times n}$.
2. For the extension of the above lemma to the case of defective eigenvalues see [@ArEr14 Lemma 4.3(i)] and [@AAS15 Prop. 2.2]. But the construction of ${\mathbf{P}}$ then involves also the generalized eigenvectors,
3. The Lyapunov inequality is a special case of a linear matrix inequality. In a standard reference of system and control theory [@Boyd94], the problem of finding the maximal positive constant $\mu$ and a positive definite matrix ${\mathbf{P}}$ satisfying is formulated as a generalized eigenvalue problem, see [@Boyd94 §5.1.3]. The optimal value for the constant $\mu$ is pointed out, but the associated matrices ${\mathbf{P}}$ (like in our construction ) are not specified.
Now we consider examples, where all eigenvalues of ${\mathbf{C}}\in{\mathbb{C}}^{n\times n}$ are non-defective and have positive real parts. Then the origin is the unique and asymptotically stable steady state $f^0=0$ of $\eqref{ODE:general:n}$: Due to Lemma \[lemma:Pdefinition\], there exists a positive definite Hermitian matrix ${\mathbf{P}}\in{\mathbb{C}}^{n\times n}$ such that ${\mathbf{C}}^* {\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}\geq 2\mu {\mathbf{P}}$ where $\mu = \min \Re(\lambda_j) >0$. Thus, the derivative of $\|f\|^2_{\mathbf{P}}= {\langle {f}\ ,\, {{\mathbf{P}}f} \rangle}$ along solutions of satisfies $$\begin{aligned}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}\|f(t)\|^2_{\mathbf{P}}&\leq -2\mu \|f(t)\|^2_{\mathbf{P}}\qquad \text{with} \quad \mu = \min \Re(\lambda_j),
\end{aligned}$$ which implies $$\label{ODE-decay}
\|f(t)\|^2_{\mathbf{P}}\leq e^{-2\mu t} \|f^I \|^2_{\mathbf{P}}\,,\qquad t\ge0\,.$$ Let $\lambda^{\mathbf{P}}_j$ $(j=1,\ldots,n)$ denote the positive eigenvalues of the positive definite Hermitian matrix ${\mathbf{P}}$ being ordered by magnitude such that $0 <\lambda^{\mathbf{P}}_1 \leq\ldots \leq\lambda^{\mathbf{P}}_n$. Then the matrix inequality $\lambda^{\mathbf{P}}_1 {\mathbf{I}}\leq {\mathbf{P}}\leq\lambda^{\mathbf{P}}_n {\mathbf{I}}$ implies the equivalence of norms $$\lambda^{\mathbf{P}}_1 \|v\|^2 \leq \|v\|_{\mathbf{P}}^2 \leq \lambda^{\mathbf{P}}_n \|v\|^2 \qquad \forall v\in{\mathbb{C}}^n \,.$$ Thus the decay in ${\mathbf{P}}$-norm translates into a decay in the Euclidean norm $$\label{ODE:exponentialDecay}
{\| {f(t)} \|}^2\leq c e^{-2\mu t} {\| {f^I} \|}^2\,,$$ with the constant $c=\lambda^{\mathbf{P}}_n /\lambda^{\mathbf{P}}_1 \ge1$, i.e. the condition number of ${\mathbf{P}}$. Note that $c=1$ if and only if ${\mathbf{P}}={\mathbf{I}}$.
In a popular textbook on linear systems theory [@Hes09], the exponential decay is obtained as follows [@Hes09 §8.5]: For a stable matrix $-{\mathbf{C}}$ (i.e. all eigenvalues of $-{\mathbf{C}}$ have negative real part) and a matrix ${\mathbf{Q}}$, the unique solution ${\mathbf{P}}$ of Lyapunov’s equation $$ {\mathbf{C}}^* {\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}= {\mathbf{Q}}$$ is given by $${\mathbf{P}}:= \int_0^\infty e^{-{\mathbf{C}}^* t}\ {\mathbf{Q}}\ e^{-{\mathbf{C}}t} {\,\operatorname{d}\!t}\ .$$ If ${\mathbf{Q}}$ is a positive definite symmetric matrix, then the unique solution ${\mathbf{P}}$ is also symmetric and positive definite. Moreover, the ${\mathbf{P}}$-norm of any solution $f(t)$ of satisfies $${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}\|f(t)\|^2_{\mathbf{P}}= -{\langle {f(t)}\ ,\, {({\mathbf{C}}^* {\mathbf{P}}+{\mathbf{P}}{\mathbf{C}})f(t)} \rangle}
= -{\langle {f(t)}\ ,\, {{\mathbf{Q}}f(t)} \rangle}
\leq -\min \lambda^{\mathbf{Q}}_j \|f(t)\|^2
\leq -\frac{\min \lambda^{\mathbf{Q}}_j}{\max \lambda^{\mathbf{P}}_j} \|f(t)\|^2_{\mathbf{P}}\ ,$$ where $\lambda^{\mathbf{Q}}_j$ and $\lambda^{\mathbf{P}}_j$ are the positive eigenvalues of the positive definite symmetric matrices ${\mathbf{Q}}$ and ${\mathbf{P}}$. This implies with $2\mu ={\min \lambda^{\mathbf{Q}}_j}/{\max \lambda^{\mathbf{P}}_j}$. However, only a suitable choice for ${\mathbf{Q}}$ would allow to recover the optimal decay rate as achieved in Lemma \[lemma:Pdefinition\].
The preceding discussion allows to characterize coercive and hypocoercive systems of linear ODEs (as well as matrices) according to the definition in the introduction: Equation with matrix ${\mathbf{C}}$ is *coercive*, if the Hermitian part of ${\mathbf{C}}$ is positive definite, i.e. $$ \exists\kappa>0 \quad \text{such that} \quad {\mathbf{C}}_H:=\tfrac12 ({\mathbf{C}}+{\mathbf{C}}^*) \ge \kappa {\mathbf{I}}\,.$$ In this case, the trivial energy method (i.e. multiplying by $\overline{f}(t)^\top$ and using $\|f\|^2$ as a Lyapunov functional) shows decay of $f(t)$ with rate $\kappa$ and $c=1$. But this exponential rate is not necessarily sharp, e.g. for some non-Hermitian matrices ${\mathbf{C}}$.
Equation with matrix ${\mathbf{C}}$ is *hypocoercive (with trivial kernel)*, if there exists $\mu>0$ such that all eigenvalues of ${\mathbf{C}}$ satisfy $$\Re(\lambda_j)\ge\mu\,, \qquad j=1,...,n\,.$$ While this notion was originally coined for operators in PDEs, such matrices are typically also called *positively stable*.
Comparing the spectrum of ${\mathbf{C}}$ and ${\mathbf{C}}_H$, it is well known that the maximum constants $\kappa$ and $\mu$ satisfy $\kappa\le\mu$. If all eigenvalues of ${\mathbf{C}}$ with $\Re(\lambda_j)=\mu$ are *non-defective*, then $f(t)$ decays at least with rate $\mu$. However, if ${\mathbf{C}}$ has a defective eigenvalue with $\Re(\lambda)=\mu$, then $f(t)$ decays “slightly slower”, i.e. with rate $\mu-{\varepsilon}$, for any ${\varepsilon}>0$ (see [@ArEr14 Proposition 4.5] and [@AAS15 Proposition 2.2] for details – applied to hypocoercive Fokker-Planck equations). Very recently this decay result has been improved as follows: In this case there is still a positive definite matrix ${\mathbf{P}}$, but it cannot be given by the simple formula , and becomes $$\label{defective-decay}
\|f(t)\|^2_{\mathbf{P}}\leq C (1+t^{2m}) e^{-2\mu t} \|f^I\|^2_{\mathbf{P}}$$ for some $C>0$, where $m$ is the maximal defect of the eigenvalues of ${\mathbf{C}}$ with $\Re(\lambda_j)=\mu$. See [@ArEiWo] for more information.
Index of hypocoercivity {#sec-hypo-index}
-----------------------
For the BGK models analyzed below we intend to construct convenient Lyapunov functionals of the form $\langle f,{\mathbf{P}}f\rangle$, where the matrix ${\mathbf{P}}$ does not necessarily have to reveal the sharp spectral gap of ${\mathbf{C}}$ (in the sense of Lemma \[lemma:Pdefinition\]). To this end we first give a definition of the structural complexity of a hypocoercive equation of the form $$\label{matrix-ODE}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}f +i\; {\mathbf{C}}_1 f = -{\mathbf{C}}_2 f\ ,\quad t\ge0\ .$$ Here we decomposed the matrix ${\mathbf{C}}\in{\mathbb{C}}^{n\times n}$ as ${\mathbf{C}}=i{\mathbf{C}}_1+{\mathbf{C}}_2$ with Hermitian matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ with ${\mathbf{C}}_2\ge0$. In the special case ${\mathbf{C}}_1=0$, $(\ker {\mathbf{C}}_2)^\perp$ corresponds to the subspace of decaying solutions $f(t)$, and $\ker {\mathbf{C}}_2$ to the non-decaying subspace. In hypocoercive equations, the semigroup generated by the skew-Hermitian matrix $i{\mathbf{C}}_1$ may turn non-decaying directions into decaying directions, hence allowing for an exponential decay of all solutions. More precisely, we assume $$\label{hypocoercive:2}
\exists \tau\in {\mathbb{N}}_0 \quad \text{and}\quad \exists \kappa>0 \ : \qquad \sum_{j=0}^\tau {\mathbf{C}}_1^j {\mathbf{C}}_2 ({\mathbf{C}}_1)^j \geq \kappa {\mathbf{I}}\ .$$
\[hyp-index\] For Hermitian matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ with ${\mathbf{C}}_2\ge0$, the *hypocoercivity index* of the matrix ${\mathbf{C}}$ (and of the ODE ) is the smallest $\tau\in{\mathbb{N}}_0$, such that holds.
Clearly, $\tau=0$ corresponds to coercive matrices ${\mathbf{C}}$; i.e., those for which all eigenvalues of its Hermitian part $\tfrac12({\mathbf{C}}+{\mathbf{C}}^*)$ are strictly positive. A simple computation shows that this definition is invariant under a change of basis. We note that condition is identical to the matrix condition in Lemma 2.3 of [@ArEr14], which characterizes the hypoellipticity of degenerate Fokker-Planck operators of the form ${\mathbf{L}}f=\operatorname{div}({\mathbf{D}}\nabla f+{\mathbf{C}}xf)$ (using the matrix correspondence ${\mathbf{D}}={\mathbf{C}}_2$, ${\mathbf{C}}={\mathbf{C}}_1$). Hence, condition for the ODE and its hypocoercivity index can be seen as an analogue of the *finite rank Hörmander condition* for hypoelliptic and degenerate diffusion equations [@Ho67 Th. 1.1]. While the hypocoercivity index of degenerate parabolic equations determines the algebraic regularization rate (e.g. from $L^2$ into $H^1$, see Theorem A.12 in [@ViH06] and Theorem 4.8 in [@ArEr14]), its role in hypocoercive ODEs is not yet clear.
### Equivalent hypocoercivity conditions
Next, we collect several statements which are equivalent to condition . They will be useful for the analysis in §\[sec:Pmatrix\].
\[prop:equivalence\] Suppose that ${\mathbf{C}}_1\in{\mathbb{C}}^{n\times n}$ and ${\mathbf{C}}_2\in{\mathbb{C}}^{n\times n}$ are Hermitian matrices. Suppose furthermore that ${\mathbf{C}}_2$ is positive semi-definite. Then the following conditions are equivalent:
1. \[cond:fullrank\] There exists $\tau\in{\mathbb{N}}_0$ such that $$\operatorname{rank}\{\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}, \ldots, {\mathbf{C}}_1^\tau \sqrt{{\mathbf{C}}_2}\} =n \ ,$$ which is often called *Kalman rank condition*.
2. \[cond:hypocoercive\] The matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ satisfy condition .
3. \[cond:invariance\] No non-trivial subspace of $\ker {\mathbf{C}}_2$ is invariant under ${\mathbf{C}}_1$.
4. \[cond:eigenvector\] No eigenvector of ${\mathbf{C}}_1$ lies in the kernel of ${\mathbf{C}}_2$.
5. \[cond:compMatrix\] There exists a skew-Hermitian matrix ${\mathbf{K}}$ such that ${\mathbf{C}}_2 +[{\mathbf{K}},{\mathbf{C}}_1] ={\mathbf{C}}_2 +({\mathbf{K}}{\mathbf{C}}_1 -{\mathbf{C}}_1{\mathbf{K}})$ is positive definite.
Moreover, the smallest possible $\tau$ in \[cond:fullrank\] and \[cond:hypocoercive\] coincides; it is the hypocoercivity index of ${\mathbf{C}}$.
The equivalence of \[cond:fullrank\] and \[cond:hypocoercive\] (with the same $\tau$) follows from [@AAS15 Lemma 2.3]. The equivalence of \[cond:hypocoercive\]–\[cond:eigenvector\] follows from [@ArEr14 Lemma 2.3]. The equivalence of \[cond:eigenvector\] and \[cond:compMatrix\] follows by the same arguments as for real symmetric matrices in [@ShKa85 Theorem 2.5].
\[rem:hypo:cond:B1’\]
1. In order to use condition \[cond:fullrank\] later on also for “infinite matrices” we give here an equivalent version:
1. \[cond:trivkernel\] There exists $\tau\in{\mathbb{N}}_0$ such that $\bigcap_{j=0}^\tau \ker\big(\sqrt{{\mathbf{C}}_2}{\mathbf{C}}_1^j\big) =\{0\}$.
2. If $\tau\in{\mathbb{N}}_0$ is such that $$\label{rank-tau1}
\operatorname{rank}\{\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}, \ldots, {\mathbf{C}}_1^\tau \sqrt{{\mathbf{C}}_2}\} = \operatorname{rank}\{\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}, \ldots, {\mathbf{C}}_1^\tau \sqrt{{\mathbf{C}}_2}, {\mathbf{C}}_1^{\tau+1} \sqrt{{\mathbf{C}}_2}\} \ ,$$ then for all $k\in{\mathbb{N}}$ $$\operatorname{rank}\{\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}, \ldots, {\mathbf{C}}_1^\tau \sqrt{{\mathbf{C}}_2}\} = \operatorname{rank}\{\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}, \ldots, {\mathbf{C}}_1^{\tau+k} \sqrt{{\mathbf{C}}_2}\} \,.$$ Condition implies that the columns of ${\mathbf{C}}_1^{\tau+1} \sqrt{{\mathbf{C}}_2}$ are linear combinations of the columns of ${\mathbf{C}}_1^j \sqrt{{\mathbf{C}}_2}$, $j\in\{0,\ldots,\tau\}$. This implies that ${\mathbf{C}}_1^{\tau+k} \sqrt{{\mathbf{C}}_2}$ are linear combinations of the columns of ${\mathbf{C}}_1^j \sqrt{{\mathbf{C}}_2}$, $j\in\{k-1,\ldots,\tau+k-1\}$. Hence, for a hypocoercive matrix we have to gain with each added term in at least one rank until we reach full rank, i.e. space dimension $n$. Thus, for hypocoercive matrices its hypocoercivity index is bounded from above by the dimension of $\ker{\mathbf{C}}_2$ (or equivalently corank of ${\mathbf{C}}_2$).
In [@ViH06 Remark 17] the connections of the above conditions to *Kawashima’s nondegeneracy condition* for the study of degenerate hyperbolic-parabolic systems [@Ka87] and *Hörmander’s rank condition* for hypoelliptic equations [@Ho67] are noted.
For real symmetric matrices ${\mathbf{C}}_1, {\mathbf{C}}_2\in {\mathbb{R}}^{n\times n}$ with ${\mathbf{C}}_2\geq 0$, condition \[cond:eigenvector\] is equivalent to the condition that ${\mathbf{C}}:=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ has only eigenvalues with positive real part, see [@ShKa85 Theorem 1.1]. And the latter statement is equivalent to the exponential stability of . Using Proposition \[prop:equivalence\], we shall now prove a similar statement for Hermitian matrices:
\[lem:hypocoercive\] Hermitian matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ with ${\mathbf{C}}_2\ge0$ satisfy condition if and only if all eigenvalues $\lambda_{{\mathbf{C}}}$ of ${\mathbf{C}}:=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ have positive real part $\Re (\lambda_{{\mathbf{C}}}) >0$.
To show Lemma \[lem:hypocoercive\] for Hermitian matrices, we will follow the proofs of [@UmKaSh84 Prop. 2.4] and [@ShKa85 Lemma 3.2] for real symmetric matrices.
First, we show that condition implies that all eigenvalues $\lambda_{{\mathbf{C}}}$ of ${\mathbf{C}}:=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ have positive real part $\Re (\lambda_{{\mathbf{C}}}) >0$: Let $\phi$ be an eigenvector of ${\mathbf{C}}$ corresponding to an eigenvalue $\lambda$, i.e. $$\label{eq:EVP}
\lambda\phi = {\mathbf{C}}\phi = (i{\mathbf{C}}_1 +{\mathbf{C}}_2) \phi \ .$$ Take the complex inner product of this equation with $\phi$, to obtain $$\overline{\lambda} \langle\phi,\phi\rangle = \langle {\mathbf{C}}\phi, \phi\rangle \ ,$$ using $\langle\phi,\psi\rangle = \overline{\phi}^\top \psi$ for all $\phi,\psi\in{\mathbb{C}}^n$. Its real part satisfies $$\label{eq:Re:lambda}
\Re (\lambda) \langle\phi,\phi\rangle = \langle{\mathbf{C}}_2 \phi,\phi\rangle \ ,$$ due to the assumptions on the matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$. Moreover, there exists a skew-Hermitian matrix ${\mathbf{K}}$ such that ${\mathbf{C}}_2 +[{\mathbf{K}},{\mathbf{C}}_1]$ is positive definite by Proposition \[prop:equivalence\]. We multiply equation with $i{\mathbf{K}}$ and take the inner product with $\phi$ such that $$\overline{\lambda} \langle i{\mathbf{K}}\phi,\phi\rangle = \langle i{\mathbf{K}}{\mathbf{C}}\phi,\phi\rangle \ .$$ Its real part satisfies $$\label{eq:Re:lambda:2}
2\Re (\lambda) \langle i{\mathbf{K}}\phi,\phi\rangle = \langle({\mathbf{C}}_1 {\mathbf{K}}-{\mathbf{K}}{\mathbf{C}}_1)\phi,\phi\rangle - i\langle({\mathbf{C}}_2 {\mathbf{K}}+{\mathbf{K}}{\mathbf{C}}_2)\phi,\phi\rangle \ ,$$ since ${\mathbf{C}}_1$, ${\mathbf{C}}_2$ and $i{\mathbf{K}}$ are Hermitian matrices. Moreover, $$\begin{gathered}
\label{est:1}
2\Re (\langle i{\mathbf{K}}{\mathbf{C}}_2 \phi,\phi\rangle)
= \langle({\mathbf{C}}_2 \ i{\mathbf{K}}+ i{\mathbf{K}}{\mathbf{C}}_2)\phi,\phi\rangle
= \langle \sqrt{{\mathbf{C}}_2}\ i{\mathbf{K}}\phi,\sqrt{{\mathbf{C}}_2}\phi\rangle +\langle \sqrt{{\mathbf{C}}_2}\phi,\sqrt{{\mathbf{C}}_2}\ i{\mathbf{K}}\phi\rangle \\
\leq 2{\| {\sqrt{{\mathbf{C}}_2}\phi} \|} {\| {\sqrt{{\mathbf{C}}_2}\ i{\mathbf{K}}\phi} \|}
\leq 2M {\| {\sqrt{{\mathbf{C}}_2}\phi} \|} {\| {\phi} \|}
\leq \epsilon {\| {\phi} \|}^2 +\frac{M^2}{\epsilon} \langle{\mathbf{C}}_2 \phi,\phi\rangle
\end{gathered}$$ for any positive $\epsilon$. Here we used $M:=\|\sqrt{{\mathbf{C}}_2}\ i{\mathbf{K}}\|$ and ${\| {\sqrt{{\mathbf{C}}_2}\phi} \|}^2 = \langle{\mathbf{C}}_2 \phi,\phi\rangle$ since ${\mathbf{C}}_2\geq 0$. Combining equations and as $2\cdot$$-\alpha\cdot$ for some constant $\alpha>0$ to be chosen later, we derive $$\label{eq1}
2\Re (\lambda) \big({\| {\phi} \|}^2 -\alpha \langle i{\mathbf{K}}\phi,\phi\rangle \big)
=\langle ({\mathbf{C}}_2 +\alpha ({\mathbf{K}}{\mathbf{C}}_1 -{\mathbf{C}}_1 {\mathbf{K}}))\phi,\phi\rangle +\langle{\mathbf{C}}_2 \phi,\phi\rangle
+i\alpha \langle({\mathbf{C}}_2 {\mathbf{K}}+{\mathbf{K}}{\mathbf{C}}_2)\phi,\phi\rangle \ .$$ There exists $\alpha_0>0$ such that $\Phi_\alpha := {\| {\phi} \|}^2 -\alpha \langle i{\mathbf{K}}\phi,\phi\rangle$ satisfies $$\label{eq2}
{\| {\phi} \|}^2/2\leq \Phi_\alpha\leq 2{\| {\phi} \|}^2\quad \forall\; \alpha\in(-\alpha_0,\alpha_0)\,,$$ since $i{\mathbf{K}}$ is a Hermitian matrix. Recall that the skew-Hermitian matrix ${\mathbf{K}}$ was chosen such that ${\mathbf{C}}_2 +[{\mathbf{K}},{\mathbf{C}}_1]$ is positive definite by Proposition \[prop:equivalence\]. Therefore, the estimate $$\label{eq3}
\langle ({\mathbf{C}}_2 +\alpha ({\mathbf{K}}{\mathbf{C}}_1-{\mathbf{C}}_1 {\mathbf{K}}))\phi,\phi\rangle \geq \alpha m {\| {\phi} \|}^2$$ holds for all $\alpha\in[0,1]$, where $m>0$ is the smallest eigenvalue of the positive definite Hermitian matrix ${\mathbf{C}}_2 +({\mathbf{K}}{\mathbf{C}}_1-{\mathbf{C}}_1 {\mathbf{K}})$. Thus we deduce from and the estimates , that $$2\Re (\lambda) \Phi_\alpha \geq \alpha (m -\epsilon) {\| {\phi} \|}^2 +(1 -\alpha\frac{M^2}{\epsilon}) \langle{\mathbf{C}}_2 \phi,\phi\rangle \ .$$ Choosing $\epsilon =m/2$ and $\alpha =\min\{1,\alpha_0,\epsilon/M^2\}$, we finally derive with $$\Re (\lambda) \geq \frac{\alpha m}8 >0 \ .$$
Finally, we show the reverse implication via a proof of its negation. If condition \[cond:eigenvector\] does not hold, then there exists a $\phi\in\ker {\mathbf{C}}_2$ and an (eigenvalue) $\mu\in{\mathbb{R}}$ such that ${\mathbf{C}}_1 \phi =\mu \phi$. This implies $(i{\mathbf{C}}_1 +{\mathbf{C}}_2)\phi =i\mu \phi$. Thus $\phi$ is an eigenvector of ${\mathbf{C}}:=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ for the purely imaginary eigenvalue $i\mu$. Thus not all eigenvalues $\lambda_{{\mathbf{C}}}$ of ${\mathbf{C}}$ have positive real part.
We conclude that, if all eigenvalues $\lambda_{{\mathbf{C}}}$ of ${\mathbf{C}}$ have positive real part $\Re (\lambda_{{\mathbf{C}}}) >0$, then condition \[cond:eigenvector\] – and equivalently – must hold.
\[kdep-estimate\] In the study of hypocoercivity for discrete velocity BGK models, a family of matrices ${\mathbf{C}}^{(k)} := ik\ {\mathbf{C}}_1 +{\mathbf{C}}_2$ $(k\in{\mathbb{N}})$ for some real symmetric matrices ${\mathbf{C}}_1, {\mathbf{C}}_2\in {\mathbb{R}}^{n\times n}$ with ${\mathbf{C}}_2\geq 0$ has to be considered, see [@AAC16 §4.1-§4.2]. Following the proof of [@UmKaSh84 Prop. 2.4], a uniform bound for the real parts of the eigenvalues $\lambda_{{\mathbf{C}}^{(k)}}$ of these matrices ${\mathbf{C}}^{(k)}$ ($k\in{\mathbb{N}}$) can be proven: $$\Re (\lambda_{{\mathbf{C}}^{(k)}}) \geq \frac{\alpha m}8 \frac{k^2}{1+k^2} >0 \qquad \forall k\in{\mathbb{N}}\ .$$
Next we relate our study of equation to the one of ${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}f + {\mathbf{L}}f =0$ in [@ViH06]. In the first part of [@ViH06], operators ${\mathbf{L}}={\mathbf{A}}^* {\mathbf{A}}+ {\mathbf{B}}$ with a skew-symmetric operator ${\mathbf{B}}$ are considered. Our operator/matrix ${\mathbf{C}}= i{\mathbf{C}}_1 +{\mathbf{C}}_2$ (for some Hermitian matrices ${\mathbf{C}}_1,{\mathbf{C}}_2 \in{\mathbb{C}}^{n\times n}$ with ${\mathbf{C}}_2\geq 0$) is of the form ${\mathbf{L}}={\mathbf{A}}^* {\mathbf{A}}+ {\mathbf{B}}$ for the choice ${\mathbf{A}}=\sqrt{{\mathbf{C}}_2}$ and ${\mathbf{B}}=i{\mathbf{C}}_1$ acting on the complex Hilbert space ${\mathbb{C}}^n$. First, we notice that ${\mathcal{K}}:=\ker {\mathbf{L}}=\ker{\mathbf{A}}\cap \ker{\mathbf{B}}$, see [@ViH06 Prop. I.2]. There, the study of hypocoercivity is based on the assumptions [@ViH06 (3.4)–(3.5)]: $$\label{cond:Villani:3.4}
\exists \tau\in{\mathbb{N}}_0 \ : \quad \ker\big(\sum_{k=0}^\tau {\mathbf{D}}_k^* {\mathbf{D}}_k \big) =\ker {\mathbf{L}}=:{\mathcal{K}}\ ,$$ or more clearly, $$\label{cond:Villani:3.5}
\exists \tau\in{\mathbb{N}}_0 \ : \quad \sum_{k=0}^\tau {\mathbf{D}}_k^* {\mathbf{D}}_k \quad\text{is coercive on } {\mathcal{K}}^\perp \ ,$$ where the iterated commutators ${\mathbf{D}}_k$ ($k\in{\mathbb{N}}_0$) are defined recursively as $$ {\mathbf{D}}_0 :={\mathbf{A}}\ , \qquad {\mathbf{D}}_k := [{\mathbf{D}}_{k-1},{\mathbf{B}}] = {\mathbf{D}}_{k-1}{\mathbf{B}}-{\mathbf{B}}{\mathbf{D}}_{k-1} \ , \quad k\in{\mathbb{N}}\ .$$ In [@ViH06 Remark 17] it is noted (without a proof) that on finite dimensional Hilbert spaces, condition is equivalent to \[cond:invariance\] in Proposition \[prop:equivalence\] (with credit to Denis Serre).
The following simple example shows that this “equivalence” needs a small modification in complex Hilbert spaces: Consider the matrices $${\mathbf{A}}= \begin{pmatrix} 1 & 0 \\ 0 & 0 \end{pmatrix}, \quad
{\mathbf{B}}= i{\mathbf{A}}= \begin{pmatrix} i & 0 \\ 0 & 0 \end{pmatrix}.$$ Matrix ${\mathbf{A}}$ has kernel $\ker{\mathbf{A}}= {\operatorname{span}}\{\binom{0}{1}\}$. Moreover ${\mathbf{D}}_0 ={\mathbf{A}}$ and ${\mathbf{D}}_k = {\mathbf{0}}$ for all $k\in{\mathbb{N}}$. Hence, ${\mathcal{K}}=\ker{\mathbf{A}}\cap\ker{\mathbf{B}}=\ker{\mathbf{A}}$ and conditions and are satisfied for all $\tau\in{\mathbb{N}}_0$. But \[cond:invariance\] does *not* hold.
Now we give a proof of a slightly modified equivalence. On finite dimensional Hilbert spaces, conditions and are obviously equivalent. Moreover, we will make use of Proposition \[prop:equivalence\] and only show the equivalence of \[cond:fullrank\] and a modified :
Let the matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2\ge0$ be Hermitian and define ${\mathbf{A}}: =\sqrt{{\mathbf{C}}_2}$ and ${\mathbf{B}}:=i{\mathbf{C}}_1$. Then $({\mathbf{A}},{\mathbf{B}})$ satisfies $$\label{cond:Villani:3.4:modified} \text{condition~\eqref{cond:Villani:3.4} together with $\ker{\mathbf{A}}\cap\ker{\mathbf{B}}=\{0\}$}$$ if and only if $({\mathbf{C}}_1,{\mathbf{C}}_2)$ satisfies \[cond:fullrank\]. Moreover, the smallest possible $\tau$ in and \[cond:fullrank\] coincides.
First, notice that for all $\tau\in{\mathbb{N}}_0$ $$\label{eq4}
\bigcap_{k=0}^\tau \ker {\mathbf{D}}_k = \ker \sum_{k=0}^\tau {\mathbf{D}}_k^* {\mathbf{D}}_k \ .$$ Defining ${\mathcal{K}}':=\bigcap_{k\geq 0} \ker {\mathbf{D}}_k$, the inclusion ${\mathcal{K}}\subset {\mathcal{K}}'$ is proven in [@ViH06 Prop I.15]. Next we prove that $$\label{Villani+fullrank}
w\in \bigcap_{k=0}^\tau \ker {\mathbf{D}}_k =: {\mathcal{K}}'_\tau \qquad \text{is equivalent to} \qquad {\mathbf{A}}{\mathbf{B}}^k w = 0 \quad \forall k\in\{0,\ldots,\tau\}$$ by induction: For $\tau=0$, $w\in \ker {\mathbf{D}}_0 =\ker {\mathbf{A}}$ holds. Assume now condition for $\tau$ and prove it for $\tau+1$. Operator ${\mathbf{D}}_{\tau+1}$ is defined as ${\mathbf{D}}_{\tau+1} = [{\mathbf{D}}_\tau,{\mathbf{B}}] = {\mathbf{D}}_\tau {\mathbf{B}}-{\mathbf{B}}{\mathbf{D}}_\tau$ and using $w\in\bigcap_{k=0}^{\tau+1} \ker {\mathbf{D}}_k$ yields $$ \begin{split}
0 &={\mathbf{D}}_{\tau+1}w = [{\mathbf{D}}_\tau,{\mathbf{B}}]w = ({\mathbf{D}}_\tau {\mathbf{B}}-{\mathbf{B}}{\mathbf{D}}_\tau)w = {\mathbf{D}}_\tau {\mathbf{B}}w \\
&=({\mathbf{D}}_{\tau-1}{\mathbf{B}}-{\mathbf{B}}{\mathbf{D}}_{\tau-1}){\mathbf{B}}w = {\mathbf{D}}_{\tau-1} {\mathbf{B}}^2 w - {\mathbf{B}}^2 {\mathbf{D}}_{\tau-1} w
={\mathbf{D}}_{\tau-1}{\mathbf{B}}^2 w = \ldots = {\mathbf{D}}_0 {\mathbf{B}}^{\tau+1} w = {\mathbf{A}}{\mathbf{B}}^{\tau+1} w \ .
\end{split}$$ The converse, $0={\mathbf{A}}{\mathbf{B}}^{\tau+1} w={\mathbf{D}}_{\tau+1}w$, is proven similarly. Thus the equivalence holds.
Finally we prove the equivalence of \[cond:fullrank\] and : If condition \[cond:fullrank\] holds for one $\tau_0$, then ${\mathbf{A}}{\mathbf{B}}^k w = 0$ for all $k\in\{0,\ldots,\tau_0\}$ implies $w=0$. Due to the equivalence in , ${\mathcal{K}}'\subset{\mathcal{K}}'_{\tau_0} =\{0\}$. Hence, $\{0\}\subset{\mathcal{K}}\subset{\mathcal{K}}'_{\tau_0}=\{0\}$. With this proves condition with $\tau=\tau_0$ and $\ker{\mathbf{L}}=\ker{\mathbf{A}}\cap\ker{\mathbf{B}}=\{0\}$.
If condition holds together with $\ker{\mathbf{A}}\cap\ker{\mathbf{B}}=\{0\}$, then ${\mathcal{K}}=\ker{\mathbf{L}}=\ker{\mathbf{A}}\cap\ker{\mathbf{B}}=\{0\}$ and there exists $\tau\in{\mathbb{N}}_0$ such that $\ker\big(\sum_{k=0}^\tau {\mathbf{D}}_k^* {\mathbf{D}}_k \big) =\ker {\mathbf{L}}={\mathcal{K}}$. Due to , $$\bigcap_{k=0}^\tau \ker {\mathbf{D}}_k =\ker \sum_{k=0}^\tau {\mathbf{D}}_k^* {\mathbf{D}}_k =\ker {\mathbf{L}}={\mathcal{K}}=\{0\} \ .$$ From the equivalence we then obtain: If some $w\in{\mathbb{C}}^n$ satisfies ${\mathbf{A}}{\mathbf{B}}^k w=0$ for all $k\in\{0,\ldots,\tau\}$, it follows that $w=0$. Therefore condition \[cond:fullrank\] holds with the same index $\tau$. This finishes the proof.
Ansatz for the transformation matrix ${\mathbf{P}}$ {#sec:Pmatrix}
---------------------------------------------------
For finite dimensional matrices with non-defective eigenvalues, an *optimal* transformation matrix ${\mathbf{P}}$ (yielding the sharp spectral gap and thus the sharp decay rate) can be constructed as stated in Lemma \[lemma:Pdefinition\]. But for “infinite matrices” the eigenfunctions $w_j$ will not be known in general. Hence, an optimal matrix ${\mathbf{P}}$ cannot be obtained from formula . Even for finite dimensional systems with $n$ large, it may not be possible to explicitly construct the matrix ${\mathbf{P}}$ defined in . However, Lemma \[lemma:Pdefinition\] still provides a guide to the construction of a non-optimal choice of ${\mathbf{P}}$ that can still be used to prove hypocoercivity and to give a quantitative decay rate. We shall exploit this in §\[sec:linBGK:1D\]–\[sec:6\] to prove hypocoercivity for BGK equations. To this end we shall only consider *minimal* matrices ${\mathbf{P}}$, i.e. matrices with a minimal number of non-zero entries in ${\mathbf{P}}-{\mathbf{I}}$, such that Lemma \[lemma:Pdefinition\] still allows to deduce hypocoercivity (but then with a suboptimal rate $\mu$).
Our focus will be to find a usable and simple ansatz for ${\mathbf{P}}$ and to prove that such an ansatz will give rise to a matrix inequality of the form . The structure of these ansatzes shall be derived from the *connectivity structure* of the matrix ${\mathbf{C}}$: We consider examples of equations , where we assume w.l.o.g. that the Hermitian matrix ${\mathbf{C}}_2$ is diagonal and hence real. Next we consider how the zero and negative diagonal elements of $-{\mathbf{C}}_2$ (or equivalently the non-decaying and decaying eigenmodes of ${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}f = -{\mathbf{C}}_2 f$) are coupled via a (non-zero) off-diagonal pair in the Hermitian matrix ${\mathbf{C}}_1$. More precisely, a non-zero off-diagonal element of ${\mathbf{C}}_1$ at $j,k$ (and hence also at $k,j$) couples, in the evolution equation, the $j$-th mode of ${\mathbf{C}}_2$ to its $k$-th mode (or diagonal element). In the sequel we shall use a simple graphical representation of such connections: there the dots $\circ$ and $\bullet$ represent, respectively, zero and negative diagonal elements of $-{\mathbf{C}}_2$, and an arrow between such dots represents their connection (or coupling). For each zero element in the diagonal of ${\mathbf{C}}_2$, we next consider a *shortest connection graph* to a non-zero element in $\operatorname{diag}({\mathbf{C}}_2)$ – realized by a sequence of non-zero off-diagonal elements of ${\mathbf{C}}_1$. This leads to a guideline to find a simple ansatz for a minimal transformation matrix of the form ${\mathbf{P}}={\mathbf{I}}+{\mathbf{A}}$: The ansatz parameters of the Hermitian matrix ${\mathbf{A}}\in{\mathbb{C}}^{n\times n}$ should be put at the positions of the non-zero off-diagonal coupling elements of ${\mathbf{C}}_1$ that are needed to establish the shortest connection graphs – choosing only one graph per zero element in $\operatorname{diag}({\mathbf{C}}_2)$.
Next we shall list some hypocoercive cases with low dimensionality of $\ker{\mathbf{C}}_2$, because these are the most important cases in kinetic equations (as discussed in §\[sec:linBGK:1D\]–\[sec:6\]). For those cases we shall then prove that the above mentioned ansatzes indeed allow to establish a spectral gap of ${\mathbf{C}}$.
### Hypocoercive matrix with $\dim(\ker{\mathbf{C}}_2)=1$
In this situation there exists only one (structurally relevant) case. For to be hypocoercive, the only zero element of the diagonal of ${\mathbf{C}}_2$ (w.l.o.g. say with index $j=1$) needs to be coupled (via ${\mathbf{C}}_1$) to a positive element of the diagonal of ${\mathbf{C}}_2$ . Due to our assumptions, $${\mathbf{C}}_2=\operatorname{diag}\{0,c_2,\ldots,c_n\} \quad \text{with }c_j>0;\,j=2,...,n;\quad \text{and} \quad {\mathbf{C}}_1=(c_{j,k})_{j,k\in\{1,\ldots,n\}} \,.$$ The matrix ${\mathbf{C}}=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ is hypocoercive if and only if \[cond:invariance\] holds. Since $\ker{\mathbf{C}}_2 ={\operatorname{span}}\{e_1\}$, Condition \[cond:invariance\] reads here ${\mathbf{C}}_1 e_1\not\in {\operatorname{span}}\{e_1\}$. Thus, we conclude from ${\mathbf{C}}_1 e_1 =(c_{1,1},\ldots,c_{n,1})^\top$ that $c_{j,1}\ne 0$ for some $j\in\{2,\ldots,n\}$. Of course, $j$ does not have to be unique, but we now fix one such index $j_0$. This means that $c_{1,j_0}= \overline{c_{j_0,1}}\ne 0$. In this case the hypocoercivity index is always $1$, since Remark \[rem:hypo:cond:B1’\](b) yields here that the hypocoercivity index is less or equal $\dim(\ker {\mathbf{C}}_2)=1$. W.l.o.g. we assume $j_0=2$. The coupling within the relevant $2\times2$-subspace (i.e. the upper left $2\times2$ block of the matrix ${\mathbf{C}}$) can then be symbolized as $\circ\!\!\longrightarrow\!\!\bullet$ . Such an example was analyzed in §4.3 of [@AAC16] (representing a linear BGK equation in 1D) using a transformation matrix with the ansatz $$\label{ansatz:Pmatrix:1D}
\sbox0{$\begin{matrix}0 & \lambda \\ \bar\lambda & 0\end{matrix}$}
{\mathbf{P}}={\mathbf{I}}+ \left(\begin{array}{c|c}
\usebox{0} & {\bf 0}\\
\hline
{\bf 0 }& {\bf 0}
\end{array}\right)\ ,$$ for some $\lambda\in{\mathbb{C}}$. Here, ${\mathbf{P}}$ and ${\mathbf{I}}$ are square matrices of the same size as ${\mathbf{C}}$, possibly even infinite. The second matrix on the r.h.s. has the same size, but only its upper left $2\times2$ block is non-zero.
While the above transformation matrix ${\mathbf{P}}$ is not optimal, this approach is important in practice: in theory, Lemma \[lemma:Pdefinition\] provides the optimal transformation matrix ${\mathbf{P}}$ to deduce the optimal ODE-decay or . But in practice, its computation is tedious, particularly when the system matrix involves a parameter, which is the case for the BGK-models to be analyzed below (cf. Remark \[kdep-estimate\]). For large systems, there is therefore a need to design a method that does not require all eigenvectors, even if the resulting decay rates are then sub-optimal. For the case $\dim(\ker {\mathbf{C}}_2)=1$, an approximate transformation matrix ${\mathbf{P}}$ of the simple structure is sufficient, and it *always* allows to prove an explicit exponential decay of the ODE : the following theorem shows that ${\mathbf{C}}$ and ${\mathbf{P}}$ satisfy a matrix inequality of form , but not necessarily with the optimal constant $\mu$. Moreover, it shows that the ansatz from §4.3 of [@AAC16] was not a “wild guess” but rather a systematic approach.\
\[lemma:ansatzP:1D\] Let ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ be Hermitian matrices with ${\mathbf{C}}_2\ge0$, $\dim(\ker {\mathbf{C}}_2)=1$ such that ${\mathbf{C}}:=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ is hypocoercive. For $|\lambda|<1$ the Hermitian matrix ${\mathbf{P}}$ in is positive definite. If a sufficiently small $\lambda\in{\mathbb{C}}$ is chosen such that $\Im(\overline{\lambda}c_{1,2})>0$, then the Hermitian matrix ${\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}$ is also positive definite.
We set ${\mathbf{P}}={\mathbf{I}}+r{\mathbf{A}}$ with $$\lambda =r e^{i\phi} \quad \text{and} \quad \sbox0{$\begin{matrix}0 & e^{i\phi} \\ e^{-i\phi} & 0\end{matrix}$}
{\mathbf{A}}= \left(\begin{array}{c|c}
\usebox{0} & {\bf 0}\\
\hline
{\bf 0 }& {\bf 0}
\end{array}\right)\ .$$ Then we consider ${\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}= 2{\mathbf{C}}_2 +r({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}})$ as a perturbation of the matrix $2{\mathbf{C}}_2$ for sufficiently small $r\geq 0$. In particular, zero is a simple eigenvalue of ${\mathbf{C}}_2$ with eigenvector $e_1$. For small $r\geq 0$, the eigenvalues of ${\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}$ are close to the eigenvalues of $2{\mathbf{C}}_2$. Therefore, we only need to study the evolution of the zero eigenvalue w.r.t. $r$. Due to [@HoJo13 Thm. 6.3.12], the lowest eigenvalue $\mu(r)$ is a continuous function satisfying $\lim_{r\to 0} \mu(r)=0$. Moreover, it is differentiable at $r=0$ with $$\begin{gathered}
\frac{{\,\operatorname{d}\!\mu}}{{\,\operatorname{d}\!r}} \big|_{r=0} = \frac{e_1^* ({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}) e_1}{e_1^* e_1} = ({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}})_{1,1}
= -i e^{-i\phi} c_{1,2} +i (e^{-i\phi} c_{1,2})^* = 2\Im (e^{-i\phi} c_{1,2}) \,.
\end{gathered}$$ Due to our assumptions, $c_{1,2}\ne 0$. Hence, we can choose $\phi$ such that $\Im (e^{-i\phi} c_{1,2})$ is positive. For such a choice, the smallest eigenvalue $\mu(r)$ of $2{\mathbf{C}}_2 +r({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}})$ will be positive. This finishes the proof.
### Hypocoercive matrix with $\dim(\ker{\mathbf{C}}_2)=2$ {#sec:dim-ker2}
Up to a change in basis of ${\mathbb{C}}^n$, we consider the Hermitian matrices $$\label{setup:DimKerC2:2}
{\mathbf{C}}_2=\operatorname{diag}\{0,0,c_3,\ldots,c_n\}\geq 0 \quad \text{and} \quad
{\mathbf{C}}_1=(c_{j,k})_{j,k\in\{1,\ldots,n\}}\in{\mathbb{C}}^{n\times n}$$ such that $c_j>0$ for $j\geq 3$ and $c_{j,j}\in{\mathbb{R}}$ for all $j\in\{1,\ldots,n\}$. We only consider hypocoercive matrices ${\mathbf{C}}=i{\mathbf{C}}_1 +{\mathbf{C}}_2$. Then, ${\mathbf{C}}_1$ cannot have a block-diagonal structure of partition size $(2,n-2)$ as, otherwise, the kernel of ${\mathbf{C}}_2$ would be invariant under ${\mathbf{C}}_1$ in contradiction to condition \[cond:invariance\]. Hence, we shall assume in the sequel w.l.o.g. that $c_{2,3}\ne0$.
In order to construct (later on) appropriate transformation matrices ${\mathbf{P}}$ we shall distinguish two cases depending on the rank of the upper right submatrix ${\mathbf{C}}_1^{ur} =(c_{j,k})_{j\in\{1,2\},\ k\in\{3,\ldots,n\}}$ of ${\mathbf{C}}_1$. These cases with appropriate ansatz for the matrix ${\mathbf{P}}$ are summarized in Table \[table:ansatzP\].\
$$\begin{aligned}
(2A) &\qquad
\sbox0{$\begin{matrix}\ast & \ast \\ \ast & \ast \end{matrix}$}
\sbox1{$\begin{matrix}\ast & \bullet & \ast & \cdots & \ast \\ \bullet & \ast & \ast & \cdots & \ast \end{matrix}$}
\sbox2{$\begin{matrix}\ast & \bullet \\ \bullet & \ast \\ \ast & \ast \\ \vdots & \vdots \\ \ast & \ast \end{matrix}$}
{\mathbf{C}}_1 = \left(\begin{array}{c|c}
\usebox{0}& \usebox{1}\\
\hline
\usebox{2}& \mbox{\bf *}
\end{array}\right)\,,
&&\quad
\sbox0{$\begin{matrix}0 & 0 & 0 & \lambda_1 \\ 0 & 0 & \lambda_2 & 0 \\ 0 & \overline{\lambda_2} & 0 & 0 \\ \overline{\lambda_1} & 0 & 0 & 0 \end{matrix}$}
{\mathbf{P}}={\mathbf{I}}+ \left(\begin{array}{c|c}
\usebox{0}& {\bf 0}\\
\hline
{\bf 0}& {\bf 0}
\end{array}\right)\,, \label{P-ansatz-2A} \intertext{\hspace{19mm} where the upper right submatrix ${\mathbf{C}}_1^{ur}\in{\mathbb{C}}^{2\times(n-2)}$ has rank 2. Here, we assume w.l.o.g. that \smallskip\newline
\hspace*{19mm} $|c_{1,4}c_{2,3}|\geq |c_{1,3}c_{2,4}|$ and $c_{1,4}\,c_{2,3} \ne c_{1,3}\,c_{2,4}$, such that $c_{2,3}\ne 0$ and $c_{1,4}\ne 0$.}
(2B) &\qquad
\sbox0{$\begin{matrix}\ast & \ast \\ \ast & \ast \end{matrix}$}
\sbox1{$\begin{matrix}\ast & \ast & \cdots & \ast \\ \bullet & \ast & \cdots & \ast \end{matrix}$}
\sbox2{$\begin{matrix}\ast & \bullet \\ \ast & \ast \\ \vdots & \vdots \\ \ast & \ast \end{matrix}$}
{\mathbf{C}}_1 = \left(\begin{array}{c|c}
\usebox{0}& \usebox{1}\\
\hline
\usebox{2}& \mbox{\bf *}
\end{array}\right)\,,
&&\quad
\sbox0{$\begin{matrix}0 & \lambda_1 & 0 \\ \overline{\lambda_1} & 0 & \lambda_2 \\ 0 & \overline{\lambda_2} & 0 \end{matrix}$}
{\mathbf{P}}={\mathbf{I}}+{\mathbf{U}}\left(\begin{array}{c|c} \usebox{0}& {\bf 0}\\ \hline {\bf 0}& {\bf 0} \end{array}\right) {\mathbf{U}}^* \,, \label{P-ansatz-2B}
\intertext{\hspace{19mm} where the upper right submatrix ${\mathbf{C}}_1^{ur}\in{\mathbb{C}}^{2\times(n-2)}$ has rank 1. Again, we assume w.l.o.g. that \smallskip\newline
\hspace*{19mm} $c_{2,3}\ne 0$. The right choice for the unitary matrix ${\mathbf{U}}$ depends on the structure of ${\mathbf{C}}_1$:}
(2B1) &\qquad
\sbox0{$\begin{matrix}\ast & \bullet \\ \bullet & \ast \end{matrix}$}
\sbox1{$\begin{matrix}0 & 0 & \cdots & 0 \\ \bullet & \ast & \cdots & \ast \end{matrix}$}
\sbox2{$\begin{matrix}0 & \bullet \\ 0 & \ast \\ \vdots & \vdots \\ 0 & \ast \end{matrix}$}
{\mathbf{C}}_1 = \left(\begin{array}{c|c}
\usebox{0}& \usebox{1}\\
\hline
\usebox{2}& \mbox{\bf *}
\end{array}\right)\,,
&&\quad
{\mathbf{U}}={\mathbf{I}}\,, \label{U-ansatz-2B1} \\
(2B2) &\qquad
\sbox0{$\begin{matrix}\ast & \ast \\ \ast & \ast \end{matrix}$}
\sbox1{$\begin{matrix}\bullet & \ast & \cdots & \ast \\ \bullet & \ast & \cdots & \ast \end{matrix}$}
\sbox2{$\begin{matrix}\bullet & \bullet \\ \ast & \ast \\ \vdots & \vdots \\ \ast & \ast \end{matrix}$}
{\mathbf{C}}_1 = \left(\begin{array}{c|c}
\usebox{0}& \usebox{1}\\
\hline
\usebox{2}& \mbox{\bf *}
\end{array}\right)\,,
&&\quad
{\mathbf{U}}= \left(\begin{array}{c|c} {\mathbf{U}}^{ul} & {\bf 0}\\ \hline {\bf 0}& {\mathbf{I}}\end{array}\right) \,, \label{U-ansatz-2B2}
\intertext{\hspace{19mm} with upper left submatrix ${\mathbf{U}}^{ul}=\tfrac1{\sqrt{|c_{1,3}|^2 +|c_{2,3}|^2}}\begin{pmatrix}\overline{c_{2,3}} & c_{1,3} \\ -\overline{c_{1,3}} & c_{2,3} \end{pmatrix}$.}
\nonumber\end{aligned}$$
In this case the upper right submatrix ${\mathbf{C}}_1^{ur}\in{\mathbb{C}}^{2\times(n-2)}$ has rank 2. Its hypocoercivity index is $1$ which can be inferred from condition \[cond:fullrank\]: Using $${\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2} = \Bigg({\mathbf{0}},{\mathbf{0}},\sqrt{c_3} \begin{pmatrix} c_{1,3} \\ \vdots \\ c_{n,3} \end{pmatrix},\ldots,
\sqrt{c_n} \begin{pmatrix} c_{1,n} \\ \vdots \\ c_{n,n} \end{pmatrix} \Bigg)$$ we see that $$\operatorname{rank}\big(\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}\big) =\operatorname{rank}\Bigg(e_3,\ldots,e_n,\sqrt{c_3} \begin{pmatrix} c_{1,3} \\ \vdots \\ c_{n,3} \end{pmatrix},\ldots, \sqrt{c_n} \begin{pmatrix} c_{1,n} \\ \vdots \\ c_{n,n} \end{pmatrix} \Bigg) \,.$$ Due to $\operatorname{rank}{\mathbf{C}}_1^{ur}=2$, we have $\operatorname{rank}\big(\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}\big)=n$. Hence, the hypocoercivity index of ${\mathbf{C}}$ is 1. Such an example (a linearized BGK equation in 1D) was analyzed in §4.4 of [@AAC16] using a transformation matrix with ansatz .
Up to a renumbering of the indices $\{j\ge3\}$, we assume $c_{1,4}\,c_{2,3} \ne c_{1,3}\,c_{2,4}$. Moreover, up to a renumbering of the indices $j\in\{3,4\}$, we assume $|c_{1,4}c_{2,3}|\geq |c_{1,3}c_{2,4}|$ such that $c_{1,4}\ne 0$ and $c_{2,3}\ne 0$. Thus, w.l.o.g. we assume that the zero in the diagonal of ${\mathbf{C}}_2$ at $j=1$ is connected to $j=4$, and the zero at $j=2$ is connected to $j=3$.
The two zeros in the diagonal of ${\mathbf{C}}_2$ are connected (via ${\mathbf{C}}_1$) to two *different* positive entries in the diagonal of ${\mathbf{C}}_2$, i.e. to two decaying modes (and possibly, in addition, also to the same). Hence, this case can occur only for $n\ge4$. Here, the two connections in the relevant (upper left) $4\times4$-subspace can be symbolized as $\circ\!\!\longrightarrow\!\!\bullet\;\;\circ\!\!\longrightarrow\!\!\bullet$.
In this case the upper right submatrix ${\mathbf{C}}_1^{ur}\in{\mathbb{C}}^{2\times(n-2)}$ has rank 1. Then $\operatorname{rank}\big(\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2}\big)=n-1$. Hence, the hypocoercivity index of ${\mathbf{C}}$ is 2 since it is bounded from above by $\dim(\ker {\mathbf{C}}_2)=2$, see Remark \[rem:hypo:cond:B1’\](b).
Let ${\mathbf{C}}_1$ be a Hermitian matrix whose upper right submatrix ${\mathbf{C}}_1^{ur}\in{\mathbb{C}}^{2\times(n-2)}$ has rank 1, and let ${\mathbf{C}}_2$ be a positive semi-definite Hermitian matrix with $\dim(\ker {\mathbf{C}}_2)=2$. Up to a change of basis, the Hermitian matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ satisfy with $c_{2,3} =\overline{c_{3,2}} \ne 0$. Then, the matrix ${\mathbf{C}}:=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ is hypocoercive if and only if $$\label{2B:condition:hypo}
c_{1,3}\ c_{2,3}\ (c_{1,1} -c_{2,2}) -c_{1,3}^2\ c_{2,1} +c_{2,3}^2\ c_{1,2} \ne 0 \,.$$
Up to a change of basis, the Hermitian matrices ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ satisfy . The upper right submatrix ${\mathbf{C}}_1^{ur}\in{\mathbb{C}}^{2\times(n-2)}$ has rank 1, therefore at least one coefficient of ${\mathbf{C}}_1^{ur}$ is non-zero. Another change of basis moves this non-zero coefficient to position $(2,3)$, hence, w.l.o.g. let $c_{2,3} =\overline{c_{3,2}} \ne 0$. To prove that condition is necessary and sufficient, we use the characterization in Proposition \[prop:equivalence\]. Condition \[cond:eigenvector\] for one-dimensional subspaces of $\ker {\mathbf{C}}_2$ reads $$\forall \,(\alpha,\beta)\in{\mathbb{C}}^2\setminus\{(0,0)\}: \qquad {\mathbf{C}}_1 (\alpha e_1 +\beta e_2)
=\alpha \begin{pmatrix} c_{1,1} \\ c_{2,1} \\ \vdots \\ c_{n,1} \end{pmatrix}
+\beta \begin{pmatrix} c_{1,2} \\ c_{2,2} \\ \vdots \\ c_{n,2} \end{pmatrix}
\notin {\operatorname{span}}\{\alpha e_1 +\beta e_2\} \,.$$ This is equivalent to the following condition: $$\begin{aligned}
\text{For all $(\alpha,\beta)\in{\mathbb{C}}^2\setminus\{(0,0)\}$, } &
(\alpha c_{1,1} +\beta c_{1,2})\beta \ne (\alpha c_{2,1} +\beta c_{2,2})\alpha \label{HC-cond:1} \\
&\text{or } \exists j\in\{3,\ldots,n\}\,:\ \alpha c_{j,1} +\beta c_{j,2} \ne 0 \text{ holds.} \label{HC-cond:2}
\end{aligned}$$ Due to the assumption $\operatorname{rank}{\mathbf{C}}_1^{ur} =1$, there exists a unique $\gamma\in{\mathbb{C}}$ (namely $\gamma = -c_{3,1} /c_{3,2}$, since $c_{2,3} =\overline{c_{3,2}} \ne 0$) such that $c_{j,1} +\gamma\ c_{j,2} = 0$ for all $j\in\{3,\ldots,n\}$. Therefore, the second condition holds if and only if $\beta\ne \alpha\gamma$. If $\beta =\alpha\gamma$ then the first condition has to hold. Inserting $\beta =\alpha\gamma$ in yields $$\label{alpha-beta}
0 \ne \alpha\beta (c_{1,1} -c_{2,2}) +\beta^2 c_{1,2} -\alpha^2 c_{2,1}
= \alpha^2 (\gamma (c_{1,1} -c_{2,2}) +\gamma^2 c_{1,2} -c_{2,1}) \,.$$ Using $\gamma = -c_{3,1} /c_{3,2}$, the r.h.s. of reads $$\frac{\alpha^2}{c_{3,2}^2} \big(-c_{3,1}\ c_{3,2} (c_{1,1} -c_{2,2}) +c_{3,1}^2 c_{1,2} -c_{3,2}^2\ c_{2,1} \big)
= -\frac{\alpha^2}{c_{3,2}^2} \overline{(c_{1,3}\ c_{2,3} (c_{1,1} -c_{2,2}) -c_{1,3}^2 c_{2,1} +c_{2,3}^2\ c_{1,2})} \,.$$ Thus, matrix ${\mathbf{C}}$ is hypocoercive if and only if condition holds.
This finishes the complete classification of the situation when $\dim(\ker {\mathbf{C}}_2)=2$. Our ansatz for matrix ${\mathbf{P}}$ depends on the structure of matrix ${\mathbf{C}}_1$. Therefore we distinguish between the subcases (2B1) and (2B2), see also Table \[table:ansatzP\]. We shall prove that these ansatzes will allow for a matrix inequality of the form and hence for an explicit exponential decay in the ODE . As in Theorem \[lemma:ansatzP:1D\] we shall construct ${\mathbf{P}}$ as a perturbation of ${\mathbf{I}}$. To verify, then, a matrix inequality of the form we shall use the following perturbation result on multiple eigenvalues:
\[lemma:ansatzP:nD\] Let ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ be Hermitian matrices with ${\mathbf{C}}_2\ge0$ and $\dim(\ker{\mathbf{C}}_2)=k\in{\mathbb{N}}_0$, such that the associated matrix ${\mathbf{C}}=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ is hypocoercive. Let $\{v_j;\,j=1,\ldots,k\}$ be an orthonormal basis of the kernel $\ker{\mathbf{C}}_2$ and let ${\mathbf{A}}$ be a Hermitian matrix (which makes ${\mathbf{P}}(r):={\mathbf{I}}+r{\mathbf{A}}$ a positive definite Hermitian matrix for sufficiently small $r\geq0$). Then, for sufficiently small $r>0$, the $k$ lowest eigenvalues $\mu_j(r)$ of the Hermitian matrix ${\mathbf{C}}^*{\mathbf{P}}(r)+{\mathbf{P}}(r){\mathbf{C}}$ satisfy $$\label{eigenvalue:expansion}
\mu_j(r) =r \xi_j +o(r)\ ,\quad j=1,...,k\,,$$ where $\xi_j$ are the eigenvalues of ${\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}$ and ${\mathbf{R}}:=(v_1,\ldots,v_k)\in{\mathbb{C}}^{n\times k}$.
We will use this result to construct perturbation matrices ${\mathbf{A}}$ and to check the admissibility of the various ansatzes for the transformation matrices ${\mathbf{P}}$ – mostly in the case $\dim(\ker{\mathbf{C}}_2)=2$. The two matrices in Lemma \[lemma:ansatzP:nD\] are related via $$\label{matrix-rel}
{\mathbf{C}}^*{\mathbf{P}}(r)+{\mathbf{P}}(r){\mathbf{C}}={\mathbf{C}}^*({\mathbf{I}}+r{\mathbf{A}})+({\mathbf{I}}+r{\mathbf{A}}){\mathbf{C}}=2{\mathbf{C}}_2 +r({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}})\,,$$ and ${\mathbf{C}}_2$ has a $k$-fold 0-eigenvalue by assumption. Now, if ${\mathbf{A}}$ is chosen such that all eigenvalues $\xi_j,\,j=1,...,k$ in are positive, then we deduce the positive definiteness of ${\mathbf{C}}^*{\mathbf{P}}(r)+{\mathbf{P}}(r){\mathbf{C}}$ for sufficiently small $r>0$.\
We remark that the positivity of $\xi_1,...,\xi_k$ is first of all a sufficient condition for the positive definiteness of ${\mathbf{C}}^*{\mathbf{P}}(r)+{\mathbf{P}}(r){\mathbf{C}}$ (for sufficiently small $r>0$). But one sees easily from that it is also necessary.\
\[th:P-admissible2D\] Let ${\mathbf{C}}_1$ and ${\mathbf{C}}_2$ be Hermitian matrices with ${\mathbf{C}}_2\ge0$ and $\dim(\ker{\mathbf{C}}_2)=2$, such that the associated matrix ${\mathbf{C}}=i{\mathbf{C}}_1 +{\mathbf{C}}_2$ is hypocoercive. Then there exists a two-parameter ansatz for a positive definite matrix ${\mathbf{P}}={\mathbf{P}}(\lambda_1,\lambda_2)$, according to Table \[table:ansatzP\], such that ${\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}$ is positive definite (for an appropriate choice of $\lambda_1,\lambda_2$).
First, one easily checks that all matrices ${\mathbf{P}}$ from Table \[table:ansatzP\] are positive definite if $|\lambda_1|^2+|\lambda_2|^2<1$. Thus, ${\mathbf{P}}(r):={\mathbf{I}}+r{\mathbf{A}}$ with ${\mathbf{A}}:={\mathbf{P}}-{\mathbf{I}}$ yields for $r\in[0,1]$ a family of positive definite Hermitian matrices ${\mathbf{P}}(r)$. Now, up to a change of basis in ${\mathbb{C}}^n$, we assume without loss of generality that ${\mathbf{C}}_2$ is a diagonal matrix of the form ${\mathbf{C}}_2 =\operatorname{diag}(0,0,c_3,\ldots,c_n)$ with $c_j>0$. Then, $\ker{\mathbf{C}}_2={\operatorname{span}}\{e_1,e_2\}$ and we choose ${\mathbf{R}}=(e_1,e_2)\in{\mathbb{R}}^{n\times 2}$. According to Lemma \[lemma:ansatzP:nD\], the positive definiteness of ${\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}$ (for sufficiently small $r>0$) can be inferred from the positive definiteness of ${\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}$.
Next we deal with each case of ${\mathbf{C}}_1$ and its corresponding ansatz ${\mathbf{P}}={\mathbf{I}}+{\mathbf{A}}$ (as listed in Table \[table:ansatzP\]) separately: we need to prove that $\lambda_1$ and $\lambda_2$ can be chosen such that ${\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}$ is indeed positive definite.
1. We consider ${\mathbf{C}}_1=(c_{j,k})_{j,k\in\{1,\ldots,n\}}$ satisfying w.l.o.g. $$\label{cond:2A:c}
|c_{1,4}\ c_{2,3}|\geq |c_{1,3}\ c_{2,4}| \quad \mbox{and} \quad c_{1,4}\ c_{2,3} \ne c_{1,3}\ c_{2,4} \,,$$ such that $c_{2,3}=\overline{c_{3,2}}\ne0$ and $c_{1,4}=\overline{c_{4,1}}\ne0$. For $${\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}= i
\begin{pmatrix}
- c_{1,4}\ \overline{\lambda_1} + \overline{c_{1,4}}\ \lambda_1 & +\overline{c_{2,4}}\ \lambda_1 - c_{1,3}\ \overline{\lambda_2} \\
- c_{2,4}\ \overline{\lambda_1} + \overline{c_{1,3}}\ \lambda_2 & -c_{2,3}\ \overline{\lambda_2}+\overline{c_{2,3}}\ \lambda_2
\end{pmatrix} \,$$ to be positive definite, all three of its minors have to be positive for appropriately chosen $\lambda_1$ and $\lambda_2$. We set $$\label{lambda-polar}
\lambda_1 := -i \ell_1 c_{1,4}\,, \quad \lambda_2 := -i \ell_2 c_{2,3} \,,$$ for some positive numbers $\ell_1$ and $\ell_2$. Then, the minors of first order satisfy $$\begin{aligned}
-i(c_{1,4}\ \overline{\lambda_1} -\overline{c_{1,4}}\ \lambda_1) &= 2\Im(c_{1,4}\ \overline{\lambda_1})=2\ell_1\ |c_{1,4}|^2 >0 \,, \\
-i(c_{2,3}\ \overline{\lambda_2} -\overline{c_{2,3}}\ \lambda_2) &= 2\Im(c_{2,3}\ \overline{\lambda_2})=2\ell_2\ |c_{2,3}|^2 >0 \,.\end{aligned}$$ The minor of second order reads (using ) $$\begin{aligned}
\det({\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}&+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}})
= 4\ell_1 \ell_2 |c_{1,4}|^2 |c_{2,3}|^2 -|\ell_1\ c_{2,4}\ \overline{c_{1,4}} +\ell_2\ \overline{c_{1,3}}\ c_{2,3}|^2 \\
&\quad = 4\ell_1 \ell_2 |c_{1,4}|^2 |c_{2,3}|^2 -|\ell_1\ c_{2,4}\ \overline{c_{1,4}}|^2 -|\ell_2\ \overline{c_{1,3}}\ c_{2,3}|^2 \\
&\qquad -\ell_1 \ell_2\ c_{2,4}\ \overline{c_{1,4}}\ c_{1,3}\ \overline{c_{2,3}} -\ell_1 \ell_2\ \overline{c_{2,4}}\ c_{1,4}\ \overline{c_{1,3}}\ c_{2,3} \\
&\quad = -\big(\ell_1 |c_{1,4}\ c_{2,4}| -\ell_2 |c_{1,3}\ c_{2,3}|\big)^2 \\
&\qquad +\ell_1 \ell_2 \Big[ 4|c_{1,4}\ c_{2,3}|^2 -2 |c_{1,4}\ c_{2,4}\ c_{1,3}\ c_{2,3}|
-c_{2,4}\ \overline{c_{1,4}}\ c_{1,3}\ \overline{c_{2,3}} -\overline{c_{2,4}}\ c_{1,4}\ \overline{c_{1,3}}\ c_{2,3} \Big] \\
&\quad = -\big(\ell_1 |c_{1,4}\ c_{2,4}| -\ell_2 |c_{1,3}\ c_{2,3}|\big)^2 \\
&\qquad +\ell_1 \ell_2 \Big[ (3 |c_{1,4}\ c_{2,3}| +|c_{1,3}\ c_{2,4}|) (|c_{1,4}\ c_{2,3}| -|c_{1,3}\ c_{2,4}|)
+|c_{1,4}\ c_{2,3} -c_{1,3} c_{2,4}|^2 \Big] \,.\end{aligned}$$ Then, the minor of second order is positive for the choice $\ell_1 =\epsilon |c_{1,3}\ c_{2,3}|$ and $\ell_2 =\epsilon |c_{1,4}\ c_{2,4}|$ with any $\epsilon>0$, due to our assumption . Finally, for sufficiently small $\epsilon>0$ the Hermitian matrix ${\mathbf{P}}$ is positive definite.
2. \[case:2B\] First, we verify that the ansatz for ${\mathbf{P}}$ in is admissible in case (2B1).
, we consider w.l.o.g. $${\mathbf{C}}_1=(c_{j,k})_{j,k\in\{1,\ldots,n\}} \quad \text{with} \quad
c_{1,2}=\overline{c_{2,1}}\ne0\,,\
c_{2,3}=\overline{c_{3,2}}\ne0\,,\
c_{1,3}=\overline{c_{3,1}}=0 \,.$$ Then, the connections in the relevant (upper left) $3\times3$-subspace can be symbolized as $\circ\!\!\longrightarrow\!\!\circ\!\!\longrightarrow\!\!\bullet$; see . To prove that the ansatz for ${\mathbf{P}}$ in with ${\mathbf{U}}={\mathbf{I}}$ is admissible, we use Lemma \[lemma:ansatzP:nD\] and we need to check the positive definiteness of $$\label{RCAR}
{\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}= i
\begin{pmatrix}
-c_{1,2}\overline{\lambda_1} +\overline{ c_{1,2}}\lambda_1 & (c_{2,2} - c_{1,1})\lambda_1 \\
(c_{1,1} -c_{2,2})\overline{\lambda_1} & c_{1,2}\overline{\lambda_1} -\overline{ c_{1,2}}\lambda_1 -c_{2,3}\overline{\lambda_2} +\overline{ c_{2,3}}\lambda_2
\end{pmatrix}$$ for appropriately chosen $\lambda_1$ and $\lambda_2$. The minors of first order are $$-i( c_{1,2}\overline{\lambda_1}-\overline{ c_{1,2}}\lambda_1) =2\Im( c_{1,2}\overline{\lambda_1}) \quad \text{and} \quad
i( c_{1,2}\overline{\lambda_1}-\overline{ c_{1,2}}\lambda_1- c_{2,3}\overline{\lambda_2} +\overline{ c_{2,3}}\lambda_2) =-2\Im( c_{1,2}\overline{\lambda_1})+2\Im( c_{2,3}\overline{\lambda_2}) \,.$$ They are positive if and only if $$\label{1st-minor-cond}
0<\Im( c_{1,2}\overline{\lambda_1})<\Im( c_{2,3}\overline{\lambda_2}) \,.$$ Due to our assumptions $ c_{1,2}\ne0$ and $ c_{2,3}\ne0$, we can choose $\lambda_1$ and $\lambda_2$ such that this condition is satisfied. The minor of second order reads $$\det({\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}})
= 4\Im( c_{1,2}\overline{\lambda_1}) \big(\Im( c_{2,3}\overline{\lambda_2})-\Im( c_{1,2}\overline{\lambda_1})\big) -|c_{1,1} -c_{2,2}|^2 |\lambda_1|^2 \,,$$ where the first summand is positive due to . First we choose $\lambda_1$ and $\lambda_2$ such that the minors of first order are positive. Then we consider $r \lambda_1$ for $r\in(0,1)$ instead of $\lambda_1$, and we choose $r\in(0,1)$ sufficiently small such that the second minor becomes positive, and hence is positive definite.
, we consider w.l.o.g. $${\mathbf{C}}_1=(c_{j,k})_{j,k\in\{1,\ldots,n\}} \quad \text{with} \quad
c_{1,3}=\overline{c_{3,1}}\ne0\,,\
c_{2,3}=\overline{c_{3,2}}\ne0\,,$$ and recall the hypocoercivity condition . The guideline to construct a simple ansatz for ${\mathbf{P}}$ at the beginning of this section would suggest to connect each non-decaying mode to the same decaying mode. However, for some examples in subcase (2B2) this ansatz is not admissible. Therefore this guideline is not universally true.
The motivation for the (alternative) ${\mathbf{P}}$-ansatz with unitary matrix ${\mathbf{U}}$ in is that the transformation $\tilde{{\mathbf{C}}}_1 ={\mathbf{U}}^{-1} {\mathbf{C}}_1 {\mathbf{U}}$ yields a matrix of form (2B1) with $\tilde c_{1,j}=0$ for $j\ge3$ (since $\operatorname{rank}({\mathbf{C}}_1^{ur})=1$), $\tilde c_{2,3}=1$ and $$\big( \widetilde{{\mathbf{C}}}_1 \big)_{1,2} = \tilde{c}_{1,2} = \tfrac1{|c_{1,3}|^2 +|c_{2,3}|^2} \big( (c_{1,1} -c_{2,2}) c_{1,3} c_{2,3} +c_{1,2} c_{2,3}^2 -\overline{c_{1,2}} c_{1,3}^2 \big) \ne 0 \,,$$ due to the hypocoercivity condition . To prove that the ansatz for ${\mathbf{P}}$ in with ${\mathbf{U}}$ in is admissible, we consider $${\mathbf{C}}^* {\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}= {\mathbf{C}}^{*} ({\mathbf{I}}+{\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* ) +({\mathbf{I}}+{\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* ){\mathbf{C}}= 2{\mathbf{C}}_2 +{\mathbf{C}}^{*} {\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* +{\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* {\mathbf{C}}\,.$$ Due to Lemma \[lemma:ansatzP:nD\], we need to check the positive definiteness of $\widetilde{{\mathbf{R}}}^*({\mathbf{C}}^{*} {\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* +{\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* {\mathbf{C}})\widetilde{{\mathbf{R}}}$ for appropriately chosen $\lambda_1$ and $\lambda_2$. Using $\widetilde{{\mathbf{R}}} ={\mathbf{U}}{\mathbf{R}}$, we deduce $$\begin{aligned}
\widetilde{{\mathbf{R}}}^*({\mathbf{C}}^{*} {\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* +{\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* {\mathbf{C}})\widetilde{{\mathbf{R}}}
&= i{\mathbf{R}}^* {\mathbf{U}}^* (-{\mathbf{C}}_1^{*} {\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* +{\mathbf{U}}{\mathbf{A}}{\mathbf{U}}^* {\mathbf{C}}_1){\mathbf{U}}{\mathbf{R}}\\
&= i{\mathbf{R}}^* \big(-({\mathbf{U}}^* {\mathbf{C}}_1^{*} {\mathbf{U}}){\mathbf{A}}+{\mathbf{A}}({\mathbf{U}}^* {\mathbf{C}}_1 {\mathbf{U}})\big){\mathbf{R}}\,.
\end{aligned}$$ Recalling that ${\mathbf{U}}^* {\mathbf{C}}_1 {\mathbf{U}}={\mathbf{U}}^{-1} {\mathbf{C}}_1 {\mathbf{U}}$ is of form (2B1), the positive definiteness of $i{\mathbf{R}}^* \big(-({\mathbf{U}}^* {\mathbf{C}}_1^{*} {\mathbf{U}}){\mathbf{A}}+{\mathbf{A}}({\mathbf{U}}^* {\mathbf{C}}_1 {\mathbf{U}})\big){\mathbf{R}}$ for suitable $\lambda_1$ and $\lambda_2$ follows as in case (2B1).
For $\dim(\ker{\mathbf{C}}_2)=1$ or 2, we just listed all possible cases. But for $\dim(\ker{\mathbf{C}}_2)=3$ we will next only consider the one situation relevant below for the linearized BGK equation in 1D, i.e. , .
### Hypocoercive matrix with $\dim(\ker{\mathbf{C}}_2)=3$ {#sec:dim-ker3}
If the three zeros in the diagonal of ${\mathbf{C}}_2$ are connected (via ${\mathbf{C}}_1$) only *consecutively* to a positive entry in the diagonal of ${\mathbf{C}}_2$, the relevant $4\times4$-subspace can be symbolized as $\circ\!\!\longrightarrow\!\!\circ\!\!\longrightarrow\!\!\circ\!\!\longrightarrow\!\!\bullet$. Proceeding as in §\[sec:dim-ker2\] one easily checks that $$ \operatorname{rank}\big(\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1^2 \sqrt{{\mathbf{C}}_2}\big) = n-1 \,,\quad
\operatorname{rank}\big(\sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1 \sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1^2 \sqrt{{\mathbf{C}}_2},{\mathbf{C}}_1^3 \sqrt{{\mathbf{C}}_2}\big) = n \,,$$ and hence the hypocoercivity index of ${\mathbf{C}}$ is 3.
With ${\mathbf{C}}_1$ of the form $$\label{C1-form-3}
\sbox0{$\begin{matrix}\ast & \bullet & 0\\ \bullet & \ast & \bullet\\ 0 & \bullet & \ast\end{matrix}$}
\sbox1{$\begin{matrix}0 & 0 & \cdots & 0 \\ 0 & 0 & \cdots & 0 \\ \bullet & \ast & \cdots & \ast \end{matrix}$}
\sbox2{$\begin{matrix}0 & 0 & \bullet \\ 0 & 0 & \ast \\ \vdots & \vdots & \vdots \\ 0 & 0 & \ast \end{matrix}$}
{\mathbf{C}}_1 = \left(\begin{array}{c|c}
\usebox{0}& \usebox{1}\\
\hline
\usebox{2}& \mbox{\bf *}
\end{array}\right)\,,$$ a natural ansatz for a simple transformation matrix is given by $$\label{P-ansatz-3}
\sbox0{$\begin{matrix}0 & \lambda_1 & 0 & 0 \\ \overline{\lambda_1} & 0 & \lambda_2 & 0 \\ 0 & \overline{\lambda_2} & 0 & \lambda_3 \\ 0 & 0 & \overline{\lambda_3} & 0 \end{matrix}$}
{\mathbf{P}}={\mathbf{I}}+ \left(\begin{array}{c|c}
\usebox{0}& {\bf 0}\\
\hline
{\bf 0}& {\bf 0}
\end{array}\right)\ ,$$ with some $\lambda_1,\,\lambda_2,\,\lambda_3\in{\mathbb{C}}$.
Indeed, this ansatz always yields a useful Lyapunov functional and hence a quantitative exponential decay rate, as we shall now show under the simplifying restriction $c_{1,1}=c_{2,2}=c_{3,3}$ (which is the relevant situation in §\[sec:linBGK:1D\]):
\[th:p-ansatz-3\] Let ${\mathbf{C}}_2=\operatorname{diag}(0,0,0,c_4,...,c_n)$ with $c_j>0$, and ${\mathbf{C}}_1$ be a Hermitian matrices of form and satisfying $c_{1,1}=c_{2,2}=c_{3,3}$. Then there exists a three-parameter ansatz for a positive definite matrix ${\mathbf{P}}={\mathbf{P}}(\lambda_1,\lambda_2,\lambda_3)$ of form , such that ${\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}$ is positive definite (for an appropriate choice of $\lambda_1,\lambda_2,\lambda_3$).
First, the matrix ${\mathbf{P}}$ is positive definite if $|\lambda_{1}|^2+|\lambda_{2}|^2+|\lambda_{3}|^2<1$. Thus, ${\mathbf{P}}(r):={\mathbf{I}}+r{\mathbf{A}}$ with ${\mathbf{A}}:={\mathbf{P}}-{\mathbf{I}}$ yields for $r\in[0,1]$ a family of positive definite Hermitian matrices ${\mathbf{P}}(r)$.
We have $\ker{\mathbf{C}}_2={\operatorname{span}}\{e_1,e_2,e_3\}$ and ${\mathbf{R}}=(e_1,e_2,e_3)\in{\mathbb{R}}^{n\times 3}$. According to Lemma \[lemma:ansatzP:nD\], the positive definiteness of ${\mathbf{C}}^*{\mathbf{P}}+{\mathbf{P}}{\mathbf{C}}$ (for sufficiently small $r>0$) can be inferred from the positive definiteness of ${\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}$.
As in the proof of Theorem \[th:P-admissible2D\] we search for conditions on $\lambda_j$ ($j=1,2,3$) such that the eigenvalues of $$\label{R-matrix-3D}
{\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}=
\begin{pmatrix}
2 \Im(c_{1,2}\overline{\lambda_1}) & 0 & i(c_{2,3} \lambda_1 -c_{1,2}\lambda_2) \\
0 & -2\Im(c_{1,2}\overline{\lambda_1}-c_{2,3}\overline{\lambda_2}) & 0 \\
i(\overline{c_{1,2}} \overline{\lambda_2} -\overline{c_{2,3}}\overline{\lambda_1}) & 0 & -2\Im(c_{2,3}\overline{\lambda_2} -c_{3,4}\overline{\lambda_3})
\end{pmatrix} \,$$ are positive. If all minors are positive, then the matrix will be positive definite (by Sylvester’s criterion). From the three minors of first order we deduce the conditions $$\label{minor1-3D}
0<\Im(c_{1,2}\overline{\lambda_1}) <\Im(c_{2,3}\overline{\lambda_2}) <\Im(c_{3,4}\overline{\lambda_3}) \,,$$ which also imply the positivity of the second minor, i.e. $$4 \Im(c_{1,2}\overline{\lambda_1}) \Im(c_{2,3}\overline{\lambda_2}-c_{1,2}\overline{\lambda_1}) >0 \,.$$ To satisfy the former conditions it is convenient to choose $$\label{lambda-polar-3D}
\arg(\lambda_1)=\arg(c_{1,2})-\frac{\pi}{2}\ ,\quad \arg(\lambda_2)=\arg(c_{2,3})-\frac{\pi}{2}\ ,\quad \arg(\lambda_3)=\arg(c_{3,4})-\frac{\pi}{2}\ ,$$ just as in . The determinant of reads $$\label{det-3D}
2 \Im(c_{2,3}\overline{\lambda_2}-c_{1,2}\overline{\lambda_1})
\left[4 \Im(c_{1,2}\overline{\lambda_1}) \Im(c_{3,4}\overline{\lambda_3}-c_{2,3}\overline{\lambda_2})
-|c_{2,3} \lambda_1-c_{1,2}\lambda_2|^2\right]\ .$$ Now the parameters $\lambda_j$ ($j=1,2,3$) can be chosen in analogy to the proof of Theorem \[th:P-admissible2D\], case \[case:2B\] to satisfy the conditions . Once $\lambda_1$, $\lambda_2$, and $\arg(\lambda_3)$ are fixed, we can choose $|\lambda_3|$ large enough to also satisfy the positivity of .
This analysis to construct appropriate matrices ${\mathbf{P}}$ could, of course, also be extended to higher dimensions of $\ker{\mathbf{C}}_2$, but this gets more cumbersome. In §\[sec:linBGK:2D\] and §\[sec:linBGK:3D\] we have $\dim(\ker{\mathbf{C}}_2)=4$ and 5, respectively.
Linearized BGK equation in 1D {#sec:linBGK:1D}
=============================
In this section we shall analyze the large time behavior of the linearized BGK equation in 1D, $$\begin{gathered}
\label{linBGK:torus:1D}
\partial_t h (x,v,t) + v\ \partial_x h(x,v,t) \\
= M_1(v) \,\left[\left( \frac32 -\frac{v^2}{2}\right)\sigma(x,t)
+v \mu(x,t)
+ \left( -\frac{1}{2}+ \frac{v^2}{2}\right)\tau(x,t)\right] -h(x,v,t)\ , \quad t\geq 0 \ ,\end{gathered}$$ for the perturbation $h(x,v,t) \approx f(x,v,t) -M_1(v)$. To prepare for the proof of Theorem \[linBGK-decay\] we shall use an expansion in $v$–modes, as in [@AAC16]. Using the probabilists’ Hermite polynomials, $$\label{hermite-polynom:1D}
H_m(v) := (-1)^m e^{\frac{v^2}{2}}\frac{{\rm d}^m}{{\rm d} v^m}e^{-\frac{v^2}{2}}\ ,\quad m\in{\mathbb{N}}_0\ ,$$ we define the normalized Hermite functions corresponding to $T=1$: $$\label{hermite-fct:1D}
g_m(v):=(2\pi m!)^{-1/2} H_m(v)\,e^{-\frac{v^2}{2}}.
$$ They satisfy $$\int_{\mathbb{R}}g_m(v) g_n(v) M_1^{-1}(v)\,{\rm d} v = \delta_{mn}\,,\quad m,n\in{\mathbb{N}}_0\,,$$ and the recurrence relation $$\label{recur:1D}
v \, g_m(v) = \sqrt{m+1}\, g_{m+1}(v) + \sqrt{m} \,g_{m-1}(v)\,,\quad m\in{\mathbb{N}}\ .
$$ The first three normalized Hermite functions $g_m(v)$ are $${g}_0(v) = M_1(v)\ , \qquad
{g}_1(v) = v M_1(v) \quad\text{and}\quad
{g}_2(v) = \frac{ v^2- 1}{\sqrt{2}}\ M_1(v)\ .$$ With this notation, reads $$\partial_t h(x,v,t) + v\ \partial_x h(x,v,t) = \left({g}_0(v) -\frac{1}{\sqrt2} {g}_2(v)\right)\sigma(x,t)
+ {g}_1(v)\mu(x,t)
+ \frac{1}{\sqrt2} {g}_2(v)\tau(x,t)-h(x,v,t)\ .$$ We start with the $x$–Fourier series of $h$: $$ h(x,v,t) = \sum_{k\in{\mathbb{Z}}} h_k(v,t)\,e^{ik\frac{2\pi}{L}x}\ .$$ Each spatial mode $h_k(v,t)$ is decoupled and evolves according to $$\label{linBGK:Fourier}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_k + ik\tfrac{2\pi}{L}v\ h_k = {g}_0(v)\sigma_k(t)
+ {g}_1(v)\mu_k(t)
+ {g}_2(v)\frac{1}{\sqrt2} \left(\tau_k(t) -\sigma_k(t)\right)-h_k\ ,\quad k\in{\mathbb{Z}};\,t\ge0\ .$$ Here, $\sigma_k$, $\mu_k$ and $\tau_k$ denote the spatial modes of the $v$–moments $\sigma$, $\mu$ and $\tau$ defined in ; hence $$\sigma_k := \int_{\mathbb{R}}h_k(v,t){\,\operatorname{d}\!v}\ , \quad
\mu_k := \int_{\mathbb{R}}v\ h_k(v,t){\,\operatorname{d}\!v}\ , \quad
\tau_k := \int_{\mathbb{R}}v^2\ h_k(v,t){\,\operatorname{d}\!v}\ .$$
Next we expand $h_k(\cdot,t)\in L^2({\mathbb{R}};M_1^{-1})$ in the orthonormal basis $\{{g}_m(v)\}_{m\in{\mathbb{N}}_0}$: $$h_k(v,t)=\sum_{m=0}^\infty \hat h_{k,m}(t)\,{g}_m(v)\ ,\quad\text{with}\quad
\hat h_{k,m} = \langle h_k(v),{g}_m(v)\rangle_{L^2(M_1^{-1})}\ .$$ For each $k\in{\mathbb{Z}}$, the “infinite vector” $\hat {\bf h}_k(t) = ( \hat h_{k,0}(t),\,\hat h_{k,1}(t),\ ...)^\top\in \ell^2({\mathbb{N}}_0)$ contains all Hermite coefficients of $h_k(\cdot,t)$. In particular we have $$\hat h_{k,0} = \int_{\mathbb{R}}h_k(v) {g}_0(v) M_1^{-1}(v) {\,\operatorname{d}\!v} = \sigma_k \ , \qquad
\hat h_{k,1} = \int_{\mathbb{R}}h_k(v) {g}_1(v) M_1^{-1}(v) {\,\operatorname{d}\!v} = \mu_k \ ,$$ and $$\hat h_{k,2} = \int_{\mathbb{R}}h_k(v) {g}_2(v) M_1^{-1}(v) {\,\operatorname{d}\!v} = \frac{1}{\sqrt2} \left(\tau_k-\sigma_k\right)\ .$$ Hence, can be written equivalently as $${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_k(v,t) + ik\tfrac{2\pi}{L} v\ h_k(v,t) = {g}_0(v) \hat h_{k,0}(t)
+ {g}_1(v) \hat h_{k,1}(t)
+ {g}_2(v) \hat h_{k,2}(t) -h_k(v,t)\ , \quad k\in{\mathbb{Z}}\ ;\ t\ge0\ .$$ Thus, the vector of its Hermite coefficients satisfies $$\label{linap4}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}\hat {\bf h}_k(t) + ik \tfrac{2\pi}{L}\,{\mathbf{L}}_1 \hat{\bf h}_k(t) = -{\mathbf{L}}_2 \hat{\bf h}_k(t)\ ,\quad k\in{\mathbb{Z}}\ ;\ t\ge0\ ,$$ where the operators ${\mathbf{L}}_1,\,{\mathbf{L}}_2$ are represented by “infinite matrices” on $\ell^2({\mathbb{N}}_0)$: $$\label{L1L2}
{\mathbf{L}}_1= \begin{pmatrix}
0 & \sqrt 1 & 0 & \cdots \\
\sqrt 1 & 0 & \sqrt 2 & 0 \\
0 & \sqrt 2 & 0 & \sqrt 3 \\
\vdots & 0 & \sqrt 3 & \ddots
\end{pmatrix}
\ ,\quad {\mathbf{L}}_2=\operatorname{diag}(0,\, 0,\, 0,\, 1,\, 1,\cdots)\ .$$
The bi-diagonal form of ${\mathbf{L}}_1$ is a direct expression of the two-term recursion relation (\[recur:1D\]). This is not special to the Hermite polynomials; a similar expression holds for the orthogonal polynomials with respect to any even reference measure.
Equation provides a decomposition of the generator in its skew-symmetric part $-ik \tfrac{2\pi}{L}{\mathbf{L}}_1$ and its symmetric part $-{\mathbf{L}}_2$, the latter introducing the decay in the evolution.
We remark that simplifies for the spatial mode $h_0$ with $k=0$. One easily verifies that, for all $d$, the flow of preserves , i.e. $\sigma_0(t)=0$, $\mu_0(t)=0$, $\tau_0(t)=0$ for all $t\ge0$. Hence, yields $$\label{h0-decay}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_0(v,t) = -h_0(v,t)\ ,\quad t\ge0\ .$$
For $k\neq 0$, we note that the linearized BGK equation is very similar to the equation specified in [@AAC16 §4.4]: The only difference is that ${\mathbf{L}}_2$ now has one more zero – at the second entry on the diagonal, which corresponds to the conservation of momentum. For $k\ne0$, has the structure of the example in §\[sec:dim-ker3\], and thus hypocoercivity index 3. This has a simple interpretation: The mass-conservation mode is coupled to the momentum-conservation mode, which is coupled to the energy-conservation mode. Finally, the latter is coupled to the decreasing mode that corresponds to $g_3(v)$. The hypocoercivity index of can also be obtained directly from Definition \[hyp-index\], in its equivalent formulation \[cond:trivkernel\] that also applies to “infinite matrices”: With ${\operatorname{corank}}{\mathbf{L}}_2=3$, $\ker{\mathbf{L}}_2={\operatorname{span}}\{e_0,\,e_1,\,e_2\}$, and the relations $${\mathbf{L}}_1 e_0=e_1\ ,\quad {\mathbf{L}}_1 e_1=e_0+\sqrt2 e_2\ ,\quad {\mathbf{L}}_1 e_2=\sqrt2 e_1+\sqrt3 e_3\ ,\quad$$ we again find $\tau=3$.\
We define the matrices ${\mathbf{C}}_k := ik \tfrac{2\pi}{L} {\mathbf{L}}_1 + {\mathbf{L}}_2$, $k\in{\mathbb{Z}}$ which determine the evolution of the spatial modes of the BGK equation in 1D, cf. . Our next goal is to establish a spectral gap of ${\mathbf{C}}_k$, uniformly in $k\ne0$. This will prove Theorem \[linBGK-decay\] in 1D. Clearly, this matrix corresponds to ${\mathbf{C}}=i{\mathbf{C}}_1+{\mathbf{C}}_2$ in §\[sec:Pmatrix\]. There, the construction of the transformation matrix ${\mathbf{P}}(r)={\mathbf{I}}+r{\mathbf{A}}$ was based on Lemma \[lemma:ansatzP:nD\], and hence on proving the positive definiteness of $${\mathbf{R}}^*({\mathbf{C}}^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}){\mathbf{R}}= i{\mathbf{R}}^*(-{\mathbf{C}}_1^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_1){\mathbf{R}}\,.$$ Here, the operator ${\mathbf{L}}_1$ carries the coefficient $ik \tfrac{2\pi}{L}$ with $k\in{\mathbb{Z}}\setminus\{0\}$. To compensate for $k$, it is natural to choose the perturbation matrix ${\mathbf{A}}$ proportional to $\frac1k$. Following §\[sec:dim-ker3\] we hence use the ansatz for the $k$–dependent transformation matrices ${\mathbf{P}}_k$: For parameters $\lambda_j;\,j=1,2,3$ to be chosen below, we define ${\mathbf{P}}_k,\, k\ne0$ to be the infinite matrix that has $$\label{P:1D}
\begin{pmatrix}
1 & \lambda_1/k & 0 & 0 \\
\overline{\lambda_1}/k & 1 & \lambda_2/k &0\\
0 & \overline{\lambda_2}/k &1 & \lambda_3/k \\
0 & 0 & \overline{\lambda_3}/k & 1
\end{pmatrix}$$ as its upper-left $4\times 4$ block, with all other entries being those of the identity. Under the assumption $|\lambda_1|^2 +|\lambda_2|^2 +|\lambda_3|^2 <1$, the matrix ${\mathbf{P}}_k$ will be positive definite for all $k\ne 0$. Recalling that ${\mathbf{L}}_1$ is an (infinite) real matrix as well as the parameter choice in , it is natural to choose also here $\arg(\lambda_j)=-\frac\pi2$. Hence turns into $$\label{P:1D'}
\begin{pmatrix}
1 & -i\alpha/k & 0 & 0 \\
i\alpha/k & 1 & -i\beta/k &0\\
0 & i\beta/k &1 & -i\gamma/k \\
0 & 0 & i\gamma/k & 1
\end{pmatrix} \ ,$$ with $\alpha:=|\lambda_1|$, $\beta:=|\lambda_2|$, $\gamma:=|\lambda_3|$.
Now, (the infinite dimensional analog of) Theorem \[th:p-ansatz-3\] asserts that the above ansatz will yield an admissible transformation matrix ${\mathbf{P}}$ and hence an exponential decay rate for , uniformly in $k$. But, as a perturbation result, it neither provides an explicit value for the decay rate $\mu$, nor does it yield a rather natural ratio between the parameters $\lambda_j$. These two aspects will be our next task.
To justify the infinite dimensional analog of Theorem \[th:p-ansatz-3\], we decompose $${\mathbf{C}}_k^*{\mathbf{P}}(r)+{\mathbf{P}}(r){\mathbf{C}}_k = 2{\mathbf{L}}_2 +r\ ({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k) = 2{\mathbf{I}}+(2{\mathbf{L}}_2 -2{\mathbf{I}}) +r\ ({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k) \,.$$ To investigate the spectrum of the Hermitian operator ${\mathbf{C}}_k^*{\mathbf{P}}(r)+{\mathbf{P}}(r){\mathbf{C}}_k$ in $\ell^2({\mathbb{N}}_0)$, it is sufficient to compute the spectrum of the compact operator $(2{\mathbf{L}}_2 -2{\mathbf{I}}) +r\ ({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k)$. The compact operators $2{\mathbf{L}}_2 -2{\mathbf{I}}=2\operatorname{diag}(-1,-1,-1,0,\ldots)$ and ${\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k$ act on a common finite-dimensional subspace of $\ell^2({\mathbb{N}}_0)$, hence we can use Lemma \[lemma:ansatzP:nD\] to analyze the restriction of the compact operators on this finite-dimensional subspace: The three lowest eigenvalues of $(2{\mathbf{L}}_2 -2{\mathbf{I}}) +r\ ({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k)$, for sufficiently small $r\geq0$, satisfy $$\widetilde{\mu_j}(r) = -2 +r\xi_j +o(r) \, ;\quad j=1,2,3\,,$$ where $\xi_j$ are the eigenvalues of ${\mathbf{R}}^*({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k){\mathbf{R}}$ and ${\mathbf{R}}=(e_1,e_2,e_3)\in{\mathbb{C}}^{n\times 3}$ (recall that $\ker({\mathbf{L}}_2) ={\operatorname{span}}\{e_1,e_2,e_3\}$). Then the three lowest eigenvalues of ${\mathbf{C}}_k^*{\mathbf{P}}(r)+{\mathbf{P}}(r){\mathbf{C}}_k$, for sufficiently small $r\geq0$, satisfy $$\mu_j(r) = r\xi_j +o(r) \,.$$
Next we search for conditions on $\alpha,\,\beta,\,\gamma>0$ such that the eigenvalues $\xi_j$ of $${\mathbf{R}}^*({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k){\mathbf{R}}= \tfrac{2\pi}{L}\
\begin{pmatrix}
2 \alpha & 0 & \sqrt2\alpha-\beta \\
0 & 2(\sqrt{2}\beta-\alpha) & 0 \\
\sqrt2\alpha-\beta & 0 & 2(\sqrt{3}\gamma -\sqrt{2}\beta)
\end{pmatrix} \,$$ are positive. If all minors are positive, then the matrix will be positive definite (by Sylvester’s criterion). We deduce the conditions $$0<\alpha<\sqrt2\beta<\sqrt3\gamma \quad \text{and} \quad
0 <4\alpha(\sqrt{3}\gamma -\sqrt{2}\beta) - |\sqrt{2} \alpha -\beta|^2 \, ,
$$ which are special cases of , . In fact, the matrix $\tfrac{L}{2\pi}\ {\mathbf{R}}^*({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k){\mathbf{R}}$ has the eigenvalues $2(\sqrt2\beta-\alpha)$ and $$\sqrt{3}\gamma-\sqrt2\beta+\alpha
\pm \sqrt{(\sqrt{3}\gamma-\sqrt2\beta-\alpha)^2 +(\sqrt{2}\alpha-\beta)^2} \,.
$$ We note that the special choice $\beta:=\sqrt{2}\alpha$ and $\gamma:=\sqrt{3}\alpha$ makes all eigenvalues of ${\mathbf{R}}^*({\mathbf{C}}_k^*{\mathbf{A}}+{\mathbf{A}}{\mathbf{C}}_k){\mathbf{R}}$ equal, which seems to be beneficial to obtain eventually a good decay estimate. Moreover, it will simplify the proof of Lemma \[lem:mu:1D\]. In the following lemma we establish an infinite dimensional analog of Lemma \[lemma:Pdefinition\] – for , the transformed linearized BGK equation in 1D. However, here we shall not aim at obtaining the optimal decay constant $\mu$ in the matrix inequality . Still, $\mu$ will be independent of the modal index $k\in{\mathbb{Z}}$, thus providing exponential decay of the full solution.
\[lem:mu:1D\] For each cell length $L>0$ we consider $\alpha^{(3)}=\alpha^{(3)}(L)>0$ defined in . If the matrices ${\mathbf{P}}_k$ are chosen with some $\alpha \in (0,\alpha^{(3)})$, $\beta =\sqrt{2} \alpha$, and $\gamma =\sqrt{3} \alpha$ uniformly for all $|k|\in{\mathbb{N}}$, then ${\mathbf{P}}_k$ from and ${\mathbf{C}}_k^*{\mathbf{P}}_k + {\mathbf{P}}_k{\mathbf{C}}_k$ are positive definite for all $k\in{\mathbb{Z}}\setminus\{0\}$. Moreover, $$\label{goodbnd}
{\mathbf{C}}_k^*{\mathbf{P}}_k + {\mathbf{P}}_k{\mathbf{C}}_k \geq 2\mu {\mathbf{P}}_k \qquad \text{ uniformly in } |k|\in{\mathbb{N}}\ ,$$ with $$\mu := \frac{\delta_3(1,\alpha)}{8(1-2\pi\alpha/L)^2 (1+\alpha \sqrt{3+\sqrt{6}})}>0\ ,$$ where $\delta_3(1,\alpha):= \det {\mathbf{D}}_{1,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}^{(3)}>0$ with the matrix ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}^{(3)}$ defined in .
The proof of this lemma is deferred to Appendix \[A1\].
\[rem3.2\]
1. Consider $$\alpha_* = \operatorname*{arg\,max}_{\alpha\in[0,\alpha^{(3)}]} \frac{\delta_3(1,\alpha)}{8(1-2\pi\alpha/L)^2 (1+\alpha \sqrt{3+\sqrt{6}})} \ .$$ Choosing ${\mathbf{P}}_k$ with $\alpha=\alpha_*$, $\beta =\sqrt{2} \alpha$, and $\gamma =\sqrt{3} \alpha$ uniformly for all $|k|\in{\mathbb{N}}$, yields with the constant $$\label{goodbnd2}
\mu_* = \frac{\delta_3(1,\alpha_* )}{8(1-2\pi\alpha_* /L)^2 (1+\alpha_* \sqrt{3+\sqrt{6}})}
= \max_{\alpha\in[0,\alpha^{(3)}]} \frac{\delta_3(1,\alpha)}{8(1-2\pi\alpha/L)^2 (1+\alpha \sqrt{3+\sqrt{6}})} \ .$$
2. In the limit $L\to \infty$, the matrix ${\mathbf{C}}_k^*{\mathbf{P}}_k + {\mathbf{P}}_k{\mathbf{C}}_k$ has zero eigenvalues, which is apparent from its upper left submatrix ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}$ defined in . Accordingly, $\alpha^{(3)} \to 0$ with $\alpha^{(3)} =O(\frac1L)$ and $\mu_*=O(\frac1{L^2})$ in the limit $L\to \infty$. It is no surprise that the exponential decay rate vanishes in this limit, as the limiting whole space problem only exhibits algebraic decay (cf. [@BDMMS17] for the large-time analysis of on ${\mathbb{R}}^d$).
In the limit $L\to 0$, again $\alpha^{(3)} \to 0$ with $\alpha^{(3)} =O(L)$. Using $$\label{alpha-limit}
\lim_{L\to0} \frac{\alpha_*(L)}{L}
= \frac{4-\sqrt{13}}{6\pi}\,,$$ we obtain $$\lim_{L\to0} \mu_*(L)= 3 (4-\sqrt{13}) \frac{(3-\sqrt{13})^2}{(1-\sqrt{13})^2} = 0.06391670961...$$ (cf. Fig. \[fig:muL\]).
![For each cell length $L$ the constant $2\mu_*(L)$ obtained from Lemma \[lem:mu:1D\] and Remark \[rem3.2\](a) yields a bound for the entropy decay rate in Theorem \[linBGK-decay\]. []{data-label="fig:muL"}](BGK-1D-plot-L-0-20-mu.eps)
Applying Lemma \[lem:mu:1D\] to each $x$-Fourier mode $\hat{\bf h}_k(t),\,k\ne0$ from allows to prove exponential decay of the linearized BGK equation in 1D:
We consider a solution $h$ of , and let the entropy functional ${\mathcal{E}}(\tilde{f})$ be defined by $$ {\mathcal{E}}(\tilde{f}) := \sum_{k\in {\mathbb{Z}}} \langle h_k(v), {\mathbf{P}}_k h_k(v)\rangle_{L^2(M_1^{-1})}\ ,$$ with $\tilde{f}(t) :=M_1 +h(t)$. Here, the “infinite matrices” ${\mathbf{P}}_0 := {\mathbf{I}}$ and ${\mathbf{P}}_k$ defined in for $k\neq 0$ are regarded as bounded operators on $L^2(M_1^{-1})$. Then $$ {\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}{\mathcal{E}}(\tilde{f}) = -\sum_{k\in {\mathbb{Z}}} \langle h_k(v), ({\mathbf{C}}_k^*{\mathbf{P}}_k + {\mathbf{P}}_k{\mathbf{C}}_k) h_k(v)\rangle_{L^2(M_1^{-1})}
\leq - 2\min\{1,\,\mu_*\} \,{\mathcal{E}}(\tilde{f}) \ ,
$$ where $1$ is the decay rate of $h_0$, cf. . This implies with $\lambda^1(L):=2\min\{1,\,\mu_*\}$ and $\mu_*$ from .
The constants $c_1$ and $C_1$ in the estimate follow from : $$\label{norm-const:1D}
c_1(L)=\left(1+\alpha_*(L)\sqrt{3+\sqrt6}\right)^{-1}\,,\quad
C_1(L)=\left(1-\alpha_*(L)\sqrt{3+\sqrt6}\right)^{-1}\,$$ and this finishes the proof of Theorem \[linBGK-decay\] in 1D.
To appreciate the above decay estimate, let us compare it to the spectral gap obtained in numerical tests for $L=2\pi$. In this case the estimate from Remark \[rem3.2\] yields the analytic bound with $\mu_*=0.041812...$. As a comparison we computed the spectrum of finite dimensional approximation matrices to ${\mathbf{L}}_2+ik{\mathbf{L}}_1$ up to the matrix size $n=500$. Apparently the spectral gap is determined by the lowest spatial modes $k=\pm1$. With increasing $n$ it grows monotonically to $\mu_{num}=0.558296...$. So, our estimate is off by a factor of about 13. Following the strategy from §4.3 in [@AAC16], i.e. by maximizing $\mu$ in the matrix inequality ${\mathbf{C}}_k^*{\mathbf{P}}_k+{\mathbf{P}}_k{\mathbf{C}}-2\mu{\mathbf{P}}_k\ge0$, the above estimate on the decay rate could be improved further. But we shall not pursue this strategy here again.
Let us briefly compare this gap to the situation in the two 1D BGK models analyzed in §4.3 and §4.4 of [@AAC16]. They only differ from the 1D model - of this section, concerning the matrix ${\mathbf{L}}_2$: there we had ${\mathbf{L}}_2=\operatorname{diag}(0,\,1,\,...)$ and ${\mathbf{L}}_2=\operatorname{diag}(0,\,1,\,0,\,1,\,...)$, resp. We recall from §\[sec-hypo-index\] that both models have hypocoercivity index 1, and their spectral gaps are 0.6974... and 0.3709717660..., resp. One might expect that removing 1 entries from ${\mathbf{L}}_2$ and hence increasing the hypocoercivity index would decrease the spectral gap. But this is obviously not always the case.
Linearized BGK equation in 2D {#sec:linBGK:2D}
=============================
Next we shall analyze the linearized BGK equation in 2D: $$\begin{gathered}
\partial_t h(x,v,t) + v\cdot \nabla_x h(x,v,t) \\
= M_1(v)\left[\big( 2 - \frac{|v|^2}{2}\big)\sigma(x,t)+v\cdot \mu(x,t)
+ \big( -\frac{1}{2}+ \frac{|v|^2}{4}\big)\tau(x,t)\right] - h(x,v,t)\ , \quad t\geq 0 \ ,\end{gathered}$$ for the perturbation $h(x,v,t) \approx f(x,v,t) -M_1(v)$ with $x\in\tilde{\mathbf{T}}^2$, $v\in{\mathbb{R}}^2$.
Again we consider the $x$–Fourier series of $h$: $$ h(x,v,t) = \sum_{k\in{\mathbb{Z}}^2} h_k(v,t)\,e^{i\frac{2\pi}{L} k\cdot x}\ ;$$ each spatial mode $h_k(v,t)$ is decoupled and evolves as $$\label{linBGK:Fourier:2D}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_k + i\tfrac{2\pi}{L} k\cdot v\ h_k =
M_1(v)\left[\big( 2 - \frac{|v|^2}{2}\big)\sigma_k(t)+v\cdot \mu_k(t)
+ \big( -\frac{1}{2}+ \frac{|v|^2}{4}\big)\tau_k(t)\right] - h_k(v,t)\ , \quad t\geq 0 \ .$$ Here, $\sigma_k$, $\mu_k$ and $\tau_k$ denote the spatial modes of the $v$–moments $\sigma$, $\mu$ and $\tau$ defined in ; hence $$\sigma_k := \int_{{\mathbb{R}}^2} h_k(v,t){\,\operatorname{d}\!v}\ , \quad
\mu_k := \int_{{\mathbb{R}}^2} v\ h_k(v,t){\,\operatorname{d}\!v}\ , \quad
\tau_k := \int_{{\mathbb{R}}^2} |v|^2\ h_k(v,t){\,\operatorname{d}\!v}\ .$$
Next we shall introduce an orthonormal basis in $v$-direction, to represent the spatial modes $h_k(\cdot,t)\in L^2({\mathbb{R}}^2;M_1^{-1})$, $k\in{\mathbb{Z}}^2$. As in 1D we shall again use Hermite functions. But their multi-dimensional generalization is not unique, and we shall present two options that seem to be practical:\
A complete set of orthogonal polynomials in $d$ variables can be formed as products of $d$ such polynomials, each in a single variable. Using the Hermite polynomials $H_m$ in 1D, i.e. $$H_0(\upsilon) =1\ , \quad H_1(\upsilon) =\upsilon\ , \quad H_2(\upsilon) =\upsilon^2 -1\ , \quad H_3(\upsilon) =\upsilon^3 -3\upsilon\ ,\ ... \quad \text{with } \upsilon\in{\mathbb{R}}\ ,$$ we construct Hermite polynomials on ${\mathbb{R}}^d$ as $$\label{hermite-polynom:2D:a}
H_m(v) := \prod_{j=1}^d H_{m_j}(v_j)\ ,\quad v\in{\mathbb{R}}^d\ ,$$ with the multi-index $m=(m_1,\ldots,m_d)\in{\mathbb{N}}_0^d$. They are also generated by a simple generalization of formula : $$ H_m(v) = (-1)^{|m|} e^{\frac{|v|^2}{2}}\frac{\partial^{|m|}}{\partial v^m}e^{-\frac{|v|^2}{2}}\ ,\quad m\in{\mathbb{N}}_0^d\ ,$$ with $|m|=\sum_{j=1}^d m_j$ (see [@Gr49], e.g.).
For $d=2$, we obtain $$\begin{aligned}
H_{0,0}(v) &=H_0(v_1) H_0(v_2) =1\ , \quad H_{1,0}(v) =H_1(v_1) H_0(v_2) =v_1\ , \quad H_{0,1}(v) =H_0(v_1) H_1(v_2) =v_2\ , \\
H_{2,0}(v) &=H_2(v_1) H_0(v_2) =v_1^2 -1\ , \quad H_{1,1}(v) =H_1(v_1) H_1(v_2) =v_1 v_2\ , \quad H_{0,2}(v) =H_0(v_1) H_2(v_2) =v_2^2 -1\ , \\
H_{3,0}(v) &=H_3(v_1) H_0(v_2) =v_1^3 -3v_1\ , \quad H_{2,1}(v) =H_2(v_1) H_1(v_2) =(v_1^2 -1) v_2\ , \\
H_{1,2}(v) &=H_1(v_1) H_2(v_2) =v_1 (v_2^2 -1)\ , \quad H_{0,3}(v) =H_0(v_1) H_3(v_2) =v_2^3 -3v_2\ .
\end{aligned}$$ Using definition of normalized Hermite functions in 1D, we define the normalized Hermite functions in $d$ dimensions as $$\label{hermite-fct:dD}
g_m(v) := \prod_{j=1}^d g_{m_j} (v_j) \qquad \text{for } m=(m_1,\ldots,m_d)\in{\mathbb{N}}_0^d \ .$$ Then, $g_m$ ($m\in {\mathbb{N}}_0^d$) form an orthonormal basis of $L^2({\mathbb{R}}^d;M_1^{-1})$ and inherit a simple recurrence relation: For $k\in\{1,\ldots,d\}$, $m\in{\mathbb{N}}_0^d$, and the Euclidean basis vectors $e_k =(\delta_{kj})_{j=1,\ldots,d}$ in ${\mathbb{R}}^d$, the recurrence relation $$\label{recur:dD1}
v_k g_m(v) = \sqrt{m_k+1}\ g_{m+e_k}(v) +\sqrt{m_k}\ g_{m-e_k}(v)$$ holds.
In order to give a vector representation of , the evolution equation of the spatial modes $h_k(v,t)$, we first need to introduce a linear ordering of the velocity basis $g_m$ ($m\in {\mathbb{N}}_0^2$). We shall use a lexicographic order, i.e. first (increasingly) with respect to the total order $|m|$, and within a set of order $|m|$ we order w.r.t. $m_1$ (decreasingly) (for $d=2$). Thus, we obtain the linearly ordered basis $$\begin{aligned}
g_0(v) &:=g_{0,0}(v) =M_1(v)\ , \quad
g_1(v) :=g_{1,0}(v) =v_1 M_1(v)\ , \quad
g_2(v) :=g_{0,1}(v) =v_2 M_1(v)\ , \\
g_3(v) &:=g_{2,0}(v) =\tfrac1{\sqrt{2}} (v_1^2 -1) M_1(v)\ , \quad
g_4(v) :=g_{1,1}(v) =v_1 v_2 M_1(v)\ , \quad
g_5(v) :=g_{0,2}(v) =\tfrac1{\sqrt{2}} (v_2^2 -1) M_1(v)\ , \\
g_6(v) &:=g_{3,0}(v) =\tfrac1{\sqrt{3!}} (v_1^3 -3v_1) M_1(v)\ , \quad
g_7(v) :=g_{2,1}(v) =\tfrac1{\sqrt{2}} (v_1^2 -1) v_2 M_1(v)\ , \\
g_8(v) &=g_{1,2}(v) =\tfrac1{\sqrt{2}} v_1 (v_2^2 -1) M_1(v)\ , \quad
g_9(v) :=g_{0,3}(v) =\tfrac1{\sqrt{3!}} (v_2^3 -3v_2) M_1(v)\ , \quad ...
$$ Given a multi-index $m\in{\mathbb{N}}_0^2$, its lexicographic index is computed as $|m| (|m|+1)/2 +m_2$ with $|m| =m_1 +m_2$.\
The second basis is a simple variant of the first one. We recall that the evolution with the BGK equation conserves the (kinetic) energy and mass. Hence, their difference is also conserved and it is related to the polynomial $\frac{|v|^2}2 -1$. In analogy to the 1D case from §\[sec:linBGK:1D\] it is thus a natural option to construct a basis of orthogonal polynomials $\tilde{H}_m(v)$, $m\in{\mathbb{N}}_0^d$, such that $\frac{|v|^2}2 -1$ is a basis element. Compared to $\{H_m(v)\}$, in fact, we only have to modify the Hermite polynomials of second order. For $d=2$ they read: $$\label{Hermite:basis:2D}
\tilde{H}_m(v) :=
\begin{cases}
H_m(v) & \text{if } |m|\ne 2\ , \\
\frac12 \big(H_{2,0}(v) +H_{0,2}(v)\big) = \frac{|v|^2}2 -1 & \text{if } m=(2,0)\ , \\
H_{1,1}(v) = v_1 v_2 & \text{if } m=(1,1)\ , \\
\frac12 \big(H_{2,0}(v) -H_{0,2}(v)\big) = \frac{v_1^2 -v_2^2}2 & \text{if } m=(0,2)\ .
\end{cases}$$ Similarly, we define normalized Hermite functions $$\tilde{g}_m(v) :=
\begin{cases}
g_m(v) & \text{if } |m|\ne 2\ , \\
\frac1{\sqrt2} \big(g_{2,0}(v) +g_{0,2}(v)\big) & \text{if } m=(2,0)\ , \\
g_{1,1}(v) & \text{if } m=(1,1)\ , \\
\frac1{\sqrt2} \big(g_{2,0}(v) -g_{0,2}(v)\big) & \text{if } m=(0,2)\ .
\end{cases}$$ The elements $\tilde{g}_m$ satisfy a recurrence relation similar to , except for identities involving $\tilde{g}_{2,0}$ or $\tilde{g}_{0,2}$. For example, $$\begin{aligned}
v_1 \tilde{g}_{2,0}(v)
&=& \frac1{\sqrt{2}} v_1\ \big(g_{2,0}(v) +g_{0,2}(v)\big)
= \frac1{\sqrt{2}} \big(\sqrt{3}\ g_{3,0}(v) +\sqrt{2}\ g_{1,0}(v) +g_{1,2}(v)\big) \\
&=& \frac1{\sqrt{2}} \big(\sqrt{3}\ \tilde{g}_{3,0}(v) +\sqrt{2}\ \tilde{g}_{1,0}(v) +\tilde{g}_{1,2}(v)\big)\ .
\end{aligned}$$ While the first basis $g_m$ $(m\in{\mathbb{N}}_0^d)$ inherits a simple recurrence formula with three elements, the recurrence formulas for the second basis $\tilde{g}_m$ $(m\in{\mathbb{N}}_0^d)$ involve four elements, when including $\tilde{g}_{2,0}(v)$ or $\tilde{g}_{0,2}(v)$.\
To derive the vector representation of , it is preferable to use the first basis with the linear ordering $g_m$ $(m\in{\mathbb{N}}_0)$. With the identity $(g_3(v) +g_5(v))/\sqrt{2} = (|v|^2/2 -1) M_1(v)$ we rewrite as $$\begin{gathered}
\label{linBGK:Fourier:2Da}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_k + i\tfrac{2\pi}{L} k\cdot v\ h_k =
\big( g_0(v) -\tfrac1{\sqrt{2}} g_3(v) -\tfrac1{\sqrt{2}} g_5(v)\big)\sigma_k(t)
+ \binom{g_1(v)}{g_2(v)}\cdot \mu_k(t) \\
+ \tfrac1{2\sqrt{2}} \big( g_3(v) +g_5(v)\big)\tau_k(t) - h_k(v,t)\ ,
\quad k\in{\mathbb{Z}}^2\ , \quad t\geq 0 \ . \end{gathered}$$ First we consider the spatial mode $h_0$ with $k=0$. With the same argument as in 1D we again obtain , i.e. ${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_0(v,t)=-h_0(v,t)$. Next we expand $h_k(\cdot,t)\in L^2({\mathbb{R}}^2;M_1^{-1})$ in the orthonormal basis $\{g_m(v)\}_{m\in{\mathbb{N}}_0}$: $$h_k(v,t)=\sum_{m=0}^\infty \hat h_{k,m}(t)\,g_m(v)\ ,\quad\text{with}\quad
\hat h_{k,m} = \langle h_k(v),g_m(v)\rangle_{L^2(M_1^{-1})}\ .$$ For each spatial mode $k\in{\mathbb{Z}}^2$, the “infinite vector” $\hat {\bf h}_k(t) = ( \hat h_{k,0}(t),\,\hat h_{k,1}(t),\ ...)^\top\in \ell^2({\mathbb{N}}_0)$ contains all 2D–Hermite coefficients of $h_k(\cdot,t)$. In particular we have $$\hat h_{k,0} = \int_{{\mathbb{R}}^2} h_k(v) g_0(v) M_1^{-1}(v) {\,\operatorname{d}\!v} = \sigma_k \ , \qquad
\binom{\hat h_{k,1}}{\hat h_{k,2}} = \int_{{\mathbb{R}}^2} h_k(v) \binom{g_1(v)}{g_2(v)} M_1^{-1}(v) {\,\operatorname{d}\!v} = \mu_k \in{\mathbb{R}}^2 \ ,$$ and $$\tfrac1{\sqrt{2}} (\hat h_{k,3} +\hat h_{k,5})
= \int_{{\mathbb{R}}^2} h_k(v) \tfrac1{\sqrt{2}} (g_3(v) +g_5(v)) M_1^{-1}(v) {\,\operatorname{d}\!v}
= \frac12\tau_k -\sigma_k\ .$$ Thus, we can rewrite as $$\begin{aligned}
\label{linBGK:Fourier2D}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_k + i\tfrac{2\pi}{L} k\cdot v\ h_k
&= g_0(v) \hat h_{k,0}
+ \binom{g_1(v)}{g_2(v)}\cdot \binom{\hat h_{k,1}}{\hat h_{k,2}}
+\frac{ g_3(v) +g_5(v)}{2} (\hat h_{k,3} +\hat h_{k,5})
- h_k(v,t)\ , \quad k\in{\mathbb{Z}}^2\ , \quad t\geq 0 \ .$$
Our next goal is to rewrite this system in the Hermite function basis as an infinite vector system – in analogy to in 1D. In that equation, the operator ${\mathbf{L}}_1$ is multiplied by the (scalar and integer) mode number $k$, which is then used in the construction of the transformation matrices ${\mathbf{P}}_k$. To extend this structure and strategy to 2D, we first have to consider the rotational symmetry of : We note that the basis functions $g_0$ and $g_3+g_5$ depend only on $|v|$, and that the interplay between the vectors $k$ and $v$ only occurs via $k\cdot v$. Hence, evolution equations from the family having the same modulus $|k|$ are identical in the following sense: Rotating the spatial mode vector $k$ and the $v$-coordinate system by the same angle, leaves invariant. Thus it suffices to consider for vectors $k=(\kappa,0)^\top$ with the discrete moduli $$\kappa\in K:=\big\{r\ge1 \, \big|\, \exists k\in{\mathbb{Z}}^2\setminus\{0\} \mbox{ with } r=|k| \big\} \ .$$ We skipped here the mode $h_0$, as it was already analyzed before. In the sequel we also denote $h_\kappa:=h_{\kappa,0}$ and $\hat {\bf h}_\kappa:=\hat {\bf h}_{\kappa,0}$. With this notation, reads $${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_\kappa + i\tfrac{2\pi}{L} \kappa v_1\ h_\kappa
= g_0(v) \hat h_{\kappa,0}
+ \binom{g_1(v)}{g_2(v)}\cdot \binom{\hat h_{\kappa,1}}{\hat h_{\kappa,2}}
+\frac{ g_3(v) +g_5(v)}{2} (\hat h_{\kappa,3} +\hat h_{\kappa,5})
- h_\kappa(v,t)\ , \quad \kappa\in K \ .$$ Then, the vector of its Hermite coefficients satisfies $$\label{linap4:2D}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}\hat {\bf h}_\kappa(t) + i\,\tfrac{2\pi}{L}\kappa\, {\mathbf{L}}_1 \hat{\bf h}_\kappa(t) = -{\mathbf{L}}_2 \hat{\bf h}_\kappa(t)\ ,\quad \kappa\in K\ ,\quad\ t\ge0\ ,$$ where the operators ${\mathbf{L}}_1,\,{\mathbf{L}}_2$ are represented by symmetric “infinite matrices” on $\ell^2({\mathbb{N}}_0)$: $$\begin{aligned}
{\mathbf{L}}_1&=& \begin{pmatrix}
0 & \sqrt 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
\sqrt 1 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & \sqrt 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & \sqrt 2 & 0 & 0 & 0 & 0 & \sqrt 3 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & \sqrt 1 & 0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 1 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & \sqrt 3 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 4 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 3 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & \sqrt 1 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 2 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 1 & \cdots \\
\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots & \ddots
\end{pmatrix}
\ ,\quad\\
{\mathbf{L}}_2&=&\operatorname{diag}(0,\, 0,\, 0,\, \tfrac12 \begin{pmatrix} 1 & 0 & -1 \\ 0 & 2 & 0 \\ -1 & 0 & 1 \end{pmatrix},\, 1,\, 1,\cdots)\ .\end{aligned}$$
To compute the hypocoercivity index of the BGK model in 2D, it is preferable to use the second basis with the linear ordering $\tilde{g}_m$ $(m\in{\mathbb{N}}_0)$. We shall give the matrix representation of the two dimensional BGK equation in the second velocity basis, again only for the spatial modes $k=(\kappa,0)^\top, \,\kappa\in K$. To obtain the corresponding matrices $\tilde{\mathbf{L}}_1$ and $\tilde{\mathbf{L}}_2$, we simply represent the linear basis transformation by the infinite matrix $${\mathbf{S}}= \operatorname{diag}(1,\ 1,\ 1,\ \tfrac1{\sqrt{2}} \begin{pmatrix} 1 & 0 & 1 \\ 0 & \sqrt{2} & 0 \\ 1 & 0 & -1 \end{pmatrix},\ 1,\ 1,\cdots)\ ,$$ which is self-inverse, i.e. ${\mathbf{S}}={\mathbf{S}}^{-1}$. Thus we compute $\tilde{{\mathbf{L}}}_1 = {\mathbf{S}}^{-1} {\mathbf{L}}_1 {\mathbf{S}}$ and $\tilde{{\mathbf{L}}}_2 = {\mathbf{S}}^{-1} {\mathbf{L}}_2 {\mathbf{S}}$, yielding $$\begin{aligned}
&&\tilde{\mathbf{L}}_1= \begin{pmatrix}
0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
1 & 0 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 1 & 0 & 0 & 0 & 0 & \sqrt{3/2} & 0 & \sqrt{1/2} & 0 & 0 & \cdots \\
0 & 0 & 1 & 0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & \cdots \\
0 & 1 & 0 & 0 & 0 & 0 & \sqrt{3/2} & 0 & -\sqrt{1/2} & 0 & 0 & \cdots \\
0 & 0 & 0 & \sqrt{3/2} & 0 & \sqrt{3/2} & 0 & 0 & 0 & 0 & 2 & \cdots \\
0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & \sqrt{1/2} & 0 & -\sqrt{1/2} & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & \cdots \\
\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots & \vdots & \vdots & \ddots
\end{pmatrix}
\ ,\\[3mm]
&&\tilde{\mathbf{L}}_2=\operatorname{diag}(0,\, 0,\, 0,\, 0,\, 1,\, 1,\cdots)\ .\end{aligned}$$
This second basis representation makes it easy of determine the hypocoercivity index of the BGK model in 2D. As for the 1D model, we use Definition \[hyp-index\] in its equivalent formulation \[cond:trivkernel\]: With ${\operatorname{corank}}\tilde{\mathbf{L}}_2=4$, $\ker\tilde{\mathbf{L}}_2={\operatorname{span}}\{e_0,\,e_1,\,e_2,\,e_3\}$, and the relations $$\label{linBGK2D:rangeL1}
\tilde{\mathbf{L}}_1 e_0=e_1\ ,\quad \tilde{\mathbf{L}}_1 e_1=e_0+e_3+e_5\ ,\quad \tilde{\mathbf{L}}_1 e_2=e_4\ ,\quad \tilde{\mathbf{L}}_1 e_3=e_1+\sqrt{3/2}\, e_6+\sqrt{1/2}\, e_8\ ,$$ we find the index $\tau=2$. At first glance this may come as a surprise, since the analogous 1D model has index 3. But in 2D, each of the two momentum-conservation modes (represented by $e_1$ and $e_2$) is directly coupled to a decreasing mode (represented by $e_5$ and $e_4$, respectively). These modes are quadratic polynomials, but orthogonal to $|v|^2$, where the latter corresponds to the (conserved) kinetic energy, cf. .\
We define the matrices ${\mathbf{C}}_\kappa := i\tfrac{2\pi}{L} \kappa \tilde{\mathbf{L}}_1 + \tilde{\mathbf{L}}_2$, $\kappa\in K\cup\{0\}$ which determine the evolution of the spatial modes of the BGK equation in 2D, cf. . Our next goal is to establish a spectral gap of ${\mathbf{C}}_\kappa$, uniformly in $\kappa\in K$. This will prove Theorem \[linBGK-decay\] in 2D. To this end we make an ansatz for the transformation matrices: Following the detailed motivation from the 1D analog in §\[sec:linBGK:1D\], let ${\mathbf{P}}_\kappa$, $\kappa\in K$ be the identity matrix whose upper-left $7\times 7$ block is replaced by $$\label{P:2D}
\begin{pmatrix}
1 & -i \alpha/\kappa & 0 & 0 & 0 & 0 & 0 \\
i \alpha/\kappa & 1 & 0 & 0 & 0 & -i \beta/\kappa & 0 \\
0 & 0 & 1 & 0 & -i \gamma/\kappa & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 & -i \omega/\kappa \\
0 & 0 & i \gamma/\kappa & 0 & 1 & 0 & 0 \\
0 & i \beta/\kappa & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & i \omega/\kappa & 0 & 0 & 1
\end{pmatrix}$$ with positive parameters $\alpha$, $\beta$, $\gamma$, and $\omega$ to be chosen below. The distribution of the non-zero off-diagonal elements follows from the pattern in matrix $\tilde{\mathbf{L}}_1$, with the following rationale: The $\alpha$-term couples the $e_0$-mode to the $e_1$-mode, which is the only choice according to . The $\beta$-term couples the $e_1$-mode to the decaying $e_5$-mode, and the $\gamma$-term couples the $e_2$-mode to the decaying $e_4$-mode. Finally, the $\omega$-term couples the $e_3$-mode to the $e_6$-mode, the first decaying mode according to .
\[lemma:decay:2D\] If the matrices ${\mathbf{P}}_\kappa$ are chosen as with $\beta =2\alpha$, $\gamma =\alpha$, and $\omega=\sqrt{6}\alpha$ uniformly for all $\kappa\in K$, then there exists $0<\alpha_+$ such that ${\mathbf{P}}_\kappa$ and ${\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa$ are positive definite for all $\alpha \in (0,\alpha_+)$ and $\kappa\in K$. Moreover, $$\label{goodbnd:2D}
{\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa \geq 2\mu {\mathbf{P}}_\kappa \qquad \text{ uniformly in } \kappa\in K \ ,$$ with $$\mu := \left( \frac{10}{14} \right)^{10} \frac{\delta_{11}(1,\alpha,2\alpha,\alpha,\sqrt{6}\alpha)}{2 \big(1+\sqrt{6} \alpha\big)}>0\ ,$$ where $\delta_{11}(1,\alpha,2\alpha,\alpha,\sqrt{6}\alpha):= \det {\mathbf{D}}_{1,\alpha,2\alpha,\alpha,\sqrt{6}\alpha}$ with ${\mathbf{D}}_{\kappa,\alpha,2\alpha,\alpha,\sqrt{6}\alpha}$ defined in .
The proof of this lemma is deferred to Appendix \[A1\].
1. Consider $$\alpha_* = \operatorname*{arg\,max}_{\alpha\in[0,\alpha_+]} \frac{\delta_{11}(1,\alpha,2\alpha,\alpha,\sqrt{6}\alpha)}{2 \big(1+\sqrt{6} \alpha\big)} \ .$$ Choosing ${\mathbf{P}}_\kappa$ with $\alpha=\alpha_*$, $\beta =2\alpha$, $\gamma =\alpha$, and $\omega=\sqrt{6}\alpha$ uniformly for all $\kappa\in K$, yields with the maximal constant $$\label{goodbnd2:2D:b}
\mu = \left( \frac{10}{14} \right)^{10} \frac{\delta_{11}(1,\alpha_*,2\alpha_*,\alpha_*,\sqrt{6}\alpha_*)}{2 \big(1+\sqrt{6} \alpha_*\big)}
= \max_{\alpha\in[0,\alpha_+]} \left( \frac{10}{14} \right)^{10} \frac{\delta_{11}(1,\alpha,2\alpha,\alpha,\sqrt{6}\alpha)}{2 \big(1+\sqrt{6} \alpha\big)} \ .$$
2. For $L=2\pi$, we compute $\alpha_+ = 0,2102380141$. Moreover, the constant $\mu$ is determined as $\mu = 0,003013362117$ with $\alpha_* = 0,1453311384$.
3. In the limit $L\to +\infty$, the matrix ${\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa$ has zero eigenvalues, which is apparent from its upper left submatrix ${\mathbf{D}}_{\kappa,\alpha,2\alpha,\alpha,\sqrt{6}\alpha}$ defined in . Accordingly, $\alpha_+ \to 0$ in the limit $L\to \infty$.
Moreover, $\alpha_+ \to 0$ in the limit $L\to 0$.
We consider a solution $h$ of , and the entropy functional ${\mathcal{E}}(\tilde{f})$ defined by $$ {\mathcal{E}}(\tilde{f}) := \sum_{k\in {\mathbb{Z}}^2} \langle h_k(v), {\mathbf{P}}_{|k|} h_k(v)\rangle_{L^2(M_1^{-1})}\ ,$$ with $\tilde{f}(t) := M_1 +h(t)$. Here the matrices ${\mathbf{P}}_0 = {\mathbf{I}}$ and ${\mathbf{P}}_\kappa$ defined in for $\kappa=|k|\neq 0$ are regarded as bounded operators on $L^2(M_1^{-1})$. Then $$ {\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}{\mathcal{E}}(\tilde{f}) = -\sum_{k\in {\mathbb{Z}}^2} \langle h_k(v), ({\mathbf{C}}_{|k|}^*{\mathbf{P}}_{|k|} + {\mathbf{P}}_{|k|}{\mathbf{C}}_{|k|}) h_k(v)\rangle_{L^2(M_1^{-1})}
\leq - 2\min\{1,\mu\} \,{\mathcal{E}}(\tilde{f}) \ ,
$$ where $1$ is the decay rate of $h_0$, cf. . This implies with $\lambda^2(L):=2\min\{1,\mu\}$ and $\mu$ from .
The constants $c_2$ and $C_2$ in the estimate follow from : $$ c_2(L)=\left(1 +\sqrt{6} \alpha_* \right)^{-1}\,,\quad
C_2(L)=\left(1 -\sqrt{6} \alpha_* \right)^{-1}\,.$$ This finishes the proof of Theorem \[linBGK-decay\] in 2D.
Linearized BGK equation in 3D {#sec:linBGK:3D}
=============================
Next we shall analyze the linearized BGK equation in 3D: $$\begin{gathered}
\partial_t h (x,v,t) + v\cdot \nabla_x h(x,v,t) \\
= M_1(v)\left[\big( \frac52 - \frac{|v|^2}{2}\big)\sigma(x,t)+v\cdot \mu(x,t)
+ \big( -\frac{1}{2}+ \frac{|v|^2}{6}\big)\tau(x,t)\right] - h(x,v,t)\ , \quad t\geq 0 \ ,\end{gathered}$$ for the perturbation $h(x,v,t) \approx f(x,v,t) -M_1(v)$ with $x\in\tilde{\mathbf{T}}^3$, $v\in{\mathbb{R}}^3$.
Again we consider the $x$–Fourier series of $h$: $$ h(x,v,t) = \sum_{k\in{\mathbb{Z}}^3} h_k(v,t)\,e^{i\frac{2\pi}{L} k\cdot x}\ .$$ Each spatial mode $h_k(v,t)$ is decoupled and evolves as $$\label{linBGK:Fourier:3D}
\partial_t h_k + i\tfrac{2\pi}{L} k\cdot v\ h_k =
M_1(v)\left[\big( \frac52 - \frac{|v|^2}{2}\big)\sigma_k(t)+v\cdot \mu_k(t)
+ \big( -\frac{1}{2}+ \frac{|v|^2}{6}\big)\tau_k(t)\right] - h_k(v,t)\ , \quad t\geq 0 \ .$$ Here, $\sigma_k$, $\mu_k$ and $\tau_k$ denote the spatial modes of the $v$–moments $\sigma$, $\mu$ and $\tau$ defined in ; hence $$\sigma_k := \int_{{\mathbb{R}}^3} h_k(v,t){\,\operatorname{d}\!v}\ , \quad
\mu_k := \int_{{\mathbb{R}}^3} v\ h_k(v,t){\,\operatorname{d}\!v}\ , \quad
\tau_k := \int_{{\mathbb{R}}^3} |v|^2\ h_k(v,t){\,\operatorname{d}\!v}\ .$$ Next we introduce an orthonormal basis in $v$-direction, to represent the spatial modes $h_k(\cdot,t)\in L^2({\mathbb{R}}^3;M_1^{-1})$, $k\in{\mathbb{Z}}^3$. Again we will use Hermite functions. As in 2D, their multi-dimensional generalization is not unique, and we present two options which seem to be practical:\
This is a straightforward generalization of the 2D case. Using and the normalized 1D-Hermite functions $g_n$ ($n\in{\mathbb{N}}_0$), we define the normalized Hermite functions in 3D as in , $$ g_m(v) := \prod_{j=1}^3 g_{m_j} (v_j) \qquad \text{for } m=(m_1,\ldots,m_3)\in{\mathbb{N}}_0^3 \ .$$ Then, $g_m$ ($m\in{\mathbb{N}}_0^3$) form an orthonormal basis of $L^2({\mathbb{R}}^3;M_1^{-1})$ and inherit a simple recurrence relation .
As in 2D, we shall use a lexicographic order, i.e. we order $\{g_m\}$ first (increasingly) with respect to the total order $|m|$, and within a set of order $|m|$, we order first (decreasingly) with respect to $m_1$, and then $m_2$. Thus, we obtain the linearly ordered basis [$$\begin{aligned}
g_0(v) &=g_{0,0,0}(v) =M_1(v)\ , \\
g_1(v) &=g_{1,0,0}(v) =v_1 M_1(v)\ , \quad
g_2(v) =g_{0,1,0}(v) =v_2 M_1(v)\ , \quad
g_3(v) =g_{0,0,1}(v) =v_3 M_1(v)\ , \\
g_4(v) &=g_{2,0,0}(v) =\tfrac1{\sqrt{2}} (v_1^2 -1) M_1(v)\ , \quad
g_5(v) =g_{1,1,0}(v) =v_1 v_2 M_1(v)\ , \quad
g_6(v) =g_{1,0,1}(v) =v_1 v_3 M_1(v)\ , \\
g_7(v) &=g_{0,2,0}(v) =\tfrac1{\sqrt{2}} (v_2^2 -1) M_1(v)\ , \quad
g_8(v) =g_{0,1,1}(v) =v_2 v_3 M_1(v)\ , \quad
g_9(v) =g_{0,0,2}(v) =\tfrac1{\sqrt{2}} (v_3^2 -1) M_1(v)\ , \\
g_{10}(v) &=g_{3,0,0}(v) =\tfrac1{\sqrt{3!}} (v_1^3 -3v_1) M_1(v)\ , \quad
g_{11}(v) =g_{2,1,0}(v) =\tfrac1{\sqrt{2}} (v_1^2 -1) v_2 M_1(v)\ , \quad
g_{12}(v) =g_{2,0,1}(v) =\tfrac1{\sqrt{2}} (v_1^2 -1) v_3 M_1(v)\ , \\
g_{13}(v) &=g_{1,2,0}(v) =\tfrac1{\sqrt{2}} v_1 (v_2^2 -1) M_1(v)\ , \quad
g_{14}(v) =g_{1,1,1}(v) = v_1 v_2 v_3 M_1(v)\ , \quad
g_{15}(v) =g_{1,0,2}(v) =\tfrac1{\sqrt{2}} v_1 (v_3^2 -1) M_1(v)\ , \\
g_{16}(v) &=g_{0,3,0}(v) =\tfrac1{\sqrt{3!}} (v_2^3 -3v_2) M_1(v)\ , \quad
g_{17}(v) =g_{0,2,1}(v) = \tfrac1{\sqrt{2}} (v_2^2 -1) v_3 M_1(v)\ , \quad
g_{18}(v) =g_{0,1,2}(v) =\tfrac1{\sqrt{2}} v_2 (v_3^2 -1) M_1(v)\ , \\
g_{19}(v) &=g_{0,0,3}(v) =\tfrac1{\sqrt{3!}} (v_3^3 -3v_3) M_1(v)\ , \quad
g_{20}(v) =g_{4,0,0}(v) =\tfrac1{\sqrt{4!}} (v_1^4 -6v_1^2 +3) M_1(v) , \quad \ldots \ .\end{aligned}$$ ]{} In analogy to the 2D case from §\[sec:linBGK:2D\], it is natural to construct a basis of orthogonal polynomials $\tilde{H}_m(v)$ ($m\in{\mathbb{N}}_0^3$) that involves the kinetic energy polynomial $|v|^2/2$ (minus a multiple of the mass); in 3D the relevant term is $(|v|^2 -3)/2$. Again, we only have to modify the Hermite polynomials of second order: $$\tilde{H}_m(v) =
\begin{cases}
\tfrac12 (H_{2,0,0}(v) +H_{0,2,0} +H_{0,0,2}(v)) = \tfrac{|v|^2}{2} -\tfrac32 & \text{if } m=(2,0,0)\ , \\
H_{2,0,0}(v) -\tfrac12 (1+\sqrt{3}) H_{0,2,0} +\tfrac12 (\sqrt{3}-1) H_{0,0,2}(v) & \text{if } m=(0,2,0)\ , \\
H_{2,0,0}(v) +\tfrac12 (\sqrt{3}-1) H_{0,2,0} -\tfrac12 (1+\sqrt{3}) H_{0,0,2}(v) & \text{if } m=(0,0,2)\ , \\
H_m(v) & \text{else.}
\end{cases}$$ Similarly, we define normalized Hermite functions $$\tilde{g}_m(v) =
\begin{cases}
\frac1{\sqrt3} \big(g_{2,0,0}(v) +g_{0,2,0}(v) +g_{0,0,2}(v)\big) & \text{if } m=(2,0,0)\ , \\
\frac1{\sqrt3} \big(g_{2,0,0}(v) -\tfrac12 (1+\sqrt{3}) g_{0,2,0}(v) +\tfrac12 (\sqrt{3}-1) g_{0,0,2}(v)\big) & \text{if } m=(0,2,0)\ , \\
\frac1{\sqrt3} \big(g_{2,0,0}(v) +\tfrac12 (\sqrt{3}-1) g_{0,2,0}(v) -\tfrac12 (1+\sqrt{3}) g_{0,0,2}(v)\big) & \text{if } m=(0,0,2)\ , \\
g_m(v) & \text{else.}
\end{cases}$$ We remark that it is most convenient to obtain $\tilde{g}_{0,2,0}$ and $\tilde{g}_{0,0,2}$ from diagonalizing the matrix ${\mathbf{L}}_2$ (see below). The elements $\tilde{g}_m$ satisfy a recurrence relation similar to ; except for identities involving $\tilde{g}_{2,0,0}$, $\tilde{g}_{0,2,0}$ or $\tilde{g}_{0,0,2}$. For example, $$\begin{aligned}
v_1 \tilde{g}_{2,0,0}(v)
&= \frac1{\sqrt3} v_1 \big(g_{2,0,0}(v) +g_{0,2,0}(v) +g_{0,0,2}(v)\big) \\
&= \frac1{\sqrt3} \big(\sqrt{3} g_{3,0,0}(v) +\sqrt{2} g_{1,0,0}(v) +g_{1,2,0}(v) +g_{1,0,2}(v)\big) \ .
\end{aligned}$$ Whereas the first basis $g_m(v)$ ($m\in{\mathbb{N}}_0^3$) inherits a simple recurrence formula with three elements; for the second basis $\tilde{g}_m(v)$ ($m\in{\mathbb{N}}_0^3$) the recurrence formula for $\tilde{g}_{2,0,0}(v)$ relates five elements.
To derive the vector representation of , it is preferable to use the first basis with the linear ordering $g_m(v)$, $m\in{\mathbb{N}}_0$. With the identity $(g_4(v) +g_7(v) +g_9(v))/\sqrt{2} = (|v|^2 -3) M_1(v)/2$, we rewrite as $$\label{linBGK:Fourier:3D:2}
\partial_t h_k + i\tfrac{2\pi}{L} k\cdot v\ h_k =
g_0(v) \sigma_k(t)
+ \begin{pmatrix} g_1(v) \\ g_2(v) \\ g_3(v) \end{pmatrix}\cdot \mu_k(t)
+ \big( g_4(v) +g_7(v) +g_9(v)\big) \frac{\tau_k(t) -3\sigma_k(t)}{3\sqrt{2}}
- h_k(v,t)
$$ for $t\geq 0$. As in 1D, the spatially homogeneous mode again satisfies ${\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}h_0(v,t)=-h_0(v,t)$, cf. . Next we expand $h_k(\cdot,t)\in L^2({\mathbb{R}}^3;M_1^{-1})$ in the orthonormal basis $g_m$ ($m\in{\mathbb{N}}_0$): $$h_k(v,t)=\sum_{m=0}^\infty \hat h_{k,m}(t)\,g_m(v) \qquad\text{with}\quad
\hat h_{k,m} = \langle h_k(v),g_m(v)\rangle_{L^2(M_1^{-1})}\ .$$ For each spatial mode $k\in{\mathbb{Z}}^3$, the “infinite vector” $\hat {\bf h}_k(t) = ( \hat h_{k,0}(t),\,\hat h_{k,1}(t),\ ...)^\top\in \ell^2({\mathbb{N}}_0)$ contains all Hermite coefficients of $h_k(\cdot,t)$. In particular, we have $$\hat h_{k,0} = \sigma_k \ , \qquad
\big(\hat h_{k,1}\,,\ \hat h_{k,2}\,,\ \hat h_{k,3} \big)^\top = \mu_k \in{\mathbb{R}}^3 \ , \qquad
\tfrac1{\sqrt{2}} (\hat h_{k,4} +\hat h_{k,7} +\hat h_{k,9}) = \frac12\tau_k -\frac32 \sigma_k\ .$$ Thus, we can rewrite as $$\label{linBGK:Fourier:3D:3}
\partial_t h_k + i\tfrac{2\pi}{L} k\cdot v\ h_k
= g_0\ \hat h_{k,0}
+ \begin{pmatrix} g_1 \\ g_2 \\ g_3 \end{pmatrix}\cdot \begin{pmatrix} \hat h_{k,1} \\ \hat h_{k,2} \\ \hat h_{k,3} \end{pmatrix}
+\tfrac1{3} \big( g_4 +g_7 +g_9\big) (\hat h_{k,4} +\hat h_{k,7} +\hat h_{k,9})
- h_k(v,t)\ ,
$$ for $k\in{\mathbb{Z}}^3$, $t\geq 0$. Since is rotationally invariant (as in 2D), it suffices to consider for vectors $k=(\kappa,0)^\top$ with the discrete moduli $$\kappa\in K:=\big\{r\ge1 \, \big|\, \exists k\in{\mathbb{Z}}^3\setminus\{0\} \mbox{ with } r=|k| \big\} \ .$$ With the notation $h_\kappa:=h_{\kappa,0}$ and $\hat {\bf h}_\kappa:=\hat {\bf h}_{\kappa,0}$, reads $$\partial_t h_\kappa + i\tfrac{2\pi}{L} \kappa v_1\ h_\kappa
= g_0 \hat h_{\kappa,0}
+ \begin{pmatrix} g_1 \\ g_2 \\ g_3 \end{pmatrix}\cdot \begin{pmatrix} \hat h_{\kappa,1} \\ \hat h_{\kappa,2} \\ \hat h_{\kappa,3} \end{pmatrix}
+\tfrac1{3} \big( g_4 +g_7 +g_9\big) (\hat h_{\kappa,4} +\hat h_{\kappa,7} +\hat h_{\kappa,9})
- h_\kappa \ , \quad
\kappa\in K\ , \quad t\geq 0 \ .$$ Then, the vector of its Hermite coefficients satisfies $$\label{linap4:3D}
\partial_t\hat {\bf h}_\kappa (t) + i\tfrac{2\pi}{L}\,\kappa {\mathbf{L}}_1 \hat{\bf h}_\kappa (t) = -{\mathbf{L}}_2 \hat{\bf h}_\kappa (t)\ ,\quad \kappa\in K\ ,\quad t\ge0\ ,$$ where the operators ${\mathbf{L}}_1,\,{\mathbf{L}}_2$ are represented by “infinite matrices” on $\ell^2({\mathbb{N}}_0)$: $$ {\mathbf{L}}_1= \begin{pmatrix}
0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
1 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 3 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & \sqrt 3 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & \sqrt 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots & \vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots & \ddots
\end{pmatrix}\ ,$$ $$\label{matrix:L2:3D}
{\mathbf{L}}_2=\operatorname{diag}(0,\, 0,\, 0,\, 0,\, \tfrac13
\begin{pmatrix}
2 & 0 & 0 & -1 & 0 & -1 \\
0 & 3 & 0 & 0 & 0 & 0 \\
0 & 0 & 3 & 0 & 0 & 0 \\
-1 & 0 & 0 & 2 & 0 & -1 \\
0 & 0 & 0 & 0 & 3 & 0 \\
-1 & 0 & 0 & -1 & 0 & 2 \\
\end{pmatrix},\, 1,\, 1,\cdots)\ .$$
To determine the hypocoercivity index of the BGK model in 3D, it is preferable to use the second basis with the linear ordering $\tilde{g}_m(v)$, $m\in{\mathbb{N}}_0$. Again, we shall give the matrix representation of the BGK equation in the second velocity basis only for the spatial modes $k=(\kappa,0)^\top$, $\kappa\in K$. To obtain the corresponding matrices $\tilde{\mathbf{L}}_1$ and $\tilde{\mathbf{L}}_2$, we simply represent the linear basis transformation by the infinite matrix $${\mathbf{S}}= \operatorname{diag}(1,\ 1,\ 1,\, 1,\
\begin{pmatrix}
1/\sqrt3 & 0 & 0 & 1/\sqrt3 & 0 & 1/\sqrt3 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
1/\sqrt3 & 0 & 0 & -(1+1/\sqrt3)/2 & 0 & (1-1/\sqrt3)/2 \\
0 & 0 & 0 & 0 & 1 & 0\\
1/\sqrt3 & 0 & 0 & (1-1/\sqrt3)/2 & 0 & -(1+1/\sqrt3)/2 \\
\end{pmatrix},\ 1,\ 1,\cdots)\ .$$ which is self-inverse, i.e. ${\mathbf{S}}={\mathbf{S}}^{-1}$. Thus we compute $\tilde{{\mathbf{L}}}_1 = {\mathbf{S}}^{-1} {\mathbf{L}}_1 {\mathbf{S}}$ and $\tilde{{\mathbf{L}}}_2 = {\mathbf{S}}^{-1} {\mathbf{L}}_2 {\mathbf{S}}$, which yields [$$ \tilde{{\mathbf{L}}}_1= \begin{pmatrix}
0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
1 & 0 & 0 & 0 & \tfrac{\sqrt{2}}{\sqrt{3}} & 0 & 0 & \tfrac{\sqrt{2}}{\sqrt{3}} & 0 & \tfrac{\sqrt{2}}{\sqrt{3}} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & \tfrac{\sqrt{2}}{\sqrt{3}} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & \tfrac1{\sqrt{3}} & 0 & \tfrac1{\sqrt{3}} & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & \tfrac{\sqrt{2}}{\sqrt{3}} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & -\tfrac{3+\sqrt{3}}{6} & 0 & \tfrac{3-\sqrt{3}}{6} & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & \tfrac{\sqrt{2}}{\sqrt{3}} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & \tfrac{3-\sqrt{3}}{6} & 0 & -\tfrac{3+\sqrt{3}}{6} & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & \cdots \\
0 & 0 & 0 & 0 & 0 & \sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & \sqrt{2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & \tfrac1{\sqrt{3}} & 0 & 0 & -\tfrac{3+\sqrt{3}}{6} & 0 & \tfrac{3-\sqrt{3}}{6} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & \tfrac1{\sqrt{3}} & 0 & 0 & \tfrac{3-\sqrt{3}}{6} & 0 & -\tfrac{3+\sqrt{3}}{6} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots & \vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\ddots
\end{pmatrix}$$ ]{} $$\tilde{{\mathbf{L}}}_2=\operatorname{diag}(0,\, 0,\, 0,\, 0,\, 0,\, 1,\, 1,\, 1,\cdots)\ .$$ To determine the hypocoercivity index of the BGK model in 3D, we use Definition \[hyp-index\] in its equivalent formulation \[cond:trivkernel\]: With $\ker\tilde{\mathbf{L}}_2={\operatorname{span}}\{e_0,\,e_1,\,e_2,\,e_3,\,e_4\}$, and the relations $$\tilde{\mathbf{L}}_1 e_0=e_1\,,\quad
\tilde{\mathbf{L}}_1 e_1=e_0+\tfrac{\sqrt{2}}{\sqrt{3}} (e_4+e_7+e_9)\,,\quad
\tilde{\mathbf{L}}_1 e_2=e_5\,,\quad
\tilde{\mathbf{L}}_1 e_3=e_6\,,\quad
\tilde{\mathbf{L}}_1 e_4=\tfrac{\sqrt{2}}{\sqrt{3}} e_1+e_{10} +\tfrac1{\sqrt{3}} (e_{13} +e_{15})\,,$$ we find the index $\tau=2$ (like in 2D). Each of the three momentum-conservation modes (represented by $e_1$, $e_2$ and $e_3$) is directly coupled to a decreasing mode.\
We define the matrices ${\mathbf{C}}_\kappa := i\tfrac{2\pi}{L} \kappa\ \tilde{\mathbf{L}}_1 + \tilde{\mathbf{L}}_2$, $\kappa\in K\cup\{0\}$, which determine the evolution of the spatial modes of the BGK equation in 3D, cf. . Our next goal is to establish a spectral gap of ${\mathbf{C}}_\kappa$, uniformly in $\kappa\in K$. This will prove Theorem \[linBGK-decay\] in 3D. To this end we make an ansatz for the transformation matrices: Following the detailed motivation from the 1D analog in §\[sec:linBGK:1D\], let ${\mathbf{P}}_\kappa$, $\kappa\in K$ be the identity matrix whose upper-left $11\times 11$ block is replaced by $$\label{P:3D}
\begin{pmatrix}
1 & -i \alpha/\kappa & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
i \alpha/\kappa & 1 & 0 & 0 & 0 & 0 & 0 & -i \beta/\kappa & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & -i \gamma/\kappa & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 & -i \omega/\kappa & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & -i \eta/\kappa \\
0 & 0 & i \gamma/\kappa & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & i \omega/\kappa & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\
0 & i \beta/\kappa & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & i \eta/\kappa & 0 & 0 & 0 & 0 & 0 & 1 \\
\end{pmatrix}$$ with positive parameters $\alpha$, $\beta$, $\gamma$, $\omega$, and $\eta$ to be chosen below. The distribution of the non-zero off-diagonal elements follows from the pattern in matrix $\tilde{\mathbf{L}}_1$.
\[lemma:decay:3D\] If the matrices ${\mathbf{P}}_\kappa$ are chosen as with $\beta =\sqrt3 \alpha$, $\gamma =\alpha$, $\omega=\alpha$, and $\eta=\alpha$ uniformly for all $\kappa\in K$, then there exists a positive $\alpha_+$ such that ${\mathbf{P}}_\kappa$ and ${\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa$ are positive definite for all $\alpha\in(0,\alpha_+)$ and for all $\kappa\in K$. Moreover, $$\label{goodbnd:3D}
{\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa \geq 2\mu {\mathbf{P}}_\kappa \qquad \text{ uniformly in } \kappa\in K \ ,$$ with $$\mu := \left( \frac{20}{32} \right)^{20} \frac{\delta_{21}(1,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha)}{2 (1+2\alpha)} > 0 \ ,$$ where $\delta_{21}(1,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha):= \det {\mathbf{D}}_{1,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha}$ with ${\mathbf{D}}_{\kappa,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha}$ defined in .
The proof of this lemma is deferred to Appendix \[A1\].
\[rem:mu:3D\]
1. Consider $$\alpha_* = \operatorname*{arg\,max}_{\alpha\in[0,\alpha_+]} \frac{\delta_{21}(1,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha)}{2 (1+2\alpha)} \ .$$ Choosing ${\mathbf{P}}_\kappa$ with $\alpha=\alpha_*$, $\beta =\sqrt3 \alpha$, $\gamma =\alpha$, $\omega=\alpha$, and $\eta=\alpha$ uniformly for all $\kappa\in K$, yields with the maximal constant $$\label{goodbnd2:3D:b}
\mu = \left( \frac{20}{32} \right)^{20} \frac{\delta_{21}(1,\alpha_*,\sqrt3 \alpha_*,\alpha_*,\alpha_*,\alpha_*)}{2 (1+2\alpha_*)}
= \max_{\alpha\in[0,\alpha_+]} \left( \frac{20}{32} \right)^{20} \frac{\delta_{21}(1,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha)}{2 (1+2\alpha)} \ .$$
2. \[rem:mu:3D:2Pi\] For $L=2\pi$, we compute $\alpha_+ = 0,214287873283229$. Moreover, the constant $\mu$ is determined as $\mu = 0,0001774540949$ with $\alpha_* = 0,1644256115$.
3. In the limit $L\to +\infty$, the matrix ${\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa$ has zero eigenvalues, which is apparent from its upper left submatrix ${\mathbf{D}}_{\kappa,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha}$ defined in . Accordingly, $\alpha_+ \to 0$ in the limit $L\to \infty$.
Moreover $\alpha_+ \to 0$ in the limit $L\to 0$.
We consider a solution $h$ of , and the entropy functional ${\mathcal{E}}(\tilde{f})$ defined by $$\label{entropy:3D}
{\mathcal{E}}(\tilde{f}) := \sum_{k\in {\mathbb{Z}}^3} \langle h_k(v), {\mathbf{P}}_{|k|} h_k(v)\rangle_{L^2(M_1^{-1})}\ ,$$ with $\tilde{f}(t) := M_1 +h(t)$. Here the matrices ${\mathbf{P}}_0 = {\mathbf{I}}$ and ${\mathbf{P}}_\kappa$ defined in for $\kappa=|k|\neq 0$ are regarded as bounded operators on $L^2(M_1^{-1})$. Then $$\label{e-inequality:3D}
{\frac{{\,\operatorname{d}\!}}{{\,\operatorname{d}\!t}}}{\mathcal{E}}(\tilde{f}) = -\sum_{k\in {\mathbb{Z}}^3} \langle h_k(v), ({\mathbf{C}}_{|k|}^*{\mathbf{P}}_{|k|} + {\mathbf{P}}_{|k|}{\mathbf{C}}_{|k|}) h_k(v)\rangle_{L^2(M_1^{-1})}
\leq - 2\min\{1,\mu\} \,{\mathcal{E}}(\tilde{f}) \ ,
$$ where $1$ is the decay rate of $h_0$, cf. . This implies with $\lambda^3(L):=2\min\{1,\mu\}$ and $\mu$ from .
The constants $c_3$ and $C_3$ in the estimate follow from : $$ c_3(L)=\left(1 +2 \alpha_* \right)^{-1}\,,\quad
C_3(L)=\left(1 -2 \alpha_* \right)^{-1}\,.$$ This finishes the proof of Theorem \[linBGK-decay\] in 3D.
Local exponential stability for the BGK equation in 3D {#sec:6}
======================================================
This analysis is an extension of §4.5 in [@AAC16]. To make is self-contained we give the complete proof and not only the modification of the key steps.
For $\gamma \ge 0$, let $H^\gamma(\tilde{\mathbf{T}}^3)$ be the Sobolev space consisting of the completion of smooth functions $\varphi$ on $\tilde{\mathbf{T}}^3$ in the Hilbertian norm $$\|\varphi\|_{H^\gamma}^2 := \sum_{k\in {\mathbb{Z}}^3} (1+|k|^2)^{\gamma}|\varphi_k|^2 \,,$$ where $\varphi_k$ ($k\in{\mathbb{Z}}^3$) is the $k$th Fourier coefficient of $\varphi$. Let $\mathcal{H}_\gamma$ denote the Hilbert space $H^\gamma(\tilde{\mathbf{T}}^3)\otimes L^2({\mathbb{R}}^3;M_1^{-1}(v){\,\operatorname{d}\!v})$, where the inner product in $\mathcal{H}_\gamma$ is given by $$\langle f,g\rangle_{\mathcal{H}_\gamma} = \int_{\tilde{\mathbf{T}}^3} \int_{{\mathbb{R}}^3} \overline{f}(x,v) \left[\left(1 - \Delta_x \right)^{\gamma} g(x,v) \right]M_1^{-1}(v) {\,\operatorname{d}\!v} {\,\operatorname{d}\!\tilde x} \,,$$ where ${\,\operatorname{d}\!\tilde x}$ denotes the normalized Lebesgue measure on $\tilde{\mathbf{T}}^3$. Then $\mathcal{H}_0$ is simply the weighted space $L^2(\tilde{\mathbf{T}}^3\times{\mathbb{R}}^3;M_1^{-1}(v){\,\operatorname{d}\!\tilde x}{\,\operatorname{d}\!v})$.
\[BGK-decay:a\] For any solution $h(t)$ to with ${\mathcal{E}}^d(h^I+M_1)<\infty$, normalized according to , we consider the function $\tilde f(t):=h(t) +f^\infty$ with $f^\infty=M_1$. We define a family of entropy functionals ${\mathcal{E}}_\gamma$ ($\gamma\geq 0$) by $$\label{entropy:gamma:3D}
{\mathcal{E}}_\gamma(\tilde f) := \sum_{k\in {\mathbb{Z}}^3} (1+|k|^2)^{\gamma} \langle h_k(v), {\mathbf{P}}_{|k|} h_k(v)\rangle_{L^2(M_1^{-1})}\ ,$$ as an extension of the entropy ${\mathcal{E}}(\tilde f)$ in . For all $\gamma\geq 0$, the estimates $$\label{equiv:norm:entropy:gamma}
\tfrac34 {\mathcal{E}}_\gamma(\tilde f)
\leq \tfrac1{1+2\alpha_*} {\mathcal{E}}_\gamma(\tilde f)
\leq \| h\|_{\mathcal{H}_\gamma}^2
\leq \tfrac1{1-2\alpha_*} {\mathcal{E}}_\gamma(\tilde f)
\leq \tfrac32 {\mathcal{E}}_\gamma(\tilde f)$$ follow from and Remark \[rem:mu:3D\]\[rem:mu:3D:2Pi\]. Moreover, the second statement in Theorem \[BGK-decay\]\[BGK-decay:a\] follows just as in the proof of Theorem \[linBGK-decay\] in 3D where the numerical values are chosen according to Remark \[rem:mu:3D\]\[rem:mu:3D:2Pi\].
\[BGK-decay:b\] Let $f$ be a solution of the BGK equation with constant temperature $T=1$ and define $h(x,v,t) := f(x,v,t) - M_1(v)$ as in the introduction. Moreover, let $\sigma$, $\mu$ and $\tau$ be defined in terms of $f$ as in . For all $\gamma\geq 0$, $\|\sigma\|_{H^\gamma}^2 = \langle \sigma M_1, f - f^\infty\rangle_{\mathcal{H}_\gamma}$ with $f^\infty =M_1$. Then by the Cauchy-Schwarz inequality, $$\label{locbound1}
\|\sigma\|_{H^\gamma}^2 \leq \|\sigma M_1\|_{\mathcal{H}_\gamma} \| f-f^\infty\|_{\mathcal{H}_\gamma} =
\|\sigma\|_{H^\gamma} \| f-f^\infty\|_{\mathcal{H}_\gamma} \ .$$ Likewise, $\|\mu\|_{H^\gamma}^2 = \langle \mu\cdot v M_1, f - f^\infty\rangle_{\mathcal{H}_\gamma}$, and hence $$\|\mu\|_{H^\gamma}^2 \leq \|\mu\cdot v M_1\|_{\mathcal{H}_\gamma} \| f-f^\infty\|_{\mathcal{H}_\gamma} \le
\sqrt{3} \|\mu\|_{H^\gamma} \| f-f^\infty\|_{\mathcal{H}_\gamma} \ ,$$ as well as, $\|\tau\|_{H^\gamma}^2 = \langle \tau |v|^2 M_1, f - f^\infty\rangle_{\mathcal{H}_\gamma}$, and hence $$\label{locbound2}
\|\tau\|_{H^\gamma}^2 \leq \|\tau |v|^2 M_1\|_{\mathcal{H}_\gamma} \| f-f^\infty\|_{\mathcal{H}_\gamma} =
\sqrt{15} \|\tau\|_{H^\gamma} \| f-f^\infty\|_{\mathcal{H}_\gamma} \ .$$ Using a Sobolev embedding (with $\gamma>3/2$) we can estimate the perturbations of the first 3 moments in $L^\infty(\tilde{\mathbf{T}}^3)$ as $$\label{locbound3}
\|\sigma\|_\infty\leq C_\gamma \| f-f^\infty\|_{\mathcal{H}_\gamma}\,, \qquad
\|\mu\|_\infty\leq C_\gamma \| f-f^\infty\|_{\mathcal{H}_\gamma} \quad
{\rm and}\qquad
\|\tau\|_\infty \leq C_\gamma \| f-f^\infty\|_{\mathcal{H}_\gamma} \ .$$ Using these estimates it is a simple matter to control the approximation in (\[taylor\]): For $s\in [0,1]$ and $(x,v) \in \tilde{\mathbf{T}}^3\times{\mathbb{R}}^3$, define (inspired by ) $$\label{def-F}
F(s,x,v) := \frac{(1+s\sigma)^{\frac{5}2}(x)}{\left(2\pi \big\{1+\frac13\big[s\tau(x)-\frac{s^2|\mu|^2(x)}{1+s\sigma(x)}\big]\big\}\right)^\frac{3}2}
\,\exp\Big\{-\frac{|v(1+s\sigma(x))-s\mu(x)|^2}{2\big(1+\frac13\big[s\tau(x)-\frac{s^2|\mu|^2(x)}{1+s\sigma(x)}\big]\big) (1+s\sigma(x))}\Big\} \ ,$$ so that the gain term in the linearized BGK equation is ${\displaystyle \partial_s F(0,x,v)}$. In this notation, $$\begin{aligned}
R_f(x,v)
&:= M_f(x,v) - M_1(v) - \left[ \left( \frac52 - \frac{|v|^2}{2}\right)\sigma(x) +v\cdot\mu(x)
+ \left( -\frac{1}{2}+ \frac{|v|^2}{6}\right)\tau(x)\right] M_1(v) \\
& = \int_0^1 \left[ \partial_s F(s,x,v) - \partial_s F(0,x,v)\right]{\rm d}s
= \int_0^1 \int_0^s \left[ \partial_s^2 F(r,x,v)\right] {\rm d}r\, {\rm d}s\ .\end{aligned}$$ To display the complicated expression for $\partial_s^2 F(s,x,v)$, we define $$\rho_s := 1+s\sigma\ , \quad u_s := \frac{s}{\rho_s} \mu\ , \quad
\mu_s := s\mu\ , \quad
P_s := 1 +\frac13 \Big(s\tau - \frac{|\mu_s|^2}{\rho_s}\Big)\ .$$ Then $\partial_s^2 F(s,x,v)$ reads $$\begin{aligned}
\frac{\partial_s^2 F(s,x,v)}{F(s,x,v)}
&=
\left( \frac{15}{4 \ \rho_s^2} -\frac5{2\ P_s \ \rho_s} \ | v -u_s|^2
\right) \sigma^2
+\left( -\frac{15}{2\ P_s \ \rho_s}
+\frac5{P_s^2} \, | v -u_s|^2
\right) \sigma \ \partial_s P_s \\
&\quad
+\left( \frac {5}{P_s \ \rho_s^2}
\right) \sigma \ \left( ( v -u_s) \cdot \mu \right)
+\left( \frac{15}{4\ P_s^2}
-\frac{5\ \rho_s}{2\ P_s^{3}}\ | v -u_s|^2
\right) \left( \partial_s P_s \right)^2 \\
&\quad
+\left( -\frac5{P_s^2\ \rho_s}
\right)\partial_s P_s \ \left( ( v -u_s) \cdot \mu \right)
+\left(
-\frac1{3 P_s^2\ \rho_s^2} | v -u_s|^2
\right) |\mu|^2 \\
&\quad
+ \left( {\frac {1}{P_s \ \rho_s} ( v -u_s) \cdot \mu }
-\tfrac12 {\frac {\sigma}{P_s} | v -u_s|^2}
+\tfrac12 {\frac {\rho_s}{P_s^2} | v -u_s|^2 \ \partial_s P_s }
\right)^2\ ,\end{aligned}$$ where $\partial_s P_s := \tfrac13 \left( \tau-2\,{\frac {s|\mu|^2}{s\sigma+1}}+{\frac {{s}^2|\mu|^2\sigma}{ ( s\sigma+1 )^2}} \right)$. One can now verify that $\partial_s^2 F(s,x,v)$ is of the order $O(\sigma^2+|\mu|^2+\tau^2)$, which will be related to $O((f-f^\infty)^2)$ due to the estimates –.
Simple but cumbersome calculations now show that if $\gamma>3/2$ and ${\| {f-f^\infty} \|}_{\mathcal{H}_\gamma}$ is sufficiently small, then there exists a finite constant $\tilde C_\gamma$ depending only on $\gamma$ such that for all $s\in [0,1]$, $$\left\Vert \partial_s^2 F(s,x,v) \right \Vert_{\mathcal{H}_\gamma}
\leq \tilde C_\gamma \| f-f^\infty\|_{\mathcal{H}_\gamma}^2\ ,$$ and hence $$\label{R-estimate}
\|R_f\|_{\mathcal{H}_\gamma}\le \tilde C_\gamma \| f-f^\infty\|_{\mathcal{H}_\gamma}^2\ .$$ \[The calculations are simplest for non-negative integer $\gamma$, in which case the Sobolev norms can be calculated by differentiation. For $\gamma>3/2$ and sufficiently small ${\| {f-f^\infty} \|}_{\mathcal{H}_\gamma}$, the estimates ensure for all $s\in [0,1]$ the boundedness of $0<\epsilon<{\| {1+s \sigma} \|}_\infty\ ,\, {\| {1+\frac13(s \tau-\frac{s^2|\mu|^2}{1+s\sigma})} \|}_\infty<\infty$ (i.e. the denominators in ) for some fixed $\epsilon>0$. They also ensure the $L^2({\mathbb{R}}^3;M_1^{-1}(v){\,\operatorname{d}\!v})$-integrability of $F(s,x,\cdot)$ by using $$\exp\big\{-\frac{|\rho_s v-\mu_s|^2}{2P_s\rho_s}\big\} \leq e^{-|v|^2/3+1} \quad \text{for all } x\in\tilde{\mathbf{T}}^3\ .$$ In , higher powers of ${\| {f-f^\infty} \|}_{\mathcal{H}_\gamma}$ (arising due to derivatives of $\sigma$, $\mu$ and $\tau$) can be absorbed into the constant of the quadratic term.\] Now define the linearized BGK operator $${\mathbf{Q}}_2 h(x,v,t) := \left[ \left( \frac52 - \frac{|v|^2}{2}\right)\sigma(x) +v\cdot\mu(x)+ \left( -\frac{1}{2}+ \frac{|v|^2}{6}\right)\tau(x)\right] M_1(v) - h(x,v,t)$$ where $\sigma$, $\mu$ and $\tau$ are determined by $h$, and hence $f$. For all $\gamma\geq 0$, ${\mathbf{Q}}_2$ is self-adjoint on $\mathcal{H}_\gamma$. Then the nonlinear BGK equation becomes $$\label{bgk:torus3}
\partial_t h(x,v,t) + v\cdot \nabla_x h(x,v,t) = {\mathbf{Q}}_2 h(x,v,t) + R_f(x,v,t)\ , \qquad t\geq 0\ ,$$ which differs from the linearized BGK equation only by the additional term $R_f$.
It is now a simple matter to prove local exponential stability. We shall use here exactly the entropy functional ${\mathcal{E}}_\gamma(f)$ defined in with $f=M_1+h$. Now assume that $h$ solves (\[bgk:torus3\]). To compute $\frac{{\rm d}}{{\rm d}t} {\mathcal{E}}_\gamma(f)$ we use an inequality like for the drift term and for ${\mathbf{Q}}_2 h$ in , as well as $\|{\mathbf{P}}_{|k|}\|\le (1+2\alpha_*)$ and for the term $R_f$. This yields $$\label{e-decay1}
\frac{{\rm d}}{{\rm d}t} {\mathcal{E}}_\gamma(f) \leq -2\mu \,{\mathcal{E}}_\gamma(f) + 2(1+2\alpha_*) \tilde C_\gamma\| h\|_{\mathcal{H}_\gamma}^3\ ,$$ (if $\|h\|_{\mathcal{H}_\gamma}$ is small enough) where we have used the fact that $h = f-f^\infty$. Due to , it is now simple to complete the proof of Theorem \[linBGK-decay\]\[BGK-decay:b\] for $L=2\pi$: In this case, the best decay rate $\mu = 0,0001774540949$ is attained for $\alpha_* = 0,1644256115$ (cf. Remark \[rem:mu:3D\](b)). Estimate shows that there is a $\delta_\gamma>0$ so that if the initial data $f^I(x,v)$ satisfies $\|f^I- f^\infty\|_{\mathcal{H}_\gamma} < \delta_\gamma$, then the solution $f(t)$ satisfies $${\mathcal{E}}_\gamma(f(t)) \leq e^{-t/2820}{\mathcal{E}}_\gamma(f^I)\ .$$ Here we used that the linear decay rate in is slightly better than $\frac{1}{2820}$, to compensate the nonlinear term.
Appendix: Deferred proofs {#A1}
=========================
We compute that ${\mathbf{C}}_k^*{\mathbf{P}}_k + {\mathbf{P}}_k{\mathbf{C}}_k$ is twice the identity matrix whose upper left $5\times 5$ block is replaced by
$${\mathbf{D}}_{k,\alpha,\beta,\gamma} =
\begin{pmatrix}
2\ell\alpha & 0 & \ell(\sqrt2 \alpha -\beta) & 0 & 0 \\
0 & 2\ell(\sqrt{2}\beta-\alpha) & 0 & \ell (\sqrt{3}\beta - \sqrt{2}\gamma) & 0\\
\ell( \sqrt{2}\alpha -\beta) & 0 & 2\ell(\sqrt{3}\gamma-\sqrt2\beta) & -i\gamma/k & 2\ell\gamma\\
0 & \ell( \sqrt{3}\beta -\sqrt{2}\gamma) & i\gamma/k & 2-2\ell\sqrt{3}\gamma & 0\\
0 & 0 & 2\ell \gamma & 0 & 2
\end{pmatrix} \ ,$$ where $\ell:= \tfrac{2\pi}{L}$. We seek to choose $\alpha$, $\beta$ and $\gamma$ to make the matrices ${\mathbf{P}}_k$ and ${\mathbf{D}}_{k,\alpha,\beta,\gamma}$ positive definite for all $k\in{\mathbb{Z}}\setminus\{0\}$. Under the assumption $|\alpha|^2 +|\beta|^2 +|\gamma|^2 <1$, the matrix ${\mathbf{P}}_k$ will be positive definite for all $k\ne0$. To simplify the analysis we shall now set $\beta =\sqrt{2} \alpha$ and $\gamma =\sqrt{3} \alpha$. On the one hand this will make the first three diagonal entries of ${\mathbf{D}}_{k,\alpha,\beta,\gamma}$ equal and annihilate four off-diagonal elements. But, on the other hand, this will then lead to a reduced decay rate. But optimal decay rates are anyhow not our goal here – due to considering only a simple ansatz for the transformation matrices ${\mathbf{P}}_k$. For $\beta =\sqrt{2} \alpha$ and $\gamma =\sqrt{3} \alpha$ we have
$$\label{Dkaaa}
{\mathbf{D}}_{k,\alpha,\beta,\gamma} =
\begin{pmatrix}
2\ell\alpha & 0 & 0 & 0 & 0 \\
0 & 2\ell \alpha & 0 & 0 & 0\\
0 & 0 & 2\ell\alpha & -i\sqrt{3}\alpha/k & 2\sqrt{3}\ell\alpha\\
0 & 0 & i\sqrt{3}\alpha/k & 2-6\ell\alpha & 0\\
0 & 0 & 2\sqrt{3}\ell\alpha & 0 & 2
\end{pmatrix} \ .$$
The positive definiteness of ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}$ will follow from Sylvester’s criterion, by choosing $\alpha$ such that all minors of ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}$ will be positive. Let $\delta_j(k,\alpha)$ denote the determinant of the lower right $j\times j$ submatrix of ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}$ for $j=1,2,3,4,5$. For our choice $\beta =\sqrt{2} \alpha$ and $\gamma =\sqrt{3} \alpha$, the first minor $\delta_1(k,\alpha) = 2$ is always positive and the second minor $\delta_2(k,\alpha) = 4(1-3\ell\alpha)$ is positive for $\alpha<1/(3\ell)$. The third minor satisfies $$\delta_3(k,\alpha)
= \alpha \big( 72\ell^3 \alpha^2 -(48\ell^2 +\tfrac{6}{k^2})\alpha +8\ell \big)
\ge \delta_3(1,\alpha)\quad \text{for all } k\ne0 \ ,$$ and the lower bound $\delta_3(1,\alpha)$ is positive if $$\label{root:3:1D}
0 < \alpha < \alpha^{(3)} := \frac{1+8\ell^2 -\sqrt{1+16\ell^2}}{24\ell^3}\,,\qquad
\ell:=\frac{2\pi}{L} \ .$$ Moreover $0<\alpha^{(3)}<1/(3\ell)$ for all $\ell>0$ and $\max_{\ell>0}\alpha^{(3)}(\ell)<0.257$. The fourth and fifth minor are multiples of the third minor: $$\delta_4(k,\alpha) = 2 \ell \alpha \delta_3(k,\alpha), \quad \delta_5(k,\alpha) = (2\ell \alpha)^2 \delta_3(k,\alpha)\,.$$ Hence, all minors are positive under assumption . Matrix ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}$ has a block diagonal structure. Thus it has a double eigenvalue $2\ell\alpha$ and the eigenvalues of its lower right $3\times 3$-submatrix $$\label{Dka3}
{\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}^{(3)} =
\begin{pmatrix}
2\ell\alpha & -i\sqrt{3}\alpha/k & 2\sqrt{3}\ell\alpha\\
i\sqrt{3}\alpha/k & 2-6\ell\alpha & 0\\
2\sqrt{3}\ell\alpha & 0 & 2
\end{pmatrix} \ .$$ Let $\{\lambda_1,\lambda_2,\lambda_3\}$ be the eigenvalues of ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}^{(3)}$ arranged in increasing order. We seek a lower bound on $\lambda_1$. As long as ${\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}^{(3)}$ is positive definite, the arithmetic-geometric mean inequality implies $$\begin{aligned}
\lambda_1(k,\alpha) = \frac{\delta_3(k,\alpha)}{\lambda_2 \lambda_3}
&\geq \delta_3(k,\alpha) \left( \frac{ \lambda_2 + \lambda_3}{2}\right)^{-2}\\
&\geq \delta_3(k,\alpha) \left( \frac{{\operatorname{Tr}}[{\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}^{(3)}]}{2}\right)^{-2}
= \frac{\delta_3(k,\alpha)}{4(1-\ell\alpha)^2} \ ,
\end{aligned}$$ since ${\operatorname{Tr}}[{\mathbf{D}}_{k,\alpha,\sqrt{2} \alpha,\sqrt{3} \alpha}^{(3)}] =4(1-\ell\alpha)$. Thus, if ${\mathbf{P}}_k$ is chosen with some $\alpha \in (0,\alpha^{(3)})$, $\beta =\sqrt{2} \alpha$, and $\gamma =\sqrt{3} \alpha$ uniformly for all $|k|\in{\mathbb{N}}$, then $$\label{Pbnd}
{\mathbf{C}}_k^*{\mathbf{P}}_k + {\mathbf{P}}_k{\mathbf{C}}_k
\geq \frac{\delta_3(1,\alpha)}{4(1-\ell\alpha)^2} \ {\bf I} \qquad \text{ uniformly in } |k|\in{\mathbb{N}}\ ,$$ since $\min\big\{2\ell\alpha, \frac{\delta_3(1,\alpha)}{4(1-\ell\alpha)^2} \big\} =\frac{\delta_3(1,\alpha)}{4(1-\ell\alpha)^2}$ for all $\alpha \in (0,\alpha^{(3)})$. A simple computation shows that the eigenvalues of ${\mathbf{P}}_k$ are $1$, $1 \pm\alpha \sqrt{3+\sqrt 6}/k$, and $1 \pm\alpha \sqrt{3-\sqrt 6}/k$. These eigenvalues are positive for all $0\leq\alpha\leq\max_{\ell>0}\alpha^{(3)}(\ell)$, $L>0$ and $k\in{\mathbb{N}}$. Hence, uniformly in $|k|$, $$\label{Pbound}
\big(1 -\alpha \sqrt{3+\sqrt 6}\big) {\bf I} \leq {\mathbf{P}}_k \leq \big(1 +\alpha \sqrt{3+\sqrt 6}\big) {\bf I}\ .$$ Combining with yields the result .
The following lemma will be needed in the proofs of Lemma \[lemma:decay:2D\] and Lemma \[lemma:decay:3D\].
\[lem:p\] Let $p(\kappa,\alpha)$ be a rational function of the form $$ p(\kappa,\alpha) = \big(p_0(\alpha) +p_1(\alpha) \tfrac1{\kappa^2}\big) \tfrac1{\kappa^2} + p_2(\alpha) \,,$$ where $p_0$, $p_1$, and $p_2$ are polynomials in $\alpha$. If there exists $\widetilde\alpha>0$ such that $$ 0\le p_1(\alpha) \quad \text{and } \quad
p_0(\alpha) +2 p_1(\alpha) \le 0 \qquad
\forall \alpha\in[0,\widetilde\alpha]\,,$$ then $p(1,\alpha) \leq p(\kappa,\alpha)$ for all $\alpha\in[0,\widetilde\alpha]$ and $1\leq \kappa$.
We want to prove $p(1,\alpha) \leq p(\kappa,\alpha)$ for all $\alpha\in[0,\widetilde\alpha]$ and $1\leq \kappa$, or equivalently, $$p_0(\alpha) +p_1(\alpha) \leq \big(p_0(\alpha) +p_1(\alpha) \tfrac1{\kappa^2}\big) \tfrac1{\kappa^2} \qquad \forall \alpha\in[0,\widetilde\alpha] \quad \forall 1\leq \kappa \,.$$ We multiply the inequality with $\kappa^2$ $$\big(p_0(\alpha) +p_1(\alpha)\big)\kappa^2 \leq p_0(\alpha) +p_1(\alpha) \tfrac1{\kappa^2}$$ and rearrange the summands $$p_0(\alpha)\ (\kappa^2-1) \leq p_1(\alpha)\ \big(\tfrac1{\kappa^2} -\kappa^2) = -p_1(\alpha) \tfrac{(\kappa^2 -1)(\kappa^2 +1)}{\kappa^2} \,.$$ For $\kappa=1$ the inequality holds trivially. Therefore, we continue with $\kappa>1$ and divide the inequality by $(\kappa^2-1)$ to obtain $$p_0(\alpha) \leq -p_1(\alpha) \tfrac{\kappa^2 +1}{\kappa^2}\ .$$ Under our assumptions this inequality holds since $$p_0(\alpha) \leq -2 p_1(\alpha) \leq -p_1(\alpha) \tfrac{\kappa^2 +1}{\kappa^2} \leq 0 \,.$$ This finishes the proof.
We compute that ${\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa$ is twice the identity matrix whose upper left $11\times 11$ block is replaced by ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ given as [$$\begin{pmatrix}
2\ell\alpha & 0 & 0 & \ell\alpha & 0 & \ell(\alpha-\beta) & 0 & 0 & 0 & 0 & 0 \\
0 & -2\ell(\alpha-\beta) & 0 & 0 & 0 & -i \beta/\kappa & \ell\tfrac{\sqrt{3}\beta-\sqrt{2}\omega}{\sqrt{2}} & 0 & -\ell\beta/\sqrt{2} & 0 & 0 \\
0 & 0 & 2\ell\gamma & 0 & -i \gamma/\kappa & 0 & 0 & \sqrt{2}\ell\gamma & 0 & 0 & 0 \\
\ell\alpha & 0 & 0 & \ell\sqrt{6}\omega & 0 & \ell(\sqrt{\tfrac32}\omega-\beta) & -i \omega/\kappa & 0 & 0 & 0 & 2\ell\omega \\
0 & 0 & i \gamma/\kappa & 0 & 2(1-\ell\gamma) & 0 & 0 & 0 & 0 & 0 & 0 \\
\ell(\alpha-\beta) & i \beta/\kappa & 0 & \ell(\sqrt{\tfrac32}\omega-\beta) & 0 & 2(1-\ell\beta) & 0 & 0 & 0 & 0 & 0 \\
0 & \ell\tfrac{\sqrt{3}\beta-\sqrt{2}\omega}{\sqrt{2}} & 0 & i \omega/\kappa & 0 & 0 & 2-\sqrt{6}\ell\omega & 0 & -\ell\omega/\sqrt{2} & 0 & 0 \\
0 & 0 & \sqrt{2}\ell\gamma & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 \\
0 & -\ell\beta/\sqrt{2} & 0 & 0 & 0 & 0 & -\ell\omega/\sqrt{2} & 0 & 2 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 \\
0 & 0 & 0 & 2\ell\omega & 0 & 0 & 0 & 0 & 0 & 0 & 2
\end{pmatrix} \ ,$$]{}
with $\ell := 2\pi/L>0$. We seek to choose $\alpha$, $\beta$, $\gamma$ and $\omega$ such that the matrices ${\mathbf{P}}_\kappa$ and ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ are positive definite for all $\kappa\in K$. The positive definiteness of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ will follow from Sylvester’s criterion, if all minors of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ are positive. This will yield restrictions on the choice of parameters $\alpha$, $\beta$, $\gamma$ and $\omega$. The analysis will simplify, if we choose $\beta$, $\gamma$ and $\omega$ as multiples of $\alpha$, because then the first four columns will depend linearly on $\alpha$ and, moreover, several terms will drop out. For $\beta =2\alpha$, $\gamma =\alpha$ and $\omega=\sqrt{6}\alpha$, we compute ${\mathbf{D}}_{\kappa,\alpha,2\alpha,\alpha,\sqrt{6}\alpha}$ as [$$\label{Dka2aa3a}
\begin{pmatrix}
2\ell\alpha & 0 & 0 & \ell\alpha & 0 & -\ell\alpha & 0 & 0 & 0 & 0 & 0 \\
0 & 2\ell\alpha & 0 & 0 & 0 & -i 2\alpha/\kappa & 0 & 0 & -\ell\sqrt{2}\alpha & 0 & 0 \\
0 & 0 & 2\ell\alpha & 0 & -i \alpha/\kappa & 0 & 0 & \sqrt{2}\ell\alpha & 0 & 0 & 0 \\
\ell\alpha & 0 & 0 & \ell 6\alpha & 0 & \ell\alpha & -i \sqrt{6}\alpha/\kappa & 0 & 0 & 0 & 2\ell\sqrt{6}\alpha \\
0 & 0 & i \alpha/\kappa & 0 & 2(1-\ell\alpha) & 0 & 0 & 0 & 0 & 0 & 0 \\
-\ell\alpha & i 2\alpha/\kappa & 0 & \ell\alpha & 0 & 2(1-\ell 2\alpha) & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & i \sqrt{6}\alpha/\kappa & 0 & 0 & 2-\ell 6\alpha & 0 & -\ell\sqrt{3}\alpha & 0 & 0 \\
0 & 0 & \ell\sqrt{2}\alpha & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 \\
0 & -\ell\sqrt{2}\alpha & 0 & 0 & 0 & 0 & -\ell\sqrt{3}\alpha & 0 & 2 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 \\
0 & 0 & 0 & 2\ell\sqrt{6}\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 2
\end{pmatrix} \ .$$ ]{} Let $\delta_j(\kappa,\alpha,\beta,\gamma,\omega)$ denote the determinant of the upper left $j\times j$ submatrix of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ for integers $j=1,2,\ldots,11$. For our choice $\beta =2\alpha$, $\gamma =\alpha$ and $\omega=\sqrt{6}\alpha$, the minors $\delta_j(\kappa,\alpha):=\delta_j(\kappa,\alpha,2\alpha,\alpha,\sqrt{6}\alpha)$ are given in Table \[table:minors:2D\]. The first four minors are positive, if $\alpha$ is positive.
$\delta_1(\kappa,\alpha)$ = $2\ell\alpha$
------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------
$\delta_2(\kappa,\alpha)$ = $4\ell^2\alpha^2$
$\delta_3(\kappa,\alpha)$ = $8\ell^3\alpha^3$
$\delta_4(\kappa,\alpha)$ = $44\ell^4 \alpha^4$
$\delta_5(\kappa,\alpha)$ = $22\ell^3 \alpha^4 (4\ell -4\ell^2 \alpha -{\alpha}/{\kappa^2})$
$\delta_6(\kappa,\alpha)$ = $\delta_5(\kappa,\alpha) p_6(\kappa,\alpha)/\ell$
with $p_6(\kappa,\alpha) :=-\tfrac{54}{11} \ell^2 \alpha +2\ell -2\alpha /{\kappa^2}$.
$\delta_7(\kappa,\alpha)$ = $\frac{2}{11\ell^2} \delta_5(\kappa,\alpha) p_7(\kappa,\alpha)$
with $p_{7}(\kappa,\alpha) = \big(p_{7,0}(\alpha) +p_{7,1}(\alpha) \frac1{\kappa^2}\big) \frac1{\kappa^2} + p_{7,2}(\alpha)$,
$p_{7,0}(\alpha) = 93\ell^2 \alpha^2 -34\ell\alpha$, $p_{7,1}(\alpha) = 12\alpha^2$, $p_{7,2}(\alpha) = 162\ell^4 \alpha^2 -120\ell^3 \alpha +22\ell^2$.
$\delta_8(\kappa,\alpha)$ = $44 \ell^3 \alpha^4 \frac{\delta_7(\kappa,\alpha)}{\delta_5(\kappa,\alpha)} p_8(\kappa,\alpha)$
with $p_8(\kappa,\alpha) = 2\ell^3 \alpha^2 -6\ell^2\alpha +4\ell -{\alpha}/{\kappa^2}$.
$\delta_9(\kappa,\alpha)$ = $8 \ell \alpha^4 p_8(\kappa,\alpha) p_9(\kappa,\alpha)$
with $p_{9}(\kappa,\alpha) = \big(p_{9,0}(\alpha) +p_{9,1}(\alpha) \frac1{\kappa^2}\big) \frac1{\kappa^2} + p_{9,2}(\alpha)$,
$p_{9,0}(\alpha) = -12\ell^3 \alpha^3 +198\ell^2 \alpha^2 -68\ell \alpha$, $p_{9,1}(\alpha) = 24\alpha^2$,
$p_{9,2}(\alpha) = -81\ell^5 \alpha^3 +411\ell^4 \alpha^2 -262\ell^3 \alpha +44\ell^2$.
$\delta_{10}(\kappa,\alpha)$ = $2 \delta_9(\kappa,\alpha)$,
$\delta_{11}(\kappa,\alpha)$ = $64 \ell \alpha^4 p_8(\kappa,\alpha) p_{11}(\kappa,\alpha)$
with $p_{11}(\kappa,\alpha) = \big(p_{11,0}(\alpha) +p_{11,1}(\alpha) \frac1{\kappa^2}\big) \frac1{\kappa^2} + p_{11,2}(\alpha)$,
$p_{11,0}(\alpha) = -72\ell^4 \alpha^4 -300\ell^3 \alpha^3 +294\ell^2 \alpha^2 -68\ell\alpha$, $p_{11,1}(\alpha) = 24\alpha^2$,
$p_{11,2}(\alpha) = 162\ell^6 \alpha^4 -909\ell^5 \alpha^3 +963\ell^4 \alpha^2 -358\ell^3 \alpha +44\ell^2$.
: Let $\delta_j(\kappa,\alpha,\beta,\gamma,\omega)$ denote the determinant of the upper left $j\times j$ submatrix of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ for integers $j=1,2,\ldots,11$. For our choice $\beta =2\alpha$, $\gamma =\alpha$ and $\omega=\sqrt{6}\alpha$, the minors $\delta_j(\kappa,\alpha)=\delta_j(\kappa,\alpha,2\alpha,\alpha,\sqrt{6}\alpha)$ are given in this table.[]{data-label="table:minors:2D"}
The fifth minor $\delta_5(\kappa,\alpha)$ satisfies for positive $\alpha$ the inequality $\delta_5(\kappa,\alpha)\geq \delta_5(1,\alpha)$ for all $\kappa\in K$. Moreover, $\delta_5(1,\alpha)$ is positive for $\alpha\in(0,\alpha_{\delta_5})$ with $\alpha_{\delta_5} :={4\ell}/{(4\ell^2 +1)}$. Thus the fifth minor $\delta_5(\kappa,\alpha)$ is positive for all $\kappa\in K$ if $\alpha\in(0,\alpha_{\delta_5})$.
The sixth minor $\delta_6(\kappa,\alpha)$ has a factorization as $\delta_6(\kappa,\alpha) = \delta_5(\kappa,\alpha) p_6(\kappa,\alpha)/\ell$. The factor $p_6(\kappa,\alpha)$ satisfies for positive $\alpha$ the inequality $p_6(\kappa,\alpha)\geq p_6(1,\alpha)$ for all $\kappa\in K$. Moreover, $p_6(1,\alpha)$ is positive for $\alpha\in(0,\alpha_{p_6})$ with $\alpha_{p_6} :={22\ell}/{(54\ell^2 +22)}$. Thus the sixth minor $\delta_6(\kappa,\alpha)$ is positive if $0<\alpha<\alpha_{\delta_6}$ with $\alpha_{\delta_6} :=\min\{\alpha_{\delta_5} , \ \alpha_{p_6} \} =\alpha_{p_6}$.
The seventh minor $\delta_7(\kappa,\alpha)$ has a factorization as $\delta_7(\kappa,\alpha) = 2 \delta_5(\kappa,\alpha) p_7(\kappa,\alpha)/{11\ell^2}$. Due to Lemma \[lem:p\], the inequality $p_{7}(\kappa,\alpha) \geq p_{7}(1,\alpha)$ holds for some positive $\widetilde\alpha_{p_7}$, and consequently $\delta_{7}(\kappa,\alpha)\geq \delta_{7}(1,\alpha)$ holds for all $0\leq \alpha \leq \min\{\widetilde\alpha_{p_7},\alpha_{\delta_5}\}$ and $\kappa\in K$. The quadratic polynomial $p_7(1,\alpha)$ has two positive roots $0<\alpha_{p_7,-}<\alpha_{p_7,+}$ and is positive for all $0<\alpha<\alpha_{p_7}$ with $\alpha_{p_7} :=\alpha_{p_7,-}$. Consequently, for $0<\alpha <\alpha_{\delta_7}$ with $\alpha_{\delta_7} := \min\{\alpha_{\delta_5}\,,\ \widetilde\alpha_{p_7}\,, \ \alpha_{p_7}\}$ the seventh minor $\delta_{7}(\kappa,\alpha)$ is positive for all $\kappa\in K$.
The eighth minor $\delta_8(\kappa,\alpha)$ has a factorization. For positive $\alpha$, factor $p_8$ satisfies the inequality $p_8(\kappa,\alpha) >p_8(1,\alpha)$ for all $\kappa\in K$. The quadratic polynomial $p_8(1,\alpha)$ has two positive roots $0<\alpha_{p_8,-}<\alpha_{p_8,+}$ and is positive for all $0<\alpha<\alpha_{p_8}$ with $\alpha_{p_8} :=\alpha_{p_8,-}$. Thus, the eighth minor $\delta_8(\kappa,\alpha)$ is positive for all $\kappa\in K$, if $0<\alpha< \alpha_{\delta_8}$ with $\alpha_{\delta_8} :=\min\{\alpha_{\delta_5}\,,\ \alpha_{\delta_7}\,,\ \alpha_{p_8} \}$.
The ninth minor $\delta_9(\kappa,\alpha)$ has a factorization as $\delta_9(\kappa,\alpha) = 8 \ell \alpha^4 p_8(\kappa,\alpha) p_9(\kappa,\alpha)$. Due to Lemma \[lem:p\], the inequality $p_{9}(\kappa,\alpha) \geq p_{9}(1,\alpha)$ holds for some positive $\widetilde\alpha_{p_9}$, and consequently $\delta_{9}(\kappa,\alpha)\geq \delta_{9}(1,\alpha)$ holds for all $0\leq \alpha \leq \min\{\widetilde\alpha_{p_9},\alpha_{p_8}\}$ and $\kappa\in K$. The cubic polynomial $p_{9}(1,\alpha)$ is positive at $\alpha=0$ and $\lim_{\alpha\to\infty} p_9(1,\alpha)=-\infty$. Hence, there exists a positive root $\alpha_{p_9}$ such that $p_9(1,\alpha)$ is positive for all $0<\alpha<\alpha_{p_9}$. Consequently, for all $\alpha\in(0,\alpha_{\delta_9})$ with $\alpha_{\delta_9} :=\min\{\alpha_{p_9}\,,\ \widetilde\alpha_{p_9}\,,\ \alpha_{p_8}\}$, the ninth minor $\delta_{9}(\kappa,\alpha)$ is positive for all $\kappa\in K$.
The tenth minor $\delta_{10}$ satisfies $\delta_{10}(\kappa,\alpha) = 2 \delta_9(\kappa,\alpha)$. Therefore the tenth minor $\delta_{10}(\kappa,\alpha)$ is positive for all $\kappa\in K$ if $\alpha\in(0,\alpha_{\delta_9})$.
The eleventh minor $\delta_{11}(\kappa,\alpha)$ has a factorization as $\delta_{11}(\kappa,\alpha) = 64 \ell \alpha^4 p_8(\kappa,\alpha) p_{11}(\kappa,\alpha)$. Due to Lemma \[lem:p\], the inequality $p_{11}(\kappa,\alpha) \geq p_{11}(1,\alpha)$ holds for some positive $\widetilde\alpha_{p_{11}}$, and consequently $\delta_{11}(\kappa,\alpha)\geq \delta_{11}(1,\alpha)$ holds for all $0\leq \alpha \leq \min\{\widetilde\alpha_{p_{11}},\alpha_{p_8}\}$ and $\kappa\in K$. The quartic polynomial $p_{11}(1,\alpha)$ is positive at $\alpha=0$. Hence, there exists a positive root $\alpha_{p_{11}}$ such that $p_{11}(1,\alpha)$ is positive for all $0<\alpha<\alpha_{p_{11}}$. Consequently, for $\alpha\in(0,\alpha_{\delta_{11}})$ with $\alpha_{\delta_{11}} := :=\min\{\alpha_{p_{11}}\,,\ \widetilde\alpha_{p_{11}}\,,\ \alpha_{p_8}\}$, the eleventh minor $\delta_{11}(\kappa,\alpha)$ is positive for all $\kappa\in K$.\
Let $\{\lambda_1,\lambda_2,\ldots,\lambda_{11}\}$ be the eigenvalues of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ arranged in increasing order. We seek a lower bound on $\lambda_1$. As long as ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}$ is positive definite, the arithmetic-geometric mean inequality implies $$\begin{aligned}
\lambda_1(\kappa,\alpha,\beta,\gamma,\omega) = \frac{\delta_{11}(\kappa,\alpha,\beta,\gamma,\omega)}{\prod_{j=2}^{11} \lambda_j}
&\geq \delta_{11}(\kappa,\alpha,\beta,\gamma,\omega) \left( \frac{10}{\sum_{j=2}^{11} \lambda_j} \right)^{10}\\
&\geq \delta_{11}(\kappa,\alpha,\beta,\gamma,\omega) \left( \frac{10}{{\operatorname{Tr}}[{\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}]} \right)^{10}
\geq \left( \frac{10}{14} \right)^{10} \delta_{11}(\kappa,\alpha,\beta,\gamma,\omega) \ ,
\end{aligned}$$ since ${\operatorname{Tr}}[{\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega}] =14$ independently of $\kappa$, $\alpha$, $\beta$, $\gamma$ and $\omega$.
A simple computation shows that the eigenvalues of ${\mathbf{P}}_\kappa$ are $1$, $1 \pm \alpha/\kappa$, $1 \pm \sqrt{5}\alpha/\kappa$, and $1\pm\sqrt{6}\alpha/\kappa$. Hence, uniformly in $\kappa\in K$, $$\label{Pbound2}
\big(1-\sqrt{6}|\alpha| \big) {\bf I} \leq {\mathbf{P}}_\kappa \leq \big(1+\sqrt{6}|\alpha| \big) {\bf I}\ .$$ Thus, all matrices ${\mathbf{P}}_\kappa$ are positive definite, if $|\alpha|<\frac1{\sqrt{6}}$.
Finally, if ${\mathbf{P}}_\kappa$ is chosen with $$ \alpha \in (0,\alpha_+)\,, \qquad \text{where }
\alpha_+ := \min\{ 1/\sqrt{6}\,,\ \alpha_{\delta_5}\,,\ \alpha_{\delta_6}\,,\ \alpha_{\delta_7}\,,\ \alpha_{\delta_8}\,,\ \alpha_{\delta_9}\,,\ \alpha_{\delta_{11}} \}\,,$$ $\beta =2\alpha$, $\gamma =\alpha$, and $\omega=\sqrt{6}\alpha$ uniformly for all $\kappa\in K$, then $$\label{Pbnd2}
{\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa \geq \left( \frac{10}{14} \right)^{10} \delta_{11}(1,\alpha,2\alpha,\alpha,\sqrt{6}\alpha) \ {\bf I} \qquad \text{ uniformly in } \kappa\in K \ .$$
Combining with yields the result.
We compute that ${\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa$ is twice the identity matrix whose upper left $21\times 21$ block is replaced by ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ given as [$$\begin{pmatrix} 2\ell\alpha & 0 & 0 & 0 & \tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & A & 0 & \tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & B & 0 & 0 & 0 & 0 & 0 & - i\beta/k & 0 & 0 & -C & 0 & 0 & \tfrac{3+\sqrt3}{6} \ell\beta & 0 & \tfrac{3-\sqrt3}{6} \ell\beta & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 2\ell \gamma & 0 & 0 & -i \gamma/k & 0 & 0 & 0 & 0 & 0 & \sqrt2 \ell \gamma & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 2\ell\omega & 0 & 0 & -i \omega/k & 0 & 0 & 0 & 0 & 0 & \sqrt2 \ell \omega & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & 0 & 2\ell\eta & 0 & 0 & D & 0 & \ell\eta & -i \eta/k & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2\ell\eta \\
0 & 0 & i \gamma/k & 0 & 0 & 2-2\ell\gamma & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & i \omega/k & 0 & 0 & 2-2\ell\omega & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
A & i \beta/k & 0 & 0 & D & 0 & 0 & 2-2\tfrac{\sqrt2}{\sqrt3} \ell\beta & 0 & -\tfrac{\sqrt2}{\sqrt3} \ell\beta & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & 0 & \ell\eta & 0 & 0 & -\tfrac{\sqrt2}{\sqrt3} \ell\beta & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & -C & 0 & 0 & i \eta/k & 0 & 0 & 0 & 0 & 0 & 2-2\ell\eta & 0 & 0 & -\tfrac{1}{\sqrt{3}} \ell\eta & 0 & -\tfrac{1}{\sqrt{3}} \ell\eta & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & \sqrt2 \ell\gamma & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & \sqrt2 \ell\omega & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & \tfrac{3+\sqrt3}{6} \ell\beta & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\tfrac{1}{\sqrt{3}} \ell\eta & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & \tfrac{3-\sqrt3}{6} \ell\beta & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\tfrac{1}{\sqrt{3}} \ell\eta & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 \\
0 & 0 & 0 & 0 & 2\ell\eta & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 \\
\end{pmatrix} \ .$$ ]{} with $\ell:=2\pi/L>0$ and $$A := \tfrac{\ell}{\sqrt3} (\sqrt2 \alpha -\sqrt3 \beta)\,,\quad B := \tfrac2{\sqrt3} \ell (\sqrt2 \beta -\sqrt3 \alpha)\,,\quad C :=
\tfrac{\ell}{\sqrt3} (\sqrt2 \eta -\sqrt3 \beta)\,,\quad D := \tfrac{\ell}{\sqrt3} (\sqrt3 \eta -\sqrt2 \beta)\,.$$ We seek to choose $\alpha$, $\beta$, $\gamma$, $\omega$ and $\eta$ such that the matrices ${\mathbf{P}}_\kappa$ and ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ are positive definite for all $\kappa\in K$. The positive definiteness of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ will follow from Sylvester’s criterion, if all minors of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ are positive. This will yield restrictions on the choice of parameters $\alpha$, $\beta$, $\gamma$, $\omega$ and $\eta$. The analysis will simplify, if we choose $\beta$, $\gamma$, $\omega$ and $\eta$ as multiples of $\alpha$, because then the first six columns will depend linearly on $\alpha$. For $\beta =\sqrt3 \alpha$, $\gamma =\alpha$, $\omega=\alpha$ and $\eta=\alpha$, we compute ${\mathbf{D}}_{\kappa,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha}$ as
$$\begin{aligned}
{\tiny \begin{pmatrix} 2\ell\alpha & 0 & 0 & 0 & \tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & \tfrac{\sqrt2 -3}{\sqrt3} \ell \alpha & 0 & \tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 2 (\sqrt2 -1) \ell\alpha & 0 & 0 & 0 & 0 & 0 & - i\sqrt3 \alpha/\kappa & 0 & 0 & \tfrac{3-\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & \tfrac{\sqrt3+1}{2} \ell\alpha & 0 & \tfrac{\sqrt3-1}{2} \ell\alpha & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 2\ell\alpha & 0 & 0 & -i \alpha/\kappa & 0 & 0 & 0 & 0 & 0 & \sqrt2 \ell \alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 2\ell\alpha & 0 & 0 & -i \alpha/\kappa & 0 & 0 & 0 & 0 & 0 & \sqrt2 \ell \alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & 0 & 2\ell\alpha & 0 & 0 & (1-\sqrt2) \ell\alpha & 0 & \ell\alpha & -i \alpha/\kappa & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2\ell\alpha \\
0 & 0 & i \alpha/\kappa & 0 & 0 & 2-2\ell\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & i \alpha/\kappa & 0 & 0 & 2-2\ell\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\tfrac{\sqrt2 -3}{\sqrt3} \ell\alpha & i \sqrt3 \alpha/\kappa & 0 & 0 & (1 -\sqrt2) \ell\alpha & 0 & 0 & 2-2\sqrt2 \ell \alpha & 0 & -\sqrt2 \ell \alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\tfrac{\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & 0 & \ell\alpha & 0 & 0 & -\sqrt2 \ell \alpha & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & \tfrac{3-\sqrt2}{\sqrt3} \ell\alpha & 0 & 0 & i \alpha/\kappa & 0 & 0 & 0 & 0 & 0 & 2-2\ell\alpha & 0 & 0 & -\tfrac{1}{\sqrt{3}} \ell\alpha & 0 & -\tfrac{1}{\sqrt{3}} \ell\alpha & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & \sqrt2 \ell\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & \sqrt2 \ell\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & \tfrac{1+\sqrt3}{2} \ell \alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\tfrac{1}{\sqrt{3}} \ell\alpha & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & \tfrac{\sqrt3-1}{2} \ell \alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\tfrac{1}{\sqrt{3}} \ell\alpha & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 & 0 \\
0 & 0 & 0 & 0 & 2\ell\alpha & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 2 \\
\end{pmatrix} }\ . \nonumber \\[3mm]
\label{Dka5a4aaa}\end{aligned}$$
Let $\delta_j(\kappa,\alpha,\beta,\gamma,\omega,\eta)$ denote the determinant of the upper left $j\times j$ submatrix of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ for integers $j=1,2,\ldots,21$. For our choice $\beta =\sqrt3 \alpha$, $\gamma =\alpha$, $\omega=\alpha$ and $\eta=\alpha$, the minors $\delta_j(\kappa,\alpha) :=\delta_j(\kappa,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha)$ are given in Tables \[table:minors:3D:1-14\]–\[table:minors:3D:15-21\].
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\delta_1(\kappa,\alpha)$ = $2\ell\alpha$
------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\delta_2(\kappa,\alpha)$ = $4(\sqrt2 -1)\ell^2\alpha^2$
$\delta_3(\kappa,\alpha)$ = $8(\sqrt2 -1)\ell^3\alpha^3$
$\delta_4(\kappa,\alpha)$ = $16(\sqrt2 -1)\ell^4 \alpha^4$
$\delta_5(\kappa,\alpha)$ = $\tfrac{80}{3} (\sqrt2 -1)\ell^5 \alpha^5$
$\delta_6(\kappa,\alpha)$ = $\tfrac{40}{3} (\sqrt2 -1)\ell^4 \alpha^5 p_6(\kappa,\alpha)$
with $p_6(\kappa,\alpha):= -4\ell^2 \alpha -\frac{\alpha}{\kappa^2} +4\ell$.
$\delta_7(\kappa,\alpha)$ = $\tfrac{20}{3} (\sqrt2 -1)\ell^3 \alpha^5 p_6(\kappa,\alpha)^2$
$\delta_8(\kappa,\alpha)$ = $12 \ell^2\ \alpha^5\ p_6(\kappa,\alpha)^2\ p_8(\kappa,\alpha)$
with $p_8(\kappa,\alpha) = \tfrac{2-3\sqrt2}{3} \ell^2 \alpha -\tfrac56 \frac{\alpha}{\kappa^2} +\tfrac{10}{9}(\sqrt2 -1)\ell$.
$\delta_9(\kappa,\alpha)$ = $2\ \delta_8(\kappa,\alpha)$
$\delta_{10}(\kappa,\alpha)$ = $\tfrac43 \ell^2\ \alpha^5\ p_6(\kappa,\alpha)^2\ p_{10}(\kappa,\alpha)$
with $p_{10}(\kappa,\alpha) = 9((\sqrt2 -1)\ell^2 +\frac1{\kappa^2})\ell \alpha^2 -6((8\sqrt2 -6)\ell^2 +\frac5{\kappa^2})\alpha +40(\sqrt2 -1)\ell$.
$\delta_{11}(\kappa,\alpha)$ = $\tfrac29 \ell\ \alpha^5\ p_6(\kappa,\alpha)^2\ p_{11}(\kappa,\alpha)$
with $p_{11}(\kappa,\alpha) = \big(p_{11,0}(\alpha) +p_{11,1}(\alpha) \frac1{\kappa^2}\big) \frac1{\kappa^2} + p_{11,2}(\alpha) \ell^2$,
$p_{11,0}(\alpha) = (54\sqrt2 -144)\ell^3 \alpha^3 +(672 -72\sqrt2)\ell^2 \alpha^2 -(216 +144\sqrt2)\ell\alpha$,
$p_{11,1}(\alpha) =18(6 -\ell\alpha) \alpha^2$,
$p_{11,2}(\alpha) =(9 -54\sqrt2)\ell^3 \alpha^3 +(456\sqrt2 -24) \ell^2 \alpha^2 +(472 -816\sqrt2) \ell \alpha + 480(\sqrt2 -1)$.
$\delta_{12}(\kappa,\alpha)$ = $\delta_{11}(\kappa,\alpha) \frac{p_{12}(\kappa,\alpha)}{p_6(\kappa,\alpha)}$ = $\tfrac29 \ell\ \alpha^5\ p_6(\kappa,\alpha)\ p_{11}(\kappa,\alpha)\ p_{12}(\kappa,\alpha)$
with $p_{12}(\kappa,\alpha) = 4\ell^3 \alpha^2 -12\ell^2 \alpha +8\ell -\frac{2 \alpha}{\kappa^2}$.
$\delta_{13}(\kappa,\alpha)$ = $\delta_{12}(\kappa,\alpha) \frac{p_{12}(\kappa,\alpha)}{p_6(\kappa,\alpha)}
= \delta_{11}(\kappa,\alpha) \Big(\frac{p_{12}(\kappa,\alpha)}{p_6(\kappa,\alpha)}\Big)^2
= \tfrac29 \ell\ \alpha^5\ p_{11}(\kappa,\alpha) p_{12}(\kappa,\alpha)^2$
$\delta_{14}(\kappa,\alpha)$ = $\tfrac1{9\ (1+\sqrt3)^2} \ell\ \alpha^5\ p_{12}(\kappa,\alpha)^2\ p_{14}(\kappa,\alpha)$
with $p_{14}(\kappa,\alpha) = \big(p_{14,0}(\alpha) +p_{14,1}(\alpha) \frac1{\kappa^2}\big) \frac1{\kappa^2} +\ell^2 p_{14,2}(\alpha)$,
$p_{14,0}(\alpha) =
(-108\sqrt6 -72\sqrt3 -180\sqrt2 -144) \ell^4 \alpha^4
+(360\sqrt6 -1824\sqrt3 +720\sqrt2 -3396) \ell^3 \alpha^3$
$+(-576\sqrt6 +5952\sqrt3 -1152\sqrt2 +11760) \ell^2 \alpha^2
+(-1152\sqrt6 -1728\sqrt3 -2304\sqrt2 -3456) \ell \alpha$,
$p_{14,1}(\alpha) = 144\ (\sqrt3 +2)\ (6 -\ell \alpha)\ \alpha^2$,
$p_{14,2}(\alpha) = (1440 -180\sqrt6 +828\sqrt3 -324\sqrt2) \ell^4 \alpha^4
-(9348 -336\sqrt6 -5400\sqrt3 -624\sqrt2) \ell^3 \alpha^3$
$+(11056 +3424\sqrt6 +6368\sqrt3 +6864\sqrt2) \ell^2 \alpha^2
+(4192 -6528\sqrt6 +1856\sqrt3 -13056\sqrt2) \ell \alpha$
$+(3840\sqrt6 -3840\sqrt3 +7680\sqrt2 -7680)$.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Let $\delta_j(\kappa,\alpha,\beta,\gamma,\omega,\eta)$ denote the determinant of the upper left $j\times j$ submatrix of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ for integers $j=1,2,\ldots,21$. For our choice $\beta =\sqrt3 \alpha$, $\gamma =\alpha$, $\omega=\alpha$ and $\eta=\alpha$, the minors $\delta_j(\kappa,\alpha)=\delta_j(\kappa,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha)$ for integers $j=1,2,\ldots,14$, are given in this table.[]{data-label="table:minors:3D:1-14"}
The first five minors are positive if $\alpha$ is positive.
The sixth minor $\delta_6(\kappa,\alpha)$ satisfies for positive $\alpha$ the inequality $\delta_6(\kappa,\alpha)\geq \delta_6(1,\alpha)$ for all $\kappa\in K$. Moreover, the factor $p_6(1,\alpha)$ is positive for $\alpha\in(0,\alpha_{p_6})$ with $\alpha_{p_6} :={4\ell}/{(4\ell^2 +1)}$. Thus the sixth minor $\delta_6(\kappa,\alpha)$ is positive for all $\kappa\in K$ if $0<\alpha<\alpha_{\delta_6}$ with $\alpha_{\delta_6} :=\alpha_{p_6}$.
Following from the analysis of factor $p_6(\kappa,\alpha)$, the seventh minor $\delta_7(\kappa,\alpha)$ is positive for all $\kappa\in K$ if $0<\alpha<\alpha_{\delta_7} :=\alpha_{\delta_6}$.
The eighth minor $\delta_8(\kappa,\alpha)$ has a factorization. For positive $\alpha$ and $\kappa\in K$, the inequalities $p_8(\kappa,\alpha) \geq p_8(1,\alpha)$ and consequently $\delta_8(\kappa,\alpha)\geq \delta_8(1,\alpha)$ hold. Moreover, the linear polynomial $p_8(1,\alpha)$ is positive for $0<\alpha<\alpha_{p_8}$ with $\alpha_{p_8} := {20}{(\sqrt2 -1)\ell}/{3 ((6\sqrt2 -4)\ell^2 +5)}$. Thus, for all $\kappa\in K$, the eighth minor $\delta_8(\kappa,\alpha)$ is positive if $0<\alpha<\alpha_{\delta_8}$ with $\alpha_{\delta_8} :=\min \big\{\alpha_{p_6}\,,\ \alpha_{p_8}\} =\alpha_{p_8}$.
The ninth minor satisfies $\delta_9(\kappa,\alpha) = 2\ \delta_8(\kappa,\alpha)$, hence, it is positive for all $0<\alpha<\alpha_{\delta_9}:=\alpha_{\delta_8}$ and $\kappa\in K$.
The tenth minor $\delta_{10}(\kappa,\alpha)$ has a factorization. The factor $p_{10}(\kappa,\alpha)$ has the $\kappa$-dependent summand $(9\ell\alpha -30)\alpha /\kappa^2$, which is negative for $0<\alpha<{10}/{3\ell}$. Under this assumption, the inequalities $p_{10}(\kappa,\alpha) \geq p_{10}(1,\alpha)$ and $\delta_{10}(\kappa,\alpha)\geq \delta_{10}(1,\alpha)$ hold for all $\kappa\in K$. The quadratic polynomial $p_{10}(1,\alpha)$ has two positive roots $0<\alpha^{(10)}_-<\alpha^{(10)}_+$ and is positive if $\alpha<\alpha_{p_{10}}$ with $\alpha_{p_{10}} :=\alpha^{(10)}_-$. Thus, the tenth minor $\delta_{10}(\kappa,\alpha)$ is positive for all $\kappa\in K$, if $0<\alpha<\alpha_{\delta_{10}}$ with $\alpha_{\delta_{10}}:=\min\big\{\alpha_{p_6}\,,\ {10}/{3\ell}\,,\ \alpha_{p_{10}}\} =\min\big\{\alpha_{p_6}\,,\ \alpha_{p_{10}}\}$.
The eleventh minor $\delta_{11}(\kappa,\alpha)$ has a factorization. Due to Lemma \[lem:p\], the inequality $p_{11}(\kappa,\alpha) \geq p_{11}(1,\alpha)$ holds for some positive $\widetilde\alpha_{p_{11}}$, and consequently $\delta_{11}(\kappa,\alpha)\geq \delta_{11}(1,\alpha)$ holds for all $0<\alpha<\widetilde\alpha_{p_{11}}$ and $\kappa\in K$. The polynomial $p_{11}(1,\alpha)$ is positive at $\alpha=0$, hence there exists a positive number $\alpha_{p_{11}}$ such that $p_{11}(1,\alpha)$ is positive for $0<\alpha <\alpha_{p_{11}}$ and all $\kappa\in K$. Consequently, for $0<\alpha <\alpha_{\delta_{11}}$ with $\alpha_{\delta_{11}} :=\min\{\alpha_{p_{11}}\,,\ \widetilde\alpha_{p_{11}}\,,\ \alpha_{p_6}\}$ the eleventh minor $\delta_{11}(\kappa,\alpha)$ is positive for all $\kappa\in K$.
The twelfth minor $\delta_{12}(\kappa,\alpha)$ has a factorization. For positive $\alpha$, the inequalities $p_{12}(\kappa,\alpha) \geq p_{12}(1,\alpha)$ and $\delta_{12}(\kappa,\alpha)\geq \delta_{12}(1,\alpha)$ hold for all $0<\alpha<\min\{\alpha_{\delta_{11}}\,,\ \alpha_{p_6}\} =\alpha_{\delta_{11}}$ and $\kappa\in K$. The quadratic polynomial $p_{12}(1,\alpha)$ has two positive roots $0<\alpha^{(12)}_- \leq\alpha^{(12)}_+$ and is positive for $0<\alpha<\alpha_{p_{12}}$ with $\alpha_{p_{12}}:= \alpha^{(12)}_-$. Thus, the twelfth minor $\delta_{12}(\kappa,\alpha)$ is positive for all $\kappa\in K$, if $0<\alpha<\alpha_{\delta_{12}}$ with $\alpha_{\delta_{12}} :=\min\{\alpha_{p_6}\,,\ \alpha_{p_{12}}\,,\ \alpha_{\delta_{11}}\} =\min\{\alpha_{p_{12}}\,,\ \alpha_{\delta_{11}}\}$.
The thirteenth minor satisfies $\delta_{13}(\kappa,\alpha) =\tfrac29 \ell\ \alpha^5\ p_{11}(\kappa,\alpha) p_{12}(\kappa,\alpha)^2$. Therefore the thirteenth minor $\delta_{13}(\kappa,\alpha)$ is positive for all $\kappa\in K$ if $0<\alpha<\alpha_{\delta_{13}}$ with $\alpha_{\delta_{13}}:=\min\{\alpha_{p_{11}}\,,\ \alpha_{p_{12}} \}$.
The fourteenth minor $\delta_{14}(\kappa,\alpha)$ has a factorization. The polynomial $p_{14,1}(\alpha)$ is positive if $0<\alpha<6/\ell$. Moreover, the quartic polynomial $p_{14,0}(\alpha) +2 p_{14,1}(\alpha)$ is zero at $\alpha=0$, having a negative derivative at $\alpha=0$. Thus there exists a positive number $\alpha^{(14,0)}$ such that $p_{14,0}(\alpha) +2 p_{14,1}(\alpha)$ is negative for $0<\alpha<\alpha^{(14,0)}$. Due to Lemma \[lem:p\], the inequality $p_{14}(\kappa,\alpha) \geq p_{14}(1,\alpha)$ holds for $0\leq \alpha\leq \widetilde\alpha_{p_{14}}:=\min\{6/\ell\,,\ \alpha^{(14,0)} \}$, and consequently $\delta_{14}(\kappa,\alpha)\geq \delta_{14}(1,\alpha)$ holds for all $0\leq \alpha\leq \widetilde\alpha_{p_{14}}$ and $\kappa\in K$. The polynomial $p_{14}(1,\alpha)$ is positive at $\alpha=0$, hence there exists a positive number $\alpha_{p_{14}}$ such that $p_{14}(\kappa,\alpha)$ is positive for $0<\alpha <\alpha_{p_{14}}$ and all $\kappa\in K$. Consequently, for $0<\alpha<\alpha_{\delta_{14}}$ with $\alpha_{\delta_{14}} :=\min\{\alpha_{p_{12}}\,,\ \widetilde\alpha_{p_{14}}\,,\ \alpha_{p_{14}}\}$ the fourteenth minor $\delta_{14}(\kappa,\alpha)$ is positive for all $\kappa\in K$.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\delta_{15}(\kappa,\alpha)$ = $2\ \delta_{14}(\kappa,\alpha)$
------------------------------ -----------------------------------------------------------------------------------------------------------------------------------------
$\delta_{16}(\kappa,\alpha)$ = $\tfrac89 \tfrac{2+\sqrt3}{(1+\sqrt3)^2} \ell\ \alpha^5\ p_{12}(\kappa,\alpha)^2\ p_{16}(\kappa,\alpha)$
with $p_{16}(\kappa,\alpha) = \big(p_{16,0}(\alpha) +p_{16,1}(\alpha) \frac1{\kappa^2}\big) \frac1{\kappa^2} +\ell^2 p_{16,2}(\alpha)$,
$p_{16,0}(\alpha) =
-36 (\sqrt2 +2) \ell^4 \alpha^4
+(144 \sqrt2 -744) \ell^3 \alpha^3$
$+(-288 \sqrt2 +2976) \ell^2 \alpha^2
+(-576 \sqrt2 -864) \ell \alpha$,
$p_{16,1}(\alpha) = 72 (6 -\alpha \ell) \alpha^2$,
$p_{16,2}(\alpha) =
27 \ell^5 \alpha^5
+(-144 \sqrt2 +216) \ell^4 \alpha^4
+(-24 \sqrt2 -2412) \ell^3 \alpha^3$
$+(1632 \sqrt2 +3104) \ell^2 \alpha^2
+(-3264 \sqrt2 +928) \ell \alpha
+1920 (\sqrt2 -1)$.
$\delta_{17}(\kappa,\alpha)$ = $2\ \delta_{16}(\kappa,\alpha)$
$\delta_{18}(\kappa,\alpha)$ = $2^2\ \delta_{16}(\kappa,\alpha)$
$\delta_{19}(\kappa,\alpha)$ = $2^3\ \delta_{16}(\kappa,\alpha)$
$\delta_{20}(\kappa,\alpha)$ = $2^4\ \delta_{16}(\kappa,\alpha)$
$\delta_{21}(\kappa,\alpha)$ = $\tfrac{256 (\sqrt3 +2) (24\sqrt2 +61)}{23121 (\sqrt3 +1)^2} \ell\ \alpha^5\ p_{12}(\kappa,\alpha)^2\ p_{21}(\kappa,\alpha)$
with $p_{21}(\kappa,\alpha) = \Big(p_{21,0}(\alpha) +p_{21,1}(\alpha) \frac1{\kappa^2}\Big) \frac1{\kappa^2} +\ell^2 p_{21,2}(\alpha)$,
$p_{21,0}(\alpha) =
(-1152 \sqrt2 +2928) \ell^5 \alpha^5
+(-468 \sqrt2 -2664) \ell^4 \alpha^4
+(75024 \sqrt2 -175272) \ell^3 \alpha^3$
$+(-130464 \sqrt2 +300768) \ell^2 \alpha^2
+(-14400 \sqrt2 -25056) \ell \alpha$,
$p_{21,1}(\alpha) = (-1728 \sqrt2 +4392) (6 -\ell \alpha) \alpha^2$,
$p_{21,2}(\alpha) =
7707 \ell^5 \alpha^5
+(-25248 \sqrt2 +95000) \ell^4 \alpha^4
+(89448 \sqrt2 -353228) \ell^3 \alpha^3$
$+(158880 \sqrt2 +38048) \ell^2 \alpha^2
+(-417216 \sqrt2 +464416) \ell \alpha
+1920 (85\sqrt2 -109)$.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Let $\delta_j(\kappa,\alpha,\beta,\gamma,\omega,\eta)$ denote the determinant of the upper left $j\times j$ submatrix of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ for integers $j=15,\ldots,21$. For our choice $\beta =\sqrt3 \alpha$, $\gamma =\alpha$, $\omega=\alpha$ and $\eta=\alpha$, the minors $\delta_j(\kappa,\alpha)=\delta_j(\kappa,\alpha,\sqrt3 \alpha,\alpha,\alpha,\alpha)$ are given in this table.[]{data-label="table:minors:3D:15-21"}
The fifteenth minor $\delta_{15}(\kappa,\alpha) = 2\ \delta_{14}(\kappa,\alpha)$ is positive for all $\kappa\in K$ if $0<\alpha< \alpha_{\delta_{15}} :=\alpha_{\delta_{14}}$.
The sixteenth minor $\delta_{16}(\kappa,\alpha)$ has a factorization. The polynomial $p_{16,1}(\alpha)$ is positive if $0<\alpha<6/\ell$. Under this assumption, the quartic polynomial $p_{16,0}(\alpha) +2 p_{16,1}(\alpha)$ is zero at $\alpha=0$, having a negative derivative at $\alpha=0$. Thus there exists a positive number $\alpha^{(16,0)}$ such that $p_{16,0}(\alpha) +2 p_{16,1}(\alpha)$ is negative for $0<\alpha<\alpha^{(16,0)}$. Due to Lemma \[lem:p\], the inequality $p_{16}(\kappa,\alpha) \geq p_{16}(1,\alpha)$ holds for all $0\leq \alpha\leq\widetilde\alpha_{p_{16}}:=\min\{6/\ell\,,\ \alpha^{(16,0)} \}$, and $\delta_{16}(\kappa,\alpha)\geq \delta_{16}(1,\alpha)$ holds for all $0\leq \alpha\leq\widetilde\alpha_{p_{16}}$ and $\kappa\in K$. The polynomial $p_{16}(1,\alpha)$ is positive at $\alpha=0$, hence there exists a positive number $\alpha_{p_{16}}$ such that $p_{16}(\kappa,\alpha)$ is positive for $0<\alpha <\alpha_{p_{16}}$ and all $\kappa\in K$. Consequently, for $0<\alpha< \alpha_{\delta_{16}}$ with $\alpha_{\delta_{16}}:=\min\{\alpha_{p_{16}}\,,\ \widetilde\alpha_{p_{16}}\,,\ \alpha_{p_{12}} \}$ the sixteenth minor $\delta_{16}(\kappa,\alpha)$ is positive for all $\kappa\in K$.
The seventeenth to twentieth minors are multiples of the sixteenth minor. Therefore, these minors are positive for all $\kappa\in K$ under the same condition $0<\alpha<\alpha_{\delta_{16}}$.
The twenty-first minor $\delta_{21}(\kappa,\alpha)$ has a factorization. The polynomial $p_{21,1}(\alpha)$ is positive if $0<\alpha<6/\ell$. The quintic polynomial $p_{21,0}(\alpha) +2 p_{21,1}(\alpha)$ is zero at $\alpha=0$, having a negative derivative at $\alpha=0$. Thus there exists a positive number $\alpha^{(21,0)}$ such that $p_{21,0}(\alpha) +2 p_{21,1}(\alpha)$ is negative for $0<\alpha<\alpha^{(21,0)}$. Due to Lemma \[lem:p\], the inequality $p_{21}(\kappa,\alpha) \geq p_{21}(1,\alpha)$ holds for $0\leq\alpha\leq \widetilde\alpha_{p_{21}}:=\min\{6/\ell\,,\ \alpha^{(21,0)} \}$, and $\delta_{21}(\kappa,\alpha)\geq \delta_{21}(1,\alpha)$ holds for all $0\leq\alpha\leq\widetilde\alpha_{p_{21}}$ and $\kappa\in K$. The polynomial $p_{21}(1,\alpha)$ is positive at $\alpha=0$, hence there exists a positive number $\alpha_{p_{21}}$ such that $p_{21}(1,\alpha)$ is positive for $0<\alpha <\alpha_{p_{21}}$. Consequently, for $0<\alpha< \alpha_{\delta_{21}}$ with $\alpha_{\delta_{21}}=\min\{\alpha_{p_{21}}\,,\ \widetilde\alpha_{p_{21}}\,,\ \alpha_{p_{12}} \}$ the twenty-first minor $\delta_{21}(\kappa,\alpha)$ is positive for all $\kappa\in K$.
Let $\{\lambda_1,\lambda_2,\ldots,\lambda_{21}\}$ be the eigenvalues of ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ arranged in increasing order. We seek a lower bound on $\lambda_1$. As long as ${\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}$ is positive definite, the arithmetic-geometric mean inequality implies $$\begin{aligned}
\lambda_1(\kappa,\alpha,\beta,\gamma,\omega,\eta) &= \frac{\delta_{21}(\kappa,\alpha,\beta,\gamma,\omega,\eta)}{\prod_{j=2}^{21} \lambda_j}
\geq \delta_{21}(\kappa,\alpha,\beta,\gamma,\omega,\eta) \left( \frac{20}{\sum_{j=2}^{21} \lambda_j} \right)^{20}\\
&\geq \delta_{21}(\kappa,\alpha,\beta,\gamma,\omega,\eta) \left( \frac{20}{{\operatorname{Tr}}[{\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}]} \right)^{20}
= \left( \frac{20}{32} \right)^{20} \delta_{21}(\kappa,\alpha,\beta,\gamma,\omega,\eta) \ ,
\end{aligned}$$ since ${\operatorname{Tr}}[{\mathbf{D}}_{\kappa,\alpha,\beta,\gamma,\omega,\eta}] =32$ independently of $\kappa$, $\alpha$, $\beta$, $\gamma$, $\omega$ and $\eta$.
A simple computation shows that the eigenvalues of ${\mathbf{P}}_\kappa$ are $1$, $1 \pm \alpha/\kappa$ (3-fold), and $1 \pm2\alpha/\kappa$. Hence for positive $\alpha$ $$\label{Pbound3}
(1-2\alpha ) {\bf I} \leq {\mathbf{P}}_\kappa \leq (1+2\alpha) {\bf I}$$ uniformly in $\kappa$. Thus, all matrices ${\mathbf{P}}_\kappa$ are positive definite, if $0<\alpha<1/2$. Finally, if ${\mathbf{P}}_\kappa$ is chosen with $$ \alpha \in (0,\alpha_+)\,, \qquad \text{where }
\alpha_+ := \min\{ 1/2\,,\ \alpha_{\delta_6}\,,\ \alpha_{\delta_7}\,,\ \ldots\,,\ \alpha_{\delta_{21}} \}\,,$$ $\beta =\sqrt3 \alpha$, $\gamma =\alpha$, $\omega=\alpha$, and $\eta=\alpha$ uniformly for all $\kappa\in K$, then $$\label{Pbnd3}
{\mathbf{C}}_\kappa^*{\mathbf{P}}_\kappa + {\mathbf{P}}_\kappa{\mathbf{C}}_\kappa \geq \left( \frac{20}{32} \right)^{20} \delta_{21}(1,\alpha) \ {\bf I} \qquad \text{ uniformly in } \kappa\in K \ .$$ Combining with yields the result.
**Acknowledgement:** The first author (FA) was supported by the FWF-funded SFB \# F65. The second author (AA) was partially supported by the FWF-doctoral school “Dissipation and dispersion in non-linear partial differential equations” and the FWF-funded SFB \# F65. The third author (EC) was partially supported by U.S. N.S.F. grant DMS 1501007.
[99.]{} F. Achleitner, A. Arnold and E. A. Carlen, On linear hypocoercive BGK models, in “From Particle Systems to Partial Differential Equations III”, Springer Proceedings in Mathematics & Statistics, **162** (2016), 1–37.
F. Achleitner, A. Arnold and D. Stürzer, *Large-time behavior in non-symmetric Fokker-Planck equations*, , **6** (2015), 1–68.
*Shuffling Cards and Stopping Times*, , **93** (1986), 333–348.
*On the rates of decay to equilibrium in degenerate and defective Fokker-Planck equations*, , arXiv:1709.10216 (2017).
*Sharp entropy decay for hypocoercive and non-symmetric [F]{}okker-[P]{}lanck equations with linear drift*, , arXiv:1409.5425 (2014).
P. Bergmann and J. L. Lebowitz, *A new approach to nonequilibrium processes*, , **99** (1955), 578–587.
, *A Model for Collision Processes in Gases. I. Small Amplitude Processes in Charged and Neutral One-Component Systems*, , **94** (1954), 511–525.
, *The [BGK]{} model with external confining potential: existence, long-time behaviour and time-periodic [M]{}axwellian equilibria*, , **136** (2009), 297–330.
, *Hypocoercivity without confinement*, , arXiv:1708.06180 (2017).
, “Linear matrix inequalities in system and control theory”, , **15** (1994), xii+193 pp. , *The cutoff phenomenon in finite [M]{}arkov chains*, , **93** (1996), 1659–1664.
, *Hypocoercivity for linear kinetic equations conserving mass*, , **367** (2015), 3807–3828. , *Hypocoercivity for kinetic equations with linear relaxation terms*, , **347** (2009), 511–516.
, *Note on [$N$]{}-dimensional [H]{}ermite polynomials*, , **2** (1949), 325–330.
, *Hypocoercivity and exponential time decay for the linear inhomogeneous relaxation [B]{}oltzmann equation*, , **46** (2006), 349–359.
, “Linear systems theory”, , (2009). , “Matrix analysis”, , 2nd ed. (2013).
L. Hörmander, *Hypoelliptic second order differential equations*, Acta Math., **119** (1969), 147–171.
, “Perturbation theory for linear operators”, , 2nd ed. (1976).
, *Large-time behaviour of solutions to hyperbolic-parabolic systems of conservation laws and applications*, , **106** (1987), 169–194.
, *Global existence to the [BGK]{} model of [B]{}oltzmann equation*, , **82** (1989), 191–205.
, *Weighted [$L^\infty$]{} bounds and uniqueness for the [B]{}oltzmann [BGK]{} model*, , **125** (1993), 289–295.
, “Information and Information Stability of Random Variables and Processes”, , (1964).
, *Contributions a l’etude mathématique des équations cinétiques*, PhD thesis, Université Paris Diderot - Paris 7, (1991).
, *Systems of equations of hyperbolic-parabolic type with applications to the discrete [B]{}oltzmann equation*, , **14** (1985), 249–275.
, *On the decay of solutions to the linearized equations of electromagnetofluid dynamics*, , **1** (1984), 435–457.
, “Hypocoercivity”, , **202** (2009), iv+141 pp.
[^1]: University of Vienna, Faculty of Mathematics, Oskar-Morgenstern-Platz 1, A-1090 Wien, Austria, [email protected]
[^2]: Vienna University of Technology, Institute of Analysis and Scientific Computing, Wiedner Hauptstr. 8-10, A-1040 Wien, Austria, [email protected]
[^3]: Department of Mathematics, Rutgers University, 110 Frelinghuysen Rd., Piscataway NJ 08854, USA, [email protected]
[^4]: An eigenvalue is [*defective*]{} if its geometric multiplicity is strictly less than its algebraic multiplicity. This difference is called [*defect*]{}.
| {
"pile_set_name": "ArXiv"
} |
---
address:
- 'Brookhaven National Lab, Upton, NY 11973'
- |
Department of Physics, University of Arizona\
Tucson, AZ 85721
author:
- 'T. Blum'
- 'J.E. Hetrick and D. Toussaint'
title: 'Simulating QCD at finite density with static quarks[^1]'
---
\#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
\#1
\#1\#2[ [[\#1]{}]{} ]{}
.5cm
Lattice QCD with a nonzero density of quarks is a difficult problem since the fermion determinant becomes complex[@ORIGINAL], thus we are led to consider approximations which hopefully capture some of the essential physics. Here we present a study of QCD at arbitrary quark density in an approximation where the dynamics of the quarks has been removed[@BHT]. This approach is analogous to the quenched approximation at zero density.
Our idea is to simultaneously take the limits of infinite quark mass and infinite chemical potential while the density of quarks remains fixed. This leaves us with quarks that can be present or absent at each lattice site, but which do not move in the spatial directions. The result is a much simpler fermion determinant such that gauge variables can be easily updated to equilibrium in the background of a prescribed density of quarks.
With a chemical potential included, the lattice Dirac operator using Kogut-Susskind quarks is M(x,y) = 2am\_q\_[x,y]{} &+& \_[=1,2,3]{} &+& Taking limits $m\rightarrow\infty$ and $\mu\rightarrow\infty$ simultaneously leaves $2ma$ along the diagonal, and the forward hopping terms, $e^{\mu a}U_t$. Each spatial point decouples from all others, and the fermion determinant is just a product of easily computed $SU(3)$ determinants: [\[detMsu3\]]{} (M) &=& \_e\^[3 a n\_t]{} (P\_+ C [**1**]{}),(P\_+C) &=& C\^3+C\^2P\_+ C P\_\^\* +1. Here $P_\xvec$ is the Polyakov loop at spatial site $\xvec$, and $n_t$ is the number of time slices. The coefficient of the unit matrix, $C$, is $(2ma/e^{\mu a})^{n_t}$, and is the fundamental parameter in our approximation, through which we fix the density. We are still left with a complex determinant, albeit a much simpler one, allowing us to generate high statistics. We estimate expectation values by taking the ratio [\[op\_over\_phase\]]{} = where $\theta$ is the phase of det($M$) and $\langle\rangle_{||}$ indicates an expectation value in the ensemble weighted by the modulus of the determinant.
The physical quark density is obtained from eq. \[detMsu3\] [\[su3den\]]{} n = = \_ where $V$ is the spatial volume. At $C=\infty$ the density is $0$; at $C=0$ the system is saturated with density 3 per site; $C=1$ represents “half-filling” and the density is $3/2$.
We have run simulations on $6^3\times 2$, $8^3\times 2$, $10^3\times
2$, and $6^3\times 4$ lattices as described in Ref. 2. The lattice spacing was set by measuring the rho mass. In figure 1 we summarize the behavior of the Polyakov loop magnitude ($|P|$) in the $T$—$\rho$ plane, on $n_t=2$ lattices, with a fitting function. At zero density, we see the strong first order transition at $T_c$. As the density increases, this transition becomes a smooth crossover for all nonzero values of the density at which we simulated (our smallest density just below the transition region was $\sim$ 0.02 quarks/fm$^3$). We are currently investigating the way in which the first order transition disappears by examining very low densities near the transition. Since we see no systematic dependence of the crossover on the spatial size, except for the expected decrease of the Polyakov loop magnitude on cold lattices, we conclude that this rounding is not a finite size effect.
= 10cm = 5cm
We use $\pbp$ evaluated for light quarks as an indicator of chiral symmetry breaking in the presence of a finite density of massive quarks. In a series of runs on $6^3\times 4$ lattices, we began with $\gbeta=5.0$, which is a fairly cold lattice, at a temperature less than half the zero density $T_c$. We find that $\pbp$ remains large at this (cold) value of $\gbeta$ even up to a density of 1.5 quarks/site (where $C=1$). Maintaining this maximum density (1.5 q’s/site) and increasing the temperature, we find a crossover to restored chiral symmetry at $\gbeta\approx 5.3$, a significantly lower temperature than the zero density $n_t=4$ transition, which occurs at $\gbeta
\approx 5.7$.
Does this static approximation have anything to do with real QCD? Certainly the nature of the high temperature transition at zero density depends strongly on the presence of dynamical quarks. However, it is not [*a priori*]{} clear to us that a deconfinement transition or chiral symmetry restoration driven by high density should depend on the quarks moving, or whether the mere presence of the quarks would be enough. In particular, we had not expected to see the zero density first order transition disappear for very small quark densities, or the signal of chiral symmetry restoration to vanish. This suggests that we might want to re-examine the conventional wisdom that a high density of quarks causes a phase transition similar to that caused by high temperature. This work was supported by DOE grant DE-FG03-95ER-40906.
[9]{}
and references therein.
[^1]: To appear in the Proceedings of [*DPF96*]{} University of Minnesota, August 10-15
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'McMullen’s $g$-vector is important for simple convex polytopes. This paper postulates axioms for its extension to general convex polytopes. It also conjectures that, for each dimension $d$, a stated finite calculation gives the formula for the extended $g$-vector. This calculation is done by computer for $d=5$ and the results analysed. The conjectures imply new linear inequalities on convex polytope flag vectors. Underlying the axioms is a hypothesised higher-order homology extension to middle perversity intersection homology (order-zero homology), which measures the failure of lower-order homology to have a ring structure.'
author:
- Jonathan Fine
bibliography:
- 'axioms-for-g-vector.bib'
- '../../2009/ehv/bayer-master.bib'
date: 18 November 2010
title: 'Axioms for the $g$-vector of general convex polytopes'
---
> How often have I said to you that when you have eliminated the impossible, whatever remains, however improbable, must be the truth?
>
> Sherlock Holmes in *The Sign of the Four* by Sir Arthur Conan Doyle
Introduction
============
This section describes the purpose and key concepts of this paper. Every simple convex polytope $X$ has a $g$-vector $g(X)$, a linear function of the face vector $f(X)$. McMullen [@mcm71:num] introduced the $g$-vector in 1971 to express his conjectured conditions on $f(X)$. In 1980 Stanley [@sta80:num] proved the necessity of these conditions, and Billera and Lee [@billee80:suf] the sufficiency. The distant goal is to find and prove the extension of these conditions to general convex polytopes.
McMullen’s conditions have three parts: (1) linear equations on $f(X)$, (2) linear inequalities $g_i(X) \geq 0$, and (3) pseudo-power growth limits $g_{i+1}\leq g_i^{\langle i \rangle}$. They correspond to there being homology groups $H(X)$ that respectively satisfy Poincaré duality, satisfy hard Lefschetz and are a ring generated by the facets. McMullen at the time was not aware of this connection.
The main results in this paper are: (1) axioms which are conjectured to determine the extended $g$-vector; (2) a conjecture that reduces this determination to a finite calculation; and (3) this calculation performed by computer in dimension $d=5$. Almost all of the calculation is using the axioms to eliminate the impossible.
Two partial extensions are already known. Bernstein, Khovanski and MacPherson independently (see [@dcml09:decomp §4.1]) found one in around 1982. From a rational convex polytope $X$ a projective algebraic variety $\mathbb{P}_X$ can be constructed (via toric geometry). The middle perversity intersection homology (mpih) Betti numbers $h(X)$ of $\mathbb{P}_X$ are by Poincaré duality palindromic. As in the simple case, $g_i(X) = h_i(X) - h_{i-1}(X)$ for $0 \leq 2i
\leq \dim X $ defines $g_i$. This extends the simple definition. By hard Lefschetz, $g_i(X) \geq 0$.
This $g$-vector is a linear function not of the face vector of $X$ (which counts the number of faces on $X$ of each dimension) but of the flag vector (which counts chains of inclusions between faces). In 1985 Bayer and Billera [@baybil85:gen] found another partial extension, which determines $g$ up to an invertible matrix. Their result is:
The vector space $\mathcal{F}_d$ spanned by the flag vectors of all $d$-dimensional convex polytopes has dimension the Fibonacci number $F_{d+1}$. The flag vectors $f(v({\mathrm{pt}}))$ are a basis for $\mathcal{F}_d$, where $v$ ranges over the $F_{d+1}$ degree $d$ words in $C$ and $IC$, as defined below.
Suppose $X$ is a convex polytope. The *cone* (or pyramid) $CX$ on $X$ is the convex hull of $X$ with a point, the *apex*, not lying in the affine linear span of $X$. The *cylinder* (or prism) $IX$ on $X$ is the Cartesian product of an interval, say $[0,
1]$, with $X$.
Thus in dimension $5$ the flag vector has $F_6 = 8$ independent components while the $g$-vector has $\lfloor 5/2 \rfloor + 3 = 3$. We wish to extend $g$ so that it encode the whole of $f$. Here are the basic hypotheses made regarding the components $g_i$ of the extended $g$-vector, besides being a linear function of the flag vector. But first we need a definition.
The difference $IC-CC$ is the $D$ operator on convex polytopes.
For all convex polytopes $X$ we have $g_i(X) \geq 0$.
$g_i(w({\mathrm{pt}})) \in \{0, 1\}$ for all words $w$ in $C$ and $D$.
By generalised Dehn-Sommerville the flag vectors $f(w({\mathrm{pt}}))$ are a basis for $\mathcal{F}_d$. The second hypothesis now implies that $g_i$ is determined by $s=\{w | g_i(w({\mathrm{pt}})) = 1\}$. For each $d$ there are finitely many (in fact $2^n$ where $n=F_{d+1}$) possibilities for $s$ and thus $g_i$. But if $g_i(X)< 0$ then $g_i$ fails the first hypothesis and so must be eliminated. The next hypothesis makes this elimination an explicit finite, but perhaps large, calculation.
The set $P_{01,d}$ of all $d$-dimensional $01$-polytopes and their polars eliminate all impossible $g_i$.
Obtaining the $g$-vector from what whatever remains requires additional steps that depend on precise concepts. The rest of this paper is organised as follows. After providing background and notation we define and motivate the $(C,D)$ basis and the zero-one on $(C,D)$ hypothesis. In particular we discuss the already known intersection homology of toric varieties, and a possible extension that encodes the remainder of the flag vector. Next we introduce axioms to give a precise statement of our assumptions and conjectures, and report on our calculations. Finally we summarise and discuss some open questions.
Background
==========
This section establishes notation, provides some definitions and provides other background. Some important matters not relevant to the core of this paper, such as the decomposition theorem, are left unexplained. Concepts used in only one section are generally not placed here.
Throughout *polytope* will mean convex polytope, $X$ will be a polytope of dimension $d$, $w$ a word of degree $d$ in $C$ and $D$ (as defined in the Introduction) and $s$ will be a set of words $w$ (also known as a *word-set*). Usually, $v$ is a word in $(C, IC)$ or in $(C, I)$.
We will use $f(X)$ to denote the *flag vector* of $X$ and $g(X)$ is our goal, the $g$-vector of $X$. It has *components* $g_i$. Suppose $Y$ is a weighted formal sum $\sum \alpha_i X_i$ of polytopes. By $f(Y)$ we will mean $\sum \alpha_if(X_i)$, and similarly for other vector- and number-valued functions, such as $g$ and $g_i$.
We use $\lambda_w = \lambda_w(X)$ to denote coefficients in the *$(C, D)$-basis*. (This should not be confused with the $cd$-index of convex polytopes.) We use $\lambda_s$ for $\sum_{w\in
s}\lambda_w$. We use $\deg_C w$ and $\deg_D w$ to denote the number of $C$’s and $D$’s in $w$, and write $\deg w = \deg_C w + 2 \deg_D w$. The *order* of a word is the number of occurences of $CD$ in it.
Every polytope $X$ has a *polar* $X^\vee$. This reverses inclusions on faces and so $f(X)$ and $f(X^\vee)$ are linear functions of each other. A polytope is *simple* if its facets (codimension 1 faces) are in general position, and *simplicial* if its vertices are in general position. Simple and simplicial correspond under polarisation.
Combinatorists usually start with simplicial polytopes but for homology simple polytopes are a better starting point. Here is an example. The polar of the $I$ operator is the bipyramid $B$, and [@baybil85:gen] used $B$ not $I$. The $IC$ equation reduces $IC -
CC$ to a Cartesian product operator. The polar of a Cartesian product not easy to describe. This paper takes the simple/homology point of view. For us, McMullen’s conditions apply to simple polytopes, even though he formulated them for simplicial polytopes.
From every rational polytope $X$ a toric variety $\mathbb{P}_X$ can be constructed. It has homology groups, which we denote by $H(X)$. The definition of $H(X)$ extends to all polytopes, rational or not. We will index homology groups by complex dimension (as the others are zero). Homology will always mean middle perversity intersection homology, which for simple polytopes is the same as ordinary homology. In §\[sect:higher\] we hypothesise the existence of *higher-order* homology groups, and call the already known part of $H(X)$ the *order-zero* homology.
We use $[a,b,c]$ to denote the polynomial $a + bx + cx^2$, and think of the order-zero part of $h(X)$ and $g(X)$ as polynomials in $x$. For example, $h(X) = \sum x^i h_i(X)$. This allows them to be multiplied. Thus, $h(\mathbb{P}_2)= [1, 1, 1]$ and $h(\mathbb{P}_2
\times \mathbb{P}_3) = [1, 1, 1][1, 1, 1, 1]$ by the Künneth formula.
The *hard Lefschetz theorem* says that the hyperplane class $\omega\in H_{d-1}(X)$ induces an isomorphism $\omega^{j-i}:H_i(X) \to
H_j(X)$ when $i\geq j$ and $i+j=d$. A class $\eta$ is *primitive* if $\omega^{j-1+1}(\eta) = 0$.
If $\mathcal{B}$ is a Boolean expression we let $(\mathcal{B})$ denote $1$ if $\mathcal{B}$ is true and $0$ otherwise. Thus, $\delta_{ij} =
(i = j)$ is the Kronecker delta and $(w \in s)$ is the characteristic function of a set $s$.
Finally, some miscellaneous definitions and notations. A $d$-cube, or *cube* for short, will mean a $d$-dimensional hypercube. We do not need in this paper the definition of the *pseudo-power* $n^{\langle i \rangle}$. The interested reader can find it in [@mcm71:num] or [@bayerlee94:aspects]. We use $\lfloor x
\rfloor$ to denote the largest integer not greater than $x$. Hypotheses, axioms and conjectures are numbered separately. Everything else is numbered together.
The $(C, D)$ basis
==================
According to generalised Dehn-Sommerville every polytope flag vector has a unique expression as a linear combination of the flag vectors of $v({\mathrm{pt}})$ where $v$ is a word in $C$ and $IC$. This applies in particular to $v({\mathrm{pt}})$ where $v$ is any word in $C$ and $I$ that is *not* a word in $C$ and $IC$. This suggests that there is a relationship between $I$ and $C$. In fact [@fine95:mvic]:
As operators on flag vectors, $I$ and $C$ satisfy the equation $$(IC - CC) \> I \> \equiv \> I \> (IC - CC)\>.$$
Because it is central to the properties of $D$, we summarise the proof given in [@fine95:mvic]. Let $X$ be a convex polytope. By definition $CCX$ has two apexes, but there is no geometric difference between the first and the second. Put another way, $CCX$ is the join of $X$ with an interval and that interval is the *apex edge* of $CCX$. Similarly, $ICX$ has an apex edge. Along their apex edges $CCX$ and $ICX$ have the same combinatorial structure and so $$ICX - CCX \equiv Y - Z$$ where $Y$ and $Z$ are respectively the truncation of $ICX$ and $CCX$ along their apex edges. However, $Y$ and $Z$ are respectively the Cartesian product of a square and a triangle with $X$. Thus, as an operator on flag vectors, $IC - CC$ is a difference of Cartesian products. But $I$ is also a Cartesian product operator, and such products commute. The $IC$ equation follows. In fact, the $IC$ equation together with $I({\mathrm{pt}}) = C({\mathrm{pt}})$ (which follows from the $IC$ equation applied to the empty set) allows any word in $I$ and $C$ applied to a point to be reduced to an equivalent combination of $C$ and $IC$ words applied to a point.
This is one motivation for $D = IC - CC$. The proof of the $IC$ equation establishes:
For any $X$ $$DX \> \equiv \> (IC - CC) X \> \equiv \> (Y - Z) \times X$$ where $Y$ and $Z$ are respectively a square and a triangle.
Every convex polytope $X$ has a unique representation $$X \equiv \sum\nolimits_w \lambda_w w({\mathrm{pt}})$$ where the sum is over all degree $d$ words $w$ in $C$ and $D$ and the coefficients $\lambda_w = \lambda_w(X)$ are linear functions of the flag vector $f(X)$.
The mapping $w \mapsto \lambda_w$ is called the *$CD$ vector* of $X$. If $s$ is a set of $(C, D)$ words we will use $\lambda_s$ to denote $\sum_{w\in s}\lambda_w$.
The zero-one hypothesis implies that each $g_i$ is equal to $\lambda_s$ for some set of words $s$.
Homology
========
In this section we look at the already known homology. In particular we state formulas for Betti numbers in terms of the $(C, D)$ basis. First simple polytopes. The product of two simple polytopes is also simple. Thus if $X$ is simple then so is $DX$ (by which we mean that its flag vector can be written as a formal sum of the flag vectors of simple polytopes). This follows because $DX$ is equivalent to $(Y -
Z)\times X$ for simple $Y$ and $Z$.
Clearly, $h({\mathrm{pt}}) = [1]$, $h(C({\mathrm{pt}})) = [1, 1]$ and similarly $h(C^j({\mathrm{pt}})) = [1, \ldots, 1]$. Similarly $h(Y) = [1,1,1]$, $h(Z)
= [1,2,1]$, and $h(DX) = [0, 1, 0]h(X)$. Thus $h(D^iC^j({\mathrm{pt}})) =
[0,1,0]^i[1, \ldots, 1]$. Also, any simple-polytope $h$-vector is a weighted sum of $D^iC^j({\mathrm{pt}})$ $h$-vectors and so the formal sums $D^iC^j({\mathrm{pt}})$ provide a basis for the flag vectors of simple polytopes. This proves:
Suppose $X$ is a simple polytope. Then $ X \equiv \sum\nolimits _{i=0}^{\lfloor d/2 \rfloor} \> \lambda_i D^iC^{d-2i}({\mathrm{pt}})
$ where $\lambda_i = g_i(X)$.
For general polytopes it follows from the known formula for intersection homology Betti numbers [@sta87:gen] that $h(IX) = [1,
1] h(X)$. It also follows that for example if $h(X) = [a, b, c, b,
a]$ then $h(CX) = [a, b, c, c, b, a]$. The general rule for $C$ is to repeat the middle term (for $d$ even) or the next to middle term (for $d$ odd). If $d$ is odd then the two next-to-middle terms are equal. Thus, if $h(X) = [a, b, b, a]$ then $h(CX) = [a, b, b, b, a]$.
From this $h(DX) = [0, 1, 0] h(X)$ follows easily. (Alternatively, the result follows from $D=Y-Z$ and the Künneth formula.) We now turn to the $g$-vector. The rule for $h(CX)$ then becomes $g(CX) =
g(X)$ while the rule for $D$ is $g(DX) = [0, 1, 0]g(X)$. Hence we have the following extension of the previous proposition:
Let $w$ be a word in $C$ and $D$. Then $$g_i(w(pt)) = (\deg_D w = i)$$ and by generalised Dehn-Sommerville this equation determines $g_i$ on all convex polytope flag vectors.
The known components $g_i$ of $g$ satisfy the zero-one on $(C,D)$ hypothesis.
The proof that $g_i(X) \geq 0$ is deep, and relies on the decomposition theorem [@dcml09:decomp].
Higher-order homology {#sect:higher}
=====================
This section assumes some knowledge of intersection homology and the topology of toric varieties. It provides both justification for the axioms and an interpretation of the outcome of calculations that follow.
For simple polytopes the homology $H(X)$ is a ring generated by the facets and from this the pseudo-power inequalities follow. For general polytopes $H(X)$ no longer has a ring structure. Our distant goal requires additional components in $H(X)$, which we will call the *higher-order homology*. The already known part we call the *order-zero homology*.
The main properties of higher-order homology are: (1) it should vanish when $X$ is simple, (2) when it vanishes there is a ring structure on $H(X)$, and (3) its Betti numbers should be a linear function of the flag vector.
The order-zero homology can be constructed by taking cycles and relations whose intersection with the strata satisfy what are called perversity conditions. This allows Poincaré duality and hard Lefschetz to hold (provided we use middle perversity). It also allows non-trivial local cycles to exist. For example there is a local cycle at each vertex of the octohedron (the polar of a $3$-cube).
The author expects the cycles and relations for higher-order homology to be subsets of those for order-zero homology, obtained by imposing locality conditions. In other words, fewer generators and fewer relations. For example, on the octohedron there are $8$ local $1$-cycles which under local equivalence are independent. However, there are $4$ independent global relations among these local $1$-cycles.
If $X=v({\mathrm{pt}})$ for $v$ a word in $(C,I)$ then the stratification is given by a single flag on $X$. This has consequences. Suppose $\eta$ and $\psi$ are cycles close to a stratum $X_i$. If $\eta$ and $\psi$ are globally equivalent then it seems that this relation can be deformed so that it too is close to $X_i$. In other words, on $X=
v({\mathrm{pt}})$ local and global equivalence are the same. This, of course, is not true in general.
Further, if $IC = D + CC$ is used to rewrite $v$ as a weighted sum of words in $(C, D)$ then the weights count the primitive cycles on $X$ with certain locality properties. Thus, each locality property gives rise to a zero-one function on the $(C,D)$ basis. (We will see, however, that for some plausible locality properties the resulting linear function will be negative on some polytope.) To summarise what this paper needs from this section:
The higher-order homology vanishes on simple polytopes.
Each higher-order homology group is defined by using a locality property.
Each locality property induces a zero-one function on the $(C,D)$ basis.
Axioms for the $g$-vector
=========================
The previously stated hypotheses have consequences for the $g$-vector, which we will call axioms. The calculation is based on the axioms. Recall that $X$ is any convex polytope, that $d$ is its dimension, that $w$ is a word of degree $d$ in $(C, D)$, and that $s$ is a set of words $w$. Here are the axioms.
$g(X)$ is a linear function of the flag vector $f(X)$.
$g(X)$ is a map $i \mapsto g_i(X)$ from $I$ to $\mathbb{Z}$, for $i$ in some index set $I = I_d$.
Each $g_i$ is called a *component* of $g$.
\[axiom:nonnegative\] $g_i(X)\geq 0$ for all $i\in I$ and all $X$.
The next axiom is crucial. It implies that $g$ is given by a choice from a known and finite set. Its motivation comes from the three previous sections. It depends upon the generalised Dehn-Sommerville equations.
For each $i$ in $I$ we have $g_i = \lambda_s$ for some set of words $s$.
Thus, we can take the index set $I$ to be a subset of the power set of all degree $d$ words in $C$ and $D$. For example, for $d=5$ we have $F_{5+1} = 8$ and so the power set has $2^8 = 256$ elements. Thus, *a priori* there are $2^{256}$ possible values for $I$. From now on we will write $s \in S$ instead of $i \in I$. The next tasks are to remove first the impossible and then the redundant. Axiom \[axiom:nonnegative\] provides some conditions.
Say that $s$ is *effective* if $\lambda_s(X) \geq 0$ for all $X$.
Let $E = E_d$ denote the set of all effective $s$ (of degree $d$). Now for the redundant. Clearly, if nonempty disjoint word-sets $s$ and $s'$ are both effective then so is their union. Such unions provide no new conditions on $f(X)$ and so are excluded from $g$. More generally:
Say that $s$ in $E_d$ is *extremal* if it cannot be expressed as a weighted sum $s = \sum_{t\neq s} \alpha_t t$ of other elements of $E_d$ with non-negative weights $\alpha_t$. (The empty set is not extremal.)
Each element of $s$ of $S$ are extremal in $E$.
The $g$-vector should embrace both zero-order and all possible higher-order homology.
The $g_i$ corresponding to order-zero homology are components of $g$.
If $s$ is effective, extremal, and $\lambda_s$ vanishes on simple polytopes, then $s \in S$.
Finally, we want $g$ to be complete and without redundancy.
The components of $g$ are a basis for all linear functions of the flag vector $f$.
Conjectures
===========
In this section we state two conjectures. If both are true then for each dimension $d$ a finite calculation will give a formula for the $g$-vector in that dimension, and hence linear inequalities $g_s(X)\geq 0$. The first conjecture is, of course:
There is a $g$-vector that satisfies the axioms stated in the previous section.
The set $E=E_d$ of effective word-sets $s$ of degree $d$ is central to this conjecture. The $g$-vector follows once we have $E$. To show that $s$ is *not* in $E$ it is enough to produce an $X$ such that $g_s(X)<0$. Therefore, a finite set $P$ of test polytopes is by elimination enough to determine $E$. We will make this more formal.
Let $P$ be a set of dimension $d$ polytopes. Define $E(P)$, the *word-sets effective on $P$*, to be the set of $s$ such that $\lambda_s(X)\geq 0$ for every $X$ in $P$.
Say that a set $P$ of polytopes is *broad* if $E(P) = E$, or in other words that $\lambda_s(X) \geq 0$ for $X\in P$ implies $\lambda_s(X)\geq 0$ for all $X$.
For each dimension $d$ there is a broad and finite set $P$ of $d$-polytopes.
In dimension $5$ there is a broad $P$ with at most $2^8 = 256$ polytopes. However, we don’t know what it is. The second conjecture, if true, provides an explicit broad set $P$. There is at present little evidence for it, other than the satisfactory output produced by the $d=5$ calculation, which is the subject of the next two sections. The set $P$ is produced from subsets of the vertices of a $d$-cube, together with polars. A cube can be represented so that each vertex component is either $0$ or $1$. We repeat the standard definition:
A *$01$-polytope* is the convex hull of a subset of the vertices of a cube.
A cube has $2^d$ vertices and $2d$ facets. The polar of a cube is the *cross-polytope*, the convex hull of the $d$ basis vectors $e_i$ and their negatives $-e_i$. The cross-polytope has $2^d$ facets (namely a choice of which $e_i$ are to have a negative sign). Every convex polytope has a *polar $X^\vee$* (which depends on the choice of a point $p$ in the interior of $X$). Each $i$-face on $X$ corresponds to a $d-i-1$ face of $X^\vee$ and vice versa. This bijection reverses inclusions and so the flag vector $f(X^\vee)$ of the polar is a linear function of $f(X)$ (and does not depend on the choice of $p$).
Let $P_{01,d}$ denote the set of all $d$-dimensional $01$-polytopes, together with their polars.
A $01$-polytope need not have the same dimension as its cube. But if of smaller dimension then there is a projection onto a face that gives an affinely equivalent polytope (which thus has the same flag vector). So without loss of generality we need only consider $X$ and $X^\vee$ obtained from the $d$-cube.
The set $P_{01,d}$ of $d$-dimensional $01$-polytopes and their polars is broad.
The $d=5$ calculation
=====================
In the previous section we conjectured that certain calculations would give a formula for the $g$-vector. In this section and the next we report on some such calculations. We applied `polymake` [@polymake] to a list of $01$-polytopes for $d=5$ supplied by Aichholzer [@aich10:personal]. This gave us a list of flag vectors. There are 1,226,525 such polytopes (up to cube symmetries) and the calculation took about 15 GHz days. Together with their polars this gives 688,298 distinct flag vectors.
We also wrote software [@fine10:bbpoly] to compute the $(C,D)$ vector for each of these flag vectors, and also to determine for each of the $256 = 2^8$ subsets of the $F_{5+1} = 8$ degree $5$ words $s$ in $(C, D)$ whether or not the sum of the corresponding coefficients in the $(C,D)$ vector was non-negative for all these flag vectors. This gives us $E(P_{01,d})$, as defined in the previous section.
Finally, we used `polymake` again to determine the extremal $s$ in $E(P_{01,d})$. Altogether there were $13$ such, of which we are expecting $8$ to be components of the $g$-vector. The meaning of the remaining $5$ is less clear. The next two results give the conjectured components of the $g$-vector.
For $P_{01,5}$ the order-zero $g_i$ (in the form ‘index $=$ deg : word-set’) $$\begin{aligned}
11111 = 0 &: CCCCC\\
2111 = 1 &: DCCC,\ CDCC,\ CCDC,\ CCCD\\
221 = 2 &: DDC,\ DCD,\ CDD\end{aligned}$$ are effective (this is already known) and extremal.
For $P_{01,5}$ the higher-order $g_s$ (in the form ‘index : word-set’) are $$\begin{aligned}
1211 &: CDCC,\ CCDC,\ CCCD,\ CDD\\
1121 &: CCDC,\ CCCD,\ DCD\\
1112 &: CCCD\\
122 &: CDD\\
212 &: DCD\end{aligned}$$
A word about the notation above, which is also used later. To each $s$ we associate an index, which is a word in the symbols $1$ and $2$. Thus, $g_{221}= g_2 = g_s$ where $s=\{DDC, DCD, CDD\}$. The index $i$, with $C$ and $D$ replacing $1$ and $2$, is the first listed element of the set $s$. This notation is concise and, with care, the context resolves any ambiguities.
We will now show that the above are a basis for all linear functions of the flag vector. Let $w$ be a word. It is enough to show that $\lambda_w$ is a linear combination of the $g_i$ above. But this is easy, because the components are listed in an upper-triangular order. In other words, every word appears first in one of the sets, and it does not appear in any subsequent set. This proves:
The above components (of the conjectured $g$-vector) are a basis for all linear functions of the flag vector.
We now turn to the remaining effective and extremal $s$. They do not vanish on simple polytopes, and they are not order-zero $g_i$ numbers. This is why they are not part of the $g$-vector.
For $P_{01,5}$ the remaining effective and extremal $s$ (in the form ‘inequality : word-set’) are $$\begin{aligned}
g_{122} \leq g_{221} + g_{2111} &\quad :\quad DDC,\ DCD,\ DCCC,\ CDCC,\ CCDC,\ CCCD\\
g_{122} \leq g_{221} + g_{1211} &\quad :\quad DDC,\ DCD,\ CDD,\ CDCC,\ CCDC,\ CCCD\\
g_{212} \leq g_{221} + g_{1112} &\quad :\quad DDC,\ CDD,\ CCCD\\
g_{212} \leq g_{221} + g_{1121} &\quad :\quad DDC,\ DCD,\ CDD,\ CCDC,\ CCCD\\
g_{212} \leq g_{221} + g_{2111} &\quad :\quad DDC,\ CDD,\ DCCC,\ CDCC,\ CCDC,\ CCCD\end{aligned}$$ where the left column is $\lambda_s \geq 0$ written in terms of the components of the $g$-vector.
These inequalities, an unanticipated by-product of the calculation, are more concisely be written as $$\begin{aligned}
g_{122} - g_{221} &\leq \min (g_{2111}, g_{1211})\\
g_{212} - g_{221} &\leq \min (g_{2111}, g_{1121}, g_{1112})\end{aligned}$$ and in this form they are similiar to McMullen’s pseudo-power inequalities $g_{i+1}\leq g_i^{\langle i \rangle}$.
A special test polytope
=======================
The $d=5$ calculation gives not only a formula for $g$ but also, implicitly, a small and broad set of test polytopes. Here is the most interesting example. The set $$s = \{ CDCC,\ CCDC,\ CCCD \}$$ has $\lambda_s = g_{1211} - g_{122}$ and is effective on the whole of $P_{01,5}$ with just one exception (up to symmetries of the cube).
In this section let $X$ be the $01$-polytope which has all the vertices of the $5$-cube except the set $V$ whose members, listed in lexicographic order, are $$\begin{aligned}
&
u_1 = (0,0,0,0,0),\quad
u_2 = (0,0,0,1,1),\quad
u_3 = (0,1,1,0,0),\quad
u_4 = (0,1,1,1,1),
\\\relax
&
v_1 = (1,0,1,0,1),\quad
v_2 = (1,0,1,1,0),\quad
v_3 = (1,1,0,0,1),\quad
v_4 = (1,1,0,1,0)\>.\end{aligned}$$
The exception is the polar $X^\vee$ of $X$. The polytope $X^\vee$ has flag vector (independent components in `polymake` order) $$f (X ) = (1, 24, 112, 152, 464, 80, 400, 696)$$ while the $(C, D)$ vector is given by $$\begin{aligned}
&(CCCC, 1),\quad (CCCD, 8),\quad (CCDC, 56),\quad (CDCC, -66),
\\
&(CDD, 20),\quad (DCCC, 20)\quad (DCD, 0),\quad (DDC, -5)\end{aligned}$$ and so $$\lambda_s(X^\vee) = \lambda_{CDCC} + \lambda_{CCDC} + \lambda_{CCCD} =
-66 + 56 + 8 = -2$$ which is, as claimed, negative. (This calculation, repeated many times, gave us $E(P_{01,5})$.
The $g$-vector of $X^\vee$, however, is $$\begin{aligned}
&g_0 = 1,
\quad g_1 = 8 + 56 - 66 + 20 = 18,
\quad g_2 = 20 + 0 - 5 = 15,\\
&g_{1211} = 8 + 56 - 66 + 20 = 18,\quad
g_{1121} = 8 + 56 + 0 = 64,\quad
g_{1112} = 8,\quad
g_{122} = 20, \quad
g_{212} = 0\end{aligned}$$ whose components are (for this polytope by construction of $g$) non-negative.
The set $V$ of missing vertices has a structure. The $u_i$ lie on a $4$-face, and the $v_i$ on the opposite $4$-face. The vertices $u_1$ and $u_4$ give a diagonal of the $4$-face, as do $u_2$ and $u_3$. The vertex pairs $\{u_1, u_2\}$, $\{u_1, u_3\}$, $\{u_2, u_4\}$ and $\{u_3, u_4\}$ all differ in two coordinates. The same is true for the $v_i$. Finally, each $u_i$ differs from each $v_j$ in three coordinates. To state this more formally, let $d_1$ denote the Hamming (or Manhattan) metric on the vertices of the cube; $d_1$ counts how many components differ. The following definition and proposition summarise these facts.
Let $V'$ be a subset of the vertices of a cube. For each vertex $v$ of the cube the $i$-th component of the *distance count of $v$ from $V'$* is the number of $v'\in V'$ with $d_1(v, v') = i$.
Let $v$ be any member of the missing vertices set $V$. Then the distance count of $v$ from $V$ is $(1,0,2,4,1,0)$.
This result has a converse. It tells us that, at least in this case, distance count can determine a vertex set. As construction of polytopes with special properties may be useful later, we state the converse and give the somewhat pedestrian proof.
Up to symmetry there is exactly one subset $V$ of the $5$-cube that has distance count $(1, 0, 2, 4, 1, 0)$.
Let $V$ be the subset above and let $V'$ be another one. Pick a vertex in $V'$. By cube symmetry without loss of generality (wlog) it is $u_1$. There’s only one vertex at distance $4$ from $u_1$ and wlog is is $u_4$. Pick a point at distance $2$ from $u_1$. Its distance from $u_4$ *a priori* is $2$ or $4$. But $d_1(u_1, u_4)=4$ and so by distance count it must be $2$ and wlog the point is $u_2$.
Now pick the other point at distance $2$ from $u_1$. As before, its first coordinate is zero. If not $u_3$ then it has distance $2$ from $u_1$, $u_2$ and $u_4$. This violates the distance count. So it is $u_3$.
Now choose a vertex $v$ at distance $3$ from all the $u_i$. We have $d_1(u_1, v) = 3$ and so it contains $3$ ones. If the first component is $0$ then $d_1(v_4, v) = 1$. So the first component is $1$. It cannot be $(1, 0, 0, 1, 1)$ or $(1, 1, 1, 0, 0)$ because of $u_2$ and $u_5$ respectively. Therefore the $v_i$ are the only vertices at distance $3$ from all the $u_i$.
Finally, we restate these results putting the focus on $V$.
The distance count $(1, 0, 2, 4, 1, 0)$ determines, by complement and polarisation, the unique up to symmetry polytope in $P_{01,5}$ that excludes $\{CDCC, CCDC, CCCD\}$.
Summary
=======
The section reviews the rest of the paper and sketches possibilities for future work.
The key new concepts are (1) that the components of the $g$-vector are zero-one on the $(C, D)$ basis, and (2) the set $E = E_d$ of effective word-sets given by $\{s | \lambda_s(X)\geq 0$ for all $X \}$. They arose as follows. The author computed $\lambda_s(X)$ for all $X$ in $P_{01,5}$ and for the $s$ in [@fine10:complete]. This was to test the $s$ for being effective. However, as we have seen, some came out negative. The author then inverted the process. He used $P_{01,5}$ to compute $E(P_{01,5})$ and hence obtained a conjecture. The structure of $E$, even if not as described by the axioms, is most likely worth studying. The conjectures make definite statements about $E(P_{01,d})$ that can be tested by finite calculation.
Two of the components of $g$ for $d=5$, namely $$\begin{aligned}
1211 &: CDCC,\ CCDC,\ CCCD,\ CDD\\
1121 &: CCDC,\ CCCD,\ DCD\end{aligned}$$ have mixed degree in $D$. The meaning of the $\deg_D=2$ terms, $CDD$ and $DCD$ respectively, is unclear. Although they may look improbable, they are an unavoidable consequence of our axioms and conjectures. The author expects that in general the lowest degree terms will be given as in [@fine10:complete].
The meaning of the conjectured inequalities $$\begin{aligned}
g_{122} - g_{221} &\leq \min (g_{2111}, g_{1211})\\
g_{212} - g_{221} &\leq \min (g_{2111}, g_{1121}, g_{1112})\end{aligned}$$ is far from clear. The analogous growth conditions in the simple case follow from $H(X)$ being generated by the facets of $X$. It is possible that [@fine98:uniform] will help here.
Further calculations would be helpful. However, the size of $P_{01,d}$ grows very rapidly with $d$. It might be practical to compute for the whole of $P_{01,6}$ but for $P_{01,7}$ only a subset can be used. Analysis of the $d=4$ and $5$ results may lead to smaller broad subsets of $P_{01,d}$. This is of course related to the construction of polytopes and thus the proof of sufficiency. Results in dimension $4$ should be compared to known results, conjectures and constructions [@bayerlee94:aspects; @bayer87:extended; @pafwern05:construction].
Simplicial poltyopes are a useful special case. Their polars are simple, and so McMullen’s conditions characterise their flag vectors. It may be instructive, say for $d=5$, to write the already known conditions on simplicial polytopes in terms of the conjectured $g$-vector.
The extension of McMullen’s conditions to general polytopes will have at least three components, namely: (1) linear equations on the flag vector, (2) linear equalities $g_s(X)\geq 0$, and (3) growth conditions on the $g_s$. The generalised Dehn-Sommerville equations of Bayer and Billera provide (1). Our first conjecture, if true, provides (2) abstractly while our second conjecture will then provide a finite calculation that gives (2) in a concrete form. It may also provide part of (3).
Here are three more distant goals: (1) produce, with supporting evidence, a conjectured formula for $g$ in all dimensions, (2) systematically produce test polytopes, particularly in higher dimensions, (3) translate the formulas and conjectures in this paper to intersection homology. This last goal is probably the only way to prove the conjectures, probably via the decomposition theorem [@dcml09:decomp].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analyse by detailed modelling the spectra observed from the sample galaxies at z$\sim$0.8 presented by Ly et al (2015), constraining the models by the \[OIII\]5007+4959/\[OIII\]4363 line ratios. Composite models (shock + photoionization) are adopted. Shock velocities $\geq$ 100 and preshock densities 0$\sim$ 200 3 characterize the gas surrounding the starburst (SB), while 0 are higher by a factor of 1.5-10 in the AGN emitting gas. SB effective temperatures are similar to those of quiescent galaxies ($\sim$4-7 10$^4$ K). Cloud geometrical thickness in the SB are $\leq$ 10$^{16}$ cm indicating major fragmentation, while in AGN they reach $>$ 10 pc. O/H are about solar for all the objects, except for a few AGN clouds with O/H= 0.3 -0.5 solar. Starburst models reproduce most of the data within the observational errors. About half of the object spectra are well fitted by an accreting AGN. Some galaxies show multiple radiation sources, such as SB+AGN, or a double AGN.'
author:
- |
M. Contini\
School of Physics and Astronomy, Tel Aviv University, Tel Aviv 69978, Israel\
title: |
Exploring AGN - starburst coexistence in galaxies at\
z$\sim$ 0.8 by the \[OIII\]4959+5007/\[OIII\]4363 line ratio
---
\[firstpage\]
[radiation mechanisms: general — shock waves — galaxies: AGN — galaxies: starburst — galaxies: high redshift]{}
Introduction
============
Starburst (SB) - active galactic nucleus (AGN) connection is generally explored in order to understand galaxy formation, structure and evolution. Processes which stimulate starbursts may lead also to AGN onset (Hopkins et al 2011). Dixon & Joseph (2011) using a variety of spectroscopic diagnostics found that $\sim$50% of the luminous infrared galaxies (LIRG) in their sample shows some evidence for an AGN. The luminosity of $\sim$ 17 % of the sample galaxies seems dominated by emission from AGN, and the remaining $\sim$ 80 % have luminosities dominated by SB. $\sim$50% of the sample galaxies indicates coupled AGN and starburst activities suggesting that AGNs and starbursts commonly coexist. Contini & Contini (2007) investigating a sample of LIRG ($>$10$^{11}$), ultra-LIRG (ULIRG, $>$10$^{12}$) and hyper-LIRG (HLIRG,$>$ 10$^{13}$) concluded that half of the LIRG contains an AGN, at least one AGN is found in all the ULIRG and none in the HLIRG. The connection between the AGN and the starburst is not direct and it is affected by a (viscous) time lag of gas flowing through the AGN accretion disc leading to AGN activity delay in star formation activity (Blank & Duschl 2013). AGN feedback may terminate star formation in the host galaxy poor gas phase and trigger it in the rich phase (Zubovas et al. 2013). SB-AGN coexistence in local galaxies has been investigated by the analysis of mid-infrared lines and of the continuum spectral energy distribution (SED) adopting diagnostic diagrams (e.g. Dixon & Joseph 2011) and by the optical- UV lines and continuum SED in relatively high z galaxies by detail modelling (Contini 2015 and references therein).
In this paper we revisit the spectra from galaxies at z$\sim$0.8, presented by the Ly et al (2015) DEEP2 survey. The observations were done by the DEIMOS multi-object spectrograph on the Keck II telescope. The observed spectra account for the \[OIII\]5007+ (the + indicates that the 5007 and 4959 lines are summed), \[OIII\]4363, \[NeIII\] 3869+, \[OII\]3727+, and lines which can be used to constrain the models. The observed spectra are generally reproduced by detailed modelling. We have noticed by modelling the spectra observed from many different galaxy types (Contini 2016 and references therein), that in some cases, even when the \[OIII\]5007+/and \[OII\]3727+/line ratios are well fitted by the models, the \[OIII\]4363/systematically disagree. For instance, \[OIII\] 4363/line ratios calculated by shock dominated models overpredict the data in super-luminous SN host galaxies (Contini 2016). Therefore, the \[OIII\]4363/line ratios, when available, are useful to constrain the models. Selecting the AGN, SB, and/or only shock dominated models best fitting the data, the galaxy types can be distinguished.
The 3 (\[OIII\]5007+4959/\[OIII\]4363) line ratios are favourite in the modelling process because they depend only indirectly on the O/H relative abundance. Actually, O/H variations affect the cooling process in the recombination zone of the emitting clouds leading to different line spectra. We use models which account for both the photoionizing flux from a primary radiation source and shocks. Shocks yield fragmentation of matter by turbulence created near the shock-front and compression of the gas downstream, leading to high densities which can trigger star formation. Composite models resolved the problem regarding 3 in AGN and LINER (low-ionization nuclear emission-line region) spectra, which indicated relatively high temperatures ($>$ 10$^5$K) and densities ($>$ 10$^5$ 3) in the emitting gas (Contini & Aldrovandi 1986). Moreover, comparing calculated with observed line ratios the gas motion direction can be determined, i.e. infalling towards the radiation source or ejected outwards (Fig. 1). This is an important issue in view of the SB - AGN feedback in galaxies. The calculation method is described in Sect. 2. Model results are compared with the data in Sect. 3. Discussion and concluding remarks appear in Sect. 4.
Calculation of the spectra
==========================
{width="8.0cm"}
![Top : log(\[OIII\]5007+/) (obs) vs log(\[OII\]3727+/) (obs); bottom : log(\[OIII\]4363)/(obs) vs log(\[OII\]3727+/) (obs) ](fg2a.eps "fig:"){width="8.0cm"} ![Top : log(\[OIII\]5007+/) (obs) vs log(\[OII\]3727+/) (obs); bottom : log(\[OIII\]4363)/(obs) vs log(\[OII\]3727+/) (obs) ](fg2b.eps "fig:"){width="8.0cm"}
Spectroscopic data provide a full physical and chemical picture for local galaxies. At high redshifts the data are reduced to a few significant lines, but the surveys contain hundreds of objects. Therefore, in order to obtain the O/H metallicity, the oxygen line ratios are generally calculated by the “direct methods” (see e.g. Ly et al 2015, Modjaz et al 2008). The “direct” or method (Seaton 1975, Pagel et al. 1992, etc) was used to obtain O/H from the observed oxygen to line ratios. By this method, the ranges of the gas physical conditions are chosen among those most suitable to the observed line ratios. The \[OIII\]4363 line derives from the upper $^1S$ level, while \[OIII\] 4959 and 5007 lines derive from the lower $^1D$ level. The relative rates of excitation of the $^1S$ and $^1D$ levels depend very strongly on the temperature T (Osterbrock 1974). Therefore, T is calculated from the \[OIII\]5007+/\[OIII\] 4363 ratio, which depends also on the gas density. This is generally obtained from the \[SII\] 6717/6730 line ratios. In the present spectra the \[SII\]6717,6731 lines were not observed. Moreover, by “direct methods” a unique temperature is adopted throughout the whole galaxy, neglecting the physical conditions less adapted to the strongest lines. This leads to discrepancies between metallicities (in terms of the O/H relative abundance) calculated by “direct methods” and detailed modelling (see, e.g. Contini 2014a), because the gas cools and recombines in the regions far from the photoionization source, as well as downstream of shock-fronts. By detailed modelling the gas at low (T$<$ 10$^4$K) temperatures contributes to the line emission. The line fluxes which result from integration throughout regions of gas at various temperatures are different from those calculated by an homogeneous temperature. If the fractional abundances of the corresponding ions are low, relatively high O/H are needed to reproduce strong oxygen observed lines. So, the metallicities calculated by detailed modelling are generally higher than those calculated by “direct methods”. A comparison between the results obtained by “direct methods” and detailed modelling (Contini 2014a) demonstrates that “direct methods” lead to O/H lower limits.
Detailed modelling of the spectra by pure photoionization models (e.g. by the CLOUDY code) gives satisfying results for intermediate ionization level lines. However, galaxies at high redshifts often originate from mergers and show a disturbed hydrodynamic structure. Collisional phenomena are critical in the calculation of the spectra. We suggest that models based on the coupled effect of photoionization and shocks are the closest approximation to the complex structure of the emitting gas. Therefore, the SUMA code (Contini 2015 and references therein) is used.
The code simulates the physical conditions in an emitting gaseous cloud under the coupled effect of photoionization from the primary radiation source ( SB or AGN) and shocks. The line and continuum emission from the gas are calculated consistently with dust-reprocessed radiation in a plane-parallel geometry. The calculations start at the shock front where the gas is compressed and thermalized adiabatically, reaching the maximum temperature in the immediate post-shock region ($T(K)\sim 1.5\times 10^5/(V_{\rm s}/100$ km s$^{-1}$$)^{2}$, where is the shock velocity). T decreases downstream following the cooling rate. The input parameters such as , the atomic preshock density 0 and the preshock magnetic field 0 (for all models 0=10$^{-4}$Gauss is adopted) define the hydrodynamical field.
The input parameters that represent the primary radiation for a SB are the effective temperature and the ionization parameter $U$. For an AGN, the primary radiation is the power-law radiation flux from the active center $F$ in number of photons cm$^{-2}$ s$^{-1}$ eV$^{-1}$ at the Lyman limit and spectral indices $\alpha_{UV}$=-1.5 and $\alpha_X$=-0.7. The primary radiation source is independent but it affects the surrounding gas. In contrast, the secondary diffuse radiation is emitted from the slabs of gas heated by the radiation flux reaching the gas and, in particular, by the shock. In our model the gas region surrounding the radiation source is not considered as a unique cloud, but as an ensemble of fragmented filaments. The geometrical thickness of these filaments is an input parameter of the code (D) which is calculated consistently with the physical conditions and element abundances of the emitting gas. Primary and secondary radiations are calculated by radiation transfer throughout the slabs downstream. The fractional abundances of the ions are calculated resolving the ionization equations. The dust-to-gas ratio ($d/g$) and the abundances of He, C, N, O, Ne, Mg, Si, S, A, Fe, relative to H, are also accounted for.
In the modelling process, we aim to reproduce the observed line ratios for each element. Each line has a different strength which translates into the different precision by the fitting process. A minimum number of significant lines (\[OIII\] 5007+, \[OII\]3727+, \[OIII\]4363, \[NII\], , ) is necessary to constrain the model but the number of the observed lines does not interfere with the modelling process. We deal with line ratios to avoid distance and morphological effects. We start adopting solar abundances by Allen (1976) in the first modelling trials because their values are between the two more recent ones by Anders & Grevesse (1989) and Asplund et al (2009). A perfect fit of the observed line ratios is not realistic because the observed data have errors, both random and systematic. The uncertainty in the calculation is due to the atomic parameters (within 10 %) which are often updated. The strongest lines are reproduced by $<$ 10%, the weakest by $\sim$ 50 %. The calculation code and our modelling method are described by Contini (2014a, 2016 and references therein).
Before starting the modelling process of the present survey galaxies, some characteristics can be guessed by comparing different line ratios in Fig. 2. The top diagram shows that excluding three objects (ID 4, 15 and 16) the slope shown by the line ratio band suggests that a radiation mechanism should be adopted. In the bottom diagram, excluding three objects (ID 4, 8 and 15), a well defined slope is not seen. This reveals large ranges of temperatures and densities in the emitting gas, indicating that other mechanisms such as e.g. shocks could be at work.
{width="8.0cm"} {width="8.0cm"} {width="8.0cm"}
Modelling results
=================
$^1$ in ; $^2$ in 10$^{-17}$ ;
Table 1-continued
\
1: in ; 2: in 3 ; 3: in 10$^4$K; 4: in 0.01 units ;5: in 10$^{16}$cm; 6: in 10$^{-4}$ units; 7: in ;
1: in ; 2: in 3; 3: in 10$^{10}$ ph cm$^{-2}$ s$^{-1}$ eV$^{-1}$ at the Lyman limit; 4: in10$^{16}$cm; 5: in 10$^{-4}$ units; 6: in;
1: in ; 2: in 3; 3: in 10$^{10}$ ph cm$^{-2}$ s$^{-1}$ eV$^{-1}$ at the Lyman limit; 4: in 10$^{16}$cm; 5: in 10$^{-4}$; 6: in
In Table 1 we compare the reddening corrected observed line ratios to with model results. Each observed spectrum is recognizable by its ID number (Ly et al 2015). We refer to the corrected observed \[OIII\]5007+/line ratios which appear in parenthesis (last column), calculated by (\[OIII\]5007+/\[OIII\]4363)$_{corr}$$\times$(\[OIII\]4363/)$_{corr}$, because the \[OIII\]5007+/line ratios reported in Table 1 from Ly et al. (2015, table 2) were not corrected. The observed corrected 3 (\[OIII\]5007+/\[OIII\]4363) line ratios (C.Ly 2015, private communication) appear in Table 1 last column.
The models are constrained by the FWHM of the line profiles (C.Ly, 2015, private communication) which roughly indicate the velocity field of the emitting gas and give an initial hint of the shock velocity.
Models MSB1-MSB28 refer to photoionization by the SB (+shocks) of clouds propagating outwards from the galaxy. Models Mpl1-Mpl28 account for photoionization by an AGN + shocks. The clouds propagate outwards from the AGN. Models Mpl01-Mpl028 refer to photoionization by an AGN + shocks, but the clouds propagate towards the active nucleus. Models MSB1-MSB28, Mpl1-Mpl28 and Mpl01-Mpl028 are described in Tables 2, 3 and 4, respectively. The physical conditions and the relative abundances presented in Tables 2-4 are regarded as the results of modelling for each galaxy. We have selected the models showing discrepancies for the strongest line ratios within 10% and for the weakest ones (e.g. \[OIII\]4363/) within 50%. However, in a few cases (e.g. for models MSB23 and MSB24) we could not find a good approximation of calculated to observed \[OIII\]4363/without destroying the good fit of \[OIII\]5007+/and \[OII\]3727+/.
The set of parameters adopted to model the SB galaxies (Table 2) shows pre-shock densities, SB effective temperatures and ionization parameters in agreement with those calculated in this redshift range for different types of galaxies (Contini 2015). The geometrical thickness of the emitting clouds are at the lower limit. The O/H and Ne/H relative abundances are close to solar. The parameters of AGN dominated models in the outflow case (Table 3) which yield the satisfactory fit of the observed line ratios are dominated by relatively high 0 and $D$ suitable to the narrow line emission region. Moreover, O/H are close to solar for most of the objects, but 0.3 solar for Mpl4, Mpl11, Mpl12 and Mpl27. Ne/H are lower than solar (10$^{-4}$) by a factor $>$ 2 for most galaxies. However, the models refer to the doublet and the data to the deblended line. Ne corresponds to the 1s2, 2s2, 2p6 closed atomic configuration, therefore Ne is less adapted to link with other species inside dust grains.
The parameters selected to reproduce the observed spectra for AGN models in the inflow case (Table 4), show relatively high densities, a few large clouds and O/H and Ne/H slightly lower than solar, except for Mpl04, Mpl07, Mpl09, Mpl022 and Mpl028 where O/H are $\sim$ half solar. High preshock densities in the inflow case are reasonable, considering that the IS clouds in the surrounding of the AGN are accreted toward the AGN.
In Fig. 3 we compare the modelling results with the Kewley et al (2001) diagnostics. \[OI\], \[NII\] and \[SII\] lines were not observed, so we use the results of models which better fit the observed line ratios. We adopt O/H solar, N/H =0.3 solar and S/H solar (Contini 2015, table 1). Fig. 3 top diagram shows that SB models are definitively in the SB zone, AGN models in the AGN zone, even if some models calculated for AGN galaxies in the outflow case, are on the border line. For the other two diagrams (\[OIII\]/vs \[NII\]/and \[OIII\]/vs \[SII\]/) most of the SB models are on the border line. This already suggests that some galaxies show a complex character of SB and AGN. The spread of the \[OI\]/line ratios and, in particular, the separation between the inflow from the outflow results, is due to the geometrical thickness large range of the emitting clouds in the AGN dominated models. In the outflow case (Fig. 1), the clouds may contain in the internal region, between the shock front edge and the edge illuminated by the AGN flux, a low temperature zone with large quantities of gas in the physical conditions suitable to the \[OI\] line. The same could explain the spread of the \[SII\] lines. However, when referring to the data instead of to the models, the S/H relative abundance also changes from object to object because S is easily trapped into dust grains. The \[NII\]/line ratios fill a more compact region throughout the diagram and the AGN inflow and outflow models overlap, because the region throughout the cloud where N$^+$ prevails is radiation dominated. We have found by detailed modelling of different types of spectra emitted from galaxies at 0.01$\leq$z$\leq$3 that N/H varies within a factor of 10 throughout all the z range (Contini 2016, fig. 5). So inserting eventual observation data instead of model results in Fig. 2 middle diagram, the picture would change.
In Fig. 4 we check the modelling result precision. The observational uncertainties are not shown for sake of clarity. The \[OIII\] 5007+/and \[OII\]3727+/calculated line ratios reproduce satisfactorily the data, indicating that the physical conditions and the relative abundances, adopted in the calculations of the spectra, are sound. The same models were used to calculate for each object the \[OIII\]4363/line ratios shown in Fig. 4, right diagram. The observed \[OIII\]4363 lines are rather weak. Even accepting discrepancies by a factor of 2, Fig. 4 (right diagram) indicates that several models referring to the AGN outflow case should be dropped.
Discussion and concluding remarks
=================================
Physical conditions, metallicities and SFR in the sample galaxies
-----------------------------------------------------------------
{width="5.8cm"} {width="5.8cm"} {width="5.8cm"}
{width="9.4cm"}
Tables 2-4 show that $\sim$ 100 and 0 $\sim$ 200 3 in most of the galaxies characterize the gaseous clouds in the SB environment, while preshock densities are higher by a factor of 1.5-10 in the AGN clouds. The SB clouds show major fragmentation ($D$ $\sim$ 10$^{16}$ cm). In the AGN, $D$ ranges from 10$^{16}$ cm to a maximum of $>$10 pc for the ejected clouds. Geometrical thick clouds up to D=3 pc were calculated for the merger galaxy NGC 6240 (Contini 2012b). Combining an average flux of 10$^{-16}$ observed at Earth (Table 1) with an average flux calculated at the nebulae (Table 4) for the SB clouds ($\sim$0.01 ) and for the clouds in the AGN environment ($\sim$0.1 ), we obtain the average distance of the emitting clouds from the SB R$_{SB}$$\sim$ 0.5 kpc and from the AGN R$_{AGN}$$\sim$ 0.16 kpc, respectively, adopting a filling factor $\sim$1. The starburst effective temperatures are similar to those found in quiescent galaxies (4-7 10$^4$ K, Contini 2014b).
Metallicities, in term of the O/H relative abundances, are a crucial issue for galaxies at high z. The O/H relative abundances calculated by detailed modelling are about solar (6.6 10$^{-4}$, Allen 1976) for all the objects, except for a few AGN with O/H= 0.3 -0.5 solar. Ly et al (2015) obtain 0.223-3.23 10$^{-4}$ by the T$_e$ based metallicity determination (cf. Contini 2014a). The discrepancies between metallicity results obtained by different modelling methods are explained in Sect. 2.
In Fig. 5 the absolute fluxes (in ) calculated at the emitting nebula and the SFR (in yr$^{-1}$) given by Ly et al. are shown as function of the redshift. Throughout a small z range, they do not show any trend, while (Contini 2016, fig. 6 and references therein) SFR in SN and GRB host galaxies were found to increase with z on a large scale for z$\geq$0.1. At lower z, SFR do not show any particular slope. On the other hand, SFR presented by Ramos Almeida et al (2013) for X-ray and mid-infrared selected galaxies at 0.4$\leq$z$\leq$1.15 exhibit an increasing trend. More objects should be considered. Moreover, SFR $\propto$ L()$_{Earth}$ (the luminosity observed at Earth) and L()$_{Earth}$ = L()$_{nebula}$ (the luminosity calculated at the nebula) which is $\propto$ R$^2$. We adopt /$\sim$3. R is the distance of the emitting cloud from the radiation source. So SFR $\propto$ , assuming an average R. Ly et al results show that SFR in AGN and SB galaxies are very similar, whereas from the flux calculated at the nebulae (Fig. 5, top diagram), higher SFR for AGN are expected. This discrepancy can be explained by the coexistence of AGN and SB in most of the sample objects.
{width="8.8cm"} {width="8.8cm"} {width="8.8cm"} {width="8.8cm"} {width="8.8cm"} {width="8.8cm"}
Multiple radiation sources in single galaxies
---------------------------------------------
By AGN and SB dominated models, we reproduced successfully the \[OIII\]5007+/and \[OII\]/observed line ratios, not always the \[OIII\]4363/from each galaxy. We investigate whether the photoionization source in each object is an AGN, a SB or both, comparing in Fig. 6 left diagrams and right diagrams, the calculated with observed corrected 3 and \[OII\]/\[OIII\]4363, respectively. The top diagrams refer to results obtained by SB dominated models, the middle ones by AGN dominated models with outflowing clouds and the bottom diagrams show the results of models dominated by AGNs which accrete the surrounding clouds. Fig. 6 shows that models MSB1-MSB28 (Table 2) reproduce the data within the observational errors, except for ID 6, 7, 23 and 24 (Fig. 6 top diagram). Most of the Mpl1-Mpl28 model results (Table 3) overpredict the data, except ID 19 (Fig. 6 , middle). ID 3, 6, 7, 15, 17, 19 and 23 3 are reproduced within $\sim$ 30%. Fig. 6 bottom diagram refers to models Mpl01-Mpl028 (Table 4). About half of the object spectra are well reproduced by the AGN accretion models. The results show that some line ratios can be explained with the same precision by both the SB and AGN dominated models. The galaxy types are schematically shown in Table 5 (S=SB, 1=AGN$_1$, 0=AGN$_0$), where AGN$_1$ refer to ejection of matter outwards and AGN$_0$ to accretion.
Summarizing, calculation results suggest that most of the objects contain a SB, about 5 objects are AGNs. Half of the galaxies show multiple radiation sources : a SB + an accreting AGN. ID 3, 15, 17 and 19 show an SB + an AGN with outflowing matter. ID 17 and ID 23 host a double AGN, but ID 17 also a SB. The AGNs show gas accretion rather than outflow, suggesting an AGN-SB correspondence. Multiple nuclei are found in local galaxies which derive from merging, e.g. in Arp 200 at z=0.018 (Graham et al 1990), where O/H relative abundances are about solar and N/H are higher than solar by a factor of $\sim$1.5 throughout the starburst region (Contini 2013). To complete our investigation we need measurements of the \[OI\], H$\alpha$, \[NII\] and \[SII\] emission lines, which are only available with near-infrared spectroscopy at z$\sim$0.8.
Acknowledgements {#acknowledgements .unnumbered}
================
I am grateful to the referee for valid comments which improved the presentation of the paper. I thank Dr. Chun Ly for preparing the spectra in the suitable format and for helpful discussions.
References {#references .unnumbered}
==========
\[A\]llen, C.W. 1976 Astrophysical Quantities, London: Athlone (3rd edition) \[A\]nders, E., Grevesse, N. 1989, Geochimica et Cosmochimica Acta, 53, 197 \[A\]splund, M., Grevesse, N., Sauval, A.J., Scott, P. 2009, ARAA, 47, 481 \[B\]lank, M. , Duschl, W.J. 2014, IAUS, 303, 379 \[C\]ontini, M. 2016, MNRAS, in press (ArXiv:1605,02438) \[C\]ontini, M. 2015, MNRAS, 452,3795 \[C\]ontini, M. 2014b, A&A, 572, 65 \[C\]ontini, M. 2014a, A&A, 564, 19 \[C\]ontini, M. 2013, MNRAS, 429, 242 \[C\]ontini, M., Contini, T. 2007, AN, 328, 953 \[C\]ontini, M., Aldrovandi, S.M.V. 1986, A&A, 168, 41 \[D\]ixon, T.G., Joseph, R.D. 2011, ApJ, 740, 99 \[E\]ngel, H. et al. 2010, A&A, 524 ,56 \[G\]raham G. R., Carico D. P., Matthews K., Neugebauer G., Soifer B. T., Wilson T. D. 1990, ApJ, 354, L5. \[H\]opkins, P.F. , Quataert, E., Murray, N. 2011, MNRAS, 417, 950 \[K\]ewley, L.J., Dopita, M.A., Sutherland, R.S., Heisler, C.A., Trevena, J. 2001, ApJ, 556, 121 \[L\]y, C., Rigby, J.R., Cooper, M., Yan, R. 2015, ApJ, 805, 45 \[O\]sterbrock, D.E. 1974 in ’Astrophysics of Gaseous Nebulae’ W.H. Freeman and Company, San Francisco \[R\]amos Almeida, C., Rodríguez Espinosa, J. M., Acosta-Pulido, J. A., Alonso-Herrero, A., Pérez García, A. M., Rodríguez-Eugenio, N. 2013, MNRAS, 429, 3449 \[Z\]ubovas, K., Nayakshin, S., King, A., Wilkinson, M. et al 2013, MNRAS, 433, 3079
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is discovered that the tokamak field geometry generates a pitch-angle scattering effect for runaway electrons. This neoclassical pitch-angle scattering is much stronger than the collisional scattering and invalidates the gyro-center model for runaway electrons. As a result, the energy limit of runaway electrons is found to be larger than the prediction of the gyro-center model and to depend heavily on the background magnetic field.'
author:
- Jian Liu
- Yulei Wang
- Hong Qin
title: 'Neoclassical Pitch-Angle Scattering of Runaway Electrons'
---
[^1]
Runaway electrons in a tokamak are energetic particles accelerated by the electric field. They cannot be braked by the collisional drag [@Drercer_REorigins_1959]. A large amount of runaway electrons are produced in tokamaks during fast shutdowns, disruptions [@Yoshino_Shutdown_RE; @Jaspers_Disruption_RE; @Helander_avalanch_2000; @Helander_2002; @Fulop_magneticThreshold4RE2009; @Gill_REref1_2000; @Jaspers_RErefs4_1993; @Nygren_RErefs5_1997; @Parks_RErefs6_1999; @Rosenbluth_RErefs7_1997; @Yoshino_REandTurbulenceDischarge2000; @Tamai_Yoshino_REtermination_JT60U_2002; @Lehnen_RMP_PRL_TEXTOR; @Finken_RElosses_2007], or aggressive current drive [@Net_Fisch_RevModP1987]. The massive energy carried by runaway beams poses great danger to plasma-facing components [@Bartels_RE_PFCs1994; @Kawamura_PFS_RE1989; @Bolt_REref0_1987; @Jaspers_REref3_2001]. Understanding the physics of runaway electrons in toroidal field configurations is thus critical. The dynamics of runaway electrons involves different timescales spanning 11 orders of magnitude, which brings difficulties to both analytical and numerical studies. Gyro-center model is often applied to tackle the multi-scale problem by averaging out the fast gyro-motion and has produced fruitful results [@Martin_Momentum_RE_1998; @Martin_Energylimit_RE_1999; @LiuJian_RE_Positron_2014; @Bakhtiari_Momentum_RE_Fb_2005; @Guan_Qin_Sympletic_RE; @Neoclassical_Drift_report; @Papp_Fulop_Helander_MPF2011; @Rax_RE_resonance1993; @Andersson_Helander2001]. According to the gyro-center model, the magnetic moment of runaway electron is an adiabatic invariant, and the parallel momentum increases due to the work by the loop electric field. There is no channel of momentum transfer from the parallel to the perpendicular direction, except for the collision with background plasmas. In general, the collisional effect is rather weak for charged particles with high velocities [@Guan_Qin_Sympletic_RE]. For a typical runaway electron, the collision time is $\tau_{col}\sim0.5\,\mathrm{s}$, which is much longer than the gyro-period $T_{ce}$ ($\sim10^{-10}\,\mathrm{s}$) and the transit period $T_{tr}$ ($\sim10^{-8}\,\mathrm{s}$). When the collisional effect can be neglected, the perpendicular momentum will monotonically decrease due to the synchrotron radiation of the gyro-motion [@Guan_Qin_Sympletic_RE], and parallel momentum will monotonically increase to its maximum limit until the electric field acceleration is finally balanced by the radiation dissipation. The pitch-angle scattering due to collisions will transfer a small amount of energy from the parallel direction to the perpendicular direction. This small collisional effect keeps the runaway electrons energetic in the perpendicular direction, but does not modify the energy limit in the parallel direction by much. The gyro-center model predicts that the energy limit of runaway electrons does not depend on collisions and the magnitude of background magnetic field [@Martin_Momentum_RE_1998].
Contrary to the common wisdom, our analysis shows that the gyro-center model is not valid for runaway electrons. This is because they move with the speed of light in the parallel direction, the local magnetic field they see changes by a large amount during one gyro-period. Therefore, the basic assumption for the gyro-center model, i.e., the magnetic field is approximately constant in one gyro-period, breaks down. In fact, the magnetic moment $\mu$ is no longer an adiabatic invariant. Similar non-conservation of $\mu$ has been observed in the presence of magnetic turbulence with wavelength comparable to the gyro-radius [@Dalena_mu_Nonconserve_2012; @Dalena_mu_Nonconserve_3Dturbulence].
In this letter, we abandon the gyro-center model and study the multi-timescale runaway dynamics by numerically solving the dynamical equations of runaway electrons directly in the six-dimensional phase space. Long-term simulation results confirm that the gyro-center model is indeed invalid, see Fig.\[fig:1-B\_delta\]. More than one hundred billion time steps are required in the simulation. To eliminate the coherent accumulation of numerical errors from each time step, which is usually a show-stopper for long-term simulations, we utilize the newly developed volume-preserving algorithm (VPA) for relativistic particles [@Ruili_VPA_2015]. The VPA can guarantee the long-term accuracy by preserving the phase-space volume of the original physical system. Its long-term conservativeness and stability have been verified.
Taking the advantage of the VPA method, we discovered that there exists a new pitch-angle scattering mechanism, which transfers momentum of runaway electrons between the parallel and perpendicular directions. It arises from the full orbit dynamics in the toroidal geometry of a tokamak, hence the name of neoclassical scattering. The neoclassical pitch-angle scattering process is about a million times faster than the collisional pitch-angle scattering, resulting in a rapid transfer of the parallel momentum, gained from the loop electric field, to the perpendicular direction. As an important result, the simulation study indicates a new energy limit for runaway electrons, which is higher than the result from the gyro-center model and varies with the magnitude of the background magnetic field. This unexpected neoclassical pitch-angle scattering effect for runaway electrons and its important consequence are the subjects of this letter.
First, we introduce the physical model. When focusing on the long-term dynamics of runaway electrons in a tokamak, we take into account the background magnetic field, the loop electric field, and the electromagnetic radiation. The dynamical equations of runaway electrons are $$\begin{aligned}
\frac{\mathrm{\mathrm{d\mathbf{x}}}}{\mathrm{d}t} & = & \mathbf{v},\label{eq:1}\\
\frac{\mathrm{\mathrm{d\mathbf{p}}}}{\mathrm{d}t} & = & -\mathrm{e}\left(\mathbf{E}+\mathbf{v}\times\mathbf{B}\right)+\mathbf{F}_{R},\label{eq:2-1}\\
\mathbf{p} & = & \gamma\mathrm{m}_{0}\mathbf{v},\end{aligned}$$ where $\mathbf{x}$, $\mathbf{v}$, $\mathbf{p}$ denote the position, velocity and mechanical momentum of a runaway electron, $\mathrm{e}$ denotes the unit charge, $\mathrm{m}_{0}$ is the rest mass of electron, $\mathbf{E}$ and $\mathbf{B}$ are the electric and magnetic field, and the Lorentz factor $\gamma$ is defined as $$\gamma=\sqrt{1+\frac{p^{2}}{\mathrm{m}_{0}^{2}\mathrm{c}^{2}}}=\frac{1}{\sqrt{1-\left(v/\mathrm{c}\right)^{2}}}\,.\label{eq:2}$$ The effective electromagnetic radiation drag force $\mathbf{F}_{R}$ is $$\mathbf{F}_{R}=-P_{R}\frac{\mathbf{v}}{v^{2}}\,,\label{eq:3}$$ where $P_{R}$ is the radiation power determined by [@Jackson_electrodynamics] $$P_{R}=\frac{q_{e}^{2}}{6\pi\mathrm{\mathrm{\epsilon}_{0}c}}\gamma^{2}\left[\left(\frac{\mathbf{a}}{\mathrm{c}}\right)^{2}-\left(\frac{\mathbf{v}}{\mathrm{c}}\times\frac{\mathbf{a}}{\mathrm{c}}\right)^{2}\right]\,.\label{eq:4}$$ Here, $\epsilon_{0}$ is the permittivity in vacuum, $\mathrm{c}$ is the speed of light in vacuum, and $\mathbf{a}=\mathrm{d}\mathbf{v}/\mathrm{d}t$ denotes the acceleration.
In order to solve Eqs.(\[eq:1\])-(\[eq:2-1\]) numerically, we have to meet the challenge brought by the multi-scale nature of the problem. Restricted by the minimal timescale, more than $10^{11}$ time-steps are required to simulate a complete runaway dynamics. Traditional algorithms, such as the 4th order Runge-Kutta method, are not qualified for this long-term simulation, because the coherent accumulation of numerical error over many time-steps leads to incorrect long-term simulation results. To overcome this difficulty, geometric algorithms which can bound the global numerical error for all time-steps should be adopted [@Qin_VariatianalSymlectic_2008; @HeYang_Spliting_2015; @Qin_Boris_2013; @Jianyuan_Multi_sympectic_2013; @Ruili_GC_canonical_2014; @Ruili_VPA_2015]. The newly developed relativistic VPA [@Ruili_VPA_2015] with radiation drag is utilized in the present study to guarantee the long-term numerical accuracy.
As a model, the background magnetic field and inductive electric field are set to be $$\begin{aligned}
\mathbf{B} & = & \frac{B_{0}R_{0}}{R}\mathbf{e}_{\xi}-\frac{B_{0}\sqrt{\left(R-R_{0}\right)^{2}+z^{2}}}{qR}\mathbf{e}_{\theta}\,,\label{eq:5}\\
\mathbf{E} & = & E_{l}\frac{R_{0}}{R}\mathbf{e}_{\xi}\,,\label{eq:6}\end{aligned}$$ where $R=\sqrt{x^{2}+y^{2}}$, $\xi$, and $z$ are radial distance, azimuth, and height of the cylindrical coordinate system respectively, $\mathbf{e}_{\xi}$ and $\mathbf{e}_{\theta}$ are the unit vectors along toroidal and poloidal directions, and $q$ denotes safety factor. Without loss of generality, we use the parameters of EAST [@Wu_East_2007]. The major radius is $R_{0}=1.7\,\mathrm{m}$, the safety factor is $q=2$, the central magnetic field is $B_{0}=3\mathrm{\, T}$, and the loop electric field is $E_{l}=0.2\,\mathrm{V}/\mathrm{m}$. The initial parallel and perpendicular momentum of a typical runaway electron are set to be $p_{\parallel0}=5\,\mathrm{m_{0}c}$ and $p_{\perp0}=1\,\mathrm{m_{0}c}$, and the initial position is $R=1.8\,\mathrm{m}$ and $\xi=z=0$. The time-step of simulation is set to $\Delta t=1.9\times10^{-12}s$, which is about 1% of the gyro-period.
![The change ratio of background magnetic field $\Delta{}_{B}$ during one gyro-period at different time of a runaway dynamics.\[fig:1-B\_delta\]](Rel_Delta_B)
To verify that the condition for gyro-center approximation is not satisfied for runaway electrons in tokamaks, we record the change ratio of the magnetic field $\Delta{}_{B}\left(t\right)$ that one runaway electron samples during one gyro-period at different time $t$, see Fig.\[fig:1-B\_delta\]. The change ratio is defined as
$$\Delta{}_{B}\left(t\right)=\frac{\left|\mathbf{B}\left(t+T_{ce}\right)-\mathbf{B}\left(t\right)\right|}{\left|\mathbf{B}\left(t\right)\right|}\,,\label{eq:7}$$
where $T_{ce}=2\pi\gamma\mathrm{m}_{0}/eB$ is the gyro-period. For the gyro-center approximation to be valid, the variation of the magnetic field a particle samples during one gyro-period should be small. However, simulation results show that the change ratio $\Delta{}_{B}$ increases monotonously with the runaway energy. Its value increases to 10% after $0.4\,\mathrm{s}$ and exceeds 30% after $1.7\,\mathrm{s}$, which can no longer be taken as a small value. This is mainly because the velocity of runaway electron approaches the speed of light, and the gyro-period $T_{ce}$ is proportional to the Lorentz factor $\gamma$. The runaway electron travels a long distance along the toroidal direction during each gyro-period, which leads to a large value of $\Delta{}_{B}$.
![The evolution of (a) parallel momentum and (b) perpendicular momentum of a runaway electron. The initial position of the electron is at $R=1.8\,\mathrm{m}$, $\xi=z=0$, and the initial momentum is $p_{\parallel}=5\,\mathrm{m_{0}c}$ and $p_{\perp}=\mathrm{m_{0}c}$. The center magnetic field is $B_{0}=3\mathrm{\, T}$ and the loop electric field is $E_{l}=0.2\,\mathrm{V}/\mathrm{m}$.\[fig:2-Moments\]](Ppara_Pperp)
The momentum evolution of runaway electrons is plotted in Fig.\[fig:2-Moments\]. The increase of the parallel momentum is similar to the results from the gyro-center model [@Guan_Qin_Sympletic_RE; @LiuJian_RE_Positron_2014], but the evolution of the perpendicular momentum is very different. The perpendicular momentum grows with rapid oscillations, even in the absence of collisions, until approaching a maximum after about $2.5\,\mathrm{s}$. Because the parallel momentum is relatively large, its oscillation is less prominent. The oscillation is an effective scattering process transferring the parallel momentum to the perpendicular direction and altering the pitch-angle. As explained later, this effective pitch-angle scattering roots in the geometric configuration of the field. It is thus a neoclassical pitch-angle scattering effect. The evolution of the perpendicular momentum exhibits four stages: (a) a rapid oscillation is developed initially; (b) then its absolute value reaches the zero point; (c) it grows quickly after passing the zero point, and (d) it saturates. At the saturation, though the pitch-angle still varies quickly due to the neoclassical scattering, the average of the perpendicular momentum does not change. The timescale of neoclassical pitch-angle scattering is $5\times10^{-7}\,\mathrm{s}$, which is about $10^{6}$ times faster than the collisional scattering, and much more momentum can be transferred to the perpendicular direction through the neoclassical scattering than the collisional effect as calculated in Ref [@Guan_Qin_Sympletic_RE].
![Motion of the momentum vector during one gyro-period at different time. The maximal and minimal perpendicular momenta, $p_{\perp max}$ and $p_{\perp min}$, during each gyro-period are marked by dashed line. \[fig:4-Snapshots-of-momentum\]](B_P_3D_split)
The seemingly complex evolution curves of the perpendicular momentum are dominated by the neoclassical scattering, which can be analyzed by looking at the variation of the momentum vector. We choose the parallel moment $p_{\parallel}$, the z-component of the perpendicular momentum $p_{\perp z}$, and the projection of perpendicular moment in the $x-y$ plane $p_{\perp xy}$ as the three coordinates for the moment. They satisfy $p_{\perp xy}^{2}+p_{\perp z}^{2}+p_{\parallel}^{2}=p^{2}$. In Fig.\[fig:4-Snapshots-of-momentum\], snapshots of the momentum vector within one gyro-period at different moments are plotted in the momentum space. During each gyro-period, the tip of the momentum vector moves approximately along a circular orbit. The minimal and maximal value of $p_{\perp z}$ are marked within each gyro-period. The circular orbit is first elongated while the center of circle shifts to $p_{\perp z}$ direction in the perpendicular plane. The elongation of the orbit corresponds to the growth of the oscillation amplitude. At $t=0.432\,\mathrm{s}$, the orbit touches the $p_{\perp z}=0$ plane, and the zero point of perpendicular momentum appears. Afterwards, the elongated orbit keeps shifting towards the larger $p_{\perp z}$ direction until approaching a steady state. It is evident that the variation of $p_{\perp}$ is mainly due to the $z$-component $p_{\perp z}$.
![Evolution of (a) $z$-component of parallel momentum $p_{\parallel z}$, (b) $z$-component of perpendicular momentum $p_{\perp z}$, and (c) $z$-component of unit vector along the magnetic field $b_{z}$.\[fig:Perp\_z\]](Ppara_z_Pperp_z_bz)
Figure \[fig:Perp\_z\] shows the evolution of the $z$-component of the parallel momentum $p_{\parallel z}$, the $z$-component of the perpendicular momentum $p_{\perp z}$, and the $z$-component of the unit vector along the magnetic field $b_{z}=\mathbf{b\cdot z}=\mathbf{B}\cdot\mathbf{z}/B$. We find that $p_{\perp z}$ increases, oscillates, and saturates around $9\, m_{0}c$. The oscillation amplitude of $p_{\parallel z}$ increases with the absolute value of $p_{\parallel}$. Meanwhile, the evolution of $p_{\parallel z}$ has the same trend as $b_{z}$, which indicates that the neoclassical scattering is closely related to the direction of the local magnetic field. Due to the neoclassical drift [@Neoclassical_Drift_report; @Guan_Qin_Sympletic_RE], the transit orbits of runaway electrons drift outwards from the magnetic axis. As a result, runaway electrons spend more and more time in the $R>R_{0}$ region. The magnetic field runaway electrons sample then tilts more towards the negative z direction on average, because of the helical configuration of the magnetic field lines. Since the parallel momentum is defined to be $p_{\parallel}=\mathbf{p\cdot b}$, the change of $\mathbf{b}$ results in a change of $p_{\parallel}$. In the $z$-direction, the time average of momentum vanishes approximately, i.e., $\left\langle p_{z}\right\rangle =\left\langle p_{\parallel z}\right\rangle +\left\langle p_{\perp z}\right\rangle =0$. Therefore, decrease of $\left\langle p_{\parallel z}\right\rangle $ corresponds to an increase of $\left\langle p_{\perp z}\right\rangle $.
![Plots of energy limit versus loop electric field $E_{l}$. The black dashed curve corresponds to the energy limit curve predicted by gyro-center model [@Martin_Momentum_RE_1998], where collisional effect is ignorable. The solid curves correspond to energy limits with different magnetic field intensities. The loop electric field is set to be radially uniform in order to compare with the gyro-center model.\[fig:5-Energy-limit\]](Comp_Martin)
The energy limit is reached at the saturation. The variation of the energy limit against the loop electric field with different magnetic field intensities are plotted in Fig.\[fig:5-Energy-limit\], which clearly shows that the energy limit depends on the intensity of the background magnetic field. This is different from the conclusion by Martín-Solís et al. using the gyro-center model [@Martin_Momentum_RE_1998]. For the case of $B_{0}=1\,\mathrm{T}$, our energy limit is about $20\%$ higher, because more energy is transferred to the perpendicular direction through the neoclassical pitch-angle scattering. If the magnetic field is extremely strong, the gyro-center approximation model will be valid, and our energy limit curve will recover the gyro-center result.
In summary, long-term simulations using the newly developed volume preserving algorithm revealed that the full orbit dynamics of a runaway electron in the tokamak field geometry generates a pitch-angle scattering effect for runaway electrons. In a typical tokamak, this neoclassical pitch-angle scattering is about 1 million times stronger than the collisional scattering and invalidates the gyro-center model for runaway electrons. As a consequence, the energy limit of runaway electrons is found to be larger than the prediction by a gyro-center model. In addition, the theoretical model developed in the present study shows that the runaway energy limit also depend heavily on the background magnetic field.
This research is supported by National Magnetic Confinement Fusion Energy Research Project (2015GB111003, 2014GB124005), National Natural Science Foundation of China (NSFC-11575185, 11575186, 11305171), JSPS-NRF-NSFC A3 Foresight Program (NSFC-11261140328), the CAS Program for Interdisciplinary Collaboration Team, and the GeoAlgorithmic Plasma Simulator (GAPS) Project.
[39]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , ().
, , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, **, vol. (, ).
, ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
[^1]: Corresponding author. [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Dimits shift is the shift between the threshold of the drift-wave primary instability and the actual onset of turbulent transport in magnetized plasma. It is generally attributed to the suppression of turbulence by zonal flows, but developing a more detailed understanding calls for consideration of specific reduced models. The modified Terry–Horton system has been proposed by St-Onge \[J. Plasma Phys. [**83**]{}, 905830504 (2017)\] as a minimal model capturing the Dimits shift. Here, we use this model to develop an analytic theory of the Dimits shift and a related theory of the tertiary instability of zonal flows. We show that tertiary modes are localized near extrema of the zonal velocity $U(x)$, where $x$ is the radial coordinate. By approximating $U(x)$ with a parabola, we derive the tertiary-instability growth rate using two different methods and show that the tertiary instability is essentially the primary drift-wave instability modified by the local $U''''$. Then, depending on $U''''$, the tertiary instability can be suppressed or unleashed. The former corresponds to the case when zonal flows are strong enough to suppress turbulence (Dimits regime), while the latter corresponds to the case when zonal flows are unstable and turbulence develops. This understanding is different from the traditional paradigm that turbulence is controlled by the flow shear $U''$. Our analytic predictions are in agreement with direct numerical simulations of the modified Terry–Horton system.'
author:
- 'Hongxuan Zhu , Yao Zhou,'
- 'I. Y. Dodin'
bibliography:
- 'Ref.bib'
title: Analytic theory of the tertiary instability and the Dimits shift within a scalar model
---
Introduction {#sec:Introduction}
============
The Dimits shift in magnetized plasmas is the shift between the threshold of drift-wave (DW) “primary” instability and the actual onset of transport that follows the scaling laws of developed turbulence [@Dimits00]. The Dimits shift is observed in both fluid and gyrokinetic simulations [@Lin98; @Rogers00; @Ricci06; @Numata07; @Mikkelsen08; @Kobayashi12; @St-Onge17] and is generally attributed to turbulence suppression by zonal flows (ZFs), which are generated by the “secondary” instability [@Rogers00; @Diamond01]. However, the Dimits shift is finite, meaning that ZFs cannot completely suppress DW turbulent transport if the primary-instability threshold is exceeded by far. Because of the detrimental effect that turbulent transport has on plasma confinement, it is important to understand this effect in detail.
After the seminal work [@Biglari90], it is widely accepted that ZFs can significantly suppress turbulence by shearing turbulent eddies. Based on this paradigm, the predator–prey model is perhaps the simplest phenomenological model that can describe how sheared flows help achieve a high-confinement regime [@Diamond94; @Malkov01; @Kim03; @Kobayashi15]. However, this paradigm may be oversimplified. For example, while direct simulations show that ZFs saturate at finite amplitude even in collisionless plasma [@Rogers00; @St-Onge17], the predator–prey model predicts otherwise. This is because the predator–prey model assumes statistically homogeneous turbulence, and this assumption is inapplicable in the Dimits regime, where strong ZFs are present and turbulence is inhomogeneous.
A more elaborate approach to understanding the Dimits shift was based on the concept of the “tertiary” instability (TI) [@Rogers00; @Rogers05]. The idea is that if ZFs are subject to the TI, then turbulence cannot be completely suppressed by ZFs and the Dimits regime ends. Despite some criticism [@Kolesnikov05], this explanation is widely accepted. However, the understanding of the TI and the Dimits shift has been largely qualitative, arguably because these effects have not been widely studied within simple enough models.
Recently, [@St-Onge17] proposed the modified Terry–Horton equation (mTHE) as a minimal model that captures the Dimits shift. @St-Onge17 calculated the TI growth rate using four-mode truncation (4MT) and derived a sufficient condition for ZFs to be stable within the mTHE. Then, this criterion was used for a “heuristic calculation” of the Dimits shift. However, that calculation is not entirely satisfactory, because deriving the actual Dimits shift takes more than a sufficient condition of ZF stability. The direct relation between @St-Onge17’s criterion and the Dimits shift is only an assumption. As a result, the agreement of @St-Onge17’s theory with numerical simulations is limited (section \[sec:DSPrediction\]). Besides, the 4MT model is only a rough approximation and cannot capture essential features of the TI in principle, as we shall discuss below. Therefore, a transparent theory of the TI and the Dimits shift within the mTHE model is yet to be developed.
In our recent letter [@Zhu20], we sketched a theory of the TI and the Dimits shift within the modified Hasegawa–Wakatani model, where the mTHE was briefly mentioned as the “adiabatic limit”. This limit is important in that the mTHE permits a detailed analytic study of the TI and an explicit quantitative prediction of the Dimits shift; thus, it deserves further investigation. Here, we present an in-depth study of the mTHE by expanding on the results presented in [@Zhu20]. We show that assuming a sufficient scale separation between ZFs and DWs, TI modes are localized at extrema of the ZF velocity $U(x)$, where $x$ is the radial coordinate. By approximating $U(x)$ with a parabola, we analytically derive the TI growth rate, $\gamma_{\rm TI}$, using two different approaches: (i) by drawing an analogy between TI modes and quantum harmonic oscillators and (ii) by using the Wigner–Moyal equation (WME). Our theory shows that the TI is essentially a primary DW instability modified by the ZF “curvature” $U''$ near extrema of $U$. (The prime denotes ${\rm d}/{\rm d}x$.) In particular, the WME helps understand how the local $U''$ modifies the mode structure and reduces the TI growth rate; it also shows that the TI is *not* the Kelvin–Helmholtz (KH) instability, or KHI. Then, depending on $U''$, the TI can be suppressed, in which case ZFs are strong enough to suppress turbulence (Dimits regime), or unleashed, so ZFs are unstable and turbulence develops. This understanding is different from the traditional paradigm [@Biglari90], where turbulence is controlled by the flow shear $U'$. Finally, by letting $\gamma_{\rm TI}=0$, we obtain an analytic prediction of the Dimits shift, which agrees with our numerical simulations of the mTHE.
Admittedly, our explicit prediction of the Dimits shift is facilitated by the fact that we use a simple enough model. Understanding of the Dimits shift is already complicated when we study the modified Hasegawa–Wakatani model in [@Zhu20], when we observed the presence of avalanche-like structures, which are not supported by the mTHE. Furthermore, the recent paper by [@Ivanov20] shows that avalanches themselves can become intricate when additional physics from finite ion temperature is taken into account. This complicates the problem even further, and more work remains to be done to understand the Dimits shift in the general case. Our paper is intended as one of the first steps in that direction.
This paper is organized as follows. In section \[sec:mTHE\] we introduce the mTHE. In section \[sec:Linear\] we describe the primary, the secondary, and the tertiary instability within the mTHE. In section \[sec:TIcalculation\] we analytically derive the TI growth rate using two different approaches mentioned above. In section \[sec:DSPrediction\] we derive an analytic prediction of the Dimits shift. Finally, a brief introduction of the WME and phase-space trajectories are presented in Appendices \[appA\] and \[appB\].
Modified Terry–Horton equation {#sec:mTHE}
==============================
The mTHE can be considered as a minimal model that simultaneously captures the primary, secondary, and tertiary instabilities. It is a two-dimensional scalar equation that describes DW turbulence in slab geometry with coordinates $\boldsymbol{x}=(x,y)$, where $x$ is the radial coordinate and $y$ is the poloidal coordinate: $$\label{eq:Equations_mTHE}
\partial_{t}w+\{\varphi,w\}-\beta\partial_{y}\varphi+\hat{\alpha}\hat{D}w=0,$$ where $$w=\nabla^{2}\varphi-n,\quad n=(\hat{\alpha}-{\rm i}\hat{\delta})\varphi.$$ Here, the system is assumed to be immersed in a uniform magnetic field perpendicular to the $(x,y)$ plane. The ions are assumed cold while the electrons are assumed to have a finite temperature $T_{{\rm e}}$. The plasma has an equilibrium density profile $n_{0}(x)$, which is parameterized by the positive constant $\beta\doteq a/L_{n}$, where $a$ is a reference length and $L_{n}\doteq(-{\rm d}\ln n_{0}/{\rm d}x)^{-1}$ is the scale length of the density gradient. (We use $\doteq$ to denote definitions.) Time is normalized by $a/c_{{\rm s}}$, where $c_{{\rm s}}\doteq\sqrt{T_{{\rm e}}/m_{{\rm i}}}$ is the ion sound speed. Length is normalized by the ion sound radius $\rho_{{\rm s}}=c_{{\rm s}}/\Omega_{{\rm i}}$, where $\Omega_{{\rm i}}$ is the ion gyro-frequency. The electrostatic potential fluctuation $\varphi$ is normalized by $T_{{\rm e}}\rho_{{\rm s}}/ea$ where $e$ is the unit charge, the electron density fluctuation $n$ is normalized by $n_{0}\rho_{\rm s}/a$, and $w$ can be considered as minus the ion guiding-center density [@Krommes00]. The Poisson bracket is defined as $$\{\varphi,w\}\doteq\boldsymbol{v}\cdot\nabla w,\quad\boldsymbol{v}\doteq\hat{\boldsymbol{z}}\times\nabla\varphi,\label{eq:Poisson}$$ which describes nonlinear advection of $w$ by the $\boldsymbol{E}\times\boldsymbol{B}$ flow with velocity $\boldsymbol{v}$. Also, $\nabla^{2}\doteq\partial_{x}^ {2}+\partial_{y}^{2}$ is the Laplacian. Finally, we note that the parameter $\beta$ can be scaled out of equation (\[eq:Equations\_mTHE\]) by replacing $(\phi,t,\hat{D})$ with $(\phi/\beta,\beta t,\hat{D}/\beta)$. Therefore, varying $\beta$ is effectively similar to varying the strength of $\hat{D}$.
The mTHE is “modified” compared to the original Terry–Horton model [@Terry82; @Terry83] in that the following operator $\hat{\alpha}$ is used: $$\hat{\alpha}\varphi=\tilde{\varphi}\doteq\varphi-\langle\varphi\rangle,\label{eq:Equations_alpha}$$ where $\langle\dots\rangle$ is the zonal average given by $$\langle\varphi\rangle\doteq\frac{1}{L_{y}}\int_{0}^{L_{y}}\varphi\,{\rm d}y$$ and $L_{y}$ is the system length along $y$. Equation (\[eq:Equations\_alpha\]) states that electrons respond only to the fluctuation (or DW) part of the potential, $\tilde{\varphi}$, but do not respond to the zonal-averaged (or ZF) part, $\langle\varphi\rangle$ [@St-Onge17; @Hammett93]. The operator $\hat{\delta}$ describes the phase difference between $n$ and $\varphi$ and determines the primary DW instability [@Terry82; @Terry83]. Note that (\[eq:Equations\_mTHE\]) reduces to the modified Hasegawa–Mima equation at $\hat{\delta}=0$ [@Hasegawa77; @Dewar07frontiers], where the total energy is conserved. The DW and the ZF part of the energy (per unit area) are given by $$E_{\rm DW}\doteq\frac{1}{2L_xL_y}\int{\rm d}x\,{\rm d}y\left[(\nabla_{\perp}\tilde{\varphi})^2+\tilde{
\varphi}^2\right],\quad E_{\rm ZF}\doteq \frac{1}{2L_x}\int{\rm d}x\,\left(\partial_x\langle\varphi\rangle\right)^2,\label{eq:Energies}$$ where $L_x$ is the system length along $x$. Various forms of $\hat{\delta}$ can be used to model different primary instabilities [@Terry82; @Tang78]. Here, we follow [@St-Onge17] and use the following simple form: $${\rm i}\hat{\delta}\doteq{\rm i}\delta_{0}\hat{k}_{y}\equiv\delta_{0}\partial_{y},\label{eq:Equations_delta}$$ with $\delta_{0}$ being a positive constant. (This can be used to model trapped-electron dynamics [@Tang78].) Finally, the operator $\hat{D}$ models damping effects such as viscosity. Following [@St-Onge17], we use $$\hat{D}=1-0.01\nabla^{2}\label{eq:Equations_D}$$ throughout this paper. (An exception is made in section \[sec:DSPrediction\], where another form of $\hat{D}$ is introduced for comparison.) Here, the first (friction) term is added in order to prevent possible energy build up at large scale, as is also done by @St-Onge17. (As will be seen from our results below, this term also increases the Dimits shift and thus facilitates its numerical observation.) Note that due to $\hat{\alpha}$ in front of $\hat{D}$ in (\[eq:Equations\_mTHE\]), the damping applies only to DWs, while ZFs are left collisionless. Then, the Dimits regime can be defined unambiguously as the regime where ZFs persist forever and the DW amplitude decreases to zero at $t \to \infty$.
Beyond the Dimits regime, DWs are not suppressed and ZFs always keep evolving in the mTHE model, as demonstrated by [@St-Onge17]. To understand the ZF dynamics, we take the zonal average of (\[eq:Equations\_mTHE\]) and obtain $$\partial_{t}U=-\partial_{x}\langle\tilde{v}_{x}\tilde{v}_{y}\rangle-\langle\tilde{v}_{x}{\rm i}\hat{\delta}\tilde{\varphi}\rangle+\mathcal{T}(t),\quad U(x,t)\doteq\partial_{x}\langle\varphi\rangle.\label{eq:Equations_Ut}$$ Here, $U$ is the ZF velocity along $y$, $(\tilde{v}_{x},\tilde{v}_{y})\doteq(-\partial_{y}\tilde{\varphi},\partial_{x}\tilde{\varphi})$ is the $\boldsymbol{E}\times\boldsymbol{B}$ velocity of DW fluctuations. The first term on the right-hand side of (\[eq:Equations\_Ut\]) is the Reynolds stress, while the second term is specific to the mTHE system. For the form of $\hat{\delta}$ given by (\[eq:Equations\_delta\]), the second term becomes $$-\langle\tilde{v}_{x}{\rm i}\hat{\delta}\tilde{\varphi}\rangle=\delta_{0}^{2}\langle\tilde{v}_{x}^{2}\rangle>0.\label{eq:Equation_T(t)}$$ Therefore, the second term will always increase the local ZF velocity $U$, and meanwhile, the value of $U$ at other locations will be adjusted by the effect of $\mathcal{T}(t)$, which is an integration constant that ensures conservation of the total momentum. Specifically, $\partial_{t}\int U{\rm d}x=0$ implies $$\quad\mathcal{T}(t)=\frac{1}{L_x}\int\langle\tilde{v}_{x}{\rm i}\hat{\delta}\tilde{\varphi}\rangle{\rm d}x.$$ Due to nonzero $\mathcal{T}$, ZFs cannot remain (quasi)stationary in the presence of fluctuations within the mTHE. In other words, either ZFs completely suppress DW turbulence, or both ZFs and DWs keep evolving indefinitely.
Primary, secondary, and tertiary instability {#sec:Linear}
============================================
We have integrated the mTHE numerically using random noise for the initial conditions. Typical simulation results are presented in figures \[fig:THhistory\] and \[fig:PI\_and\_SI\]. It is seen that the primary instability of DWs arises and is followed by ZF generation through the secondary instability. Then, at the fully nonlinear stage, DW turbulence becomes inhomogeneous, exhibiting signatures of the TI. In the following, we study these stages in detail.
![Snapshots from numerical simulations of the mTHE (\[eq:Equations\_mTHE\]) with $\delta_0=1.5$ (see (\[eq:Equations\_delta\])) at (a) $\beta=4.5$ (first row) and (b) $\beta=6.5$ (second row). The simulation domain size is $L_{x}=L_{y}=20\upi$, with the corresponding numbers of grid points being $N_x=128$ and $N_y=64$, respectively. Periodic boundary conditions are used in both directions, and the nonlinear term is treated using the pseudospectral method with 2/3 dealiasing rule [@boyd2001chebyshev]. The initial conditions are random noise with a small amplitude. Shown are the fluctuations $\tilde{w}$ (colorbar) and the ZF velocity $U$ (green curve) at three different moments of time. It is seen that at $\beta=4.5$, the DW amplitude decreases down to zero (Dimits regime), while at $\beta=6.5$, fluctuations remain strong and ZFs keep evolving.[]{data-label="fig:THhistory"}](figures/THhistory){width="1\columnwidth"}
Primary instability
-------------------
It is straightforward to show that $\lbrace \varphi, w \rbrace = 0$ for Fourier eigenmodes of the form $$\varphi=\varphi_{\boldsymbol{k}}{\rm e}^{{\rm i}\boldsymbol{k}\cdot\boldsymbol{x}-{\rm i}\Omega_{\boldsymbol{k}}t}+{\rm c.c.},$$ where $\boldsymbol{k}=(k_{x},k_{y})$. Therefore, a Fourier eigenmode is an exact solution of the system provided that $\Omega_{\boldsymbol{k}}$ satisfies the following relation: $$\label{eq:Linear_Primary_dispersion}
\omega_{\boldsymbol{k}}\doteq\Real\,\Omega_{\boldsymbol{k}}=\frac{\beta k_{y}(1+k^{2})}{(1+k^{2})^{2}+\delta_{0}^{2}k_{y}^{2}},\quad
\gamma_{\boldsymbol{k}}\doteq\Imag\,\Omega_{\boldsymbol{k}}=\frac{\beta\delta_{0}k_{y}^{2}}{(1+k^{2})^{2}+\delta_{0}^{2}k_{y}^{2}}-\alpha_{\boldsymbol{k}}D_{\boldsymbol{k}}.\label{eq:Linear_Primary_Omegak}$$ Here, $k^{2}\doteq k_{x}^{2}+k_{y}^{2}$, $D_{\boldsymbol{k}}=1+0.01k^{2}$, and we have used (\[eq:Equations\_delta\]). Also, $\alpha_{\boldsymbol{k}}=1$ for $k_{y}\neq0$ and $\alpha_{\boldsymbol{k}}=0$ for $k_{y}=0$, and hence a ZF ($k_{y}=0$) corresponds to $\Omega_{\boldsymbol{k}}=0$, [[i.e., ]{}]{}to a stationary state. From (\[eq:Linear\_Primary\_Omegak\]), it is seen that when $D_{\boldsymbol{k}}=0$, $\gamma_{\boldsymbol{k}}$ is maximized at $(k_{x},k_{y})=(0,1)$. A nonzero $D_{\boldsymbol{k}}$ can modify the value of $\boldsymbol{k}$ that maximizes $\gamma_{\boldsymbol{k}}$, but for the chosen form of $\hat{D}$, (\[eq:Equations\_D\]), this modification is very small. Therefore, if one numerically simulates (\[eq:Equations\_mTHE\]) with small random noise as the initial conditions, then nonlinear interactions can be neglected at first and coherent DW structures will grow exponentially with typical wavenumber $\boldsymbol{k}\approx(0,1)$, as seen in figure \[fig:THhistory\].
Secondary instability {#subsec:Linear_Secondary}
---------------------
When many Fourier modes are present and have grown to a finite amplitude, the nonlinear term in (\[eq:Equations\_mTHE\]) becomes important. This can be seen from the Fourier representation, $\varphi=\sum_{\boldsymbol{k}}\varphi_{\boldsymbol{k}}(t)\exp({\rm i}\boldsymbol{k}\cdot\boldsymbol{x})$, where (\[eq:Equations\_mTHE\]) is written as $$\frac{{\rm d}\varphi_{\boldsymbol{k}}}{{\rm d}t}=-{\rm i}\Omega_{\boldsymbol{k}}\varphi_{\boldsymbol{k}}+\frac{1}{2}\sum_{\boldsymbol{k}_{1},\boldsymbol{k}_{2}}T(\boldsymbol{k},\boldsymbol{k}_{1},\boldsymbol{k}_{2})\delta_{\boldsymbol{k},\boldsymbol{k}_{1}+\boldsymbol{k}_{2}}\varphi_{\boldsymbol{k}_{1}}\varphi_{\boldsymbol{k}_{2}}\label{eq:Linear_Primary_Fourier}$$ and $\delta_{\boldsymbol{k}_1, \boldsymbol{k}_2}$ is the Kronecker symbol. Also, $$T(\boldsymbol{k},\boldsymbol{k}_{1},\boldsymbol{k}_{2})\doteq-\frac{\bar{k}_{1}^{2}-\bar{k}_{2}^{2}}{\bar{k}^{2}}(\boldsymbol{k}_{1}\times\boldsymbol{k}_{2})\cdot\hat{\boldsymbol{z}}$$ are the coefficients that govern the nonlinear mode coupling, $\bar{k}^{2}$ is defined as $$\bar{k}^{2}\doteq\alpha_{\boldsymbol{k}}+k^{2}-{\rm i}\delta_{0}k_{y},\label{eq:Equations_kbar}$$ and similarly for $\bar{k}_{1}^{2}$ and $\bar{k}_{2}^{2}$.
Due to nonlinear interactions, ZFs can be generated from DWs, which process is known as the secondary instability. Here, we use the 4MT model to analyze this instability, namely, by considering a primary DW with $\boldsymbol{k}=(0,k_{y})$, a ZF with $\boldsymbol{q}=(q_{x},0)$, and two DW sidebands with $\boldsymbol{k}_{\pm}=(\pm q_{x},k_{y})$. Assume that the ZF is small, so the exponential growth of the primary DW is unaffected; [[i.e., ]{}]{}$\varphi_{\boldsymbol{k}}=\varphi_{0}\exp(-{\rm i}\Omega_{\boldsymbol{k}}t$), with $\varphi_{0}$ being a constant. Then, from (\[eq:Linear\_Primary\_Fourier\]), the equations that describe the ZF and the sidebands are as follows [@St-Onge17]: $$\begin{aligned}
\label{eq:Linear_Secondary_4MT}
{\rm d}_{t}\varphi_{\boldsymbol{q}}=\frac{k_{y}{\rm e}^{\gamma_{\boldsymbol{k}} t}}{q_{x}}\left[\left(q_{x}^{2}-{\rm i}\delta_{+}\right)\varphi_{\boldsymbol{k}_{+}}\varphi_{0}^{*}{\rm e}^{{\rm i}\omega_{\boldsymbol{k}}t}-\left(q_{x}^{2}+{\rm i}\delta_{+}\right)\varphi_{\boldsymbol{k}_{-}}^{*}\varphi_{0}{\rm e}^{-{\rm i}\omega_{\boldsymbol{k}}t}\right],\label{eq:Linear_Secondary_4MT-a}\\
{\rm d}_{t}\varphi_{\boldsymbol{k}_{+}}=-{\rm i}\Omega_{\boldsymbol{k}_{+}}\varphi_{\boldsymbol{k}_{+}}+T(\boldsymbol{k}_{+},\boldsymbol{k},\boldsymbol{q})\varphi_{0}\varphi_{\boldsymbol{q}}{\rm e}^{-{\rm i}\Omega_{\boldsymbol{k}}t},\\
{\rm d}_{t}\varphi_{\boldsymbol{k}_{-}}=-{\rm i}\Omega_{\boldsymbol{k}_{-}}\varphi_{\boldsymbol{k}_{-}}+T(\boldsymbol{k}_{-},\boldsymbol{k},-\boldsymbol{q})\varphi_{0}\varphi_{\boldsymbol{q}}^{*}{\rm e}^{-{\rm i}\Omega_{\boldsymbol{k}}t},\end{aligned}$$ where $\delta_{+}\doteq\delta_{\boldsymbol{k}}+\delta_{\boldsymbol{k}_{+}}=2\delta_{0}k_{y}$. We have also used $\Omega_{\boldsymbol{k}}=\omega_{\boldsymbol{k}}+{\rm i}\gamma_{\boldsymbol{k}}$. These equations can be combined to yield a single time-evolution equation for the ZF amplitude $\varphi_{\boldsymbol{q}}$: $$\frac{{\rm d}^{3}\varphi_{\boldsymbol{q}}}{{\rm d}t^{3}}-A\,\frac{{\rm d}^{2}\varphi_{\boldsymbol{q}}}{{\rm d}t^{2}}+(B-C)\,\frac{{\rm d}\varphi_{\boldsymbol{q}}}{{\rm d}t}-D\varphi_{\boldsymbol{q}}=0.\label{eq:Linear_Secondary_phiqt}$$ Here, $A=2\gamma_{+}$, $B=\omega_{-}^{2}+\gamma_{+}^{2}$, $C,D\propto|\varphi{}_{0}{\rm e}^{\gamma_{\boldsymbol{k}}t}|^{2}$, $\gamma_{+}\doteq\gamma_{\boldsymbol{k}}+\gamma_{\boldsymbol{k}_{+}}$, and $\omega_{-}\doteq\omega_{\boldsymbol{k}}-\omega_{\boldsymbol{k}_{+}}$. The derivation of (\[eq:Linear\_Secondary\_phiqt\]) can be found in @St-Onge17. Expressions for $C$ and $D$ can also be found there but will not be important for our discussion; but note that compared to @St-Onge17, we have absorbed the coefficient ${\rm e}^{\gamma_{\boldsymbol{k}}t}$ into the definitions of $C$ and $D$.
When $C$ and $D$ are much larger than $A$ and $B$, $\varphi_{\boldsymbol{q}}$ can grow “super-exponentially” [@Rogers00; @St-Onge17], [[i.e., ]{}]{}as an exponential of an exponential. This is also known as the secondary KH instability [@Rogers00]. In the opposite case, when $A$ and $B$ dominate over $C$ and $D$, the non-constant solution of (\[eq:Linear\_Secondary\_phiqt\]) is approximately $$\varphi_{\boldsymbol{q}}\propto{\rm e}^{(\gamma_{+}\pm{\rm i}\omega_{-})t}.$$ Since $\gamma_{+}$ decreases as $|q_{x}|$ increases (see (\[eq:Linear\_Primary\_Omegak\])), the growth rate is maximized at the lowest ZF wavenumber $|q_{x}|=2\upi/L_{x}$. In other words, the box-scale ZF grows fastest, with the growth rate given by $\gamma_{+}\approx2\gamma_{\boldsymbol{k}}$, i.e., twice the growth rate of the primary DW instability.
![The time history of the DW and ZF energies (\[eq:Energies\]) corresponding to figure \[fig:THhistory\]. The primary and secondary instabilities are clearly seen, with the secondary-instability growth rate being twice the primary-instability growth rate. The black dashed line is the ZF energy calculated from (\[eq:Linear\_Secondary\_Ec\]) that corresponds to the critical ZF amplitude $\varphi_{{\rm c}}$ from (\[eq:Linear\_Secondary\_phic\]). It is seen that this energy roughly corresponds to the onset of the fully nonlinear regime. This value is reached by both $E_{\rm DW}$ and $E_{\rm ZF}$ at approximately the same time.[]{data-label="fig:PI_and_SI"}](figures/PI_and_SI){width="1\columnwidth"}
In the following, we show that this exponential growth of the ZF at the box scale is more common than the super-exponential growth, provided that the characteristic amplitude $\varphi_0$ of the initial random noise is small enough. At first, both the primary DW and the sidebands grow exponentially, $$|\varphi_{\boldsymbol{k}}|\sim|\varphi_{\boldsymbol{k}_{\pm}}|\sim|\varphi_{{\rm 0}}|{\rm e}^{\gamma_{\boldsymbol{k}}t},$$ while the ZF amplitude remains at the noise level. Then, DWs grow for some time $t_{{\rm p}}$ before they begin to affect ZFs. Assume that at $t=t_{{\rm p}}$, the box-scale ZF with the amplitude $\varphi_{\boldsymbol{q}}\sim\varphi_{0}$ starts to grow with the growth rate $\gamma_{+}\approx2\gamma_{\boldsymbol{k}}$; then, $\delta_{+}=2\delta_{0}k_{y}\gg q_{x}^{2}$, and we have from (\[eq:Linear\_Secondary\_4MT-a\]) that $$|\partial_{t}\varphi_{\boldsymbol{q}}|\sim 2\gamma_{\boldsymbol{k}}|\varphi_{{\rm 0}}|\approx\frac{2|\varphi_{\boldsymbol{k}}||\varphi_{\boldsymbol{k}_{+}}|k_{y}\delta_{+}}{q_{x}}\approx\frac{2k_{y}\delta_{+}|\varphi_{{\rm 0}}{\rm e}^{\gamma_{\boldsymbol{k}}t_{{\rm p}}}|^{2}}{q_{x}}.\label{eq:Linear_Secondary_phi0}$$ This leads to $$C,D\propto\frac{q_{x}\gamma_{\boldsymbol{k}}|\varphi_{0}|}{2\delta_{0}k_{y}^{2}}.$$ Therefore, $C$ and $D$ are small when the initial noise level $|\varphi_{0}|$ is small enough; hence, the assumptions made above are self-consistent, namely, $A$ and $B$ are indeed much larger than $C$ and $D$, and the box-scale ZF with wavenumber $q_{x}=2\pi/L_{x}$ grows fastest with the growth rate $2\gamma_{\boldsymbol{k}}$.
The secondary instability will persist for some time $t_{{\rm s}}$ until ZFs grows up to a finite amplitude that is enough to significantly distort the DW structure. Using the result from [@Zhu18b], this amplitude can be estimated as follows (also see (\[eq:appB\_U\_critical\])): $$\varphi_{{\rm c}}=\frac{\beta/q_{x}}{2(1+k_{y}^{2})-q_{x}^{2}}.\label{eq:Linear_Secondary_phic}$$ At $\varphi_{\boldsymbol{q}}\ll\varphi_{{\rm c}}$, DWs do not “see” the ZF and hence keep growing exponentially, while at $\varphi_{\boldsymbol{q}}\gtrsim\varphi_{{\rm c}}$ the system enters the fully nonlinear regime. Therefore, $t_{{\rm s}}$ is the time when the ZF amplitude grows from $\varphi_{0}$ to $\varphi_{{\rm c}}$, and it can be estimated as follows: $$t_{{\rm s}}=\frac{1}{2\gamma_{\boldsymbol{k}}}\ln\frac{\varphi_{{\rm c}}}{\varphi_{0}}.\label{eq:ts}$$ Note that (\[eq:Linear\_Secondary\_phic\]) is obtained from the modified Hasegawa–Mima system, so it is based on the assumption that $\delta_0 = 0$. For nonzero $\delta_0$, it is modified accordingly (see (\[eq:appB\_U\_critical\])), but the above estimate is sufficient for our qualitative description.
By the time when the system enters the fully nonlinear regime, the DW amplitude becomes $|\varphi_{\boldsymbol{k}}|\sim\varphi_0\exp{\gamma_{\boldsymbol{k}}(t_{\rm s}+t_{\rm p})}$, which can be estimated from (\[eq:Linear\_Secondary\_phi0\]) and (\[eq:ts\]) as $$|\varphi_{\boldsymbol{k}}|\sim\sqrt{\frac{q_{x}\gamma_{\boldsymbol{k}}\varphi_{\rm c}}{2\delta_{0}k_{y}^{2}}}.$$ From (\[eq:Energies\]), the corresponding DW and ZF energies are as follows: $$E_{{\rm ZF}}\sim\frac{\beta^{2}}{8(1+k_{y}^{2})^{2}},\quad
E_{{\rm DW}}\sim\frac{\beta\gamma_{\boldsymbol{k}}}{8\delta_{0}k_{y}^{2}},\label{eq:Linear_Secondary_Ec}$$ where we assumed $q_x^2\ll 1+k_y^2$. Using (\[eq:Linear\_Primary\_Omegak\]) for $\gamma_{\boldsymbol{k}}$ and assuming $D_{\boldsymbol{k}}=0$ for simplicity, we obtain $$\frac{E_{{\rm ZF}}}{E_{{\rm DW}}}\sim 1 + \frac{\delta_0^2k_y^2}{(1+k_y^2)^2}\,.$$ This shows that the ZF energy and the DW energy are roughly equal to each other when the system enters the fully nonlinear regime, since $\delta_0$ and $k_y$ are of order unity. This conclusion will be used to estimate the ZF curvature in section \[sec:DSPrediction\].
These predictions are in agreement with numerical simulations (figure \[fig:PI\_and\_SI\]). This indicates that the 4MT captures the basic dynamics of the primary and the secondary instabilities. However, as shown below, the 4MT does not capture essential features of the TI, and thus more accurate models are needed to describe the TI and the Dimits shift.
Tertiary instability {#subsec:Linear_Tertiary}
--------------------
![The first four tertiary eigenmodes found numerically using the ZF velocity profile (\[eq:Linear\_Tertiary\_sinusoidalU\]). The ordering is such that $\gamma_{\rm TI}$ decreases from left to right. The first two eigenmodes are runaway and trapped modes, respectively. The parameters are $\beta=6$, $\delta_{0}=1.5$, $q_{x}=0.4$, and $u=10$. The first row shows the eigenmode structures $\tilde{w}(x,y)=\Real[w(x){\rm e}^{{\rm i}k_{y}y}]$ ((\[eq:Linear\_Tertiary\_eigenmode\]), color), the ZF velocity $U$ (green curve), and the analytic mode structure $\tilde{w}=\Real[\mathsf{H}_m(x)\exp(S+{\rm i}k_y y)]$ ((\[eq:AnalyticTI\_S\]), dashed contour), where $m=0$ for (a1) and (b1), $m=1$ for (c1), and $m=2$ for (d1). The second row shows the corresponding Wigner function $W(x,k_x)$ ((\[eq:appA\_Wigner\]), color) and the isosurfaces of the drifton Hamiltoninan $\mathcal{H}$ ((\[eq:appB\_Drifton\_Hamiltonian\]), dashed contour). The striped structure of $W$ away from the actual location of DW quanta is a signature of a quantumlike “cat state” [@Weinbub18].[]{data-label="fig:TI_eigenvalue"}](figures/eigen){width="100.00000%"}
In the fully nonlinear regime, DW turbulence becomes inhomogeneous and localized at the extrema of the ZF velocity $U$ (figure \[fig:THhistory\]). To understand the DW dynamics in this case, let us linearize (\[eq:Equations\_mTHE\]) to obtain $$\partial_{t}\tilde{w}+U\partial_{y}\tilde{w}-(\beta+U'')\partial_{y}\tilde{\varphi}+\hat{D}\tilde{w}=0,\label{eq:Linear_Tertiary_linearDW}$$ where $$\tilde{w}=(\nabla^{2}-1+{\rm i}\hat{\delta})\tilde{\varphi},\quad U''\doteq{\rm d}^{2}U(x)/{\rm d}x^{2}.$$ For given boundary conditions in $x$, eigenmodes of (\[eq:Linear\_Tertiary\_linearDW\]) can be searched for in the form $$\tilde{w}=w(x){\rm e}^{{\rm i}(k_{y}y-\omega t)},\quad\tilde{\varphi}=\left(\frac{{\rm d^{2}}}{{\rm d}x^{2}}-k_{y}^{2}-1+{\rm i}\delta_{0}k_{y}\right)^{-1}\tilde{w},\label{eq:Linear_Tertiary_eigenmode}$$ which leads to the following equation for $w(x)$: $$\omega w=\hat{H}w,\quad \hat{H}(\hat{x},\hat{k}_x)\doteq k_{y}\hat{U}+k_{y}(\beta+\hat{U}'')\hat{\bar{k}}^{-2}-{\rm i}\hat{D},\label{eq:Linear_Tertiary_eigen}$$ where $$\hat{U}=U(\hat{x}),\quad \hat{k}_x=-{\rm i}\,{\rm d}/{\rm d}x,\quad\hat{\bar{k}}^{2}=1+k_{y}^{2}+\hat{k}_x^2-{\rm i}\delta_{0}k_{y}.\label{eq:Linear_Tertiary_khat}$$ If an eigenvalue $\omega$ exists and $$\gamma_{{\rm TI}}\doteq\Imag\,\omega>0,$$ then the perturbation grows exponentially. This is the TI.
Equation (\[eq:Linear\_Tertiary\_linearDW\]) does not have an analytic solution for an arbitrary profile $U$, but a general understanding can be developed by considering special cases. In [@Zhu18c], we considered the ZF velocity profile $$U(x)=u\cos q_{x}x,\label{eq:Linear_Tertiary_sinusoidalU}$$ with $\hat{\delta} = \hat{D} = 0$. In this case, the system exhibits an instability of the KH type provided that $q_{x}^{2}>1$ and $q_x^2u>\beta$. In [@Zhu18c], we also discussed a generalization to periodic nonsinusoidal profiles. However, generalizing those results to nonzero $\hat{\delta}$ and $\hat{D}$ is challenging. The common approach is to adopt the 4MT again, [[i.e., ]{}]{}to assume a DW perturbation with $\boldsymbol{k}=(0,k_{y})$ and two sidebands with $\boldsymbol{k}_{\pm}=(\pm q_{x},k_{y})$ as small perturbations [@Kim02; @St-Onge17; @Rath18; @Zhu18a]. In particular, [@St-Onge17] derived $\gamma_{{\rm TI}}$ within the 4MT and estimated the Dimits shift by finding a sufficient condition for $\gamma_{{\rm TI}}=0$. However, the 4MT-based approach is not entirely satisfactory, because the ZF is typically far from sinusoidal, as seen in simulations. Even more importantly, the 4MT approach ignores the fact that there are multiple TI modes with different growth rates. As we show below, understanding the variety of these modes is essential for understanding the Dimits shift.
Let us assume the same sinusoidal ZF profile (\[eq:Linear\_Tertiary\_sinusoidalU\]) as in @St-Onge17 for now, and let us calculate the corresponding eigenmodes (\[eq:Linear\_Tertiary\_eigen\]) numerically, assuming periodic boundary conditions $x$. In this case, we can search for solutions in the form $$w(x)=\sum_{n=-N}^{N}w_{n}{\rm e}^{{\rm i}nq_{x}x},$$ where $N$ is some large enough integer. In other words, we truncate the Fourier series by keeping only the first $2N+1$ Fourier modes. This turns (\[eq:Linear\_Tertiary\_eigen\]) into a vector equation for $\{w_{-N}, \dots w_0, \dots w_N\}$, where $\hat{H}$ becomes a $(2N+1)\times(2N+1)$ matrix. Then, one finds $2N+1$ eigenmodes with complex eigenfrequencies. Typical numerical eigenmodes are illustrated in figure \[fig:TI\_eigenvalue\]. It is seen that the TI-mode structure is localized at the maximum ($x=0$) or minimum ($x=-\upi/q_{x}$) of the ZF velocity and has either even or odd parity because of the symmetry of $U$. Within the figure, the eigenmodes localized at the ZF minimum can be labeled by the integer $m=0,1,2,\dots$, which also indicates the parity of $w(x)$. Eigenmodes localized near the ZF maximimum can be labeled similarly. Note that in order for a mode to be localized, the ZF must be large-scale, namely, $q_x^2\ll 1+k_y^2$, which is consistent with numerical simulations.
Apart from the eigenmode structures, we also show in figure \[fig:TI\_eigenvalue\] their corresponding Wigner functions $W(x,k_x)$ (\[eq:appA\_Wigner\]) and contour plots of the drifton Hamiltonian $\mathcal{H}$ (\[eq:appB\_Drifton\_Hamiltonian\]). The Wigner function can be understood as the distribution function of “driftons” (DW quanta) in the $(x,k_{x})$ phase space [@Smolyakov99; @Ruiz16; @Zhu18c], and its shape is expected to align with the contours of $\mathcal{H}$ . Then, eigenmodes are naturally centered at phase-space equilibria of $\mathcal{H}$, namely, $$\partial_{x}\mathcal{H}=\partial_{k_{x}}\mathcal{H}=0\quad\Rightarrow\quad U' = k_x=0.\label{eq:Linear_Tertiary_localize}$$ This explains eigenmode localization near extrema of $U$. \[Strictly speaking, (\[eq:Linear\_Tertiary\_localize\]) stems from our approximation of sinusoidal flow (\[eq:Linear\_Tertiary\_sinusoidalU\]), which ensures that $U'$ and $U'''$ become zero at same locations. Nevertheless, (\[eq:Linear\_Tertiary\_localize\]) remains a good approximation as long as ZFs are large-scale, [[i.e., ]{}]{}$|U'''/\bar{k}^2|\ll U'$.\] Maxima of $U$ (even $n$) correspond to phase-space islands encircled by “trapped” trajectories, and minima of $U$ (odd $n$) correspond to saddle points passed by the “runaway” trajectories [@Zhu18a; @Zhu18b; @Zhu18c]. Hence, we call the modes localized near maxima and minima of $U$ trapped and runaway modes, respectively. (See Appendix \[appB\] for more discussions on drifton phase-space trajectories.) In the next section, we provide analytic calculation of the TI growth rates based on the above observations.
Tertiary-instability growth rate {#sec:TIcalculation}
================================
Analogy with a quantum harmonic oscillator {#subsec:Analytic_oscillator}
------------------------------------------
As seen in figure \[fig:TI\_eigenvalue\], tertiary modes are centered at the phase-space equilibria. Based on this, let us expand the Hamiltonian up to the second order both in $x$ and in $\hat{k}_x$. Specifically, we approximate the ZF velocity with a parabola: $$U\approx U_0+\frac{1}{2}\,\mathcal{C}x^2,\label{eq:AnalyticTI_oscillator_Uexpand}$$ where $U_0$ is the local ZF velocity and $\mathcal{C}\doteq U''(0)$ is the local ZF curvature. For the sinusoidal velocity (\[eq:Linear\_Tertiary\_sinusoidalU\]), this corresponds to $U_0=\pm u$ and $\mathcal{C}=\mp q_x^2 u$. We also make the approximation that $\hat{D}\approx D_0\doteq D_{\boldsymbol{k}=(0,k_y)}$ and $$\quad \hat{k}^{-2}\approx k_0^{-2}+k_0^{-4}\frac{\rm d^2}{{\rm d}x^2},\quad k_0^2\doteq 1+k_y^2-{\rm i}\delta_0 k_y.\label{eq:AnalyticTI_oscillator_Kexpand}$$ Then, the Hamiltonian operator $\hat{H}$ (\[eq:Linear\_Tertiary\_eigen\]) is approximated as $$\hat{H}\approx k_y U_0 +\frac{1}{2}k_y\mathcal{C}\hat{x}^2+k_y(\beta+\mathcal{C})\left(k_0^{-2}+k_0^{-4}\frac{\rm d^2}{{\rm d}x^2}\right)-{\rm i}D_0,\label{eq:AnalyticTI_oscillator_Hexpand}$$ and the corresponding eigenmode equation (\[eq:Linear\_Tertiary\_eigen\]) becomes $$\left(-\tau^2\frac{\rm d^2}{{\rm d}x^2}+x^2\right)w=\varepsilon w.\label{eq:AnalyticTI_oscillator}$$ It is the same equation that describes a quantum harmonic oscillator, except that here the coefficients are complex; specifically, $$\tau^2\doteq-\frac{2}{k_0^4}\left(1+\frac{\beta}{\mathcal{C}}\right),\quad\varepsilon\doteq\frac{2[\omega-k_y U_0+{\rm i}D_0-k_y(\beta+\mathcal{C})/k_0^2]}{k_y\mathcal{C}}.\label{eq:AnalyticTI_oscillator_coefficients}$$ Note that the coefficients are different at minima and maxima of $U$, as they depend on the sign of $\mathcal{C}$. Also note that for runaway modes, we have shifted the coordinate as $x\to x+\upi/q_{x}$ to recenter the ZF minimum at $x=0$.
Following the standard procedure known from quantum mechanics [@Sakurai94book], one can show that the asymptotic behavior of the solution at large $|x|$ is $$w(x)\sim {\rm e}^{S(x)},\quad S(x)=-\frac{x^2}{2\tau}=-\left[\frac{{\rm i}(1+k_y^2)+\delta_0k_y}{2\sqrt{2(1+\beta/\mathcal{C})}}\right]x^2.\label{eq:AnalyticTI_S}$$ To ensure that $w\to 0$ at large $|x|$, we require ${\rm Im}\sqrt{1+\beta/\mathcal{C}}\,>0$ if $1+\beta/\mathcal{C}<0$. We also assumed that $\delta_0,k_y>0$. Then, letting $w=\phi(x)\exp S(x)$, we obtain $$\phi''-\frac{2x}{\tau}\phi'+\frac{\varepsilon-\tau}{\tau^2}\phi=0.\label{eq:Hermite}$$ Solutions are $\phi=\mathsf{H}_m(x/\sqrt{\tau})$, where $\mathsf{H}_m$ are Hermite polynomials, $m=0,1,2,\dots$, and $$\varepsilon=(2m+1)\tau.$$ Therefore, for each sign of $\mathcal{C}$, eigenmodes are labeled by $m$. In figure \[fig:TI\_eigenvalue\], these approximate solutions are compared with numerical solutions of (\[eq:Linear\_Tertiary\_eigen\]). In the following, we shall focus on the two modes with $m = 0$, since they are most unstable. In this case, $\phi=\mathsf{H}_0$ is constant and $\varepsilon=\tau$. This corresponds to $\tilde{w}=\Real[\exp(S+{\rm i}k_y y)]$, and the eigenfrequencies are found from (\[eq:AnalyticTI\_oscillator\_coefficients\]) to be $$\omega=\bar{\Omega}+k_yU_0-\frac{{\rm i}k_y\mathcal{C}\sqrt{(1+\beta/\mathcal{C})/2}}{1+k_y^2-{\rm i}\delta_0 k_y},\quad \bar{\Omega}=\frac{k_y(\beta+\mathcal{C})}{k_0^2}-{\rm i}D_0.\label{eq:AnalyticTI_oscillator_omega}$$ Here, $\bar{\Omega}$ is the primary-mode eigenfrequency $\Omega_{\boldsymbol{k}}$ (\[eq:Linear\_Primary\_Omegak\]) modified by $\mathcal{C}$, $k_yU_0$ is the local Doppler shift, and the remaining term in $\omega$ vanishes at zero $\mathcal{C}$. Note that at $\mathcal{C}=0$, $\omega$ reduces to the primary-mode frequency $\Omega_{\boldsymbol{k}}$ at $\boldsymbol{k}=(0,k_y)$. Hence, TI modes found here can be interpreted as standing primary modes modified by ZFs. Accordingly, the TI growth rate $\gamma_{\rm TI}$ approaches the primary-instability growth rate in the limit $\mathcal{C} \to 0$.
Let us examine the validity of our approximation in (\[eq:AnalyticTI\_oscillator\_Hexpand\]). First, the parabolic approximation of $U$ is valid if the mode spatial width in $x$, which is determined by $\mathcal{C}$, is much smaller than $q_x^{-1}$, which is the characteristic scale of ZFs. Specifically, for the sinusoidal ZF (\[eq:Linear\_Tertiary\_sinusoidalU\]), we have $\mathcal{C}=q_x^2 u$, so the parabolic approximation (\[eq:AnalyticTI\_oscillator\_Uexpand\]) is valid at small enough $q_x$ and large enough $u=\mathcal{C}/q_x^2$. Second, the expansion of $\hat{k}^{-2}$ in (\[eq:AnalyticTI\_oscillator\_Kexpand\]) is valid at $|k_x^2|\ll |k_0^2|$, where $k_x$ is the characteristic mode wavenumber in $x$. From (\[eq:AnalyticTI\_oscillator\]), $k_x$ can be estimated as $k_x=1/\sqrt{\tau}$. Then, the requirement $|k_x^2|\ll |k_0^2|$ leads to $|\sqrt{2(1+\beta/\mathcal{C})}|\gg 1$, or equivalently, $|\mathcal{C}|\ll\beta$, and one expects that the approximation in (\[eq:AnalyticTI\_oscillator\_Kexpand\]) becomes invalid as $|\mathcal{C}|$ approaches $\beta$. Therefore, in the following, we restrict our consideration to the parameter regime $|\mathcal{C}|<\beta$, which is also the regime relevant to our numerical simulations.
The TI growth rate $\gamma_{\rm TI}$ is obtained by taking the imaginary part of $\omega$. Within the regime $|\mathcal{C}|<\beta$, let us introduce the notation $$\bar{\gamma}\doteq{\rm Im}\,\bar{\Omega}=\frac{\delta_0k_y^2(\beta+\mathcal{C})}{(1+k_y^2)^2+\delta_0^2k_y^2}-D_0,$$ which is the primary-instability growth rate $\gamma_{\boldsymbol{k}}$ (\[eq:Linear\_Primary\_Omegak\]) modified by $\mathcal{C}$. Then, for the runaway mode (labeled with superscript “R”), which corresponds to $\mathcal{C}>0$, one has $$\gamma_{\rm TI}^{\rm R}=\bar{\gamma}-\sqrt{\frac{1}{2}\left(1+\frac{\beta}{\mathcal{C}}\right)}\,\frac{(1+k_y^2)k_y\mathcal{C}}{(1+k_y^2)^2+\delta_0^2k_y^2}. \label{eq:Analytic_gammaR}$$ For the trapped mode (labeled with superscript “T”), which corresponds to $\mathcal{C}<0$, one has $$\gamma_{\rm TI}^{\rm T}=\bar{\gamma}+\sqrt{\frac{1}{2}\left(\left|\frac{\beta}{\mathcal{C}}\right|-1\right)}\,\frac{\delta_0k_y^2\mathcal{C}}{(1+k_y^2)^2+\delta_0^2k_y^2}.\label{eq:Analytic_gammaT}$$ Figure \[fig:TIgrowthRate\](c) shows that these formulas are in good agreement with our numerical calculations of the eigenvalues. Also, we have verified that (not shown) the results in figure \[fig:TIgrowthRate\](c) are insensitive to $q_x$ as long as $q_x$ is small, more specifically, $q_x^2\ll 1+k_y^2$.
Notably, while the trapped-mode growth rate always decreases with $|\mathcal{C}|$, the runaway-mode growth rate can increase at large $\mathcal{C}$ if $\delta_0$ is large. In fact, at $\mathcal{C}\gg\beta$, (\[eq:Analytic\_gammaR\]) becomes $$\gamma_{\rm TI}^{\rm R}\approx \left[\frac{\delta_0k_y^2}{(1+k_y^2)^2+\delta_0^2k_y^2}-\sqrt{\frac{1}{2}}\,\frac{(1+k_y^2)k_y}{(1+k_y^2)^2+\delta_0^2k_y^2}\right]\mathcal{C},$$ which predicts that $\gamma_{\rm TI}^{\rm R}$ increases with $\mathcal{C}$ if $\delta_0>(k_y + k_y^{-1})/\sqrt{2}$. This is verified by our numerical calculations for $k_y=1$ (not shown), which predict a similar condition, $\delta_0\gtrsim 1.7$. Therefore, it is possible that the TI can develop in strong ZFs, but the physical mechanism is very different from the KH mode, as will be discussed in section \[subsec:KHI\].
Alternative approach {#subsec:Analytic_WME}
--------------------
![The structure of the tertiary modes with $m=0$ in the zonal-velocity profile (\[eq:Linear\_Tertiary\_sinusoidalU\]). The parameters are $\delta_{0}=1.6$, $\beta=5$, $q_{x}=0.2$, $u=50$ (hence, $\mathcal{C}=\pm 2$), and $\hat{D}$ given by (\[eq:Equations\_D\]). These parameters result in $\gamma_{{\rm TI}}^{{\rm R}}=-0.276$ and $\gamma_{{\rm TI}}^{{\rm T}}=-0.587$. (a) The Wigner function $W(x,k_x)$ of the runaway mode (color), the local $U$ (magenta curve), and the runaway trajectory (dashed curve; see (\[eq:GammaTI\_runaway\_traj\])). Note that we have shifted the coordinates as $x\to x+\pi/q_x$ to recenter the ZF minimum at $x=0$. (b) The structure of each term in (\[eq:GammaTI\_xintegration\]) calculated from $W$ of the runaway mode in (a). (c) The Wigner function $W(x,k_x)$ of the trapped mode (color), the local $U$ (magenta curve), and isosurfaces of $\mathcal{H}$ (dashed contours; see (\[eq:GammaTI\_Hamiltonian\])). In this figure, $\Delta x$ and $\Delta k_x$ denote the characteristic widths of the mode in the $x$ and $k_x$ directions, correspondingly. (d) Same as (b) but for the trapped mode. \[fig:TIstructure\]](figures/modeStructure){width="1\columnwidth"}
![(a) The empirical factor $\eta_{\rm R}$ (\[eq:GammaTI\_runaway\_eta\]) as a function for $\delta_0$: numerical values of a sinusoidal ZF (\[eq:Linear\_Tertiary\_sinusoidalU\]) (blue circles) versus the fitting formula (black curve). The parameters are $\beta=6$, $q_x=0.4$, $u=10$, and $\hat{D}$ given by (\[eq:Equations\_D\]). It is found that $\eta_{\rm R}$ is not sensitive to $u$. (b) Same as (a) except for $\eta_{\rm T}$ (\[eq:GammaTI\_trapped\_eta\]). It is found that $\eta_{\rm T}$ is not sensitive to $u$ at $u< \beta/q_x^2$. (c) The TI growth rates versus $|\mathcal{C}|=q_x^{2}u$ at $\delta_{0}=1.5$, $\beta=6$, $q_{x}=0.4$, and varying $u$. Black curves: numerical solutions of (\[eq:Linear\_Tertiary\_eigen\]) indicated by the superscript “N”. Multiple branches are shown, with the two most unstable branches being the runaway mode and the trapped mode. Blue dashed curve and red dash-dotted curve: analytic formulas (\[eq:Analytic\_gammaR\]) and (\[eq:Analytic\_gammaT\]). The superscript “1” corresponds to predictions made using the approach described in section \[subsec:Analytic\_oscillator\]. Blue circles and red squares: analytic formulas (\[eq:GammaTI\_runaway\_GrowthRate\]) with $\eta_{{\rm R}}=0.595$ and (\[eq:GammaTI\_trapped\_GrowthRate\]) with $\eta_{{\rm T}}=1.2$. The superscript “2” corresponds to predictions made using the approach described in section \[subsec:Analytic\_WME\].[]{data-label="fig:TIgrowthRate"}](figures/growthRate){width="1\columnwidth"}
An alternative formula for $\gamma_{\rm TI}$ can be obtained using the Wigner–Moyal equation (WME) for the Wigner function $W$ of the fluctuations $\tilde{w}$ (Appendix \[appA\]). This approach is somewhat more accurate because the Hamiltonian is expanded only in $x$ but not in $k_x$. As in section \[subsec:Analytic\_oscillator\], let us assume $U=U_0+\mathcal{C}x^2/2$. Then, $U''=\mathcal{C}$ is constant, $U'''$ vanishes, and the drifton Hamiltonian is simplified down to (Appendix \[appA\]) $$\mathcal{H}=k_{y}U_{0}+\frac{1}{2}k_{y}\mathcal{C}x^{2}+k_{y}(\beta+\mathcal{C})\Real\left(\frac{1}{\bar{k}^{2}}\right),\quad\Gamma=k_{y}(\beta+\mathcal{C}){\rm Im}\left(\frac{1}{\bar{k}^{2}}\right)-D_{\boldsymbol{k}},\label{eq:GammaTI_Hamiltonian}$$ where $\bar{k}^2=1+k_x^2+k_y^2-{\rm i}\delta_0k_y$. Then, the WME (\[eq:Drifton\_WignerMoyal\]) acquires the form $$\frac{\partial W}{\partial t} = k_{y}\mathcal{C}x\,\frac{\partial W}{\partial k_{x}}-V_{{\rm g}}\,\frac{\partial W}{\partial x}+2\Gamma W+\frac{\partial Q}{\partial x},\label{eq:GammaTI_WME}$$ where $$V_{{\rm g}}(k_{x})\doteq\frac{\partial\mathcal{H}}{\partial k_{x}}=k_{y}(\beta+\mathcal{C})\frac{\partial\Real(1/\bar{k}^{2})}{\partial k_{x}}$$ is the drifton group velocity. (Details of drifton dynamics are discussed in Appendix \[appB\].) The value of $Q$ is given by (\[eq:Drifton\_Q\]), but it is not important for our calculations, because we are interested only in the spatial integral of (\[eq:GammaTI\_WME\]). Since $V_{{\rm g}}$ and $\Gamma$ are independent of $x$, integrating (\[eq:GammaTI\_WME\]) over $x$ leads to $$2\gamma_{{\rm TI}}f_{1}=k_{y}\mathcal{C}\,\frac{\partial f_{2}}{\partial k_{x}}+2\Gamma f_{1},\label{eq:GammaTI_xintegration}$$ where we have replaced $\partial_{t}$ with $2\gamma_{{\rm TI}}$ and introduced $$f_{1}(k_{x})\doteq\int W{\rm d}x,\quad f_{2}(k_{x})\doteq\int xW{\rm d}x.$$ The functions are shown in figure \[fig:TIstructure\] for the runaway mode and the trapped mode, respectively. Note that from comparing (\[eq:GammaTI\_xintegration\]) with (\[eq:AnalyticTI\_oscillator\_omega\]), it is seen that $f_1$ is associated with the modified frequency $\bar{\Omega}$, namely, $\Gamma=\bar{\gamma}=\Imag\bar{\Omega}$; meanwhile, $\partial f_2/\partial k_x$ is associated with the additional term in (\[eq:AnalyticTI\_oscillator\_omega\]) that vanishes at $\mathcal{C}=0$.
To obtain $\gamma_{\rm TI}$ from (\[eq:GammaTI\_xintegration\]), one needs to find the relation between $f_1$ and $f_2$. Let us first consider the runaway mode. As shown in figure \[fig:TIstructure\](a), the Wigner function of this mode peaks along $x=x_{{\rm R}}(k_{x})$, which is the runaway trajectory that passes through the saddle point of $\mathcal{H}$ at $x=k_x=0$, and is given by (\[eq:GammaTI\_runaway\_traj\]) below. Therefore, let us adopt $f_{2}\approx x_{{\rm R}}f_{1}$; then, $$\frac{\partial f_{2}}{\partial k_{x}}\approx\frac{\partial x_{{\rm R}}}{\partial k_{x}}f_{1}+x_{{\rm R}}\frac{\partial f_{1}}{\partial k_{x}}.\label{eq:GammaTI_runaway_f2f3}$$ With this assumption, let us evaluate (\[eq:GammaTI\_xintegration\]) at $k_{x}=0$ where $\partial f_{1}/\partial k_{x}=0$ ($f_1$ is even in $k_x$ due to the symmetry of $\hat{H}$; see figure \[fig:TIstructure\]), we find $$\gamma_{{\rm TI}}^{{\rm R}}=\left(\Gamma+\frac{k_{y}\mathcal{C}}{2\eta_{{\rm R}}}\frac{\partial x_{{\rm R}}}{\partial k_{x}}\right)\bigg|_{k_{x}=0}.\label{eq:GammaTI_runaway_GrowthRate}$$ Here, the first term $\Gamma$ is given by (\[eq:GammaTI\_Hamiltonian\]). The second term is negative because $\partial x_{{\rm R}}/\partial k_{x}<0$ (see (\[eq:GammaTI\_runaway\_slope\]) below). The coefficient $\eta_{{\rm R}}>0$ is an empirical factor that compensates for the inaccuracy of (\[eq:GammaTI\_runaway\_f2f3\]). We proceed to determine $x_{{\rm R}}(k_{x})$ and $\eta_{{\rm R}}$. The runaway trajectory $x_{{\rm R}}$ is determined from (\[eq:GammaTI\_Hamiltonian\]) by equating $\mathcal{H}$ to its value at the origin $(x,k_{x})=(0,0)$ and solving $x$ as a function of $k_{x}$. This gives $$x_{{\rm R}}(k_{x})=\pm\sqrt{2\left(1+\frac{\beta}{\mathcal{C}}\right)}
\sqrt{\frac{1+k_{y}^{2}}{(1+k_{y}^{2})^{2}+\delta_{0}^{2}k_{y}^{2}}-\frac{1+k_{y}^{2}+k_{x}^{2}}{(1+k_{y}^{2}+k_{x}^{2})^{2}+\delta_{0}^{2}k_{y}^{2}}},\label{eq:GammaTI_runaway_traj}$$ where the plus sign is for $k_x<0$ and the minus sign is for $k_x>0$. Figure \[fig:TIstructure\](a) demonstrates that this solution indeed correlates well with the actual runaway-mode structure. Also note that $x_{{\rm R}}$ is finite, namely, $$x_{{\rm R}}(k_{x}=\infty)=-\sqrt{2\left(1+\frac{\beta}{\mathcal{C}}\right)\frac{1+k_{y}^{2}}{(1+k_{y}^{2})^{2}+\delta_{0}^{2}k_{y}^{2}}}.\label{eq:GammaTI_xR}$$ From (\[eq:GammaTI\_runaway\_traj\]), we obtain $$\frac{\partial x_{{\rm R}}}{\partial k_{x}}\bigg|_{k_{x}=0}=-\frac{\sqrt{2(1+\beta/\mathcal{C})\left[(1+k_{y}^{2})^{2}-\delta_{0}^{2}k_{y}^{2}\right]}}{(1+k_{y}^{2})^{2}+\delta_{0}^{2}k_{y}^{2}}.\label{eq:GammaTI_runaway_slope}$$ Notably, $\partial x_{{\rm R}}/\partial k_{x}$ becomes zero at $\delta_{0}=|k_{y}+k_{y}^{-1}|$, which corresponds to the transition from runaway to trapped trajectory at the ZF minimum, as shown in figure \[fig:THtraj\].
Now, let us consider the correction factor $\eta_{\rm R}$, which can be formally defined as $$\eta_{{\rm R}}\doteq\left(\frac{\partial x_{{\rm R}}}{\partial k_{x}}\,\frac{f_{1}}{\partial f_{2}/\partial k_{x}}\right)\bigg|_{k_{x}=0}.\label{eq:GammaTI_runaway_eta}$$ We determine $\eta_{{\rm R}}$ numerically from the eigenmode structures obtained in section \[subsec:Linear\_Tertiary\]. It can be shown that if $\hat{D}=0$, then rescaling $t\to {k_y\mathcal{C}t}/{(1+k_y^2)}$, $x\to x\sqrt{1+k_y^2}$, and $k_x\to {k_x}/{\sqrt{1+k_y^2}}$ leaves only two parameters in the WME (\[eq:GammaTI\_WME\]), namely, $\mathcal{C}/\beta$ and $\delta_0k_y/(1+k_y^2)$; hence, $\eta_{{\rm R}}$ mainly depends on these two parameters. Numerically, we see that $\eta_{{\rm R}}$ changes little as $\mathcal{C}/\beta$ varies from zero to unity. Meanwhile, the dependence of $\eta_{{\rm R}}$ on $\delta_{0}k_{y}/(1+k_{y}^{2})$ is shown in figure \[fig:TIgrowthRate\](a), which suggests the following approximation: $$\eta_{{\rm R}}\approx\sqrt{1-\left(\frac{\delta_{0}k_{y}}{1+k_{y}^{2}}\right)^{2}}\,.\label{eq:GammaTI_runaway_fitting}$$ Then, (\[eq:GammaTI\_runaway\_GrowthRate\]) is simplified as $$\gamma_{{\rm TI}}^{{\rm R}}\approx\Gamma|_{k_{x}=0}-\sqrt{\frac{1}{2}\left(1+\frac{\beta}{\mathcal{C}}\right)}\frac{(1+k_{y}^{2})k_{y}\mathcal{C}}{(1+k_{y}^{2})^{2}+\delta_{0}^{2}k_{y}^{2}}.\label{eq:GammaTI_runaway_connection}$$ Remarkably, this formula is identical to (\[eq:Analytic\_gammaR\]) that was obtained in section \[subsec:Analytic\_oscillator\] by drawing an analogy with a quantum harmonic oscillator.
The above approach can also be applied to the trapped mode. Similarly to (\[eq:GammaTI\_runaway\_GrowthRate\]), the trapped-mode growth rate can be expressed as follows: $$\gamma_{{\rm TI}}^{{\rm T}}=\left(\Gamma+\frac{k_{y}\mathcal{C}}{2\eta_{{\rm T}}}\,\frac{\Delta x}{\Delta k_{x}}\right)\bigg|_{k_{x}=0},\label{eq:GammaTI_trapped_GrowthRate}$$ where $$\frac{\Delta x}{\Delta k_{x}}\doteq\frac{\sqrt{2(|\beta/\mathcal{C}|-1)\left[(1+k_{y}^{2})^{2}-\delta_{0}^{2}k_{y}^{2}\right]}}{(1+k_{y}^{2})^{2}+\delta_{0}^{2}k_{y}^{2}},\quad\eta_{{\rm T}}\doteq\left(\frac{\Delta x}{\Delta k_{x}}\,\frac{f_{1}}{\partial f_{2}/\partial k_{x}}\right)\bigg|_{k_{x}=0}.\label{eq:GammaTI_trapped_eta}$$ Here, $\mathcal{C}<0$, and we consider the regime $\beta/\mathcal{C}<-1$. Also, $\Delta x/\Delta k_{x}$ is not the slope of the runaway trajectory but the ratio of the $x$-axis radii and the $k_{x}$-axis radii of the elliptic trapped trajectories near $(x,k_{x})=(0,0)$ in figure \[fig:TIstructure\](c). ($\Delta x/\Delta k_{x}$ becomes zero at $\delta_0=|k_y+k_y^{-1}|$, which corresponds to the transition from a single island to two islands, as shown in Fig. \[fig:THtraj\].) The coefficient $\eta_{\rm T}$ is determined numerically. As shown in figure \[fig:TIgrowthRate\](b), $\eta_{\rm T}$ can be approximated as $$\eta_{\rm T}\approx\sqrt{\left(\frac{1+k_y^2}{\delta_0k_y}\right)^2-1}\label{eq:GammaTI_trapped_fitting}$$ at $\beta/\mathcal{C}<-1$, when the mode is well localized in phase space. In this case, (\[eq:GammaTI\_runaway\_GrowthRate\]) becomes identical to (\[eq:Analytic\_gammaT\]).
These results show that the alternative approach adopted here is in agreement with the one we used in section \[subsec:Analytic\_oscillator\] if we use the fitting formula (\[eq:GammaTI\_runaway\_fitting\]) for $\eta_{\rm R}$ and (\[eq:GammaTI\_trapped\_fitting\]) for $\eta_{\rm T}$. If these factors are calculated numerically instead, then the alternative approach is slightly more accurate, as seen in figure \[fig:TIgrowthRate\](c).
Connection with the Kelvin–Helmholtz instability {#subsec:KHI}
------------------------------------------------
![Numerical solutions of (\[eq:Linear\_Tertiary\_eigen\]) illustrating the relation between the runaway mode and the KH mode at $\beta = 6$, $u = 10$, and various $q_x$ and $\delta_0$. (a) At $\delta_0 = 0$ and $q_x = 1.6$, the unstable mode is the KH mode, which has a global structure as discussed in [@Zhu18a; @Zhu18c]. (b) The KH mode transitions to an “intermediate” mode as $\delta_0$ is increased from $\delta_0=0$ to $\delta_0 = 1.5$ while keeping $q_x = 1.6$ fixed. (c) The corresponding evolution of $\gamma$ with $\delta_0$ at constant $q_x = 1.6$. Blue curves show multiple branches of eigenmodes, but only one branch (the KH mode) is unstable. (d) $\gamma$ as a function of $q_x$ at constant $\delta_0 = 1.5$. As $q_x$ decreases, the intermediate mode analytically continues into the runaway TI mode in figure \[fig:TI\_eigenvalue\]. See the main text for details. []{data-label="fig:KHI"}](figures/KHIplot){width="1\columnwidth"}
The above analysis shows that the TI can be considered as a primary instability modified by ZFs. As seen from figure \[fig:TIgrowthRate\], the growth rate $\gamma_{\rm TI}$ decreases with $|\mathcal{C}|$ in general. Therefore, the TI is very different from the KHI, which develops only in strong ZFs. To study the relation between the TI and the KHI, we numerically solve (\[eq:Linear\_Tertiary\_eigen\]) for various $q_{x}$ and $\delta_{0}$ and explore how the mode structure changes with these parameters. The results are shown in figure \[fig:KHI\].
First, consider figure \[fig:KHI\](a), which shows a global (not localized) KH mode that corresponds to $q_x = 1.6$ and $\delta_0 = 0$. This KH mode has been discussed in [@Zhu18a]; it is global because the ZF is small-scale, specifically, $q_x^2 > 1$. Next, let us increase $\delta_0$ from zero up to $\delta_0=1.5$ while keeping $q_x=1.6$ fixed. Then, the original KH mode transforms into an “intermediate” mode shown in figure \[fig:KHI\](b). It is not a pure KHI, because dissipation ([[i.e., ]{}]{}nonzero $\delta_0$) is now important, but it is not quite the TI either, because $q_x^2$ is large and the mode localization is less pronounced. Our theory does not apply to such modes, but we have calculated the growth rate numerically as a function of $\delta_0$, as shown in figure \[fig:KHI\](c). Finally, with $\delta_0 = 1.5$ fixed, let us reduce $q_x$. The mode localization improves and the instability rates goes down at first, as seen in figure \[fig:KHI\](d). But eventually, when $q_x$ has become small enough ($q_x \sim 0.6$), the mode transforms into the runaway mode that we introduced earlier (figure \[fig:TIstructure\]) and our theory becomes applicable.
This shows that in principle, the KH mode can be continuously transformed into the runaway mode. However, the KHI and TI are fundamentally different in physical mechanisms, because the TI is due to dissipation and $\gamma_{\rm TI}^{\rm R}$ is determined by $\delta_0$, while the KHI requires a strongly sheared flow and has $\gamma_{\rm KHI} \sim k_y u$. Since typical large-scale ZFs seen simulations have $q_x^2 \ll 1$, the TI is more relevant to them than the KHI.
Dimits shift {#sec:DSPrediction}
============
![The Dimits shift obtained by simulating the mTHE (\[eq:Equations\_mTHE\]) numerically (colored markers) versus analytic theory (black curves) for two different choices of the damping operator: (a) $\hat{D}=1-0.01\nabla^2$ and (b) $\hat{D}=0.3|k_y|+10^{-4}\nabla^4$. Green circles indicate the Dimits regime, in which the system saturates in a state with ZFs and no turbulence. Red crosses correspond to the situation where the system remains in a turbulent state indefinitely. Dot-dashed curve: the linear threshold of the primary instability. Solid curve: our prediction of the Dimits shift, $\Delta_{\rm DS}$ (\[eq:DSpredction\_DeltaDS\]), with (a) $\varrho=0.05$ and (b) $\varrho=0.025$. Ideally, the curve $\beta_{\rm c}=\beta_{\rm lin}+\Delta_{\rm DS}$ is supposed to separate regions with green circles and with red crosses. Dashed curve (denoted $\beta_{\rm ZF}^*$): the prediction of $\beta_{\rm c}$ from [@St-Onge17].[]{data-label="fig:DSprediction"}](figures/DSprediction){width="1\columnwidth"}
As seen from the previous sections, the TI is nothing but the primary instability modified by nonzero ZF curvature $\mathcal{C}$. The nonzero $\mathcal{C}$ modifies the growth rate by $\Delta\gamma = \gamma_{\rm TI}(\mathcal{C}) - \gamma_{\rm TI}(0)$. We take $\gamma_{\rm TI}=\gamma_{\rm TI}^{\rm R}$ (\[eq:Analytic\_gammaR\]), since the runaway mode usually has the largest growth rate in the mTHE model. Letting $\gamma_{\rm TI}(\mathcal{C})=0$, we obtain an implicit expression for the critical value of $\beta$, denoted $\beta_{\rm c}$: $$\beta_{\rm c}=\frac{\beta_{\rm lin}}{(1+\varrho)-\delta_0^{-1}(k_y+k_y^{-1})\sqrt{(\varrho+\varrho^2)/2}}, \quad\beta_{\rm lin}\doteq\frac{D_0[(1+k_y^2)^2+\delta_0^2k_y^2]}{\delta_0k_y^2}.$$ Here, $\varrho\doteq\mathcal{C}/\beta_{\rm c}$, and $\beta_{\rm lin}$ is the linear threshold of the primary instability, which is obtained by letting $\gamma_{\boldsymbol{k}}=0$ (see (\[eq:Linear\_Primary\_Omegak\])). Due to nonzero $\mathcal{C}$, the value of $\beta_{\rm c}$ differs from $\beta_{\rm lin}$ by a finite value $\Delta_{\rm DS}$, which represents the Dimits shift: $$\Delta_{\rm DS}=\beta_{\rm c}-\beta_{\rm lin}.\label{eq:DSpredction_DeltaDS}$$ Note that the chosen formula (\[eq:Analytic\_gammaR\]) for $\gamma_{\rm TI}^{\rm R}$ is not as accurate as its counterpart (\[eq:GammaTI\_runaway\_GrowthRate\]); nevertheless, we choose (\[eq:Analytic\_gammaR\]) because it does not involve the fitting parameter $\eta_{\rm R}$.
In section \[subsec:Linear\_Secondary\], we discussed the evolution of the secondary instability, where we found that the system enters a fully nonlinear stage when the ZF amplitude $u$ reaches $q_x\varphi_{\rm c}\sim\beta$ (see (\[eq:Linear\_Secondary\_phic\])). Therefore, we adopt the assumption that $\mathcal{C}\sim q_x^2u$ is proportional to $\beta$; hence, $\varrho$ is assumed constant and is treated as a fitting parameter. Then, for each value of $\delta_0$, $\Delta_{\rm DS}$ can be obtained by minimizing it over $k_y$. The results are in good agreement with numerical simulation of the mTHE (figure \[fig:DSprediction\]). A similar figure can be found in figure 7 of [@St-Onge17], where simulation results are compared with a different theory.
Note that the assumption of constant $\varrho$ is not a rigorous result but only a rough approximation. In section \[subsec:Linear\_Secondary\] we showed that the ZF with $q_x=2\pi/L_x$ grows fastest as the secondary instability. However, at the fully nonlinear stage, the ZF shape is changed by additional DW–ZF interactions, and $q_x$ is no longer determined by $L_x$. As a result, the Dimits shift is insensitive to $L_x$ as long as $L_x$ is large enough. From numerical simulations, we found that the ZF shape differs from one realization to another, but in general, $q_x$ (and hence $\varrho$) is larger at smaller $\delta_0$. In fact, we also tried $\varrho=\varrho(\delta_0)$ such that $\varrho$ gets larger at smaller $\delta_0$, but the improvements in predicting the Dimits shift were not significantly compared to the simpler assumption of constant $\varrho$.
For comparison, the prediction of $\beta_{\rm c}$ made by [@St-Onge17] is also plotted in figure \[fig:DSprediction\], where it is denoted $\beta_{\rm ZF}^*$. As a reminder, @St-Onge17 obtained $\beta_{\rm ZF}^*$ from a sufficient condition for the ZF to be stable based on the 4MT approximation and considered $\beta_{\rm ZF}^*$ as a “heuristic calculation” of the Dimits shift. Since the 4MT method misses essential features of TI modes such as mode localization, @St-Onge17’s model is less accurate than ours. Besides, the direct relation between @St-Onge17’s criterion and the Dimits shift is only an assumption. In contrast, our calculation provides an explicit formula for the Dimits shift, namely, (\[eq:DSpredction\_DeltaDS\]). Note that our (\[eq:DSpredction\_DeltaDS\]) predicts infinite $\beta_{\rm c}$ at $\delta_0= |k_y+k_y^{-1}|\sqrt{\varrho/2}$, [[i.e., ]{}]{}small $\delta_0$ (assuming $\varrho\ll 1$), which is in agreement with simulation results. In contrast, $\beta_{\rm ZF}^*$ is still finite in this region. Also, @St-Onge17’s criterion does not have a solution at $\delta_0>|k_y+k_y^{-1}|$, suggesting zero $\Delta_{\rm DS}$; however, our theory gives nonzero $\Delta_{\rm DS}$ in this region, which is in agreement with numerical simulations.
Conclusions
===========
In conclusion, this paper expands on our recent theory [@Zhu20], where the TI and the Dimits shift were studied within reduced models of drift-wave turbulence. Here, we elaborate on a specific limit of that theory where turbulence is governed by the scalar mTHE model and the problem becomes analytically tractable. We show that assuming a sufficient scale separation between ZFs and DWs, TI modes are localized at extrema of the ZF velocity $U(x)$, where $x$ is the radial coordinate. By approximating $U(x)$ with a parabola, we analytically derive the TI growth rate, $\gamma_{\rm TI}$, using two different approaches: (i) by drawing an analogy between TI modes and quantum harmonic oscillators and (ii) by using the WME. Our theory shows that the TI is essentially a primary DW instability modified by the ZF curvature $U''$ near extream of $U$. In particular, the WME allows us to understand how the local $U''$ modifies the mode structure and reduces the TI growth rate; it also shows that the TI is *not* the KHI. Then, depending on $U''$, the TI can be suppressed, in which case ZFs are strong enough to suppress turbulence (Dimits regime), or unleashed, so ZFs are unstable and turbulence develops. This understanding is different from the traditional paradigm [@Biglari90], where turbulence is controlled by the flow shear $U'$. Finally, by letting $\gamma_{\rm TI}=0$, we obtain an analytic prediction of the Dimits shift, which agrees with our numerical simulations of the mTHE.
The authors thank W. D. Dorland, N. R. Mandell, D. A. St-Onge, P. G. Ivanov, A. A. Schekochihin for helpful discussions, and the anonymous reviewers for providing numerous valuable comments. This work was supported by the US DOE through Contract No. DE-AC02-09CH11466. Digital data can also be found in DataSpace of Princeton University (<https://dataspace.princeton.edu/jspui/handle/88435/dsp015425kd34n>).
Wigner–Moyal equation for the mTHE model {#appA}
========================================
Here, we present the WME for the mTHE model following the same method that was originally used by [@Ruiz16] for the modified Hasegawa–Mima model. We start with the linearized DW dynamics described by (\[eq:Linear\_Tertiary\_linearDW\]). Because the flow velocity $U(x,t)$ does not depend on $y$, we assume that the wave is monochromatic in $y$, namely, $$\tilde{w}=w(x,t){\rm e}^{{\rm i}k_{y}y}.$$ Then, equation (\[eq:Linear\_Tertiary\_linearDW\]) can be written symbolically as $${\rm i}\partial_t w=\hat{H}w,\quad\hat{H}(x,\hat{k}_{x},t)=k_{y}\hat{U}+k_{y}(\beta+\hat{U}'')\hat{\bar{k}}^{-2}-{\rm i}\hat{D},\label{eq:Drifton_Shrodinger}$$ where $$\hat{U}=U(\hat{x},t),\quad\hat{k}_x=-{\rm i}\,{\rm d}/{\rm d}x,\quad\hat{\bar{k}}^{2}=1+k_{y}^{2}+\hat{k}_x^2-{\rm i}\delta_{0}k_{y}.$$ This can be considered as a linear Schrödinger equation with an non-Hermitian Hamiltonian. From here, we derive the following WME using the same phase-space formulation that is used in quantum mechanics [@Moyal49]: $$\partial_{t}W(x,k_{x},t)=\{\!\{ \mathcal{H},W\}\!\}
+\left[\left[\Gamma,W\right]\right].\label{eq:Drifton_WignerMoyal}$$ Here, $W$ is the Wigner function defined as $$W(x,k_{x},t)\doteq\int{\rm d}s\,{\rm e}^{-{\rm i}k_x s}w^*(x-s/2,t)w(x+s/2,t)\label{eq:appA_Wigner}$$ ($^*$ denotes complex conjugate), and $\mathcal{H}$ and $\Gamma$ are the Hermitian and anti-Hermitian parts of the Hamiltonian:
=k\_y U+()+(U”|[k]{}\^[-2]{}+|[k]{}\^[\*-2]{}U”),\
=[Im]{}()+(U”|[k]{}\^[-2]{}-|[k]{}\^[\*-2]{}U”)-D\_,
where $\bar{k}^2\doteq 1+k_y^2+k_x^2-{\rm i}\delta_0 k_y$. The symbol $\star$ is the Moyal star product: $$A\star B\doteq A\exp({\rm i}\hat{\mathcal{L}}/2) B,\quad \hat{\mathcal{L}}\doteq \frac{\overleftarrow{\partial}}{\partial x}\,\frac{\overrightarrow{\partial}}{\partial k_x}-\frac{\overleftarrow{\partial}}{\partial k_x}\,\frac{\overrightarrow{\partial}}{\partial x},$$ where the overhead arrows in $\hat{\mathcal{L}}$ indicate the directions in which the derivatives act on, and $\lbrace\!\lbrace.,.\rbrace\!\rbrace$ and $\left[\left[.,.\right]\right]$ are the Moyal brackets: $$\begin{aligned}
\lbrace\!\lbrace A,B \rbrace\!\rbrace\doteq -{\rm i}(A\star B - B\star A),\quad [[A,B]]\doteq A\star B + B\star A.\end{aligned}$$ Equation (\[eq:Drifton\_WignerMoyal\]) is mathematically equivalent to (\[eq:Drifton\_Shrodinger\]), and the corresponding equation for TI eigenmodes is obtained by replacing $\partial_{t}W$ with $2\gamma_{{\rm TI}}W$.
If we adopt the parabolic approximation of the ZF velocity, $U=U_0+\mathcal{C}x^2/2$, then $U''=\mathcal{C}$ is constant and $$\mathcal{H}=k_{y}U_{0}+\frac{1}{2}k_{y}\mathcal{C}x^{2}+k_{y}(\beta+\mathcal{C})\Real\left(\frac{1}{\bar{k}^{2}}\right),\quad\Gamma=k_{y}(\beta+\mathcal{C}){\rm Im}\left(\frac{1}{\bar{k}^{2}}\right)-D_{\boldsymbol{k}}.\label{eq:appA_Drifton_Hamiltonian}$$ Then, the $x$-dependent part and the $k_{x}$-dependent part in $\mathcal{H}$ are separated, and $\Gamma$ is independent of $x$. This greatly simplifies the WME (\[eq:Drifton\_WignerMoyal\]), such that it acquires the form (\[eq:GammaTI\_WME\]), which we repeat here: $$\frac{\partial W}{\partial t} = k_{y}\mathcal{C}x\,\frac{\partial W}{\partial k_{x}}-V_{{\rm g}}\,\frac{\partial W}{\partial x}+2\Gamma W+\frac{\partial Q}{\partial x}.$$ Here, $Q$ is given by a lengthy expression, $$Q=\sum_{n=1}^{\infty}\frac{(-1)^{n+1}}{(2n+1)!\times2^{2n}}\frac{\partial^{2n+1}f}{\partial k_{x}^{2n+1}}\frac{\partial^{2n}W}{\partial x^{2n}}+\sum_{n=1}^{\infty}\frac{(-1)^{n}}{(2n)!\times2^{2n-1}}\frac{\partial^{2n}\Gamma}{\partial k_{x}^{2n}}\frac{\partial^{2n-1}W}{\partial x^{2n-1}},\label{eq:Drifton_Q}$$ with $f(k_x)\doteq k_y U_0 +k_y(\beta+\mathcal{C})\Real(\bar{k}^{-2})$. However, $\partial_x Q$ does not contribute to the integral of (\[eq:Drifton\_Q\]) over $x$ that we are interested in. Therefore, the WME provides a transparent description of the TI under the assumption of parabolic $U$.
Wave-kinetic equation and phase-space trajectories {#appB}
==================================================
Here, we briefly overview the derivation and the structure of drifton phase-space trajectories from the wave-kinetic equation (WKE). This discussion helps clarify the terms “runaway mode” and “trapped mode” used in the main text. It also illustrates how the TI-mode structures change with the parameter $\delta_0$.
The WKE is an approximation of the WME in the limit when, roughly speaking, the characteristic ZF scales are much larger than the typical DW wavelength. Since a parabolic $U$ does not have a well-defined spatial scale, we switch to the sinusoidal ZF velocity, $$U=u\cos q_x x,$$ in which case the ZF scale is characterized by $q_x^{-1}$. For large enough ZF scale, the WME reduces to the WKE: $$\frac{\partial W}{\partial t}=\frac{\partial \mathcal{H}}{\partial x}\frac{\partial W}{\partial k_x}-\frac{\partial \mathcal{H}}{\partial k_x}\frac{\partial W}{\partial x}+2\Gamma W,\label{eq:Drifton_iWKE}$$ where $$\mathcal{H}= k_{y}\left[1-\Real\left(\frac{q_x^2}{\bar{k}^{2}}\right)\right]u\cos q_x x+k_{y}\beta\Real\left(\frac{1}{\bar{k}^{2}}\right),\label{eq:appB_Drifton_Hamiltonian}$$ while $\Gamma$ is not important for the following discussions. The form of the WKE (\[eq:Drifton\_iWKE\]) indicates that $W$ can be considered as the distribution function of DW quanta, or driftons, in the $(x,k_{x})$ phase space. The driftons trajectories are governed by Hamilton’s equations, $$\frac{{\rm d}x}{{\rm d}t}=\frac{\partial\mathcal{H}}{\partial k_x},\quad \frac{{\rm d}k_x}{{\rm d}t}=-\frac{\partial\mathcal{H}}{\partial x},$$ where $\mathcal{H}$ serves as the Hamiltonian. However, unlike true particles, driftons are not conserved. Instead, $\Gamma$ determines the rate at which $W$ evolves along the ray trajectories.
If ZFs are stationary, as is the case for our calculation of the TI, then $\mathcal{H}$ is independent of time and driftons move along curves that satisfy $\mathcal{H}(x, k_x) = \mathcal{E}$, where $\mathcal{E}$ is a constant. In [@Zhu18b], we systematically studied these trajectories for the modified Hasegawa–Mima system ($\hat{\delta}=0$), and three types of trajectories have been identified, which we called passing, trapped, and runaway trajectories. Although the mTHE has nonzero $\hat{\delta}$, it corresponds to similar drifton dynamics unless $\hat{\delta}$ is too large. Note that $\mathcal{H}$ depends on $\Real(1/\bar{k}^{2})$, which is $$\Real\left(\frac{1}{\bar{k}^{2}}\right)=\frac{1+k_{x}^{2}+k_{y}^{2}}{(1+k_{x}^{2}+k_{y}^{2})^{2}+\delta_{0}^{2}k_{y}^{2}}.$$ Therefore, $\Real(1/\bar{k}^{2})$ is a monotonically decreasing function of $k_{x}^{2}$ if $\delta_{0}^{2}k_{y}^{2}<(1+k_{y}^{2})^{2}$, [[i.e., ]{}]{}when $\delta_{0}<|k_{y}+k_{y}^{-1}|$. However, $\Real(1/\bar{k}^{2})$ has a maximum at nonzero $k_{x}^{2}$ if $\delta_{0}>|k_{y}+k_{y}^{-1}|\geq2$. In the following, we discuss the two situations separately.
![Contour plots of the drifton Hamiltonian $\mathcal{H}$ (\[eq:appB\_Drifton\_Hamiltonian\]) at (a) small $\delta_{0}$ and (b) large $\delta_{0}$; the color marks the corresponding value of $\mathcal{H}$. The parameters are $\beta=k_{y}=1$, $q_x=0.4$; also, (a) $\delta_{0}=1.5$ and $u=0.5$, and (b) $\delta_{0}=3$ and $u=0.1$. At small $\delta_{0}$, trapped trajectories are found near the ZF maximum $x=0$ and runaway trajectories are found near the ZF minimum $x=-\upi/q_{x}$. At large $\delta_{0}$, two separate trapped islands form at $x=0$ and trapped trajectory replace runaway trajectories at $x=-\upi/q_x$.\[fig:THtraj\]](figures/THtraj){width="1\columnwidth"}
First, consider $\delta_{0}<|k_{y}+k_{y}^{-1}|$. Then, letting $\mathcal{H}=\mathcal{E}$ leads to $$k_{x}^{2}(x,\mathcal{E})=(1+k_{y}^{2})\frac{\mathcal{H}^{0}(x,\mathcal{E})-\mathcal{E}}{\mathcal{E}-\mathcal{H}^{\infty}(x)},\label{eq:Drifton_px(x)}$$ where $$\begin{aligned}
&\mathcal{H}^{\infty}(x)\doteq k_{y}u\cos q_{x}x,
\\
&\mathcal{H}^{0}(x,\mathcal{E})=k_{y}u\cos q_{x}x+\frac{k_{y}}{1+k_y^2}(\beta-q_{x}^{2}u\cos q_{x}x)
\left(1-\frac{\lambda}{2}\right),\end{aligned}$$ and $$\lambda=\lambda_{\pm}(x,\mathcal{E})\doteq1\pm\sqrt{1-\frac{4\delta_{0}^{2}(\mathcal{H}^{0}-\mathcal{E})^2}{(\beta-q_{x}^{2}u\cos q_{x}x)^{2}}}.$$ This shows that at given $x$, there are two solutions for $k_{x}^{2}$ depending on whether $\lambda=\lambda_{+}$ or $\lambda=\lambda_{-}$. However, it turns out that $\lambda=\lambda_{+}$ corresponds to negative $k_{x}^{2}$ and hence can be ignored, which is consistent with the fact that $\mathcal{H}$ is a monotonic function of $k_{x}^{2}$ at small $\delta_{0}$. Therefore, only $\lambda=\lambda_{-}$ is possible, and one could use (\[eq:Drifton\_px(x)\]) to identify passing, trapped, and runaway trajectories as in [@Zhu18b]. At very small $u$, ZFs do not matter, so all trajectories are passing. However, when $u$ exceeds a certain critical amplitude $u_{\rm c}$, passing trajectories disappear, which indicates that DWs are strongly affected by ZFs in this case. The critical ZF amplitude is obtained by letting $$\max\mathcal{H}^{\infty}=\min\mathcal{H}^{0}.$$ This leads to $$u_{{\rm c}}=\frac{\beta}{2(1+k_{y}^{2})-q_{x}^{2}}\left[1-\frac{\lambda_{0}}{2}\left(1+\frac{q_{x}^{2}u_{\rm c}}{\beta}\right)\right],\label{eq:appB_U_critical}$$ where $$\lambda_{0}\doteq\lambda_{-}(x=0,\mathcal{E}=k_{y}u_{\rm c}).$$ Therefore, $u_{{\rm c}}$ is smaller than that in the modified Haseagawa–Mima system, where $\lambda_{0}=0$ [@Zhu18b]. Phase-space trajectories at $u>u_{\rm c}$ are shown in figure \[fig:THtraj\](a).
At $\delta_{0}>|k_{y}+k_{y}^{-1}|\geq2$, $\lambda=\lambda_{-}$ still gives passing and runaway trajectories as before. However, because $\mathcal{H}$ becomes non-monotonic with respect to $k_{x}^{2}$, the other solution $\lambda=\lambda_{+}$ can also give positive $k_{x}^{2}$ for some values of $\mathcal{E}$. As a result, runaway trajectories are replaced with trapped trajectories near the ZF minimum, and two separate trapped islands are formed near the ZF maximum. The corresponding phase-space trajectories are shown in figure \[fig:THtraj\](b).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove that the lower bounds for Betti numbers of the rack, quandle and degeneracy cohomology given in [@cjks] are in fact equalities. We compute as well the Betti numbers of the twisted cohomology introduced in [@ces]. We also give a group-theoretical interpretation of the second cohomology group for racks.'
address:
- 'Pavel Etingof: MIT Math. Dept. Of. 2-176 77 Mass. Ave. 02139, Cambridge, MA, USA.'
- 'Matías Graña: MIT Math. Dept. Of. 2-155 77 Mass. Ave. 02139, Cambridge, MA, USA. Permanent: Depto de Matemática - FCEyN Universidad de Buenos Aires Pab. I - Ciudad Universitaria 1428 - Buenos Aires - Argentina'
author:
- 'P. Etingof & M. Graña'
title: On rack cohomology
---
Introduction
============
A *rack* is a pair $(X,{\triangleright})$ where $X$ is a set and ${\triangleright}:X\times X\to X$ is a binary operation such that:
1. The map $\phi_x:X\to X, \quad \phi_x (y) = x{\triangleright}y$, is a bijection for all $x\in X$, and
2. $x{\triangleright}(y{\triangleright}z)=(x{\triangleright}y){\triangleright}(x{\triangleright}z)\ \forall x,y,z\in X$.
It is easy to show that $(X,{\triangleright})$ is a rack if and only if the map $R:X^2\to X^2$ given by $R(x,y)=(x,x{\triangleright}y)$ is an invertible solution of the quantum Yang-Baxter equation $R^{12}R^{13}R^{23}=
R^{23}R^{13}R^{12}$.
Racks have been studied by knot theorists in order to construct invariants of knots and links and their higher dimensional analogs (see [@cssurvey] and references therein). A basic example of a rack is a group with the operation $x{\triangleright}y=xyx^{-1}$ (or, more generally, a conjugation invariant subset of a group).
Several years ago, Fenn, Rourke and Sanderson [@frs] proposed a cohomology theory of racks. Namely, for each rack $X$ and an abelian group $A$, they defined cohomology groups $H^n(X,A)$. This cohomology is useful for knot theory and also, as was recently found, for the theory of pointed Hopf algebras [@g]. There have been a number of results about this cohomology [@ln; @m; @cjks], in particular it was shown in [@cjks] that for a finite rack $X$ and a field $k$ of characteristic zero, the Betti numbers $\dim H^n(X,k)$ are bounded below by $|X/\sim|^n$, where $\sim$ is the equivalence relation on $X$ generated by the relation $z{\triangleright}y\sim y$ $\forall y,z\in X$. The equality was anticipated in [@cjks], and proved in a number of cases [@ln; @m], but not in general.
The main result of this paper implies that the Betti numbers of a finite rack are always equal to $|X/\sim|^n$. The proof is based on a group-theoretical approach to racks, originating from the works [@lyz], [@s] on set-theoretical solutions of the quantum Yang-Baxter equation. Namely, we use the structure group $G_X$ and the reduced structure group $G_X^0$ of a rack $X$ considered in [@lyz; @s].
We also give a group-theoretic interpretation of the second cohomology group $H^2(X,A)$, which is used in the theory of Hopf algebras. Namely, we show that this group is isomorphic to the group cohomology $H^1(G_X,{\operatorname{Fun}}(X,A))$, where ${\operatorname{Fun}}(X,A)$ is the group of functions from $X$ to $A$. This is a relatively explicit description, since it is shown by Soloviev [@s] that for a finite rack $X$, the group $G_X$ is a central extension of the finite group $G_X^0$ by a finitely generated abelian group. Thus the cohomology of $G_X$ can be studied using the Hochschild-Serre sequence.
[**Acknowledgments.**]{} The work of P.E. was supported by the NSF grant DMS-9988796. The work of M.G. was supported by Conicet.
Definitions and notation
========================
The *structure group* of a rack $X$ is the group $G_X$ with generators being the elements of $X$ and relations $x\cdot y=(x{\triangleright}y)\cdot x\quad \forall x,y\in X$. [^1]
The group $G_X$ acts on $X$ from the left by ${\triangleright}$. Consider the quotient $G_X^0$ of $G_X$ by the kernel of this action, i.e. the group of trasformations of $X$ generated by $x{\triangleright}$. This group is called the reduced structure group of $X$.
The groups $G_X,G_X^0$ were studied by Soloviev [@s] (we note that in his work, racks are called “derived solutions"). In particular, he showed that the category of racks is equivalent to the category of quadruples $(G,X,\rho,\pi)$, where $G$ is a group, $X$ a set, $\rho: G\times X\to X$ a left action, and $\pi: X\to G$ an equivariant mapping (where $G$ acts on itself by conjugation), such that $\pi(X)$ generates $G$ and the $G$-action on $X$ is faithful. Namely, the quadruple corresponding to $X$ is simply $(G_X^0,X,\rho,\pi)$, where $\rho$ and $\pi$ are obvious.
Now let us define rack cohomology. Let $X$ be a rack. Let $G_X$ be its structure group. Let $M$ be a right $G_X$-module. We define a cochain complex $(C^{\bullet}(X,M),d)$, where $C^n(X,M)={\operatorname{Fun}}(X^n,M)$, $n\ge 0$, with differential $$df(x_1,\ldots,x_{n+1})
=\sum_{i=1}^{n+1}(-1)^{i-1}
\Big(f(x_1,\ldots,x_{i-1},x_{i+1},\ldots,x_{n+1})
-f(x_1,\ldots,x_{i-1},x_i{\triangleright}x_{i+1},
\ldots,x_i{\triangleright}x_{n+1})\cdot x_i\Big)$$ (Here $X^0$ is a set of one element, and ${\operatorname{Fun}}(Y,Z)$ is the set of functions from $Y$ to $Z$ for any sets $Y,Z$).
The cohomology of $C^\bullet(X,M)$ is called the rack cohomology of $X$ with coefficients in $M$.
This includes the ordinary rack cohomology with coefficients in an abelain group $A$, introduced in [@frs] (this corresponds to taking $M=A$ with the trivial action of $G_X$), as well as the twisted rack cohomology introduced in [@ces] (in this case one needs to take a ${\mathbb Z}[T,T^{-1}]$ module $M$, and define a right action of $G_X$ on it by $vx=Tv$, $x\in X$).
One can also define the dual notion of rack homology. As usual, it is completely analogous to cohomology, so we will not consider it.
In [@ag] there is a more general definition of cohomology, with coefficients in objects of a wider category than that of $G_X$-modules. When restricted to $G_X$-modules, the definition there takes as differential the map $d'$, defined by $$\begin{aligned}
d'f(x_1,\ldots,x_{n+1})
&=\sum_{i=1}^{n+1}(-1)^{i-1}\Big(f(x_1,\ldots,x_{i-1},x_{i+1},\ldots,x_{n+1})
(x_1{\triangleright}(x_2{\triangleright}(\cdots x_i)))^{-1} \\
&\hspace*{4cm} -f(x_1,\ldots,x_{i-1},x_i{\triangleright}x_{i+1},\ldots,x_i{\triangleright}x_{n+1})\Big)\end{aligned}$$ This complex is isomorphic to the one we consider here, by means of the map $$T:(C^{\bullet}(X,M),d) \to (C^{\bullet}(X,M),d'),$$ defined by $(Tf)(x_1,\ldots,x_n)=f(x_1,\ldots,x_n)(x_1\cdots x_n)^{-1}$.
The structure of rack cohomology
================================
Let $M$ be a right $G_X$-module. Then $C^n(X,M)={\operatorname{Fun}}(X^n,M)$ is also a right $G_X$-module, with the action defined on the generators by $$(f\cdot y)(x_1,\ldots,x_n)=
f(y{\triangleright}x_1,\ldots,y{\triangleright}x_n)\cdot y.$$
\[gr-act-in-H\]
1. The coboundary operator $d:C^n(X,M)\to C^{n+1}(X,M)$ is a map of $G_X$-modules. In particular, there is a natural right action of $G_X$ on the groups of cocycles $Z^n(X,M)$, coboundaries $B^n(X,M)$, and cohomology $H^n(X,M)$.
2. $H^n(X,M)$ is a trivial $G_X$-module.
<!-- -->
1. Straightforward.
2. Let $f\in Z^n(X,M)$ and consider $f_y\in C^{n-1}(X,M)$, defined by the formula $$f_y(x_2,\ldots,x_n)=f(y,x_2,\ldots,x_n).$$ Notice that $$\label{eq:dfy}
d(f_y)(x_1,\ldots,x_n)
=(f-f\cdot y)(x_1,\ldots,x_n)-(df)(y,x_1,\ldots,x_n)
=(f-f\cdot y)(x_1,\ldots,x_n).$$ Then $f\cdot y=f$ in $H^n(X,M)$.
The action $f\cdot y$ and the assignments $f\mapsto f_y$, as well as , appear in [@ln].
By Lemma \[gr-act-in-H\] we can consider the subcomplex ${C_{\textup{inv}}}^{\bullet}(X,M)=C^{\bullet}(X,M)^{G_X}$. We define the *invariant rack cohomology* ${H_{\textup{inv}}}^{\bullet}(X,M)=H^{\bullet}({C_{\textup{inv}}}^{\bullet}(X,M))$. Clearly, we have a natural map $$\xi:{H_{\textup{inv}}}^{\bullet}(X,M)\to H^{\bullet}(X,M),$$ induced by the inclusion of complexes.
\[rm:fy\] If $f\in{Z_{\textup{inv}}}^n(X,M)$, by the proof of Lemma \[gr-act-in-H\] part 2, it is clear that $f_y\in Z^{n-1}(X,M)$ $\forall y\in X$.
For $M,N$ right $G_X$-modules, consider the natural multiplication map $$C^a(X,M)\times C^b(X,N)\to C^{a+b}(X,M\otimes N).$$ This map will be denoted by $f,g\to f\otimes g$.
\[lm:dfg\] Suppose that $A$ is a trivial $G_X$-module. Then for any $f\in C^i(X,A)$, $g\in {C_{\textup{inv}}}^j(X,N)$, one has $$d(f\otimes g)=df\otimes g+(-1)^if\otimes dg.$$
The proof is straightforward. We note that the statement becomes false if $A$ is nontrivial as a $G_X$-module or $g$ is not invariant.
Lemma \[lm:dfg\] shows that if $f\in Z^i(X,A)$ and $g\in {Z_{\textup{inv}}}^j(X,N)$ then $f\otimes g\in Z^{i+j}(X,A\otimes N)$. Furthermore, by the same Lemma, the cohomology class of $f\otimes g$ depends only of the cohomology classes of $f$ and $g$. Thus, we have a product $$H^{\bullet}(X,A)\times {H_{\textup{inv}}}^{\bullet}(X,N)\to H^{\bullet}(X,A\otimes N).$$ In particular, if $R$ is a (unital) ring with the trivial $G_X$-action, then ${H_{\textup{inv}}}^\bullet(X,R)$ is a graded algebra, and for any left $R$-module $M$ with a compatible $G_X$-action, ${H_{\textup{inv}}}^\bullet(X,M)$ is a graded left ${H_{\textup{inv}}}^\bullet(X,R)$-module.
Cohomology of finite racks
==========================
In this section we will assume that $X$ is a finite rack.
Let $M$ be a right $G_X$-module, such that the kernel $K$ of the action of $G_X$ on $M$ has finite index. Let $L$ be the intersection of $K$ with the kernel $\Gamma$ of the action of $G_X$ on $X$, and let $G=G_X/L$ (notice that $G$ is finite). Assume that the multiplication by $|G|$ is an isomorphism $M\to M$.
\[lm:tc\]Under these conditions the map $\xi:{H_{\textup{inv}}}^{\bullet}(X,M)\to H^{\bullet}(X,M)$ is an isomorphism.
The complex $C^\bullet(X,M)$ is a complex of $G$-modules. On each term of this complex we have a projector given by $P=\frac 1{|G|}\sum_{g\in G}g$, which projects to $G_X$-invariants. This projector commutes with the differential, so the complex $C^\bullet(X,M)$ is representable as a direct sum of complexes: $$C^{\bullet}(X,M)={C_{\textup{inv}}}^\bullet(X,M)\oplus C^\bullet(X,M)(1-P).$$ By Lemma \[gr-act-in-H\], the second summand is acyclic: indeed, any cohomology class in it satisfies $cP=0$, while the lemma says that $cP=c$, hence $c=0$. This implies the desired statement.
In particular, for any ring $R$ with trivial $G_X$-action, such that $N=|G_X^0|$ is invertible in $R$ (for example, $R={\mathbb Z}[1/N]$ or $R={\mathbb Q}$), the cohomology $H^\bullet(X,R)$ is an algebra, and if $M$ is an $R$-module with a compatible $G_X$ action then $H^\bullet(X,M)$ is a left module over this algebra.
Let ${\operatorname{Orb}}(X)=X/G_X$ be the set of $G_X$-orbits on $X$, and $m=|{\operatorname{Orb}}(X)|$. The main result in this section is
\[th1\] Under the conditions of Lemma \[lm:tc\], we have $$H^{\bullet}(X,R)\simeq T_R^{\bullet}(H^1(X,R))
\simeq T_R^{\bullet}({\operatorname{Fun}}({\operatorname{Orb}}(X),R))\simeq{\operatorname{Fun}}({\operatorname{Orb}}(X)^\bullet,R)$$ as an algebra (where $T_R^\bullet(B)$ denotes the tensor algebra of an $R$-bimodule $B$), and if $M$ is an $R$-module with a compatible $G_X$ action then $$H^{\bullet}(X,M)\simeq T_R^{\bullet}(H^1(X,R))\otimes_R M^{G_X}
\simeq T_R^{\bullet}({\operatorname{Fun}}({\operatorname{Orb}}(X),R))\otimes_R M^{G_X}
\simeq{\operatorname{Fun}}({\operatorname{Orb}}(X)^\bullet,M^{G_X})$$ as a left module over the algebra $H^{\bullet}(X,R)$.
Before proving the theorem, we will derive a corollary.
\[cbn\] The Betti numbers of $X$ are $\dim H^i(X,{\mathbb Q})=m^i$. Furthermore, the only primes which can appear in the torsion of $H^{\bullet}(X,{\mathbb Z})$ are those dividing $N$.
The first assertion is clear taking $R={\mathbb Q}$. For the second one, take $R={\mathbb Z}[\frac 1N]$ (or $R={\mathbb Z}/p$, $p\nmid N$) and apply the universal coefficent theorem.
This, together with the lower bounds for the Betti numbers of the quandle and degeneracy cohomology in [@cjks] and the splitting result of [@ln], implies that those lower bounds are in fact equalities.
(of Theorem \[th1\]). Since $M^{G_X}=H^0(X,M)$, for any $M$ we have an obvious multiplication mapping $\mu: T^\bullet(H^1(X,R))\otimes_R M^{G_X}\to H^\bullet(X,M)$, which is compatible with the algebra and module structures. Thus, all we have to show is that $\mu$ is an isomorphism.
Let us first show that $\mu$ is injective. This is in fact the lower bound of [@cjks], but we will give a different proof. The proof is by induction in degree. The base of induction is clear. Assume the statement is known in degrees $<n$, and $c\in {\operatorname{Fun}}({\operatorname{Orb}}(X)^n,M^{G_X})$ is such that $\mu(c)=0$. This means that the pullback $f: X^n\to M$ of the function $c$ is a coboundary: $f=dg$. Because $f$ is invariant (under the diagonal action of $G_X$), and $C^\bullet={C_{\textup{inv}}}^\bullet\oplus C^\bullet(1-P)$, we can assume that $g$ is invariant. This means that for any $y\in X$, we have $(dg)_y=d(g_y)$ (we recall that $g_y(x_1,\ldots,x_l):=g(y,x_1,\ldots,x_l)$). Thus, $f_y=dg_y$. But $f_y$ is a pullback of a function $c_y\in {\operatorname{Fun}}({\operatorname{Orb}}(X)^{n-1},M^{G_X})$, so by the induction assumption $c_y=0$. Hence $c=0$.
Now let us prove that $\mu$ is surjective. For this it suffices to show that $H^n(X,M)\subset H^1(X,R)H^{n-1}(X,M)$. Let $c\in H^n(X,M)$. By Lemma \[lm:tc\], the element $c$ can be represented by an invariant cycle, $f\in {Z_{\textup{inv}}}^n(X,M)$. By remark \[rm:fy\], $f_y\in Z^{n-1}(X,M)$ for all $y\in X$. For each $y\in X$, decompose $f_y$ as $f_y=(f_y)^++(f_y)^-$, where $$(f_y)^+=f_y\cdot P\in{Z_{\textup{inv}}}^{n-1}(X,M)\text{\ \ and\ \ }(f_y)^-=f_y\cdot(1-P)\in Z^{n-1}(X,M).$$ These functions give rise to unique functions $f^+,f^-\in C^n(X,M)$ such that $(f^\pm)_y=(f_y)^\pm$ $\forall y\in X$. Moreover, it is clear that $f=f^++f^-$. Since $(f^+)_y\in{Z_{\textup{inv}}}^{n-1}(X,M)$ $\forall y$, it is easy to see that $f^+\in Z^n(X,M)$. Thus, also $f^-\in Z^n(X,M)$. Let us see now that $f^\pm$ are invariant: for any $h\in C^n(X,M)$, $g\in G_X$, we have the equality $h_y\cdot g=(h\cdot g)_{g^{-1}y}$, which implies that $$f^+_{gy}=f_{gy}\cdot P=f_{gy}\cdot g^{-1}P=(f\cdot g^{-1})_y\cdot P=f_y^+,$$ and thus $(f^+\cdot g)_y=(f^+_{gy})\cdot g=(f^+)_y$. Since this equality holds $\forall y\in X$, we have $f^+\in{Z_{\textup{inv}}}^n(X,M)$ as claimed. Since $f\in{Z_{\textup{inv}}}^n(X,M)$, we also have $f^-\in{Z_{\textup{inv}}}^n(X,M)$. Now, as $G_X$ acts trivially on cohomology, there exists $h\in C^{n-1}(X,M)$ such that $d(h_y)=f^-_y$ for each $y\in X$. Take $\tilde h=hP$. We have $$d((h\cdot g)_y)=d(h_{gy}\cdot g)
=d(h_{gy})\cdot g=f^-_{gy}\cdot g=(f^-\cdot g)_y=f^-_y,$$ and thus, by , $(d\tilde h)_y=d(\tilde h_y)=f^-_y$, whence $d\tilde h=f^-$. Thus, $f^-$ is a coboundary, and we can assume that $f=f^+$. In other words, $f\in {\operatorname{Fun}}({\operatorname{Orb}}(X),Z^{n-1}(X,M)^{G_X})$. This means that $f=\sum_{s\in {\operatorname{Orb}}(X)}1_s\otimes f(s)$, where $1_s$ is the characteristic function of $s$ with values in $R$. Since $1_s$ is a cocycle, we have proved that $c\in H^1(X,R)H^{n-1}(X,M)$, as desired.
Now let $M$ be a semisimple finite dimensional $G_X$-module over a field $k$ of characteristic zero (but we do not require the image of $G_X$ to be finite). In this case, we have
\[th2\] Lemma \[lm:tc\] and Theorem \[th1\] are true for such $M$.
By a Chevalley’s theorem [@c], the representations $C^n(X,M)={\operatorname{Fun}}(X,k)^{\otimes n}\otimes M$ are semisimple (as tensor products of semisimple representations). Therefore, there exists an invariant projector $P: C^\bullet\to (C^\bullet)^{G_X}$. The rest of the proof is the same as in the previous case.
Recall [@s] that $G_X$ is a central extension of the finite group $G_X^0$ with kernel being the finitely generated abelian group $\Gamma$.
If $M$ is a finite dimensional ${\mathbb Q}[G_X]$-module and $M(1)$ the generalized eigenspace for the trivial character of $\Gamma$, then $H^\bullet(X,M)=H^\bullet(X,M(1))$.
Write $M=\oplus_{\chi}M(\chi)$, where $\chi$ runs over the characters of $\Gamma$. We have $H^\bullet(X,M)=\oplus_{\chi}H^{\bullet}(X,M(\chi))$. Now, we prove by induction on the dimension of $M(\chi)$ that if $\chi$ is non-trivial then $H^\bullet(X,M(\chi))=0$. If $\dim M(\chi)=0$, the cohomology clearly vanishes. Suppose now that $\dim M(\chi)=n>0$ and for smaller dimensions the statement is known. Let $M_0$ be a simple submodule of $M(\chi)$. We have then the short exact sequence of complexes $$0\to C^{\bullet}(X,M_0)\to C^{\bullet}(X,M(\chi))
\to C^{\bullet}(X,M(\chi)/M_0) \to 0.$$ The first complex is acyclic by Theorem \[th2\], the third one is acyclic by the induction assumption, so by the long exact sequence in cohomology, the complex in the middle is also acyclic. The induction step and the corollary are proved.
Let $M$ be a finite dimensional ${\mathbb Q}[T^{\pm 1}]$-module. Then the twisted rack cohomology $H^i_T(X,M)$ equals the twisted rack cohomology $H^i_T(X,M(1))$, where $M(1)$ is the generalized eigenspace of $T$ in $M$ with eigenvalue $1$.
To compute the Betti numbers of twisted cohomology, the only lacking case is that in which the elements of the rack $X$ act on $M$ by a Jordan block with $1$ on the diagonal.
\[pr:tcj\] Let $M$ be an ${\mathbb Q}G_X$-module with basis $\{v_1,\ldots,v_k\}$ on which the elements of $X$ act by $v_i\mapsto v_{i-1}+v_i$ ($v_0:=0$). Then $\dim H^n(X,M)=m^n$, where $m=|{\operatorname{Orb}}(X)|$.
Before proving the Proposition we state two easy lemmas:
\[lm:j1\] Let $(C^{\bullet},d)$ be a complex and suppose that $C^\bullet=C_1^\bullet\oplus C_2^\bullet$ and that the differential $d$ has the form ${\left(\begin{array}{cc}d_1 & \alpha \\ 0 & d_2\end{array}\right)}$ for this decomposition. Then $\alpha$ induces a map $\alpha_*^n:H^{n-1}(C^{\bullet}_2)\to H^n(C^{\bullet}_1)$. Consider then the short exact sequence of complexes $$0\longrightarrow C^{\bullet}_1 \stackrel{i}{\longrightarrow}
C^{\bullet} \stackrel{p}{\longrightarrow} C^{\bullet}_2
\longrightarrow 0$$ and let $\beta^n:H^{n-1}(C^{\bullet}_2)\to H^n(C^{\bullet}_1)$ be the connecting homomorphism. Then $\beta^n=\alpha_*^n$.
Since $d^2=0$, we have $d_1\alpha=-\alpha d_2$, whence it induces a map in cohomology. The second assertion follows in a straightforward way from the definition of the connecting homomorphism.
\[lm:j2\] Let $C^{\bullet}=C^{\bullet}_1\oplus C^{\bullet}_2$ be as in Lemma \[lm:j1\]. Suppose that $(C^{\bullet\prime}_2,d'_2)$ is a complex and that $f:C^{\bullet\prime}_2\to C^{\bullet}_2$ is a quasi-isomorphism. Then $({\operatorname{id}}\oplus f):C^{\bullet}_1\oplus C^{\bullet\prime}_2\to C^{\bullet}$ is a quasi-isomorphism, where the first complex has differential given by ${\left(\begin{array}{cc}d_1 & \alpha f \\ 0 & d'_2\end{array}\right)}$.
This follows easily from the $5$-lemma.
The proof is by induction on $k$. If $k=1$ the assertion is Corollary \[cbn\]. Assume that the result is true for dimensions $<k$. Let us decompose $C^{\bullet}=C^{\bullet}(X,M_1)\oplus C^{\bullet}(X,M_2)$, where $M_1$ is generated by $v_1,\ldots,v_{k-1}$ and $M_2$ is generated by $v_k$. Notice that the differential $d$ in $C^{\bullet}$ can be written as ${\left(\begin{array}{cc}d_1 & \alpha \\ 0 & d_2\end{array}\right)}$, where $d_i:C^{\bullet}(X,M_i)\to C^{\bullet}(X,M_i)$ are the differentials of the same complex we are considering for $M$ of dimension $k-1$ and $1$ respectively.
Let us take $C^{\bullet\prime}_2=T^{\bullet}({\operatorname{Fun}}({\operatorname{Orb}}(X),{\mathbb Q}))$. By Theorem \[th1\], the inclusion $i:C^{\bullet\prime}_2\to C^{\bullet}_2$ is a quasi-isomorphism, and thus by Lemma \[lm:j2\] we can work with $C^{\bullet}(X,M_1)\oplus T^{\bullet}({\operatorname{Fun}}({\operatorname{Orb}}(X),{\mathbb Q}))$. We consider the long exact sequence $$\label{eq:lsl}
\to H^{n-1}(C^{\bullet\prime}_2) \stackrel{\beta^n}{\longrightarrow}
H^n(C^{\bullet}_1) \stackrel{i^n}{\longrightarrow}
H^n(C^{\bullet}_1\oplus C^{\bullet\prime}_2) \stackrel{p^n}{\longrightarrow}
H^n(C^{\bullet\prime}_2) \stackrel{\beta^{n+1}}{\longrightarrow}
H^{n+1}(C^{\bullet\prime}_2) \to$$ Let $\bar\alpha=\alpha|_{C^{\bullet\prime}_2}$ and consider the induced map in cohomology $\bar\alpha_*$, i.e., $$\bar\alpha_*^n:H^{n-1}(C^{\bullet\prime}_2)
=T^{n-1}({\operatorname{Fun}}({\operatorname{Orb}}(X),{\mathbb Q}))\to H^n(C^{\bullet}_1)=H^n(X,M_1).$$ By Lemma \[lm:j1\], $\beta^n=\bar\alpha_*^n$. We claim that ${\operatorname{rk}}\bar\alpha_*={\operatorname{rk}}\bar\alpha$. To see this, it suffices to prove that ${\operatorname{Im}}\bar\alpha^n\cap B^n(C^{\bullet}_1)=0$. Suppose that $\bar\alpha^n(f)\in B^n(C^{\bullet}_1)$, then it has the form $\bar\alpha^n(f)=\sum_{i=1}^{k-1}b_iv_i$, where $b_i\in C^n(X,{\mathbb Q})$. Furthermore, it is clear that $b_{k-1}\in B^n(X,{\mathbb Q})$. On the other hand, if $\pi:X\to{\operatorname{Orb}}(X)$ is the canonical projection, we have $$\alpha^n(f)(x_1,\ldots,x_n)=
\sum_{i=1}^n(-1)^if(\pi(x_1),\ldots,\pi(x_{i-1}),\pi(x_{i+1}),
\ldots,\pi(x_n)) v_{k-1},$$ which shows that $b_{k-1}\in T^n({\operatorname{Fun}}({\operatorname{Orb}}(X),{\mathbb Q}))$. But it is shown in the injectivity part of the proof of Theorem \[th1\] that $T^n({\operatorname{Fun}}({\operatorname{Orb}}(X),{\mathbb Q}))\cap B^n(X,{\mathbb Q})=0$, and the claim is proved.
Then, ${\operatorname{rk}}\beta^n={\operatorname{rk}}\bar\alpha^n$. But the latter is not difficult to compute: if we consider the complex $(D^{\bullet},\hat d)$, where $D^n={\operatorname{Fun}}(({\operatorname{Orb}}(X))^n,{\mathbb Q})$ and $\hat d$ is given by $$\hat d(f)(a_1,\ldots,a_n)=
\sum_{i=1}^n(-1)^if(a_1,\ldots,a_{i-1},a_{i+1},
\ldots,a_n),$$ then it is clear that $\bar\alpha^n$ and $\hat d^n$ have the same rank. Furthermore, it is well known that $D^{\bullet}$ is acyclic (it gives the reduced cohomology of a simplex of dimension $m-1$). It is easy then to compute the rank of $\hat d$; we have ${\operatorname{rk}}\hat d^n=m^{n-1}-m^{n-2}+m^{n-3}-\cdots\pm 1$. We add this computation to the long exact sequence and we are done: we have ${\operatorname{rk}}\beta^n=m^{n-1}-m^{n-2}+\cdots\pm 1$, and since by the inductive assumption $\dim H^n(C^{\bullet}_1)=m^n$, then ${\operatorname{rk}}i^n=m^n-m^{n-1}+\cdots\pm 1$. Also, we have ${\operatorname{rk}}\beta^{n+1}=m^n-m^{n-1}+\cdots\pm 1$ and since $\dim H^n(C^{\bullet\prime}_2)=m^n$, we get ${\operatorname{rk}}p^n=m^{n-1}-m^{n-2}+\cdots\pm 1$. Thus, $\dim H^n(C^{\bullet})={\operatorname{rk}}i^n+{\operatorname{rk}}p^n=m^n$, proving the inductive step.
Since for $M$ as above we have $\dim M^{G_X}=1$, we have proved:
Let $M$ be a right ${\mathbb Q}G_X$-module on which all the elements of $X$ act by the same operator. Then $\dim H^n(X,M)=m^n\times\dim M^{G_X}$.
It is interesting to study the graded algebra $H_{\text{inv}}^{\bullet}(X,\mathbf{k})$, where $\mathbf{k}$ is a field of characteristic $p$ dividing $|G_X^0|$, to which Theorem \[th1\] does not apply. One may ask the following questions about this ring:
- Is it finitely generated?
- What is its Poincaré series? Is it a rational function?
A relation with group cohomology
================================
In this section, for any rack $X$, we want to give a group theoretical interpretation of the group $H^2(X,A)$ (where $A$ is a trivial $G_X$-module). This group is useful in the theory of pointed Hopf algebras [@g].
We start with the following obvious, but useful proposition.
\[shift\] Let $A$ be a trivial $G_X$-module. Then one has a natural isomorphism of complexes$J: C^n(X,A)\to C^{n-1}(X,{\operatorname{Fun}}(X,A))$, $n\ge 1$, where we consider the action of $G_X$ on ${\operatorname{Fun}}(X,A)$ given by $(hy)(x)=h(y{\triangleright}x)$. It is given by $(Jf)(x_1,\ldots,x_{n-1})(x_n)=f(x_1,\ldots,x_n)$. In particular, it induces an isomorphism $H^n(X,A)\to H^{n-1}(X,{\operatorname{Fun}}(X,A))$.
We note that this proposition becomes false if the action of $G_X$ on $A$ is not trivial.
Now we give the main result of this section. Let $M$ be a right $G_X$-module.
\[xgx\] $H^1(X,M)\simeq H^1(G_X,M)$.
Propositions \[shift\] and \[xgx\] imply
\[cor\] If $A$ is a trivial $G_X$-module, then $H^2(X,A)\simeq H^1(G_X,{\operatorname{Fun}}(X,A))$.
(of Proposition \[xgx\]) Let $C^\bullet(G,M)$ be the standard complex of a group $G$ with coefficient in a right $G$-module $M$. Let $\eta: C^1(G_X,M)\to C^1(X,M)$ be the homomorphism induced by the natural map $X\to G_X$. It is easy to show that this homomorphism maps cocycles to cocycles and coboundaries to coboundaries. Thus, it induces a homomorphism $\eta: H^1(G_X,M)\to H^1(X,M)$. Thus, our job is to show that any $f\in Z^1(X,M)$ lifts uniquely to a $1$-cocycle on $G_X$.
To do this, recall that a map $\pi: G_X\to M$ is a $1$-cocycle iff the map $\hat \pi: G_X\to G_X\ltimes M$ given by $g\to (g,\pi(g))$ is a homomorphism. On the other hand, we have a map $\xi_f: X\to G_X\ltimes M$ given by $\xi_f(x)=(x,f(x))$. So we need to show that $\xi_f$ extends to a homomorphism $G_X\to G_X\ltimes M$. But the group $G_X$ is generated by $X$ with relations $xy=(x{\triangleright}y)x$. Thus, we only need to check that $\xi_f(x)$ satisfy the same relations. But it is easy to check that this is exactly the condition that $df=0$. We are done.
Another, more conceptual, proof runs as follows: let $N$ be a right $X$-module (i.e, a right $G_X$-module) and consider on $X\times N$ the following structure: $$(x,n){\triangleright}(y,m)=(x{\triangleright}y,n(1-(x{\triangleright}y)^{-1})+mx^{-1}).$$ It is easy to verify that this is a rack structure on the product; we shall denote it by $(X\ltimes N,{\triangleright})$ (it is actually the same structure as in [@ag] for the left $X$-module $N$ with $x\cdot n=nx^{-1}$). We have then, with a straightfoward proof,
Let $\omega:X\to N$ and define $\hat \omega:X\to X\ltimes N$ by $\hat \omega(x)=(x,\omega(x)x^{-1})$. Then $\hat \omega$ is a rack homomorphism if and only if $\omega\in Z^1(X,N)$.
Take $\al:X\ltimes N\to G_X\ltimes N$, $\al(x,n)=(x,nx)$. One can check that in the square $$\begin{CD}
X @>{\hat\omega}>> X\ltimes N \\
@VVV @VV\alpha V \\
G_X @>{\hat\pi}>> G_X\ltimes N
\end{CD}$$ each of $\omega$, $\pi$ determines uniquely the other in such a way that the diagram is commutative.
Corollary \[cor\] holds also when $A$ is nonabelian. In this case $H^2(X,A)$ is the quotient of the set $Z^2(X,A)=\{f:X\times X\to A\ |\
f(x{\triangleright}y,x{\triangleright}z)f(x,z)=f(x,y{\triangleright}z)f(y,z)\}$ by the equivalence relation $f\sim f'$ if there is a $\gamma:X\to A$ such that $f'(x,y)=\gamma(x{\triangleright}y)f(x,y)\gamma(y)^{-1}$. The proof is the same as in the abelian case.
[ZEUS]{} N. Andruskiewitsch & M. Graña, *From racks to pointed Hopf algebras*, Adv. Math, to appear. Also in `math.QA/0202084`.
C. Chevalley, *Théorie des groupes de Lie. Tome III. Théorèmes généraux sur les algèbres de Lie*, Hermann & Cie, Paris, 1955.
J.S. Carter, M. Elhamdadi & M. Saito, *Twisted Quandle Cohomology Theory and Cocycle Knot Invariants*, [math.GT/0108051]{}.
J.S. Carter, D. Jelsovsky, S. Kamada & M. Saito, *Quandle Homology Groups, Their Betti Numbers, and Virtual Knots*, J. Pure Applied Algebra **157** (2001), 135–155.
J.S. Carter & M. Saito, *Quandle Homology Theory and Cocycle Knot Invariants* `math.GT/0112026`.
R. Fenn, C. Rourke & B. Sanderson, *James bundles and applications*, preprint available at <http://www.maths.warwick.ac.uk/~bjs>.
M. Graña, *On Nichols algebras of low dimension*, in “New Trends in Hopf Algebra Theory" (ed. Andruskiewitsch, Ferrer Santos, Schneider); Contemp. Math. [**267**]{} (2000), 111–136.
D. Joyce, *A Classifying Invariant of knots, The knot Quandle*, J. Pure Appl. Alg. **23** (1982), 37–65.
R.A. Litherland & S. Nelson, *The Betti numbers of some finite racks*, `math.GT/0106165`.
Jiang-Hua Lu, Min Yan & Yong-Chang Zhu, *On set-theoretical Yang–Baxter equation*, Duke Math. J. **104** (2000), 1–18.
T. Mochizuki, *Some calculations of cohomology groups of finite Alexander quandles*, preprint available at <http://math01.sci.osaka-cu.ac.jp/~takuro>.
A. Soloviev, *Non-unitary set-theoretical solutions to the quantum Yang–Baxter equation*, Math. Res. Lett. **7** (2000), no. 5-6, 577–596.
[^1]: This group appears already in the work of Joyce [@j], who pointed out that the functor $X\to G_X$ is adjoint to the functor assigning to a group the underlying rack (with the conjugation operation). Thus the group $G_X$ can be viewed as the “enveloping group” of $X$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this review phenomenological consequences of the Standard Model extension by means of new spin-1 chiral fields with the internal quantum numbers of the electroweak Higgs doublets are summarized. The prospects for resonance production and detection of the chiral vector $Z^*$ and $W^{*\pm}$ bosons at the LHC energies are considered on the basis of quantitative simulations within the CompHEP/CalcHEP package. The $Z^*$ boson can be observed as a Breit-Wigner resonance peak in the invariant dilepton mass distributions in the same way as the well-known extra gauge $Z'$ bosons. However, the $Z^*$ bosons have unique signatures in transverse momentum, angular and pseudorapidity distributions of the final leptons, which allow one to distinguish them from other heavy neutral resonances.
In 2010, with 40 pb$^{-1}$ of the LHC proton-proton data at the energy 7 TeV, the ATLAS detector was used to search for narrow resonances in the invariant mass spectrum of $e^+e^-$ and $\mu^+\mu^-$ final states and high-mass charged states decaying to a charged lepton and a neutrino. No statistically significant excess above the Standard Model expectation was observed. The exclusion mass limits of 1.15 TeV$/c^2$ and 1.35 TeV$/c^2$ were obtained for the chiral neutral $Z^*$ and charged $W^*$ bosons, respectively. These are the first direct limits on the $W^*$ and $Z^*$ boson production.
Based on the above, a novel strategy for the chiral boson search in the LHC dijet data is discussed. For almost all currently considered exotic models the relevant signal is expected in the central dijet rapidity region $y_{1,2}\simeq 0$ and $|y_1-y_2|\simeq 0$. On the contrary, the chiral bosons do not contribute to this region but produce an excess of dijet events far away from it. In particular, for these bosons the appropriate kinematical restrictions lead to a dip in the centrality ratio distribution over the dijet invariant mass instead of a bump expected in the most exotic models.
author:
- 'M. V. Chizhov'
- 'V. A. Bednyakov'
- 'I. R. Boyko'
- 'J. A. Budagov'
- 'M. A. Demichev'
- 'I. V. Yeletskikh'
date: '14. 10. 2011'
title: Anomalously interacting new extra vector bosons and their first LHC constraints
---
Introduction
============
Gauge interactions are the only well-established fundamental interactions in Nature. The method of the covariant derivatives leads to the unique minimal form of the gauge boson couplings to the matter fermions. Nevertheless, the Yukawa interactions of the Higgs bosons are also necessary for self-consistent construction of the Standard Model (SM). Furthermore, although the gauge symmetry allows anomalous interactions in the initial Lagrangian, all known fundamental spin-1 bosons (photon, $W^\pm$, $Z$ boson and gluons) possess only renormalizable minimal interactions with the known fermions. The anomalous interactions are considered as effective ones and are generated at the level of the quantum loop corrections. They are usually proportional to the additional square of a small coupling constant and can be neglected in the first-order approximation.
New heavy neutral gauge bosons are predicted in many extensions of the SM. They are associated with additional U(1)$'$ gauge symmetries and are generically called $Z'$ bosons. The minimal gauge interactions of these bosons with matter lead to the well-known angular distribution of outgoing leptons (the $Z'$ decay product) in the dilepton center-of-mass reference frame $$\label{sV}{
\frac{{\rm d}\sigma_{Z'}}{{\rm d}\cos\theta^*}\propto 1+
A_{\rm FB}\cdot\cos\theta^*+\cos^2\theta^*},$$ which at present is interpreted as a canonical signature for the intermediate vector (spin-1) bosons. The coefficient $A_{\rm FB}$ defines the backward-forward asymmetry, depending on $P$-parity of $Z'$ boson couplings to matter.
In addition, another type of spin-1 bosons may exist, which leads to a different signature in the angular distribution. This follows from the presence of different types of relativistic spin-1 fermion currents $\bar{\psi}\gamma^\mu(1\pm\gamma^5)\psi$ and $\partial_\nu[\bar{\psi}\sigma^{\mu\nu}(1\pm\gamma^5)\psi]$, which can couple to the corresponding bosons. A clear example of such kind of interactions is provided by the hadron physics of the quark-antiquark mesons, which is considered as a low-energy QCD effective theory, where gluon and quark degrees of freedom are substituted by physical hadronic states. It was pointed out [@Chizhov:1996ky] that [*three different*]{} quantum numbers J$^{PC}$ of existing neutral spin-1 mesons, ${1^{--}}$, ${1^{++}}$ and ${1^{+-}}$, cannot be assigned just to [*two*]{} vector $\bar{q}\gamma^\mu q$ and axial-vector $\bar{q}\gamma^\mu \gamma^5 q$ quark states, which possess quantum numbers ${1^{--}}$ and ${1^{++}}$, respectively. The additional quark states $\partial_\nu(\bar{q}\sigma^{\mu\nu}q)$ and $\partial_\nu(\bar{q}\sigma^{\mu\nu}\gamma^5 q)$ are required, which also describe vector and axial-vector mesons, but with different transformation properties with respect to the Lorentz group and with different quantum numbers $1^{--}$ and $1^{+-}$, respectively. This example demonstrates that the pure tensor state, ${b_1}$ meson, exists. Furthermore, due to strong dynamics, vector $\rho$ and $\rho'$ mesons have minimal and anomalous couplings with vector $\bar{\psi}\gamma^\mu\psi$ and tensor $\partial_\nu(\bar{\psi}\sigma^{\mu\nu}\psi)$ interpolating currents comparable in magnitude [@Ball:1996tb; @Braun:2003jg; @Becirevic:2003pn; @Chizhov:2003qy]. The both currents have the same quantum numbers J$^{PC}=1^{--}$ and mix. Since the parity and charge conjugation are conserved in QCD, they define the quantum numbers of the mesons.
The mesons assigned to the tensor quark states are some types of “excited” states as far as the only orbital angular momentum with $L=1$ contributes to the total angular momentum, while the total spin of the system is zero. This property manifests itself in their derivative couplings to matter and a different chiral structure of the anomalous interactions in comparison with the minimal gauge ones. In contrast with the minimal gauge couplings, where either only left-handed or right-handed fermions participate in the interactions, the tensor currents mix both left-handed and right-handed fermions. Therefore, like the Higgs particles, the corresponding bosons carry a nonzero chiral charge. To our knowledge, such bosons were first introduced by Kemmer [@Kemmer:1938zz] and they naturally appear in the extended conformal supergravity theories [@Salam:1989fm]. In fact, this QCD feature can be realized in electroweak physics as well like the technicolor models. This analogy gives us arguments in favor of existence of anomalously interacting vector bosons.
Up to now no search for excited bosons has been done, but regular searches for the excited lepton and quark states $f^*$ have been carried out at the modern colliders, such as LEP [@Barate:1998ce; @Abbiendi:2002wf; @Achard:2003hd], HERA [@Adloff:2002dy; @Chekanov:2001xk] and Tevatron [@Acosta:2004ri; @Abazov:2008hw]. These excited fermions have magnetic moment (Pauli) type couplings to ordinary matter $$\label{f*}
{\cal L}_{\rm excited}^{f^*}=
\frac{g}{\Lambda}{\bar{f}^*}\sigma^{\mu\nu}\!f
\left(\partial_\mu Z_\nu-\partial_\nu Z_\mu\right) +{\rm h.c.},$$ where the parameter $\Lambda$ is connected to the compositeness mass scale of the new physics. There are no objections to interpreting the interactions (\[f\*\]) from a different point of view, introducing excited boson states instead of fermionic ones (“shifting the $*$-sign to the right from $f$ to $Z$”) $$\label{Z*}
{\cal L}_{\rm excited}^{Z^*}=
\frac{g}{\Lambda}\bar{f}\,\sigma^{\mu\nu}\!f
\left(\partial_\mu{ Z^*_\nu}
-\partial_\nu{ Z^*_\mu}\right).$$ This “symmetry” between excited fermions and bosons further supports our interest in consideration of the anomalously interacting vector bosons.
This review paper summarizes our attempts [@Chizhov:2008tp; @Chizhov:2008jw; @Chizhov:2010hh; @Chizhov:2010jg; @Chizhov:2011wt] to fill the gap in the consideration of experimental properties of heavy chiral bosons and continues discussions of possibilities of disentangling them from other particles.
The material below is given in the following order. In section II a simple chiral boson model is formulated for further consideration. Section III contains our numerical estimations of the boson properties within the framework of the CompHEP/CalcHEP package [@Pukhov:1999gg; @Boos:2004kh; @Pukhov:2004ca]. To this end a new model has been implemented, which includes additional new bosons and their corresponding interactions. In section IV the first experimental constraints on the masses of both chiral vector bosons are given. Section V briefly describes new unique signatures of the bosons in dijet final states. Our conclusions are given in section VI.
The chiral boson model
======================
Let us assume that the electroweak gauge sector of the SM is extended by a doublet of new spin-1 [*chiral*]{} bosons $\mbox{\boldmath$W$}^*_\mu$ with the internal quantum numbers of the SM Higgs boson. There are at least three different classes of theories, all motivated by the Hierarchy problem, which predict new vector weak doublets with masses not far from the electro-weak scale. In particular, they can originate from the extensions of the SM such as Gauge-Higgs unification, larger gauge groups or technicolor models [@Chizhov:2009fc]. However, due to the lack of fully realistic models, the collider expectations for signals from these chiral bosons have not yet been studied in detail. Nevertheless, it is possible to point out several model-independent and unique signatures which allow one to identify production of such bosons at the hadron colliders [@Chizhov:2008tp].
Since the tensor current mixes the left-handed and right-handed fermions, which in the SM are assigned to different representations, the gauge doublet should have only anomalous interactions $$\label{master}
{\cal L}^*=\frac{g}{M}\left(
\partial_\mu W^{*-}_\nu\; \partial_\mu \overline{W}^{*0}_\nu
\right)\cdot
\overline{D_R}\;\sigma^{\mu\nu}\left(\hspace{-0.2cm}
\begin{array}{c}
U_L \\
D_L
\end{array}\hspace{-0.2cm}
\right)+
\frac{g}{M}\left(
\overline{U_L}\; \overline{D_L}
\right)
\sigma^{\mu\nu}D_R\cdot
\left(\hspace{-0.2cm}
\begin{array}{c}
\partial_\mu W^{*+}_\nu \\
\partial_\mu W^{*0}_\nu \\
\end{array}\hspace{-0.2cm}\right),$$ where $M$ is the boson mass, $g$ is the coupling constant of the SU(2)$_{\rm W}$ weak gauge group, and $U$ and $D$ generically denote up-type and down-type leptons and quarks. This choice of couplings makes identical all partial fermionic decay widths of the well-known hypothetical $W'$ boson with the SM-like interactions $$\label{W'}
{\cal
L}'_{CC}=\frac{g}{\sqrt{2}}\,W'^-_\mu\cdot\overline{D_L}\gamma^\mu
U_L
+\frac{g}{\sqrt{2}}\,\overline{U_L}\gamma^\mu D_L\cdot
W'^+_\mu$$ and the charged $W^{*\pm}$ boson with the same mass. Here we also assume universality of lepton and quark couplings with different flavors. In full analogy with the above-mentioned mesons these bosons, coupled to the tensor quark currents, can be considered as [*excited*]{} states. This property manifests itself in their derivative couplings to fermions and in the different chiral structure of the interactions in contrast to the minimal gauge interactions.
For simplicity, in (\[master\]) we have introduced only interactions with the down-type right-handed singlets, $D_R$. In particular, in order to allow a possibility of detecting the neutral $CP$-even $Z^*=(W^{*0}+\overline{W}^{*0})/\sqrt{2}$ bosons via their decays into charged leptons (the Drell–Yan-like process) they should couple to the [*down*]{} type of fermions $$\label{Z*ed}{
{\cal L}^{*}_{NC}=\frac{g}{2\sqrt{2} M}
\left(\bar{\ell}\sigma^{\mu\nu}\ell+\bar{d}\sigma^{\mu\nu}d\right)
\left(\partial_\mu Z^*_\nu-\partial_\nu Z^*_\mu\right)}.$$ Since we have introduced the complex $\mbox{\boldmath$W$}^*_\mu$ doublet, there is an additional neutral $CP$-odd $\tilde{Z}^*=(W^{*0}-\overline{W}^{*0})/\sqrt{2}$ boson. However, in the case of light final states it is impossible to discriminate the multiplicative quantum numbers of the neutral bosons, namely $P$ and $C$. Therefore, in the following calculations we will consider only one of them, for instance, the $Z^{*}$ boson.
For comparison we will consider topologically analogous gauge interactions of the $Z'$ boson $$\label{Z'ed}
{\cal L}'_{NC}=\frac{g}{2}
\left(\bar{\ell}\gamma^{\mu}\ell+\bar{d}\gamma^{\mu}d\right)Z'_\mu$$ with the same mass $M$. The coupling constants are chosen in such a way that all fermionic decay widths in the Born approximation of the both neutral bosons are identical. It means that their total production cross sections at the hadron colliders are nearly equal up to next-to-leading order corrections. Their total fermionic decay width $$\label{Gl}
\Gamma=\frac{g^2}{4\pi}M\approx 0.034~M$$ is sufficiently narrow so that they can be identified as resonances at the hadron colliders in the Drell–Yan process.
Furthermore, as several Higgs doublets are introduced in many of the SM extensions, the realistic model could include several gauge doublets. Using the charge-conjugated doublet $$\label{W*c}
\mbox{\boldmath $W$}^{*\,{\rm c}}_\mu=\left(
\begin{array}{c}
\overline{W}^{*0}_\mu \\
-W^{*-}_\mu \\
\end{array}
\right)$$ (or new ones with the hypercharges opposite to the $\mbox{\boldmath
$W$}^*_\mu$ doublet) one can construct more complicated models including up-type right-handed singlets, $U_R$, as well.
Numerical simulations of the chiral bosons
==========================================
Up to now, any excess in the yield of the Drell–Yan process with high-energy invariant mass of the lepton pairs remains the clearest indication of possibile production of a new heavy neutral boson at the hadron colliders. Therefore, we will first concentrate on consideration of the production and decay of neutral bosons, where full kinematics is experimentally reconstructible. In what follows we will use the CompHEP/CalcHEP package [@Pukhov:1999gg; @Boos:2004kh; @Pukhov:2004ca] for the numerical calculations of various distributions for the inclusive processes $pp\rightarrow\gamma/Z/Z'+X\to\ell^+\ell^-+X$ and $pp\rightarrow\gamma/Z/Z^*+X\to\ell^+\ell^-+X$ with a CTEQ6M choice for the proton parton distribution set at $\sqrt{s}=10$ TeV. For both final leptons we impose angular restrictions (cuts) on the pseudorapidity range $|\eta_\ell|<2.5$ and the transverse momentum $p_{\rm T} > 20$ GeV$/c$, which are relevant to the general LHC detectors.
Let us choose $M=1$ TeV/$c^2$ as a reference mass for new heavy bosons. For the high dilepton masses the cross sections of the new boson production with this mass at the peak is about two orders of magnitude higher (in our model) than the corresponding Drell–Yan background, being the SM $\gamma$ and $Z$ boson tails in the invariant dilepton mass distributions (Fig. \[fig:1-Mee\], left). Therefore, the peak(s) should be clearly visible.
![ \[fig:1-Mee\] \[fig:2\] The invariant dilepton mass distributions (left) for the $Z'$ boson (blue) and the excited chiral $Z^*$ boson (red) with mass 1 TeV$/c^2$ together with the Drell–Yan SM background (from the photon and the $Z$ boson) at the LHC for $\sqrt{s}=10$ TeV. The differential cross-sections (right) for the gauge $Z'$ boson (blue) and the excited chiral $Z^*$ boson (red) with the Drell–Yan SM background as functions of the lepton transverse momentum at the CERN LHC.](figures/Z300i10.eps "fig:"){width="48.00000%"} ![ \[fig:1-Mee\] \[fig:2\] The invariant dilepton mass distributions (left) for the $Z'$ boson (blue) and the excited chiral $Z^*$ boson (red) with mass 1 TeV$/c^2$ together with the Drell–Yan SM background (from the photon and the $Z$ boson) at the LHC for $\sqrt{s}=10$ TeV. The differential cross-sections (right) for the gauge $Z'$ boson (blue) and the excited chiral $Z^*$ boson (red) with the Drell–Yan SM background as functions of the lepton transverse momentum at the CERN LHC.](figures/Z300t10.eps "fig:"){width="48.00000%"}
The peaks in the dilepton invariant mass distributions originate from the Breit–Wigner propagator form, which is [*the same*]{} for both the gauge and chiral [*neutral*]{} bosons in the Born approximation.
Concerning discovery of the [*charged*]{} heavy boson at the hadron colliders one believes that the cleanest method is detection of its subsequent leptonic decay into an isolated high transverse-momentum charged lepton (better without a prominent associated jet activity). In this case the heavy new boson can be observed through the Jacobian peak in the transverse $p^{}_{\rm T}$ or $m^{}_{\rm T}$ distribution. It has become proverbial (see, for example, the textbook [@Barger:1987nn]) that the Jacobian peak is an inevitable characteristic of any two-body decay. However, it is not the case for decays of the new chiral bosons [@Chizhov:2006nw]. It has been found in [@Chizhov:2000vt] that tensor interactions lead to a new angular distribution of the outgoing fermions $$\label{GLR} \frac{{\rm d} \sigma(q\bar{q}\to Z^*\!/W^*\to f\bar{f})}
{{\rm d} \cos\theta} \propto \cos^2\theta,$$ in comparison with the well-known vector interaction result $$\label{GLL}
\frac{{\rm d} \sigma(q\bar{q}\to Z'\!/W'\to f\bar{f})}
{{\rm d} \cos\theta} \propto 1+\cos^2\theta \, .$$ It was realized later [@Chizhov:2006nw] that this property ensures a distinctive signature for the detection of the new interactions at the hadron colliders. At first sight, the small difference between the distributions (\[GLR\]) and (\[GLL\]) seems unimportant. However, the absence of the constant term in the first case results in very new experimental signatures.
The angular distribution for vector interactions (\[GLL\]) includes a nonzero constant term, which leads to the kinematical singularity in the $p^{}_{\rm T}$ distribution of the final fermion $$\label{1/cos}
\frac{1}{\cos\theta}\propto\frac{1}{\sqrt{(M/2)^2-p^2_T}}$$ in the narrow width approximation $\Gamma <\!\!\!< M$ $$\label{narrow}
\frac{1}{(s-M^2)^2+M^2\Gamma^2}\approx\frac{\pi}{M\Gamma}\delta(s-M^2).$$ This singularity is transformed into a well-known Jacobian peak due to a finite width of the resonance. In contrast, the pole in the decay distribution of the $Z^*/W^*$ bosons is canceled out and the fermion transverse momentum $p_{\rm T}$ distribution even reaches zero at the kinematical endpoint $p_{\rm T}=M/2$, rather than the Jacobian peak (at the kinematical endpoint $M/2$) for the gauge bosons (Fig. \[fig:2\], right). Therefore, even the lepton transverse momentum distribution demonstrates a difference between the gauge and chiral bosons.
According to (\[GLR\]), there exists a characteristic plane, perpendicular to the beam axis in the parton rest frame, where emission of final-state pairs is forbidden. The nonzero probability in the perpendicular direction in the laboratory frame is due to the longitudinal boosts of colliding partons. So, at the Fermilab Tevatron the production of such heavy bosons occurs almost at the threshold with approximately zero longitudinal momenta. Hence, the lepton pseudorapidity distribution for the chiral bosons has a minimum at $\eta_\ell=0$ (Fig. \[fig:3\], left). On the other hand, the CERN LHC is sufficiently powerful to produce heavy bosons with the mass $M=1$ TeV/$c^2$ with high longitudinal boosts. Therefore, the pseudorapidity distributions for the gauge and chiral bosons at the LHC look similar (Fig. \[fig:4\], right).
![ \[fig:3\] \[fig:4\] The differential cross sections for the gauge $Z'$ boson (blue) and the excited chiral $Z^*$ boson (red) decaying to a lepton pair with the invariant mass 800 GeV/$c^2$ $<M_{\ell\ell}<$ 1200 GeV/$c^2$ as functions of the lepton pseudorapidity at the Fermilab Tevatron $\sqrt{s}=1.96$ TeV (left) and at the CERN LHC $\sqrt{s}=14$ TeV (right).](figures/etaTev.eps "fig:"){width="48.00000%"} ![ \[fig:3\] \[fig:4\] The differential cross sections for the gauge $Z'$ boson (blue) and the excited chiral $Z^*$ boson (red) decaying to a lepton pair with the invariant mass 800 GeV/$c^2$ $<M_{\ell\ell}<$ 1200 GeV/$c^2$ as functions of the lepton pseudorapidity at the Fermilab Tevatron $\sqrt{s}=1.96$ TeV (left) and at the CERN LHC $\sqrt{s}=14$ TeV (right).](figures/etaLHC.eps "fig:"){width="48.00000%"}
In order to make more substantial and experiment-looking conclusions, let us investigate signal distributions selecting only “on-peak” events with the invariant dilepton masses in the range 800 GeV$/c^2<M_{\ell\ell}<$ 1200 GeV$/c^2$. To this end (as a test example) in [@Chizhov:2010hh] for these dilepton masses and kinematical restrictions $|\eta_\ell|<2.5$ and the transverse lepton momentum $p_{\rm T} > 20$ GeV$/c$ we have simulated production of dilepton events via $pp\rightarrow\gamma/Z/Z'+X\to\ell^+\ell^-+X$ and $pp\rightarrow\gamma/Z/Z^*+X\to\ell^+\ell^-+X$ for the LHC integrated luminosity of 100 pb$^{-1}$ and $\sqrt{s}=10$ TeV. As far as the center-of-mass energy for the 2010–2011 runs was 7 TeV, at which the cross sections are roughly half as large, 200 pb$^{-1}$ of data will be equivalent to the case with $\sqrt{s}=10$ TeV. For the $Z'$ and $Z^*$ bosons (with mass 1 TeV$/c^2$) production cross-sections $\sigma_{Z'}=0.45$ pb and $\sigma_{Z^*}=0.41$ pb were obtained. The values respectively transform into 44.9 and 41.2 dilepton events in the mass window of 800 GeV$/c^2<M_{\ell\ell}<$ 1200 GeV$/c^2$ (see Fig. \[fig:M\_eeZZ\]). Under the same kinematical conditions the SM gives 0.6 events with the production cross-section 5.75 fb.
![ \[fig:M\_eeZZ\] The invariant mass dilepton distributions for the gauge $Z'$ boson (left) and the excited chiral $Z^*$ boson (right) simulated with the common mass 1 TeV$/c^2$. They consist of 44.9 and 41.2 dilepton events in the mass window of 800 GeV$/c^2<M_{\ell\ell}<$ 1200 GeV$/c^2$ for $Z'$ and $Z^*$ bosons. ](figures/Vi.eps "fig:"){width="48.00000%"} ![ \[fig:M\_eeZZ\] The invariant mass dilepton distributions for the gauge $Z'$ boson (left) and the excited chiral $Z^*$ boson (right) simulated with the common mass 1 TeV$/c^2$. They consist of 44.9 and 41.2 dilepton events in the mass window of 800 GeV$/c^2<M_{\ell\ell}<$ 1200 GeV$/c^2$ for $Z'$ and $Z^*$ bosons. ](figures/Zi.eps "fig:"){width="48.00000%"}
As already mentioned, the peaks in the invariant mass distributions originate from the Breit–Wigner propagator form, which is the same for both $Z'$ and $Z^*$ bosons in the leading Born approximation. Therefore, in order to discriminate them we need to investigate additional distributions selecting only “on-peak” events with the invariant dilepton masses in the optimal window size $[M-2\Gamma,M+2\Gamma]$. According to [@Chizhov:2008tp], a crucial difference between the neutral chiral bosons and other resonances should come from the analysis of the angular distribution of the final-state leptons with respect to the boost direction of the heavy boson in the rest frame of the latter (the Collins–Soper frame [@Collins:1977iv]) (Fig. \[fig:CS\]).
![\[fig:CS\] The differential lepton angular distributions of the gauge $Z'$ boson (left) and the excited chiral $Z^*$ boson (right) as functions of $\cos\theta^*_{\rm CS}$ for $M=1$ TeV/$c^2$. ](figures/fig04.eps "fig:"){width="48.00000%"} ![\[fig:CS\] The differential lepton angular distributions of the gauge $Z'$ boson (left) and the excited chiral $Z^*$ boson (right) as functions of $\cos\theta^*_{\rm CS}$ for $M=1$ TeV/$c^2$. ](figures/fig05.eps "fig:"){width="48.00000%"}
Instead of a smoother angular distribution for the gauge interactions (left), a peculiar “swallowtail” shape of the chiral boson distribution (right) occurs with a dip at $\cos\theta^*_{\rm CS}=0$. It will indicate the presence of the new interactions. Neither scalars nor other particles possess such a type of angular behavior. Indeed, the angular distribution of outgoing leptons for the $Z^*$ bosons will lead to the large negative value of the so-called centre-edge asymmetry $A_{\rm CE}$: $$\label{ACE}
\sigma\times A_{\rm CE}=\hspace{-0.2cm}\int^{+\frac{1}{2}}_{-\frac{1}{2}}
\hspace{-0.2cm}\frac{\drm\sigma}{\drm\cos\theta^*_{\rm CS}}\,\drm\cos\theta^*_{\rm CS}-
\hspace{-0.1cm}\left[\int^{+1}_{+\frac{1}{2}}
\hspace{-0.2cm}\frac{\drm\sigma}{\drm\cos\theta^*_{\rm CS}}\,\drm\cos\theta^*_{\rm CS}+
\hspace{-0.2cm}\int^{-\frac{1}{2}}_{-1}
\hspace{-0.3cm}\frac{\drm\sigma}{\drm\cos\theta^*_{\rm CS}}\,\drm\cos\theta^*_{\rm CS}
\right]\!,$$
![\[fig:eta\] The differential distributions for the gauge $Z'$ boson (left) and the excited chiral $Z^*$ boson (right) as functions of the difference of the lepton pseudorapidities for $M=1$ TeV/$c^2$.](figures/fig06.eps "fig:"){width="48.00000%"} ![\[fig:eta\] The differential distributions for the gauge $Z'$ boson (left) and the excited chiral $Z^*$ boson (right) as functions of the difference of the lepton pseudorapidities for $M=1$ TeV/$c^2$.](figures/fig07.eps "fig:"){width="48.00000%"}
while the distributions of other known resonances (even with different spins) possess positive or near-zero asymmetries. Using this asymmetry, one can strongly reduce the systematic uncertainties from the hadron structure [@Gounaris:1992kp]. Another “unexpected” consequence of the new form of angular distribution (\[GLR\]) is a very different shape of the event distribution over pseudorapidity difference ($\eta_1-\eta_2$) between both outgoing charged leptons. It is shown in Fig. \[fig:eta\]. Combining these distributions one will have a possibility of differentiating these bosons for higher resonance masses.
We would like again to consider the distributions of the lepton transverse momentum $p_{\rm T}$. As mentioned before and demonstrated for our simulated sample in Fig. \[fig:pT\],
![\[fig:pT\] The differential distributions for the $Z'$ boson (left) and the chiral excited $Z^*$ boson (right) as functions of the lepton transverse momentum $p_{\rm T}$ for $M=1$ TeV/$c^2$ the LHC integrated luminosity of 100 pb$^{-1}$ and $\sqrt{s}=10$ TeV.](figures/fig08.eps "fig:"){width="48.00000%"} ![\[fig:pT\] The differential distributions for the $Z'$ boson (left) and the chiral excited $Z^*$ boson (right) as functions of the lepton transverse momentum $p_{\rm T}$ for $M=1$ TeV/$c^2$ the LHC integrated luminosity of 100 pb$^{-1}$ and $\sqrt{s}=10$ TeV.](figures/fig09.eps "fig:"){width="48.00000%"}
the relevant $Z^*$ and $W^*$ boson decay distributions have a broad smooth hump with the maximum below the kinematical endpoint, instead of an “expected” sharp Jacobian peak. Therefore, in contrast to the usual procedure of the direct and precise determination of the $W'$ resonance mass, the new distribution does not allow doing it for $W^*$ bosons. Moreover, even a relatively small decay width of the chiral bosons has a wide $p_{\rm T}$ distribution that obscures their identification as resonances at hadron colliders.
The first experimental constraints on the chiral bosons
=======================================================
The first direct experimental search for the excited chiral vector bosons was performed by the ATLAS collaboration [@Aad:2008zzm; @Collaboration:2010knc; @Aad:2011dr] in 2010. At the LHC energy of 7 TeV with the integral luminosity around 40 pb$^{-1}$ the ATLAS detector was used for searching for narrow resonances in the invariant mass spectrum above 110 GeV$/c^2$ of $e^+e^-$ and $\mu^+\mu^-$ final states. The main physical results of the relevant paper “Search for high mass dilepton resonances in $pp$ collisions at $\sqrt{s}=7$ TeV with the ATLAS experiment” [@Aad:2011xp] are presented in Fig. \[fig:init\_res3\] together with main backgrounds and expected $Z'$ decay signals for three masses around 1 TeV$/c^2$. Expected signals from the $Z^*$ boson decays considered in this paper (are shown in Fig. \[fig:ZstarPlots\]) have similar shapes and approximately 40% larger cross sections. Three “interesting” events in the vicinity of $m_{ee} = 600$ GeV$/c^2$ and a single event at $m_{\mu\mu} = 768$ GeV$/c^2$ are observed. All details of the data selection and the physical data analysis can be found in [@Aad:2011xp].
![Dielectron (left) and dimuon (right) invariant mass distribution measured by the ATLAS collaboration in 2010 [@Aad:2011xp]. They are compared with all expected backgrounds and three example $Z'$ signals. []{data-label="fig:init_res3"}](figures/InvariantMass_All_Log.eps "fig:"){width="0.5\columnwidth"} ![Dielectron (left) and dimuon (right) invariant mass distribution measured by the ATLAS collaboration in 2010 [@Aad:2011xp]. They are compared with all expected backgrounds and three example $Z'$ signals. []{data-label="fig:init_res3"}](figures/ZLogmass_NoRatio.eps "fig:"){width="0.5\columnwidth"}
\[fig:moneyPlot\]
![Dielectron (left) and dimuon (right) invariant mass distribution after final selection, compared to the stacked sum of all expected backgrounds, with three example $Z^*$ signals overlaid. []{data-label="fig:ZstarPlots"}](figures/ZstarEle.eps "fig:"){width="0.5\columnwidth"} ![Dielectron (left) and dimuon (right) invariant mass distribution after final selection, compared to the stacked sum of all expected backgrounds, with three example $Z^*$ signals overlaid. []{data-label="fig:ZstarPlots"}](figures/ZstarMu.eps "fig:"){width="0.5\columnwidth"}
It is seen that both the dielectron and dimuon invariant mass distributions are well described by the prediction from SM processes. From these figures one can conclude that no statistically significant excess above the SM expectation is observed with these data samples. Nevertheless, these distributions were for the first time used to obtaine a lower direct mass limit (1.152 TeV$/c^2$, see below) for the neutral chiral $Z^*$ boson described in this paper. To this end, 95% C.L. exclusion limits on $Z^*$ production and its dilepton decay, ${\sigma B}$, for the combination of the electron and muon $Z^*$ boson decay channels were used. The combination was performed by defining the likelihood function in terms of the total number of $Z^*$ events produced in both channels.
In the three cases (dielectron, dimuon and combined channels), the 95% C.L. ${\sigma B}$ limit was used to set mass limits for each of the considered models. The observed combined mass limit for the Sequential Standard Model is 1.048 TeV$/c^2$. The limits on the E6-motivated $Z'$ bosons are in the range 0.738–0.900 TeV$/c^2$. Although the angular lepton decay distributions are not the same for $Z'$ and $Z^*$ bosons, it was found that the difference in geometrical acceptance is negligible for the boson pole masses above 750 GeV$/c^2$. Therefore, the same procedure as for the $Z'$ bosons is used to calculate the limit on $\xbr (\zsll )$ and on the $Z^*$ boson mass in each channel and for their combination. Finally the observed combined lower mass limit for the $Z^*$ boson is 1.152 TeV$/c^2$. This is the first direct mass limit on this particle. The $Z^*$ limits are about 100–200 GeV/$c^2$ more stringent than the corresponding limits on all considered $Z'$ bosons.
Furthermore, in 2010 the ATLAS collaboration searched for high-mass states, such as heavy charged gauge bosons ($W'$, $W^*$), decaying to a charged lepton and a neutrino (see Fig. \[fig:final\_mt\]). The relevant paper “Search for high-mass states with one lepton plus missing transverse momentum in proton-proton collisions at $\sqrt{s}= 7$ TeV with the ATLAS detector” contains all details of event selection and physical data analysis [@Aad:2011fe].
{width="50.00000%"} {width="43.00000%"} {width="50.00000%"} {width="43.00000%"} {width="50.00000%"} {width="43.00000%"}
The search for heavy charged resonances inclusively produced at the LHC looks more complicated than the search for neutral states due to the absence of the second decay particle — the undetectable neutrino. In this case the kinematic variable used to identify the is the transverse mass $$\mt = \sqrt{ 2 \pt \met (1 - \cos \phi_{l\nu})}$$ which displays a Jacobian peak that, for $W'\to \ell\nu $, falls sharply above the resonance mass. Here is the lepton transverse momentum, is the magnitude of the missing transverse momentum (), and $\phi_{l\nu}$ is the angle between the and vectors. In the analysis [@Aad:2011fe], “transverse” refers to the plane perpendicular to the colliding beams, “longitudinal” means parallel to the beams, $\theta$ and $\phi$ are the polar and azimuthal angles with respect to the longitudinal direction, and pseudorapidity is defined as $\eta = -\ln(\tan(\theta/2))$.
The main physical results obtained in [@Aad:2011fe] and relevant to our consideration are given in Fig. \[fig:final\_mt\]. The left panel of Fig. \[fig:final\_mt\] shows the , missing , and spectra measured in the muon decay channel for the data, for the expected background, and for three examples of signals at different masses as open histograms. The $W^*$ boson signals are not shown. The QCD background is estimated from the data. The signal and other background samples are normalized using the integrated luminosity of the data and the NNLO (near-NNLO for t-tbar) cross sections. Furthermore, the uncertainties for the $W'$ boson are obtained by varying renormalization and factorization scales and by varying PDFs. Only the latter are employed for the $W^*$ boson search.
One can see from the figures that the agreement between the data and the expected background is rather good. No excess beyond the Standard Model expectations is observed. The lower mass limits expected and obtained from these measurements are depicted in the right panel of Fig. \[fig:limits\_xbr\]. The figure also shows the expected limits and the theoretical as a function of for both channels and their combination. The intersection between the central theoretical prediction and the observed limits provides the 95% C.L. lower limit on the mass. It was found that the charged chiral $W^*$ boson considered in the paper was excluded for masses below 1.350 TeV$/c^2$. These are the first direct limits on the $W^*$ boson production.
Dijet signals of the chiral bosons
==================================
In what follows we will extend the set of possible experimental observables of the chiral bosons to the rich field of hadron final states, in particular we consider peculiarity of the decay of these bosons into two hadronic jets. In fact we will draw extra attention to a novel signal of new physics in the dijet data at the hadron colliders. It is usually accepted that all exotic models predict that these two jets populate the central (pseudo)rapidity region where $y_{1,2}\simeq 0$. Contrariwise, the excited bosons do not contribute into this region, but produce an excess of dijet events over the almost flat QCD background in $\chi=\exp|y_1-y_2|$ away from this region.
At the hadron colliders, inclusive dijet production has one of the largest cross sections and allows data-driven background estimation at the early stage of the collider operation. The feature can be used to search for a signal of new physics in the very early data. In particular, a possible bump in the dijet invariant mass spectrum would indicate the presence of a resonance decaying into two energetic partons. Nevertheless, we could say nothing about its nature, because this bump stems from the Breit–Wigner propagator form, which is characteristic of any type of resonance regardless of its other properties, like spin, internal quantum number, etc. Therefore, other observables are necessary in order to confirm the bump and to reveal the resonance properties.
As in the lepton case (discussed above), this kind of observable could be the dijet distribution over the polar angle $\theta$, which is an angle between the axis of the jet pair and the beam direction in the dijet rest frame. This distribution is directly sensitive to the resonance spin and the dynamics of the underlying process [@Collins:1977iv].
While the QCD processes are dominated by $t$-channel gluon exchanges, which lead to the Rutherford-like dijet distribution $1/(1-\cos\theta)^2$, exotic physics processes proceed mainly through the $s$-channel, where the spin of the resonance uniquely defines the angular distribution. For high-mass resonances and practically massless partons it is convenient to use the helicity formalism since the helicity is a good quantum number for massless particles.
In fact, in the center-of-momentum frame of a particle with spin $s$ and helicity $\lambda$ ($-s\leq\lambda\leq s$) decaying into two massless particles with helicities $\lambda_1$ and $\lambda_2$ the angular distribution of the outgoing particle can be written as [@Haber:1994pe] $$\label{ds/dcos}
\frac{{\rm d}\Gamma_s}{{\rm d}\cos\theta\;{\rm d}\phi}=
\frac{1}{64\pi^2 M}\,
\left\vert\, \sqrt{\frac{2s+1}{4\pi}}\,e^{i(\lambda-\delta)\phi}\,
d^s_{\lambda\delta}(\theta)\,{\cal M}^s_{\lambda_1\lambda_2}
\right\vert^2,$$ where $\delta\equiv\lambda_1-\lambda_2$ with $-s\leq\delta\leq s$. The reduced decay amplitude ${\cal M}^s_{\lambda_1\lambda_2}$ is only a function of $s$ and the helicities of the outgoing particles. It does not dependent on the azimuthal $\phi$ and polar $\theta$ angles. The $\theta$ dependence is concentrated only in the well-known $d$-functions $d^s_{\lambda\delta}(\theta)$. Furthermore, the absolute value of the dijet rapidity difference is related to the polar scattering angle $\theta$ with respect to the beam axis by the formula $\Delta y\equiv |y_1-y_2|=\ln[(1+|\cos\theta|)/(1-|\cos\theta|)]\ge 0$ and is invariant under boosts along the beam direction. The choice of the other variable $\chi\equiv\exp(\Delta y)=(1+|\cos\theta|)/(1-|\cos\theta|)\ge 1$ is motivated by the fact that the distribution of the Rutherford scattering is flat in this variable. These variables allow systematic consideration of angular decay distributions of resonances with different spins and different interactions with partons.
The simplest case of the resonance production of a (pseudo)scalar particle $h$ with spin 0 in the $s$-channel leads to a uniform decay distribution in the scattering angle $$\label{G0}
\frac{{\rm d} \Gamma_0(h\to q\bar{q})}
{{\rm d} \cos\theta} \propto \vert d^{\,0}_{00}\vert\,^2\sim 1.$$ The spin-1/2 fermion resonance, like an excited quark $q^*$, leads to asymmetric decay distributions for the given spin parton configurations $$\label{G12}
\frac{{\rm d} \Gamma_{1/2}(q^*\to qg)} {{\rm d} \cos\theta}\propto \vert d^{1/2}_{1/2,\pm 1/2}\vert\,^2\sim 1 \pm \cos\theta
.$$ However, the choice of the variables which depend on the absolute value of $\cos\theta$ cancels out the apparent dependence on $\cos\theta$. In other words, both distributions (\[G12\]) for dijet events look like uniform distributions in $\Delta y$ and $\chi$ According to the simple formula $$\label{trans}
\frac{{\rm d}\Gamma}{{\rm d}(\Delta y/\chi)}=
\frac{{\rm d}\cos\theta}{{\rm d}(\Delta y/\chi)}\;
\frac{{\rm d}\Gamma}{{\rm d}\cos\theta}\,,$$ the uniform distribution leads to kinematical peaks at the small values $\Delta y=0$ (the dotted curve in the left panel of Fig. \[fig:angular\]) and $\chi=1$ (the dotted curve in the right panel of Fig. \[fig:angular\]) $$ \frac{{\rm d}\Gamma_0}{{\rm d}\Delta y}\propto
\frac{{\rm e}^{\Delta y}}{({\rm e}^{\Delta y}+1)^2}
\qquad {\rm and} \qquad \label{chi0}
\frac{{\rm d}\Gamma_0}{{\rm d}\chi}\propto
\frac{1}{(\chi+1)^2}.$$
There are two different possibilities for spin-1 resonances. The gauge bosons, which are associated with additional $U(1)'$ gauge symmetry (or transform under the adjoint representation of the extra gauge group), are generally called $Z'(\boldsymbol{W}')$ particles. They have minimal gauge interactions with the known light fermions (see eqs. (\[Z’ed\]) and (\[W’\])), which preserve the fermion chiralities and possess maximal helicities $\lambda=\pm 1$. At a symmetric $pp$ collider, like the LHC, such interactions lead to the symmetric angular distribution (\[GLL\]) of the decay products over the polar angle $\theta$: $$\label{GLL11}
\frac{{\rm d} \Gamma_1(Z'\to q\bar{q})}
{{\rm d} \cos\theta} \propto
\vert d^1_{11}\vert^2+\vert d^1_{-11}\vert^2 \sim
1+\cos^2\theta \, .$$ Similar to the uniform distribution (\[G0\]), this one also leads to kinematical peaks at small values of $\Delta y=0$ (the dash-dotted curve in the left panel of Fig. \[fig:angular\]) and $\chi=1$ (the dash-dotted curve in the right panel of Fig. \[fig:angular\]) $$\label{dy1prime}
\frac{{\rm d}\Gamma'_1}{{\rm d}\Delta y}\propto
\frac{{\rm e}^{\Delta y}({\rm e}^{2\Delta y}+1)}{({\rm e}^{\Delta y}+1)^4}
\qquad {\rm and } \qquad \frac{{\rm d}\Gamma'_1}{{\rm d}\chi}\propto
\frac{\chi^2+1}{(\chi+1)^4}.$$ Another possibility is the resonance production and decay of new longitudinal spin-1 bosons with helicity $\lambda=0$. These bosons arise in many extensions [@Chizhov:2009fc] of SM which solve the Hierarchy problem. They are transformed as doublets $(Z^*\;W^*)$ under the fundamental representation of the SM $SU(2)_W$ group like the SM Higgs boson. They are above-mentioned extra chiral bosons (see (\[master\])). While the $Z'$ bosons with helicities $\lambda=\pm 1$ are produced in left(right)-handed quark and right(left)-handed antiquark fusion, the longitudinal $Z^*$ bosons can be produced through the anomalous chiral couplings with the ordinary light fermions in left-handed or right-handed quark-antiquark fusion [@Chizhov:2008tp]. As already noted before, these anomalous couplings lead to a different angular distribution of the resonance decay $$\label{GLR11}
\frac{{\rm d} \Gamma^*_1(Z^*\to q\bar{q})}
{{\rm d} \cos\theta} \propto
\vert d^1_{00}\vert\,^2\sim\cos^2\theta.$$ As has already been noted for the dilepton case the absence of the constant term in (\[GLR11\]) results in novel experimental signatures. First of all, the uniform distribution (\[G0\]) for scalar and spin-1/2 particles and the distribution (\[GLL11\]) for gauge vector bosons with minimal coupling include a nonzero constant term, which leads to a kinematic singularity in the transverse momentum distribution of the final parton (as for charged leptons, see (\[1/cos\]) and (\[narrow\])). After smearing of the resonance finite width the singularity is transformed into the well-known Jacobian peak (the dash-dotted curve in the left panel of Fig. \[fig:pt\]). The analytic expression of the $p_{\rm T}$ distribution describing the Jacobian peak with finite width can be found in [@Barger:1983wf]. Using the same method one can derive an analogous distribution for the excited bosons (the solid curve in the left panel of Fig. \[fig:pt\]). $$\label{pTanal}
\frac{{\rm d}\Gamma^*_1}{{\rm d}p_{\rm T}}\propto
p_{\rm T}\sqrt{\sqrt{\left(4p^2_{\rm T}-M^2\right)^2+\Gamma^2 M^2}-4p^2_{\rm
T}+M^2}.$$ In contrast to the previous case, the pole in the decay distribution of the excited bosons is canceled out and the final parton $p_{\rm
T}$ distribution has a broad smooth hump [@Chizhov:2006nw] with a maximum at $p_{\rm T}=\sqrt{(M^2+\Gamma^2)/8}\simeq M/\sqrt{8}$ below the kinematic endpoint $p_{\rm T}=M/2$ instead of a sharp Jacobian peak, which obscures their experimental identification as resonances. Therefore, the transverse jet momentum is not the appropriate variable for the excited boson search.
Another striking feature of the distribution (\[GLR11\]) is the forbidden decay direction perpendicular to the boost of the excited boson in the rest frame of the latter (the Collins–Soper frame [@Collins:1977iv]). It leads to a profound dip at $\cos\theta=0$ in the Collins–Soper frame [@Chizhov:2008tp] in comparison with the gauge boson distribution (the right panel in Fig. \[fig:pt\]). Similar dips also occur at the small values $\Delta y=0$ [@Chizhov:2010hh] (the solid curve in the left panel of Fig. \[fig:angular\]) and $\chi=1$ (the solid curve in the right panel of Fig. \[fig:angular\]) $$\label{dy1star}
\frac{{\rm d}\Gamma^*_1}{{\rm d}\Delta y}\propto
\frac{{\rm e}^{\Delta y}({\rm e}^{\Delta y}-1)^2}{({\rm e}^{\Delta y}+1)^4}
\qquad {\rm and } \qquad
\frac{{\rm d}\Gamma^*_1}{{\rm d}\chi}\propto
\frac{(\chi-1)^2}{(\chi+1)^4}.$$ It can be seen from Fig. \[fig:angular\] that the excited bosons have a unique signature in the angular distributions. They manifest themselves through the absolute minima at the small values $\Delta y=0$ and $\chi=1$ and absolute maxima right away from the origin. So, the rapidity difference distribution reaches the absolute maximum at $\Delta y=\ln(3+\sqrt{8})\approx 1.76$ and at $\chi=3$ for the angular distribution in the dijet variable $\chi$.
These features will be considered below in more detail. In order to have more practical analysis, it is convenient to use equidistant binning in $\log\chi$ [@Collaboration:2010eza], which corresponds to periodic cell granularity of the calorimeter in $\eta$. In this case the smooth $\chi$-spectra (see eqs. (\[dy1prime\]) and (\[dy1star\])) are transformed into histograms with the maximum in the lowest bin for the gauge bosons with the minimal coupling and with the maximum in the bin containing the value $\chi=3+\sqrt{8}\approx 5.8$ for the excited bosons (Fig. \[fig:chi\]).
Using distributions in (pseudo)rapidity and $\chi$ one can construct two useful ratios of numbers of events $N$ measured under specified experimental constraints schematically given below in brackets. The first is the wide-angle to small-angle ratio $$\label{wide-small}
R_\chi(a,b)=\frac{N(1<\chi<a)}{N(a<\chi<b)}$$ and the second is centrality or the $\eta$-ratio of both jets $$\label{centrality}
R_\eta(a,b)=\frac{N(|\eta_{1,2}|<a)}{N(a<|\eta_{1,2}|<b)}.$$
The ratios are less affected by the systematic uncertainties and can be used for searching for new physics in dijet data. To understand how they work, let us suppose that one has found some bump in the experimental dijet invariant mass distribution. Then we can compare the angular distributions for “on-peak” events (or events comprising the bump) and “off-peak” events (or events far away from the bump), using the aforementioned ratios. At the moment we ignore the complications of experimental separation of “on-peak” events from the “off-peak” ones. Since the QCD background is dominated by the Rutherford-like distribution, we can consider, as an approximation, a simple case where the QCD dijet $\chi$-distribution is flat. It means that for the selected equal kinematical regions $b-a=a-1$ ($b=2a-1$), the ratio $R_\chi$ for the “off-peak” events should be approximately one and does not depend on the dijet mass. When the “on-peak” events originate from the new physics and have the angular distribution different from the one predicted in QCD, the ratio $R_\chi$ should deviate from one. Due to an excess at small $\chi$ values irrespectiveof the maximal $\chi$-value (in our case equals to $a$) one has $R_\chi>1$ for all known exotic models, except the excited bosons, when one expects $R_\chi<1$.
In order to emphasize the effect of the excited bosons (to increase sensitivity to these bosons), we need to choose a value of $a$, which makes the ratio as small as possible $$\label{minChi}
R_\chi(a,2a-1)=
\frac{N_{\rm QCD}+N_{\rm new}(<a)}
{N_{\rm QCD}+N_{\rm new}(>a)}
\approx 1+
\frac{N_{\rm new}(<a)-N_{\rm new}(>a)}
{N_{\rm QCD}}<1.$$ Here $N_{\rm new}(<a)$ and $N_{\rm new}(>a)$ denotes the number of events generated by the new physics in the regions $1<\chi<a$ and $a<\chi<2a-1$, respectively. Simple integration gives the QCD contribution $N_{\rm QCD}(1<\chi<a) \propto \int^{a}_{1} d\chi
=N_{\rm QCD}(a<\chi<2a-1) \propto \int^{2a-1}_{a} d\chi = (a-1)$. Simialr integrations of the $\chi$-distribution (\[dy1star\]) in $1<\chi<a$ and $a<\chi<2a-1$ give $N_{\rm new}(<a) \propto (a-1)^3/(a+1)^3$ and $N_{\rm new}(>a) \propto (a-1)^3/a^3-N_{\rm new}(<a)$, respectively. Therefore, due to the monotonic increase of the distribution (\[dy1star\]) to the maximum at $\chi=3$ it is possible to reach the minimal value for (\[minChi\]) with the parameters $a\approx 1.87$ and $b\approx 2.74$. In our “ideal” case these parameters are optimal for the search for excited bosons with the ratio (\[wide-small\]). The larger value $a\ge
1/(\sqrt[3]{2}-1)\approx 3.85$ will lead to a compensation of the contributions from the low and high $\chi$-parts ($N_{\rm new}(<a) \approx N_{\rm new}(>a)$), and $R_\chi \ge 1$ again. For the QCD generated dijets the centrality ratio $R_\eta$ (\[centrality\]) is also almost constant and should not depend on the dijet invariant mass when the parameters $a$ and $b$ are fixed. When a dijet new-physics signal takes place, this ratio could deviate from its constant value. Signal events for almost all exotic models are expected in the central (pseudo)rapidity region. Therefore, one could see a bump in the $R_\eta$ distribution as a function of the dijet mass. Contrariwise, the signal from the excited bosons could lead to a novel signature: instead of the bump one will have a dip in the distribution at the resonance mass.
To investigate quantitatively this possibility we have again used the CompHEP package [@Boos:2004kh; @Pukhov:1999gg], which was extended with the excited bosons model of [@Chizhov:2010ry]. In particular, two dimensional pseudorapidity distributions were generated for “$\,2\to 2\,$” processes proceeding through the gauge $W'$ and excited $W^*$ boson resonances with the same mass 550 GeV$/c^2$ at $\sqrt{s}=7$ TeV $pp$ collider (Fig. \[fig:eta2dim\]). The CTEQ6L parton distribution functions were used. For both final jets we impose cuts on the pseudorapidity $|\eta|<2.5$ and the transverse momentum $p_{\rm T} > 30$ GeV/$c$.
To minimize the potential differences in jet response and efficiency of jet registration between the [*inner*]{} and [*outer*]{} dijet events one can choose the central region of the calorimeter $|\eta|<1$. Using calculations given in the scatter plots (Fig. \[fig:eta2dim\]) one can estimate the centrality ratios (\[centrality\]) for the gauge and excited bosons $$\label{etaRprime}
R'_\eta(0.5,1.0)\simeq 1.08 \qquad {\rm and} \qquad
R^*_\eta(0.5,1.0)\simeq 0.29.$$ The dramatic difference between these two numbers is clearly seen, it should lead to the corresponding experimental signature. Since the QCD ratio $R^{\rm QCD}_\eta(0.5,1.0)\simeq 0.6$ is located [*right between*]{} the numbers, the gauge bosons with the minimal coupling will lead to an increase in the QCD ratio at the resonance mass, while the excited bosons should decrease the ratio. It is interesting to notice that a hint of this type of the novel signature can be seen in the low-statistics distribution of the $\eta$-ratio versus the dijet invariant mass in the ATLAS data in approximately the same mass range, $450~{\rm GeV}<M_{jj}<600$ GeV, as for the resonance bump in the dijet events [@:2010bc].
Unfortunately, one should stress that the extensions of the “signal region” up to $|\eta|<1.3$ and the central region up to $|\eta|<0.7$ do not change drastically the QCD ratio $R^{\rm QCD}_\eta(0.7,1.3)\simeq 0.55$, but dilute the signal from the excited bosons since $R^*_\eta(0.7,1.3)\simeq 0.68$. In order to increase the sensitivity to the excited bosons one can consider the centrality ratio only for the dijet events with the opposite pseudorapidities $\left.R_\eta\right|_{(\eta_1\cdot\eta_2)\le 0}={\cal R}_\eta$ (the hatched regions in Fig. \[fig:eta2dim\]). In this case the difference between ${\cal R}'$ and ${\cal R}^*$ increases $$\label{etaR0prime}
{\cal R}'_\eta(0.5,1.0)\simeq 1.12
\qquad {\rm and} \qquad
{\cal R}^*_\eta(0.5,1.0)\simeq 0.25,$$ but we lose half of the statistics. Therefore, it is convenient to consider the distribution in $\Delta\eta\equiv|\eta_1-\eta_2|\ge0$ for the events in the rectangle region $\Delta\eta<b$ and $\eta^{}_B\equiv|\eta_1+\eta_2|<c$. The cut $\eta_B<c$ is necessary to reduce the effect of the parton distribution functions on different $\Delta\eta$ bins. The corresponding centrality ratio $R_{\Delta\eta}$ is defined as $$\label{deltaEtaR}
R_{\Delta\eta}(a,b,c)=
\left.\frac{N(\Delta\eta<a)}{N(a<\Delta\eta<b)}\right|_{\eta^{}_B<c}.$$ The normalized histograms of the $\Delta\eta$-spectra and the theoretical curves are shown in Fig. \[fig:eta1\] for the following parameters values: $b=3.5$ and $c=1.5$.
It can be seen from the figure that the theoretical distributions describe very well the above-mentioned simulation data. In the same way as was done for the $\chi$-distribution of the excited bosons one can maximize the deviation of $R_{\Delta\eta}(a,b,c)$ from the QCD ratio $R^{\rm QCD}_{\Delta\eta}=N_{\rm QCD}^{(a)}/N_{\rm QCD}^{(b-a)}$. One finds the desired minimum of $R_{\Delta\eta}(a,b,c)$, which corresponds to the maximal deviation from QCD, for $c=1.5$, $b=1$ and $a=0.67$ [@Chizhov:2010jg].
Conclusion
==========
In this review phenomenological consequences of the Standard Model extension by means of new spin-1 chiral fields with the internal quantum numbers of the electroweak Higgs doublets are summarized. It is worth stressing that the new type of spin-1 chiral bosons can exist. They are well motivated from the point of view of the Hierarchy problem and are predicted by at least three different classes of theories that represent different approaches to explanation of the relative lightness of the Higgs doublets. The decay distributions of the chiral bosons differ drastically from the distributions of the known gauge bosons and can be distinguished from others. The discovery of this type of distributions will point to existence of compositeness, new symmetry and even extra dimensions.
The prospects for resonance production and detection of the chiral vector $Z^*$ and $W^{*\pm}$ bosons at the LHC energies are considered on the basis of quantitative simulations within the CompHEP/CalcHEP package. The experimental signatures of the excited chiral heavy $Z^*$ bosons are considered and compared with those of the gauge $Z'$ bosons.
The $Z^*$ boson can be observed as a Breit-Wigner resonance peak in the invariant dilepton mass distributions in the same way as the well-known extra gauge $Z'$ bosons. This naturally puts the chiral bosons on the list of very interesting objects for early searches with the first LHC data. Moreover, $Z^*$ bosons have unique signatures in transverse momentum, angular and pseudorapidity distributions of the final decay products, which allow one to distinguish them from other heavy neutral resonances. In particular, there is no Jacobian peak in the transverse momentum distribution of the decay products, and the angular distribution (in the Collins-Soper frame for high on-peak invariant masses) has a peculiar “swallowtail” shape.
In 2010, with 40 pb$^{-1}$ of the LHC proton-proton data at energy 7 TeV, the ATLAS detector was used to search for narrow resonances in the invariant mass spectrum of $e^+e^-$ and $\mu^+\mu^-$ final states and high-mass charged states decaying to a charged lepton and a neutrino. No statistically significant excess above the Standard Model expectation was observed. Therefore, low mass limits of 1.15 TeV$/c^2$ and 1.35 TeV$/c^2$ were obtained for the neutral chiral $Z^*$ and charged $W^*$ bosons, respectively. These are the first direct limits on the $W^*$ and $Z^*$ boson production.
Finally, a novel strategy for the neutral chiral boson search in the LHC dijet data is discussed. For almost all currently considered exotic models the relevant signal is expected in the central dijet rapidity region $y_{1,2}\simeq 0$ and $|y_1-y_2|\simeq 0$. Contrariwise, the excited bosons do not contribute to this region but produce an excess of dijet events far away from it. In particular, for these bosons the appropriate kinematic restrictions can lead to a dip in the centrality ratio distribution over the dijet invariant mass instead of a bump expected in the most exotic models.
We expect that the experimental results, presented here and based on low statistics of 2010 will be very soon improved with much higher statistics of 2011. Furthermore, we plan to extend our search for excited bosons by means of thorough investigation of the dijet data from the ATLAS detector at the LHC.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are very grateful to O. Fedin and his colleagues for fruitful cooperation, V.G. Kadyshevsky and N.A. Russakovich for support and interest in this research. The work of M.V. Chizhov was partially supported by the grant of the Plenipotentiary of the Republic of Bulgaria in JINR for the year 2011.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Qing-Guo Huang [^1]\'
title: 'Scale dependence of $f_{NL}$ in N-flation'
---
Introduction
============
In the single-component slow-roll inflation the different Fourier modes of the curvature perturbation are roughly uncorrelated and their distribution is almost Gaussian [@Maldacena:2002vr]. Using $\delta N$ formalism [@Starobinsky:1986fxa], the curvature perturbation can be expanded to the non-linear orders =N=\_i N\_[,i]{}\_i+\_[i,j]{} N\_[,ij]{} \_i\_j+[16]{} \_[i,j,k]{} N\_[,ijk]{} \_i\_j\_k+ ... , \[deltaN\] where the subscript ${,i}$ denotes the derivative with respect to $\phi_i$. The above formula indicates that the curvature perturbation is determined by the evolution of the homogeneous universe corresponding both to the classical inflation trajectory and to nearby trajectories. In principle, an important difference for the multi-field model from the single-component case is that the classical trajectory is not uniquely specified by the potential, but rather has to be given as a separate piece of information. It provides an opportunity to generate a large non-Gaussianity in multi-field slow-roll inflation even at the inflationary epoch [@Byrnes:2008wi; @Byrnes:2008zy; @Kim:2010ud], not at the end of inflation [@Lyth:2005qk; @Sasaki:2008uc; @Huang:2009vk]. See some other relevant references in [@Vernizzi:2006ve; @Kim:2006ys; @Piao:2006nm; @Kim:2006te; @Battefeld:2006sz; @Battefeld:2007en; @Langlois:2008vk; @Wang:2010si; @Mulryne:2010rp; @Suyama:2010uj].
On the other hand, not only the size of bispectrum $f_{NL}$ is measurable, but also its scale dependence characterized by $n_{f_{NL}}$ is possibly measured by the accurate cosmological observations, such as Planck and CMBPol. In [@Sefusatti:2009xu] the authors concluded that Planck and CMBPol are able to provide a 1-$\sigma$ uncertainty on the spectral index of $f_{NL}$ for local form bispectrum as follows n\_[f\_[NL]{}]{}0.1 [50f\_[NL]{}]{}[1]{}, and n\_[f\_[NL]{}]{}0.05 [50f\_[NL]{}]{}[1]{}. Recently the scale dependence of $f_{NL}$ for inflation and curvaton model is discussed in [@Byrnes:2009pe; @Byrnes:2010ft; @Byrnes:2010xd; @Huang:2010cy; @Riotto:2010nh] and a large parameter space for generating a detectable $n_{f_{NL}}$ in curvaton model is illustrated in [@Byrnes:2010xd; @Huang:2010cy]. The value of $n_{f_{NL}}$ will be an important discriminator to distinguish different models.
In this paper we focus on a class of multi-field inflation, so-called N-flation, in which there are no cross couplings among the inflaton fields or the cross couplings are negligibly small. For an example, we consider the axion N-flation with a detectably large $f_{NL}$. But we find that $n_{f_{NL}}$ is negligibly small when all inflatons have the same potential. However, a negative $n_{f_{NL}}$ with detectably large absolute value can be obtained in the model with different decay constants for different inflatons.
Our paper is organized as follows. In Sec. 2 we derive the spectral index of $f_{NL}$ for general N-flation. We explore the observables, including $n_{f_{NL}}$, in the axion N-flation in Sec. 3. Some discussions are contained in Sec. 4.
Scale dependence of $f_{NL}$ in general N-flation
=================================================
In this section we will derive the spectral index of $f_{NL}$ in general N-flation with $N_f$ uncoupled inflaton fields. The potential of inflatons is given by V=\_[i=1]{}\^[N\_f]{} V\_i(\_i). The dynamics of inflation is governed by H\^2&=&[13]{}\_[i=1]{}\^[N\_f]{} $\half {\dot \phi}_i^2+V_i(\phi_i)$,\
\_i+3H\_i&=&-V\_i’, where $V_i'= \p V_i/\p \phi_i$. In this paper we work on the unit of $M_p=1$. In the slow-roll limit, the above equations are simplified to be H\^2&&[V3]{}, \[friedm\]\
3H\_i&& -V\_i’. Similarly, we introduce some ‘slow-roll’ parameters, such as \_i$V_i'\over V_i $\^2. From Eq.(\[friedm\]), we find H=-[16]{} \_i V\_i’\^2/V, and then \_H-[HH\^2]{}=, or equivalently, \_H=\_i ([V\_iV]{})\^2 \_i. Inflation happens when $\epsilon_H<1$. The multi-field slow-roll inflation can be achieved even when some inflaton fields have steep potentials because the friction term for each inflaton is contributed from all of fields.
Following [@Lyth:1998xn], the number of e-folds before the end of inflation is given by N= \_[i=1]{}\^[N\_f]{} \_[\_i\^[end]{}]{}\^[\_i]{} [V\_iV\_i’]{}d\_i. \[efolds\] Therefore we obtain N\_[,i]{}=[V\_iV\_i’]{}, and N\_[,ij]{}=$1-{V_iV_i''\over V_i'^2}$\_[ij]{}. Using the $\delta N$ formalism [@Starobinsky:1986fxa], the power spectrum of scalar perturbation is given by P\_=${H\over 2\pi}$\^2 \_i N\_[,i]{}\^2. A slow-roll inflation model predicts a near scale-invariant power spectrum of scalar perturbation. The deviation from exact scale invariance is described by the spectral index $n_s$ which is defined by n\_s1+[dP\_dk]{} =1-2\_H-2[\_[i,j]{}\_[ij]{}N\_[,i]{}N\_[,j]{}\_k N\_[,k]{}\^2]{}, where \_[ij]{}-[V\_i”V]{}\_[ij]{}. See [@Lyth:1998xn; @Sasaki:1995aw] in detail. In this paper, we assume, without loss of generality, $\dot \phi_i<0$, so that $V_i'>0$ for $\forall \ i$. On the other hand, the gravitational wave perturbation is also generated during inflation and the amplitude of its power spectrum is determined by the inflation scale P\_T=[H\^2\^2/2]{}, and then the tensor-scalar ratio $r_T$ becomes r\_T=16/\_i [1\_i]{} . . The amplitude of gravitational perturbation is also near scale-invariant and its tilt is defined by n\_T=-2\_H. In single-field inflation, $r_T=16\epsilon_H$ which leads to a consistency relation $n_T=-r_T/8$. For $N$-field inflation, it is modified to be n\_T-r\_T/8. Here we used the inequality \_H1/\_i[1\_i]{}. , where the equality is satisfied when $\epsilon_i=\epsilon_H V/V_i$ for $\forall \ i$.
From Eq.(\[deltaN\]), the non-Gaussianity parameter $f_{NL}$ can be written by f\_[NL]{}=[56]{}\_[i,j]{}[N\_[,i]{}N\_[,j]{}N\_[,ij]{}(\_k N\_[,k]{}\^2)\^2]{}. Since $N_{,ij}\propto \delta_{ij}$ for N-flation, the above formula is simplified to be f\_[NL]{}=\_[i]{}f\_[NL]{}\^[i]{}, \[sfnli\] where f\_[NL]{}\^[i]{}=[56]{} [N\_[,i]{}\^2N\_[,ii]{}(\_k N\_[,k]{}\^2)\^2]{}=[56]{}[r\_T\^2128]{} [1\_i]{}$1-{V_iV_i''\over V_i'^2}$. \[fnli\] Following [@Byrnes:2010ft], the spectral index of $f_{NL}$ for N-flation is n\_[f\_[NL]{}]{}=[1f\_[NL]{}]{}\_i f\_[NL]{}\^i (2n\_[multi,i]{}+n\_[f,i]{}), \[nfnl\] where n\_[multi,i]{}&=& 2\_[k,l]{} \_[kl]{}${N_{,k}N_{,l}\over \sum_j N_{,j}^2}-\delta_{il}{N_{,k}\over N_{,i}}$,\
n\_[f,i]{}&=&-\_k [N\_[,k]{}F\_[kii]{}\^[(2)]{}N\_[,ii]{}]{}, and F\_[kii]{}\^[(2)]{}=-2([V\_i’ V]{})\^2 [V\_k’V]{}+[V\_i”V]{}[V\_k’V]{}+2 [V\_i’V]{}[V\_i”V]{}\_[ki]{}-[V\_i”’V]{}\_[ki]{}. Considering $N_{,i}=V_i/V_i'$, we obtain n\_[multi,i]{}&=&(1-n\_s-2\_H)-4[V\_iV]{}\_i $1-{V_iV_i''\over V_i'^2}$, \[nmulti\]\
n\_[f,i]{}&=&4 ([V\_iV]{})\^2\_i-[V\_iV]{}${V_i'' \over V_i}-{V_i'''\over V_i'}$/$1-{V_iV_i''\over V_i'^2}$ . . \[nf\] Using Eqs.(\[sfnli\]), (\[fnli\]), (\[nfnl\]), (\[nmulti\]) and (\[nf\]), we can easily calculate the spectral index of $f_{NL}$ in general N-flation. The first term in (\[nmulti\]) contribute $2(1-n_s-2\epsilon_H)$ to $n_{f_{NL}}$. If $\epsilon_H\ll1$, $2(1-n_s-2\epsilon_H)\simeq 0.08$ for $n_s=0.96$.
In addition, we are also interested in the trispectrum in this case. See the Appendix. \[trispec\] in detail.
Axion N-flation
===============
In this section we focus on axion N-flation model [@Kim:2010ud] in which there are $N_f$ inflaton fields and the potential of $\phi_i$ is given by V\_i=\_i\^4 (1-\_i), where $\alpha_i=2\pi \phi_i/f_i$ and $f_i$ is the i-th axion decay constant. The ‘slow-roll’ parameters $\epsilon_i$’s are given by \_i= [2\^2 f\_i\^2]{}[1+\_i1-\_i]{}, and N\_[,ii]{}=1-[V\_iV\_i”V\_i’\^2]{}=[11+\_i]{}. Now the number of e-folds before the end of inflation becomes N\_[i=1]{}\^[N\_f]{} ${f_i\over 2\pi}$\^2 . Quoting the results in [@Kim:2010ud], the observables of axion N-flation are P\_&=& [H\^28\^2]{}\_i [1\_i]{},\
n\_s&=&1- 2\_H-[8\^23H\^2]{}\_i[\_i\^4f\_i\^2]{}[1\_i]{}/ \_j [1\_j]{} . ,\
r\_T&=&16/ \_i [1\_i]{} . ,\
f\_[NL]{}&=& [56]{} [r\_T\^2128]{}\_i [1\_i]{}[11+\_i]{}. Using the results in the previous section, we can calculate the spectral index of $f_{NL}$ for the axion N-flation, namely n\_[multi,i]{}&=&(1-n\_s-2\_H)-[8\^23H\^2]{}[\_i\^4f\_i\^2]{},\
n\_[f,i]{}&=&4([V\_iV]{})\^2\_i-2[V\_iV]{}\_i, and then n\_[f\_[NL]{}]{}&=&2(1-n\_s-2\_H)-[16\^23H\^2]{} \_i [\_i\^4 f\_i\^2]{} [1\_i]{}[11+\_i]{} / \_j [1\_j]{}[11+\_j]{} .\
&+& 2 \_i $2({V_i\over V})^2-{V_i\over V}$[11+\_i]{} / \_j [1\_j]{}[11+\_j]{} . . \[sfnl\] The first and second lines in Eq.(\[sfnl\]) are contributed by $n_{multi,i}$ and $n_{f,i}$ respectively. Since the second term in $n_{multi,i}$ is large compared to $n_{f,i}$ for $\alpha_i\simeq \pi$ which corresponds to a large $f_{NL}$, the spectral index of $f_{NL}$ is mainly determined by $n_{multi,i}$.
Axion N-flation with the same potential for different inflatons {#ans}
---------------------------------------------------------------
In this subsection we discuss the model with $\Lambda_i=\Lambda$ and $f_i=f$ for $\forall \ i$ in detail. Now the tensor-scalar ratio and $f_{NL}$ become r\_T&=&[32 \^2f\^2]{}/ \_i[1-\_i1+\_i]{}. ,\[rtff\]\
f\_[NL]{}&=& [10\^23f\^2]{} \_i[1-\_i(1+\_j)\^2]{} /$ \sum_j{1-\cos\alpha_j\over 1+\cos\alpha_j}$\^2 . , \[fnlff\] and the spectral indices of scalar power spectrum and $f_{NL}$ are simplified to be n\_s&=&1- 2\_H-[8\^2 \^43H\^2f\^2]{},\
n\_[f\_[NL]{}]{}&=& [f\^28\^2]{}(1-n\_s-2\_H)\^2 \_i [1\_i]{}(1-\_i) / \_j [1\_j]{}[11+\_j]{} .\
&-&[5r\_T96]{} [1f\_[NL]{}]{} (1-n\_s-2\_H), where = 1/ \_i(1-\_i). , and \_H= [2\^2f\^2]{} \_i [1-2\_i2]{} / $\sum_j (1-\cos\alpha_j)$\^2 . . We see that $n_{multi,i}=0$ and $n_{f_{NL}}$ is expected to be small in this simple setup. We will show the smallness of $n_{f_{NL}}$ and extend the discussions in [@Kim:2010ud] in the following two subsubsections.
### $N_*\simeq N_{tot}$ {#nnt}
Assuming the initial conditions for $\alpha_i$ satisfy a uniform distribution, the number of e-folds before the end of inflation is roughly given by N\_[tot]{}f\^2 N\_f. A large enough total number of e-folds, for example not less than 60, can be achieved as long as the number of inflatons is large enough, namely $N_f\gtrsim 1.7\times 10^3/f^2$.
Similar to [@Kim:2010ud], we consider the case with $N_*\simeq N_{tot}$ in this subsubsection, where $N_*$ denotes the number of e-folds corresponding to the CMB scale. So we have N\_f f\^2, and $\Lambda^4/(3H^2)\simeq 1/N_f$, $\epsilon_H\simeq \pi^2/(N_f f^2)$. The spectral index of the scale power spectrum is related to $N_*$ by n\_s1-[52 N\_\*]{}, which is the same as that in [@Kim:2010ud]. For $N_*\simeq 60$, $n_s\simeq 0.942$. The numerical results for $N_*\simeq N_{tot}\simeq 60$ and $f=1/2$ are illustrated in Fig. \[fig:random\].
![The observables of axion N-flation with the same potential for different inflatons. Here $f=1/2$ and the initial conditions of $\alpha_i$ is assumed to be the uniform distribution.[]{data-label="fig:random"}](randomrnsfnl.eps "fig:"){width="10cm"}\
![The observables of axion N-flation with the same potential for different inflatons. Here $f=1/2$ and the initial conditions of $\alpha_i$ is assumed to be the uniform distribution.[]{data-label="fig:random"}](randomnfnl.eps "fig:"){width="10cm"}
The spectral index of scalar power spectrum is nicely compatible with the estimation, and the tensor-scalar ratio is roughly ${\cal O}(10^{-5}\sim 10^{-3})$. As long as $N_f$ is large enough, from Eqs. (\[rtff\]) and (\[fnlff\]), both $r_T$ and $f_{NL}$ are proportional to $1/f^2$. From the numerical results, we find that $f_{NL}$ is bounded from above, namely f\_[NL]{}. \[bfnlx\] For $f\lesssim 1$, the value of $f_{NL}$ can be of order ${\cal O} (10)$ which is large enough to be detected in the near future, but its tilt $n_{f_{NL}}$ $(-{\cal O}(10^{-7})\sim -{\cal O}(10^{-14}))$ is too small to be detected in this case.
### A few inflatons stays around the hilltop {#hilltop}
In [@Kim:2010ud], the authors proposed an alternative way to achieve a large $f_{NL}$. Roughly speaking, the summation of the function which is proportional to $1/\epsilon_i$ is dominated by those fields with the smallest $\epsilon_i$, or equivalently staying around the hilltop. Suppose some number $\bar N$ of fields have roughly comparable $\epsilon_i$, of order $\bar \epsilon$ which is related to the angle $\bar \alpha$ by |(1+) (-[|]{})\^2, with $(\pi-{\bar \alpha})\ll1$. In this case, $P_\zeta$, $r_T$, and $f_{NL}$ becomes P\_&&[H\^28\^2]{} [[|N]{}]{}, \[palpha\]\
r\_T&&[16[|]{}]{}(-[|]{})\^2,\
f\_[NL]{}&& [5\^23[|N]{} f\^2]{}. \[fnlb\] Considering ${\bar N}\geq 1$, f\_[NL]{}, which is the same as that in Sec. \[nnt\]. Now the spectral index of $f_{NL}$ is given by n\_[f\_[NL]{}]{}-[1+2]{} (1-n\_s-2\_H), which is much smaller than $(1-n_s-2\epsilon_H)$. So the spectral index of $f_{NL}$ is undetectable in this case as well.
Because the potential become very flat around the hilltop, one may worry that the classical motions of those inflatons staying around the hilltop are not reliable. Requiring that the classical motion of inflaton ${\bar \phi}=f{\bar \alpha}/2\pi$ per Hubble time be not less than its quantum fluctuation $H/2\pi$ yields (-[|]{})=2\^[-1]{}[Hf]{}, where $\xi=8\pi^2\Lambda^4/(3H^2f^2)=(1-n_s-2\epsilon_H)\ll 1$. [^2] On the other hand, from (\[palpha\]), $(\pi-{\bar \alpha})$ can be estimated by (-[|]{})=[ Hf2\^2 \_[R]{}]{}, where $\Delta_{\cal R}=\sqrt{P_\zeta}$. Therefore f\^2. Taking Eq.(\[fnlb\]) into account, $f_{NL}$ is bounded from above by f\_[NL]{}/ . . WMAP normalization [@Komatsu:2010fb] implies $P_\zeta=\Delta_{\cal R}^2=2.461\times 10^{-9}$, and then $f_{NL}\lesssim 8\times 10^3 \xi/\sqrt{\bar N}$. Since $\xi\lesssim {\cal O}(10^{-2})$, $f_{NL}\lesssim 80/\sqrt{\bar N}$. The value of $f_{NL}$ cannot be very large in this case.
Axion N-flation with different potential for different inflatons
----------------------------------------------------------------
In this subsection, we consider a class of axion N-flation in which $\Lambda_i=\Lambda$ for $\forall \ i$, but the decay constants can be different for different inflatons. For simplicity, we only focus on the cases with only two different decay constants.
### One inflaton with different decay constant stays around the hilltop {#dnnt}
In this subsubsection we consider a model with potentials V(\_i)&=&\^4 (1-\_i),\
V(|)&=&\^4(1-), where $\alpha_i=2\pi \phi_i/f$ for $i=1,2,...,N_f$ and $\bar\alpha=2\pi {\bar \phi}/{\bar f}$. Here $f$ is assumed to be different from $\bar f$. We consider $N_f\gg 1$ and the initial values of $\alpha_i$ satisfy a uniform distribution. [^3] The total number of e-folds before the end of inflation is given by N\_[tot]{}f\^2 N\_f+[[|f]{}\^24\^2]{}. We assume $\bar\phi$ stays around the hilltop of its potential. Requiring that the classical motion of inflaton ${\bar \phi}$ is reliable leads to (-[|]{}\_[ini]{}), \[dbalpha\] where ${\bar \alpha}_{ini}$ denotes the initial value of ${\bar \alpha}$.
In this model, the Hubble parameter is roughly related to $\Lambda$ by $H^2\simeq N_f \Lambda^4/3$. Because the inflaton staying around the hilltop almost does not move, we also have \_H /$N_{tot}-{{\bar f}^2\over 2\pi^2}\ln {2\over \pi-\bar \alpha}$ . . The spectral index of power spectrum and $f_{NL}$ become n\_s&& 1-[2N\_[tot]{}-[[|f]{}\^22\^2]{}]{}$$1+4 {c_r(N_f)N_f +{4\over (\pi-\bar\alpha)^2}\over c_r(N_f)N_f +(\bar f/f)^2{4\over (\pi-\bar\alpha)^2}}$$, \[nss\]\
n\_[f\_[NL]{}]{}&& [8 2N\_[tot]{}-[[|f]{}\^22\^2]{}]{}$${c_r(N_f)N_f +{4\over (\pi-\bar\alpha)^2}\over c_r(N_f)N_f +(\bar f/f)^2{4\over (\pi-\bar\alpha)^2}}-{d_r(N_f)N_f +{8\over (\pi-\bar\alpha)^4}\over d_r(N_f)N_f +(\bar f/f)^2{8\over (\pi-\bar\alpha)^4}}$$,\
where c\_r(N)=[1N]{}\_[i=1]{}\^N [1-\_i1+\_i]{}, d\_r(N)=[1N]{}\_[i=1]{}\^N [1-\_i(1+\_i)\^2]{}, which depend on the total number of points and the random choice of the series ‘$r$’ for $\alpha_i\in[0,\pi]$.
Similar to Sec. \[nnt\], we consider that the total number of e-folds corresponds to the CMB scale, namely $N_*\simeq N_{tot}\simeq 60$. Considering ${\bar f}\lesssim 1$, the total number of e-fold is mainly contributed by $\phi_i$ for $i=1,2,...,N_f$. If $f={\bar f}$, $n_s\simeq 1-5\ln 2/N_*$ and $n_{f_{NL}}\simeq 0$ which are consistent with the results in Sec. \[nnt\].
For $\bar f<f$, the second term in the bracket of Eq. (\[nss\]) is larger than 4 and then $n_s\lesssim 1-5\ln 2/N_*$. For $\bar f>f$, we find $n_s\gtrsim 1-5\ln 2/N_*$. In general, it is difficult to get the analytical estimation for the model with $f\neq \bar f$, but we can expect that the results should be quite different the model with the same decay constants. The numerical results are showed in Fig. \[fig:drandom\].
![The observables of axion N-flation with different decay constants. Here we adopt $f=1/2$ and $(\pi-{\bar \alpha}_{ini})/\pi=2\times 10^{-4}$. The blue and green dots correspond to ${\bar f}=1/4$ and the red dots corresponds to ${\bar f}=1$. The green dots illustrate the case with detectable scale dependence of $f_{NL}$.[]{data-label="fig:drandom"}](drnsfnl.eps "fig:"){width="10cm"}\
![The observables of axion N-flation with different decay constants. Here we adopt $f=1/2$ and $(\pi-{\bar \alpha}_{ini})/\pi=2\times 10^{-4}$. The blue and green dots correspond to ${\bar f}=1/4$ and the red dots corresponds to ${\bar f}=1$. The green dots illustrate the case with detectable scale dependence of $f_{NL}$.[]{data-label="fig:drandom"}](dnfnl.eps "fig:"){width="10cm"}
Here we keep $f=1/2$ and $(\pi-{\bar \alpha}_{ini})/\pi=2\times 10^{-4}$ fixed and explore the results for two choices of $\bar f$. From Fig. \[fig:drandom\], we see that a detectable negative $n_{f_{NL}}$ is possibly generated for ${\bar f}=1/4$. We also check the requirement in (\[dbalpha\]) which is satisfied as long as $(\pi-{\bar \alpha})/\pi\gtrsim 2\times 10^{-5}$ for ${\bar f}=1/4$. Unfortunately, the spectral index $n_s\ (\lesssim 0.91)$ in the case with detectable $n_{f_{NL}}$ is small compared to WMAP 7yr data [@Komatsu:2010fb] $(n_s=0.968\pm 0.012)$. See the green dots in Fig. \[fig:drandom\]. For ${\bar f}=1$, the spectral index can fit the WMAP 7yr data nicely, but the scale dependence of $f_{NL}$ becomes negligibly small. See the red dots in Fig. \[fig:drandom\]. This setup indicates that a large $n_{f_{NL}}$ can be achieved in axion N-flation.
### A general axion N-flation with two kinds of inflatons {#dnntNN}
In this subsubsection we consider a more general axion N-flation model in which there are two kinds of inflatons who have different decay constants, V\_1(\_[1,i]{})&=&\^4 (1-\_[1,i]{}),\
V\_2(\_[2,j]{})&=&\^4(1-\_[2,j]{}), where $\alpha_{1,i}=2\pi \phi_{1,i}/f_1$ for $i=1,2,...,N_1$, and $\alpha_{2,j}=2\pi \phi_{2,j}/f_2$ for $j=1,2,...,N_2$. Here $f_1$ is assumed to be different from $f_2$. We also consider $N_1,\ N_2\gg 1$ and the initial values of $\alpha_{1,i}$ and $\alpha_{2,j}$ satisfy a uniform distribution respectively. In this model the total number of e-folds before the end of inflation is given by N\_[tot]{}(f\_1\^2 N\_1+f\_2\^2N\_2). In this model, the Hubble parameter is related to $\Lambda$ by $H^2=(N_1+N_2)\Lambda^4/3$ and the slow-roll parameter is given by \_H=[\^2(N\_1+N\_2)\^2]{} $N_1/f_1^2+N_2/f_2^2$. Similarly, we can also derive the spectral index of power spectrum and $f_{NL}$, namely n\_s&& 1-[2\^2(N\_1+N\_2)\^2]{} ${N_1\over f_1^2}+{N_2\over f_2^2}$-[8\^2N\_1+N\_2]{}[c\_[r\_1]{}(N\_1)N\_1 +c\_[r\_2]{}(N\_2)N\_2 c\_[r\_1]{}(N\_1)f\_1\^2N\_1 +c\_[r\_2]{}(N\_2)f\_2\^2N\_2 ]{},\
n\_[f\_[NL]{}]{}&& [16\^2N\_1+N\_2]{} $${c_{r_1}(N_1)N_1 +c_{r_2}(N_2)N_2 \over c_{r_1}(N_1)f_1^2N_1 +c_{r_2}(N_2)f_2^2N_2 } - {d_{r_1}(N_1)N_1 +d_{r_2}(N_2)N_2 \over d_{r_1}(N_1)f_1^2N_1 +d_{r_2}(N_2)f_2^2N_2 }$$. Again, if $f_1=f_2$, the scale dependence of $f_{NL}$ is negligibly small and our results recover those in Sec. \[nnt\].
For the case with $f_1\neq f_2$, we need to adopt numerical method to explore it. Considering $N_*\simeq N_{tot}\simeq 60$, our numerical results are illustrated in Fig. \[fig:drandomNN\].
![The observables of a general axion N-flation with different decay constants. Here we adopt $f_1=1/2$ and $f_2=1/4$. The green dots illustrate the case with detectable scale dependence of $f_{NL}$.[]{data-label="fig:drandomNN"}](drnsfnlNN.eps "fig:"){width="10cm"}\
![The observables of a general axion N-flation with different decay constants. Here we adopt $f_1=1/2$ and $f_2=1/4$. The green dots illustrate the case with detectable scale dependence of $f_{NL}$.[]{data-label="fig:drandomNN"}](dnfnlNN.eps "fig:"){width="10cm"}
Here we adopt $f_1=1/2$, $N_1=4800$, $f_2=1/4$ and $N_2=8000$. We see that a detectable scale dependence of $f_{NL}$ can be obtained.
Discussions
===========
In this paper we work out the spectral index of $f_{NL}$ in general N-flation model by adopting the horizon-crossing approximation. For an instance, we focus on the axion N-flation proposed in [@Kim:2010ud] and we find that $f_{NL}$ can be large enough to be detected in the near future.\
$\bullet$ All inflatons have the same potential. In the case with the total number of e-folds corresponding to the CMB scale, $f_{NL}$ is roughly of order ${\cal O}(10)$ for a uniform distribution of initial angles if the decay constant $f$ is around the Planck scale. Our numerical results indicate that $f_{NL}$ is bounded from above by $5\pi^2/3f^2$, and $\tau_{NL}\simeq (2\pi/f)^{2/3}f_{NL}^{5/3}$. Another way to obtain a large $f_{NL}$ is that a few inflatons stay around the hilltop and the summations in the formula of $f_{NL}$ is mainly contributed by these inflatons. Typically $f_{NL}$ is less than one hundred in the second case. In both cases the tilt of $f_{NL}$ is negligibly small. The smallness of $n_{f_{NL}}$ in these two case comes from the accident cancellation in $n_{multi,i}$ because all inflatons have the same potential.\
$\bullet$ Different inflatons have different decay constants. In this kind of model, the accident cancellation in $n_{multi,i}$ does not happen and a negative detectable $n_{f_{NL}}$ can be obtained. It implies that a scale independent $f_{NL}$ is not generic prediction of N-flation. A more general axion N-flation is left to be studied in the future.
In addition, we also study the trispectrum for N-flation in the Appendix. \[trispec\]. We find that there is a universal relation between $g_{NL}$ and $\tau_{NL}$, namely $g_{NL}\simeq 25 \tau_{NL}/27$, for the cases with large $f_{NL}$. But the relation between $f_{NL}$ and $\tau_{NL}$ depends on detail of the model.
If there is only one inflaton field, the axion N-flation becomes natural inflation. The slow-roll parameters are roughly given by $M_p^2/f^2$ and then the slow-roll conditions are satisfied if $f>M_p$. However such a large decay constant cannot be achieved in some fundamental theories, such as string theory [@ArkaniHamed:2003wu; @Banks:2003sx]. In [@Dimopoulos:2005ac] the N-flation is supposed to realize a slow-roll inflation even when $f<M_p$ by calling multi inflaton fields. On the other hand, in [@Huang:2007st] we proposed a conjecture that the effective gravity scale should be $\Lambda_G=M_p/\sqrt{N_f}$ for the case with multi scalar fields, not $M_p$. If the decay constant of axion field is required to be smaller than $\Lambda_G$, namely $N_f f^2<1$, the total number of e-folds is less than one and the axion N-flation fails.
[**Acknowledgments**]{}
We would like to thank S. A. Kim and A. R. Liddle for useful discussions. QGH is supported by the project of Knowledge Innovation Program of Chinese Academy of Science and a grant from NSFC.
The trispectrum in N-flation {#trispec}
============================
From Eq. (\[deltaN\]), the size of trispectrum is characterized by two parameters $\tau_{NL}$ and $g_{NL}$ which are defined by \_[NL]{}&=&\_[i,j,k]{} N\_[,ij]{}N\_[,ik]{}N\_[,j]{}N\_[,k]{}/ (\_l N\_[,l]{}\^2)\^3 . ,\
g\_[NL]{}&=& [2554]{} \_[i,j,k]{} N\_[,ijk]{}N\_[,i]{}N\_[,j]{}N\_[,k]{}/ (\_l N\_[,l]{}\^2)\^3 . . For N-flation, the above two equations are simplified to be \_[NL]{}&=&\_i N\_[,ii]{}\^2 N\_[,i]{}\^2 / (\_l N\_[,l]{}\^2)\^3 . ,\
g\_[NL]{}&=& [2554]{} \_i N\_[,iii]{}N\_[,i]{}\^3 / (\_l N\_[,l]{}\^2)\^3 . , where N\_[,iii]{}=2[V\_i V\_i”\^2V\_i’\^3]{}-[V\_i”V\_i’]{}-[V\_iV\_i”’V\_i’\^2]{}. We remind the readers that the value of $\tau_{NL}$ is bounded from below by $({6\over 5}f_{NL})^2$. For the axion N-flation, we have N\_[,iii]{}=[2f\_i]{} [\_i (1+\_i)\^2]{}. Therefore \_[NL]{}&=&\_i [f\_i\^24\^2]{} [1-\_i(1+\_i)\^3]{} / (\_j [f\_j\^24\^2]{} [1-\_j1+\_j]{})\^3 . ,\
g\_[NL]{}&=&[2554]{} \_i [f\_i\^24\^2]{} [(1-\_i)\^2 (1+\_i)\^3]{} / (\_j [f\_j\^24\^2]{} [1-\_j1+\_j]{})\^3 . .
For the case with $f_i=f$ for $\forall \ i$, both $\tau_{NL}$ and $g_{NL}$ are proportional to $1/f^4$. Similar to Sec. \[nnt\], we consider $N_*\simeq N_{tot}\simeq 60$ and the numerical results for $f=1/2$ are showed in Fig. \[fig:gtfnl\].
![The non-Gaussianity parameters of axion N-flation for the uniform distribution of $\alpha_i$ when $f=1/2$. Here we consider the case with $N_*\simeq N_{tot}\simeq 60$. []{data-label="fig:gtfnl"}](randomgtfnl.eps){width="10cm"}
Fitting the numerical results, we find that $\tau_{NL}$ and $g_{NL}$ are roughly related to $f_{NL}$ and $f$ by \_[NL]{}${2\pi\over f}$\^[2/3]{} f\_[NL]{}\^[5/3]{}, g\_[NL]{} \_[NL]{}. \[fft\] Roughly speaking, $f_{NL}\sim 1/f^2$ and then $g_{NL}\simeq {25\over 27}\tau_{NL}\sim 1/f^4$ which is consistent with our estimation. On the other hand, combining (\[fft\]) and $\tau_{NL}\geq ({6\over 5}f_{NL})^2$, we get $f_{NL}\lesssim 4\pi^2/3f^2$ which is roughly the same as our results in Sec. \[nnt\].
The non-Gaussianity parameters of trispectrum for the case in which a few inflatons stay around the hilltop were worked out in [@Kim:2010ud]: $\tau_{NL}=({6\over 5}f_{NL})^2$ and $g_{NL}= {25\over 27} \tau_{NL}$, which are different from the previous case unless $f_{NL}\simeq 4\pi^2/3f^2$.
The non-Gaussianity parameters for the model in Sec. \[dnnt\] are showed in Fig. \[fig:dgtfnl\].
![The non-Gaussianity parameters in the axion N-flation where $N_f$ inflatons have the same decay constant $f$, and the decay constant of another inflaton is $\bar f$. Here we consider the case with $N_*\simeq N_{tot}\simeq 60$ and adopt $f=1/2$ and $(\pi-{\bar \alpha}_{ini})/\pi=2\times 10^{-4}$. The blue and green dots correspond to ${\bar f}=1/4$ and the red dots corresponds to ${\bar f}=1$. The green dots illustrate the case with detectable scale dependence of $f_{NL}$. []{data-label="fig:dgtfnl"}](dgtfnl.eps){width="10cm"}
Fitting the green dots corresponding to a detectable scale dependence of $f_{NL}$, we find \_[NL]{}7.78 f\_[NL]{}\^[1.77]{}, g\_[NL]{}\_[NL]{}. Fitting the blue and red dots respectively, we reach the same results \_[NL]{}&8.28 f\_[NL]{}\^[1.59]{},g\_[NL]{}\_[NL]{}. Here the relations between $\tau_{NL}$ and $f_{NL}$ depend on the choice of the parameters, such as $f$, $\bar f$ and $\bar \alpha$, in the model.
Similarly we also investigate the non-Gaussianity parameters for the model in Sec. \[dnntNN\]. See the numerical results in Fig. \[fig:dgtfnlNN\].
![The non-Gaussianity parameters in the general axion N-flation with two kinds of inflatons. Here we consider the case with $N_*\simeq N_{tot}\simeq 60$. The green dots illustrate the case with detectable scale dependence of $f_{NL}$. []{data-label="fig:dgtfnlNN"}](dgtfnlNN.eps){width="10cm"}
Fitting the blue and green dots respectively, we find \_[NL]{}7.16 f\_[NL]{}\^[1.72]{}, g\_[NL]{}\_[NL]{}.
To summarize, there is a universal relation between $g_{NL}$ and $\tau_{NL}$, namely $g_{NL}\simeq {25\over 27}\tau_{NL}$. The value of $g_{NL}$ is the same order of $\tau_{NL}$. But the relations between $\tau_{NL}$ and $f_{NL}$ are different for different cases.
[99]{}
J. M. Maldacena, “Non-Gaussian features of primordial fluctuations in single field inflationary models,” JHEP [**0305**]{}, 013 (2003) \[arXiv:astro-ph/0210603\].
A. A. Starobinsky, “Multicomponent de Sitter (Inflationary) Stages and the Generation of Perturbations,” JETP Lett. [**42**]{} (1985) 152;\
M. Sasaki and E. D. Stewart, “A General Analytic Formula For The Spectral Index Of The Density Perturbations Produced During Inflation,” Prog. Theor. Phys. [**95**]{}, 71 (1996) \[arXiv:astro-ph/9507001\];\
M. Sasaki and T. Tanaka, “Super-horizon scale dynamics of multi-scalar inflation,” Prog. Theor. Phys. [**99**]{}, 763 (1998) \[arXiv:gr-qc/9801017\];\
D. H. Lyth, K. A. Malik and M. Sasaki, “A general proof of the conservation of the curvature perturbation,” JCAP [**0505**]{}, 004 (2005) \[arXiv:astro-ph/0411220\];\
D. H. Lyth and Y. Rodriguez, “The inflationary prediction for primordial non-gaussianity,” Phys. Rev. Lett. [**95**]{}, 121302 (2005) \[arXiv:astro-ph/0504045\].
C. T. Byrnes, K. Y. Choi and L. M. H. Hall, “Conditions for large non-Gaussianity in two-field slow-roll inflation,” JCAP [**0810**]{}, 008 (2008) \[arXiv:0807.1101 \[astro-ph\]\].
C. T. Byrnes, K. Y. Choi and L. M. H. Hall, “Large non-Gaussianity from two-component hybrid inflation,” JCAP [**0902**]{}, 017 (2009) \[arXiv:0812.0807 \[astro-ph\]\].
S. A. Kim, A. R. Liddle and D. Seery, “Non-gaussianity in axion Nflation models,” arXiv:1005.4410 \[astro-ph.CO\].
D. H. Lyth, “Generating the curvature perturbation at the end of inflation,” JCAP [**0511**]{}, 006 (2005) \[arXiv:astro-ph/0510443\].
M. Sasaki, “Multi-brid inflation and non-Gaussianity,” Prog. Theor. Phys. [**120**]{}, 159 (2008) \[arXiv:0805.0974 \[astro-ph\]\].
Q. G. Huang, “A geometric description of the non-Gaussianity generated at the end of multi-field inflation,” JCAP [**0906**]{}, 035 (2009) \[arXiv:0904.2649 \[hep-th\]\].
F. Vernizzi and D. Wands, “Non-Gaussianities in two-field inflation,” JCAP [**0605**]{}, 019 (2006) \[arXiv:astro-ph/0603799\].
S. A. Kim and A. R. Liddle, “Nflation: Multi-field inflationary dynamics and perturbations,” Phys. Rev. D [**74**]{}, 023513 (2006) \[arXiv:astro-ph/0605604\].
Y. S. Piao, “On perturbation spectra of N-flation,” Phys. Rev. D [**74**]{}, 047302 (2006) \[arXiv:gr-qc/0606034\];\
R. G. Cai, B. Hu and Y. S. Piao, “Entropy Perturbations in N-flation Model,” Phys. Rev. D [**80**]{}, 123505 (2009) \[arXiv:0909.5251 \[astro-ph.CO\]\].
S. A. Kim and A. R. Liddle, “Nflation: Non-gaussianity in the horizon-crossing approximation,” Phys. Rev. D [**74**]{}, 063522 (2006) \[arXiv:astro-ph/0608186\].
T. Battefeld and R. Easther, “Non-gaussianities in multi-field inflation,” JCAP [**0703**]{}, 020 (2007) \[arXiv:astro-ph/0610296\].
D. Battefeld and T. Battefeld, “Non-Gaussianities in N-flation,” JCAP [**0705**]{}, 012 (2007) \[arXiv:hep-th/0703012\].
D. Langlois, F. Vernizzi and D. Wands, “Non-linear isocurvature perturbations and non-Gaussianities,” JCAP [**0812**]{}, 004 (2008) \[arXiv:0809.4646 \[astro-ph\]\].
D. J. Mulryne, D. Seery and D. Wesley, “Moment transport equations for the primordial curvature perturbation,” arXiv:1008.3159 \[astro-ph.CO\].
T. Wang, “Note on Non-Gaussianities in Two-field Inflation,” arXiv:1008.3198 \[astro-ph.CO\].
T. Suyama, T. Takahashi, M. Yamaguchi and S. Yokoyama, “On Classification of Models of Large Local-Type Non-Gaussianity,” arXiv:1009.1979 \[astro-ph.CO\].
E. Sefusatti, M. Liguori, A. P. S. Yadav, M. G. Jackson and E. Pajer, “Constraining Running Non-Gaussianity,” JCAP [**0912**]{}, 022 (2009) \[arXiv:0906.0232 \[astro-ph.CO\]\].
C. T. Byrnes, S. Nurmi, G. Tasinato and D. Wands, “Scale dependence of local $f_{NL}$,” JCAP [**1002**]{}, 034 (2010) \[arXiv:0911.2780 \[astro-ph.CO\]\].
C. T. Byrnes, M. Gerstenlauer, S. Nurmi, G. Tasinato and D. Wands, “Scale-dependent non-Gaussianity probes inflationary physics,” arXiv:1007.4277 \[astro-ph.CO\].
C. T. Byrnes, K. Enqvist and T. Takahashi, “Scale-dependence of Non-Gaussianity in the Curvaton Model,” arXiv:1007.5148 \[astro-ph.CO\].
Q. G. Huang, “Negative spectral index of $f_{NL}$ in the axion-type curvaton model,” arXiv:1008.2641 \[astro-ph.CO\].
A. Riotto and M. S. Sloth, “Strongly Scale-dependent Non-Gaussianity,” arXiv:1009.3020 \[astro-ph.CO\].
D. H. Lyth and A. Riotto, “Particle physics models of inflation and the cosmological density perturbation,” Phys. Rept. [**314**]{}, 1 (1999) \[arXiv:hep-ph/9807278\].
M. Sasaki and E. D. Stewart, “A General Analytic Formula For The Spectral Index Of The Density Perturbations Produced During Inflation,” Prog. Theor. Phys. [**95**]{}, 71 (1996) \[arXiv:astro-ph/9507001\].
E. Komatsu [*et al.*]{}, “Seven-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Cosmological Interpretation,” arXiv:1001.4538 \[astro-ph.CO\].
N. Arkani-Hamed, H. C. Cheng, P. Creminelli and L. Randall, “Extranatural inflation,” Phys. Rev. Lett. [**90**]{}, 221302 (2003) \[arXiv:hep-th/0301218\].
T. Banks, M. Dine, P. J. Fox and E. Gorbatov, “On the possibility of large axion decay constants,” JCAP [**0306**]{}, 001 (2003) \[arXiv:hep-th/0303252\].
S. Dimopoulos, S. Kachru, J. McGreevy and J. G. Wacker, “N-flation,” JCAP [**0808**]{}, 003 (2008) \[arXiv:hep-th/0507205\].
Q. G. Huang, “Weak Gravity Conjecture for the Effective Field Theories with N Species,” Phys. Rev. D [**77**]{}, 105029 (2008) \[arXiv:0712.2859 \[hep-th\]\].
[^1]: [email protected]
[^2]: In [@Kim:2010ud], the authors consider $|\phi_{hilltop}-{\bar \phi}|\gtrsim H/2\pi$, namely $(\pi-{\bar \alpha})\gtrsim H/f$, which is much looser than what we require, where $\phi_{hilltop}\equiv f/2$.
[^3]: Thank C. T. Byrnes for the helpful discussion.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We calculate the static and dynamic properties of single crystal, single molecule magnets consisting of equal spin $S=1/2$ or 5/2 dimers. The spins in each dimer interact with each other via the Heisenberg exchange interaction and with the magnetic induction ${\bf B}$ via the Zeeman interaction, and interdimer interactions are negligible. For antiferromagnetic couplings, the static magnetization and specific heat exhibit interesting low temperature $T$ and strong ${\bf B}$ quantum effects. We calculate the frequency spectrum of the Fourier transform of the real part of the time autocorrelation function ${\cal C}_{11}(t)$ for arbitrary $T$, ${\bf B}$, and compare our results with those obtained for classical spins. We also calculate the inelastic neutron magnetic dynamical structure factor $S({\bf q},\omega)$ at arbitrary $T$, ${\bf B}$.'
author:
- 'Dmitri V. Efremov'
- 'Richard A. Klemm'
title: Heisenberg Dimer Single Molecule Magnets in a Strong Magnetic Field
---
Introduction
============
Recently, there has been a considerable interest in the physics of single molecule magnets (SMM’s), or magnetic molecules.[@Mn12; @Fe8; @Fe6; @Fe10; @V2; @V2neutron; @V2P2O9; @Cu2; @Yb2; @Cr2; @Fe2; @Fe2structure] These consist of small clusters of magnetic ions embedded within a non-magnetic ligand group, which may crystallize into large, well-ordered single crystals of sufficient quality for neutron scattering studies. Usually, the spins within a single molecule interact mainly via the Heisenberg exchange interaction. In the simplest SMM’s, V2, Cu2, two examples of Yb2, Cr2, and four examples of Fe2, the magnetic cores of the molecules consist of dimers of spin $S=1/2$ V$^{4+}$, Cu$^{2+}$, or Yb$^{3+}$, $S=3/2$ Cr$^{3+}$, or $S=5/2$ Fe$^{3+}$ spins, respectively.[@V2; @Cu2; @Yb2; @Cr2; @Fe2] Low-field magnetization, nuclear magnetic resonance (NMR) and electron paramagnetic resonance (EPR) experiments, and zero-field inelastic neutron scattering experiments were made on some of these dimers. [@V2; @V2neutron; @V2P2O9; @Cu2; @Yb2; @Cr2; @Fe2; @Fe2structure] Theoretically, zero-field results for the time dependence of the autocorrelation function of the quantum $S=1/2$ and $S=5/2$ dimers were presented.[@MSL]
Here we study the simplest model of interacting Heisenberg spins in a magnetic field ${\bf H}$. We assume only two spins, which interact with each other via the ordinary Heisenberg exchange interaction $J$, and also with a constant magnetic induction ${\bf B}$ induced by the application of ${\bf H}$. For simplicity, we limit our discussion to the spin values $S=1/2$ and $S=5/2$.
We first evaluate the static magnetization ${\bf M}$ and specific heat $C_V$ as functions of $T$ and ${\bf B}$. We find that for ferromagnetic (FM) exchange couplings $J>0$, both ${\bf M}$ and $C_V$ behave at low $T$ as do a single paramagnetic ion with spin $2S$, qualitatively similar to that expected from a classical treatment. For antiferromagnetic (AFM) exchange couplings $J<0$, however, for $k_BT\ll|J|$, the low-$T$ results are very non-classical, even for $S=5/2$. ${\bf M}({\bf B})$ for the AFM spin-$S$ dimer exhibits $2S$ discrete steps, reminiscent of the transverse conductivity in the integer quantum Hall effect. In addition, $C_V(T,B)$ exhibits $2S$ doublet peaks centered about the corresponding magnetization step fields, the splitting of which is proportional to $T$.
At arbitrary $T, {\bf B}$, we then evaluate the real part of the time autocorrelation function ${\cal
C}_{11}(t)$, and focus upon its Fourier transform $\tilde{\cal
C}_{11}(\omega)$, which is applicable to inelastic neutron scattering experiments. For ${\bf B}=0$, $\tilde{\cal
C}_{11}(\omega)$ for the spin-$S$ dimer has $4S+1$ equally-spaced modes with frequencies $\omega^0_{S,n}=n|J|$, where $n=-2S,-2S+1,...,2S-1,2S$. Depending upon $T$ and the sign of $J$, the relative importance of these modes varies significantly. For ${\bf B}\ne0$, each of these $\omega^0_{S,n}$ modes is split into 3 modes. We study the $T$ and ${\bf B}$ dependence of the most important of these modes for $S=1/2$ and $5/2$, for both FM and AFM exchange couplings. For the FM case, only a few modes are important at low $T$, and their relative strength is nearly independent of $|B|$. For the AFM case, however, the situation is more complicated, as many modes can be important at rather low $T$, and their relative importance shifts with $|B|$. For comparison, we also present the analogous results for classical spins. Finally, at arbitrary $T, {\bf B}$, we evaluate the magnetic dynamical structure factor $S({\bf q},\omega)$ measurable with inelastic neutron scattering, and identify a method by which it can measure $\tilde{\cal C}_{11}(\omega)$.
Thermodynamic Properties
========================
Partition Function
------------------
Here we derive the partition function $Z$ for the quantum dimer for ${\bf
B}\ne0$ , with spin $i=1,2$ represented by the operator ${\bf S}_i$, and set $\hbar=1$ for convenience. The Hamiltonian is $$\begin{aligned}
{\cal H}&=&-J{\bf S}_1\cdot{\bf S}_2-\gamma {\bf B}\cdot({\bf S}_1+{\bf S}_2),\end{aligned}$$ where $\gamma=g\mu_B$ is the gyromagnetic ratio. Letting the total spin operator ${\bf s}={\bf S}_1+{\bf S}_2$, ${\cal H}$ is rewritten as $$\begin{aligned}
{\cal H}&=&-J({\bf s}^2-{\bf S}_1^2-{\bf S}_2^2)/2-\gamma Bs_z.\end{aligned}$$ The dimer quantum states are then indexed by the quantum numbers $s$ and $m$, where $$\begin{aligned}
{\bf s}^2|sm\rangle&=&s(s+1)|sm\rangle,\\
s_z|sm\rangle&=&m|sm\rangle,\label{m}\end{aligned}$$ and since ${\bf S}_i^2|sm\rangle=S(S+1)|sm\rangle$, where $S(S+1)$ is a constant for all measurable quantities, we drop the terms proportional to ${\bf S}_1^2$ and ${\bf S}_2^2$ in ${\cal H}$ for convenience.
![Plots of the static magnetization $M$ normalized to $2Sg\mu_B$ versus $\overline{B}=g\mu_BB/|J|$ for the AFM $S=1/2$ and $S=5/2$ dimers at $\alpha=J/(2k_BT)=-0.1, -1, -10$, as indicated.[]{data-label="fig1"}](ekfig1.eps){width="45.00000%"}
Letting $\beta=1/(k_BT)$, $\alpha=\beta J/2$, $b=\beta\gamma B$, and $\overline{B}=g\mu_BB/|J|=b/(2|\alpha|)$, the partition function $Z={\rm
Tr}\exp(-\beta{\cal H})$ for the Heisenberg dimer is simply $$\begin{aligned}
Z&=&\sum_{s=0}^{2S}\sum_{m=-s}^se^{\alpha s(s+1)+bm}=\sum_{s=0}^{2S}{\cal
D}_s(\alpha,b),\label{pf}\\
\noalign{\rm where}\nonumber\\
{\cal D}_S(\alpha,b)&=&e^{\alpha
s(s+1)}{{\sinh[(2s+1)b/2]}\over{\sinh(b/2)}}.\label{dos}\end{aligned}$$
Magnetization
-------------
The magnetization ${\bf M}=\beta^{-1}\overrightarrow{\nabla}_{\bf
B}(\ln{Z})$ for the spin-$S$ dimer in a field is then easily found to be $$\begin{aligned}
{\bf M}(\alpha,b)&=&{{\gamma\hat{\bf B}}\over{Z}}\sum_{s=0}^{2S}{\cal D}_s(\alpha,b)B_s(b)\label{magnetization}\\
&\equiv&
\gamma\hat{\bf B}\langle B_s(b)\rangle,\label{average}\\
\noalign{\rm where}\nonumber\\
B_s(b)&=&(s+1/2)\coth[(s+1/2)b]-{1\over{2}}\coth(b/2)\nonumber\label{bsofb}\\\end{aligned}$$ is the standard Brillouin function for a spin $s$ paramagnet in a magnetic field. We note that $\langle\ldots\rangle$ represents a thermodynamic average over the dimer quantum $s$ values, evaluated using the weighting function ${\cal D}_s(\alpha,b)$.
In Fig. 1, we plotted $M/(2Sg\mu_B)$ as a function of $\overline{B}$ for the AFM $S=5/2$ and $S=1/2$ dimers at $\alpha=-0.1, -1, -10$, respectively. At high $k_BT/|J|$ ($|\alpha|<<1$), the dotted curves for $\alpha=-0.1$, $S=1/2$, 5/2 are smooth functions of $\overline{B}$, not too different from the analogous results obtained for classical spins. At $\alpha=-1$, the dashed curves remain smooth, but are shifted over to smaller field values, with hints of a kink just below the saturation magnetization value. Most interesting are the low-$T$ effects. At $\alpha=-10$, $M$ has $2S$ steps at integral values of $\overline{B}$, as shown by the solid curves for $S=1/2,5/2$ in Fig. 1. Since these steps are thermodynamic, they are reversible. However, a measurement at fixed $B$ near to a step value can lead to interesting, non-monotonic behavior of $M(T)$ at low $T$.
These thermodynamic steps in $M$ are a consequence of quantum level crossing due to the strong $B$.[@Fe10] The energy $E_{sm}$ of the state $|sm\rangle$ is given by $$\begin{aligned}
E_{sm}&=&-Js(s+1)-m\tilde{B}.\end{aligned}$$ where $\tilde{B}=\gamma B$ and $\overline{B}=\tilde{B}/|J|$. For the AFM case, $J=-|J|$, the lowest energy state for each $s$ value is $E_{sm,\rm min}=E_{ss}$. The difference in energy between the lowest energy state with quantum number $s$ and that with the next highest quantum number, $s+1$, is then $$\begin{aligned}
\Delta E_{ss}&=E_{s+1,s+1}-E_{ss}=|J|(s+1-|\overline{B}|),\label{deltaE}\end{aligned}$$ which vanishes at $|\overline{B}|=s+1$. For $S=1/2$, there will only be one step, as this crossing can only occur between states corresponding to the $s=0$ and $s=1$ quantum numbers. Similarly, for $S=5/2$, there will be 5 level crossings, corresponding to $s=0,\ldots,4$.
For the FM case, the situation is rather boring by comparison. We found that for both $S=1/2, 5/2$, $M(\alpha,b)/\gamma$ is closely approximated by $B_{2S}(b)$. For $\alpha=0.1$, this approximation is accurate to a few percent, but for $\alpha=10$, the corresponding curves for $M$ and $B_{2S}$ are indistinguishable. This is because the $s=2S$ term in both the numerator and the denominator of $\langle B_s\rangle$ is dominant for $J>0$.
![Plots of $C_V/k_B$ versus $\overline{B}$ at $\alpha=-0.1,-1,-3,-10$, as indicated, for the $S=1/2$ AFM dimer.[]{data-label="fig2"}](ekfig2.eps){width="45.00000%"}
![Plots of $C_V/k_B$ versus $\overline{B}$ at $\alpha=-0.1,-1,-3,-10$, as indicated, for the $S=5/2$ AFM dimer.[]{data-label="fig3"}](ekfig3.eps){width="45.00000%"}
![Plots of $C_V/k_B$ versus $k_BT/|J|$ at $\overline{B}=0, 0.4, 0.8, 1.0, 1.2$, as indicated, for the $S=5/2$ AFM dimer. []{data-label="fig4"}](ekfig4.eps){width="45.00000%"}
Specific Heat
-------------
The specific heat $C_V$ is given by $$\begin{aligned}
k_BT^2C_V&=&{{\partial^2\ln{Z}}\over{\partial\beta^2}},\end{aligned}$$ leading to $$\begin{aligned}
C_V/k_B&=& b^2\Bigl[\langle s(s+1)\rangle-\coth(b/2)\langle B_s(b)\rangle-\langle
B_s(b)\rangle^2\Bigr]\nonumber\\
& &+2\alpha b\Bigl[\langle s(s+1)B_s(b)\rangle-\langle B_s(b)\rangle\langle
s(s+1)\rangle\Bigr]\nonumber\\
& &+\alpha^2\Bigl[\langle [s(s+1)]^2\rangle-\langle
s(s+1)\rangle^2\Bigr].\label{heatcap}\end{aligned}$$
In numerical evaluations of $C_V$, we checked that Boltzmann’s law, $\int_0^{\infty}dTC_V/T=2k_B\ln(2S+1)$, is satisfied. As for the FM $M$, $C_V$ for the FM spin-$S$ dimer is rather boring. It is closely approximated by the $s=2S$ terms in each of the sums. In this approximation, only the first term proportional to $b^2$ survives, so $C_V$ reduces to the standard Schottky result for a spin $2S$ paramagnet, $$\begin{aligned}
C_V/k_B&{{\rightarrow}\atop{\alpha\gg1}}&C_{2S}(b),\\
&=&b^2\Bigl(2S(2S+1)+{1\over{4}}\coth^2(b/2)\nonumber\\
& &-(2S+1/2)^2\coth^2[(2S+1/2)b]\Bigr).\end{aligned}$$ Numerically, for $S=5/2$, the FM $C_V$ is nearly indistinguishable from $C_{2S}(b)$ for $\alpha\ge1$. We note that $C_{2S}(b)$ grows as $2S(2S+1)b^2/3$ for $b\ll1$, decays as $b^2\exp(-b)$ for $b\gg1$, and has a maximum for $b\approx1$. For $S=5/2$, the maximum occurs at $b\approx0.78$.
On the other hand, $C_V$ for the AFM spin-$S$ dimer is much more interesting. In Fig. 2 and 3, we plotted $C_V/k_B$ versus $\overline{B}$ at $\alpha = -0.1, -1, -3, -10$, for dimers of spin $S=1/2$ and 5/2, respectively. The curves for $\alpha=-0.1$ have broad maxima at fields too high to appear in these figures. But, as $T$ is lowered to $\alpha=-1$, Fig. 2 illustrates that this broad maximum develops into two peaks centered at $\overline{B}\approx 0,2$, respectively. Then, as $T$ is lowered further, $C_V$ for $S=1/2$ becomes two well-defined peaks symmetrically centered about $\overline{B}=1$, with a splitting between them proportional to $T$.
In Fig. 3, the analogous AFM $C_V$ for $S=5/2$ is shown. For $\alpha=-1$, instead of two broad peaks, as in Fig. 2, for $S=5/2$ there is an irregular pattern resulting from many accessible energy levels. The situation becomes clearer at $\alpha=-3$, with six rather symmetric peaks roughly centered at half-integral values of $\overline{B}$. However, as for $S=1/2$, the low-$T$ limiting behavior is not reached until $\alpha=-10$. For $S=5/2$, Fig. 3 shows that $C_V$ consists of five double peaks symmetrically centered about $\overline{B}=s+1$ for $s=0,\ldots,4$, and nearly vanishes at those points.
We note that this multiplicity of double peaks at low $T$ is also a consequence of quantum level crossing. At low $T$, the energy difference between the two lowest states is given by Eq. (\[deltaE\]). At the exact level crossing, these levels are degenerate, leading to an exponentially small value of $C_V$, \[$\propto\exp(2\alpha)$\] Just away from these points, excitations from the ground state to the first excited state can occur for $k_BT\agt|J||s+1-|\overline{B}||$. This should lead to a peak on each side of the level crossing points, with a splitting $\Delta\overline{B}$ of the double peaks approximately equal to $1/|\alpha|$. From the data in Figs. 2 and 3 for $|\alpha|=3,10$, the splitting is $\approx2.4/|\alpha|$, which is $\propto T$, but 2.4 times as large as in the above crude estimate.
We remark that the $2S$ double peak structures in $C_V$ shown in Figs. 2 and 3 comprise a new quantum effect. Although it may be difficult to increase $B$ to 40 T while holding $T$ fixed at $\alt1$ K, it might be easier to lower $T$ in a fixed, strong $B$. From Figs. 2 and 3, for $\overline{B}$ fixed at 0.8 or 1.2, non-monotonic $C_V(T)$ behavior upon lowered $T$ is predicted. However, at $\overline{B}=1$, the behavior at low $T$ is very different, as $C_V$ at $\alpha=-10$ is vanishingly small. This striking sensitivity to the precise value of $B$ is pictured in Fig. 4 for $S=5/2$. In Fig. 4, for $k_BT/|J|>1$, the $T$, $B$ dependencies of $C_V$ are monotonic. Aside from the rather ordinary peak for $B=0$, the unusual behavior is illustrated by comparing the curves for $\overline{B}=0.8, 1.0, 1.2$ for $k_BT/|J|<0.5$ The curves for $\overline{B}=0.8,
1.2$ both have peaks at low $T$, but the curve for the intermediate value $\overline{B}=1.0$ does not. This highly sensitive dependence of $C_V$ upon $T, B$ is a new quantum level crossing effect.
Spin Dynamics
=============
We now evaluate the time autocorrelation function for both quantum and classical spin dimers for ${\bf B}\ne0$.
Quantum Spin Dynamics
---------------------
The time evolution of quantum spins is given by the commutator of the spin operator with the Hamiltonian, $$\begin{aligned}
i\dot{\bf S}_i(t)&=&[{\bf S}_i,{\cal H}],\\
{\bf S}_i&=&\exp(it{\cal H}){\bf S}_i(0)\exp(-it{\cal H}).\end{aligned}$$ It is easiest in the quantum case just to keep the time dependence in this form, and then to let ${\cal H}$ operate on the eigenstates in the matrix elements of the correlation function. The total spin operators ${\bf s}^2$ and $s_z$ are independent of $t$.
Classical Spin Dynamics
-----------------------
The two classical spins ${\bf S}_i(t)$ each precess according to classical Heisenberg dynamics, $$\begin{aligned}
\dot{\bf S}_i&=&J{\bf S}_i\times{\bf S}+\gamma{\bf S}_i\times{\bf B},\label{S1oft}\\
\noalign{\rm where $|{\bf S}_i(t)|=1$, for $i=1, 2$, leading to}\nonumber\\
\dot{\bf S}&=&\gamma{\bf S}\times{\bf B}.\label{Soft}\end{aligned}$$ Hence, the total spin ${\bf S}$ precesses about ${\bf B}$, and the ${\bf
S}_i$ precess about both ${\bf B}$ and ${\bf S}$. The solution to these equations is given in the Appendix.
Time Correlation Functions
==========================
Quantum Spins
-------------
In this section, we evaluate the time correlation functions for quantum spins with general spin value $S$. Here we calculate the time autocorrelation function for the quantum dimer. For easy comparison with the classical results, we normalize the correlation functions such that ${\rm Re}\>\>{\cal C}_{11}(0)=1$.[@MSL] Then ${\cal C}_{11}(t)=\langle {\bf S}_1(t)\cdot{\bf S}_1(0)\rangle$, or $$\begin{aligned}
{\cal C}_{11}(t)&=&\sum_{s=0}^{2S}\sum_{m=-s}^s\langle sm|e^{i{\cal H}t}{\bf
S}_1(0)e^{-i{\cal H}t}\cdot{\bf S}_1(0)\nonumber\\
& &\qquad\times e^{-\beta{\cal H}}|sm\rangle/D,\label{c11oft}\\
&=&\sum_{s,m}\sum_{s',m'}\sum_{\alpha=1}^3e^{it(E_{sm}-E_{s'm'})}e^{-\beta
E_{sm}}\nonumber\\
& &\qquad\times\Bigl|\langle sm|S_{1\alpha}|s'm'\rangle\Bigr|^2/D,\label{quantum}\\
-\beta E_{sm}&=&\alpha s(s+1)+bm,\\
D&=&S(S+1)Z,\end{aligned}$$ where $Z$ is given by Eq. (\[pf\]). In Eq. (\[quantum\]), the expectation values are related to Clebsch-Gordon coefficients.
For quantum spins, the near-neighbor correlation function and the autocorrelation function are related by the sum rule, $$\begin{aligned}
\langle s_z^2\rangle+\langle{\bf s}^2-s_z^2\rangle\cos(\tilde{B}t)&=&
{1\over{2}}\langle{\bf s}(t)\cdot{\bf s}(0)+{\bf s}(0)\cdot{\bf s}(t)\rangle
\nonumber\\
&=&2S(S+1)\nonumber\\
& &\times{\rm Re}[{\cal C}_{12}(t)+{\cal C}_{11}(t)],\label{sumrule}\end{aligned}$$ where ${\cal C}_{12}(t)$ is normalized as is ${\cal C}_{11}(t)$. The thermal averages $\langle{\bf s}^2\rangle$ and $\langle s_z^2\rangle$ can be written in the notation of Eqs. (\[dos\])-(\[bsofb\]) as $\langle
s(s+1)\rangle$ and $\langle s(s+1)\rangle-\coth(b/2)\langle B_s(b)\rangle$, respectively.
Classical Spins
---------------
It is interesting to compare the quantum time correlation functions with those obtained from classical spin dynamics. For classical spin dimers, one solves the equations of motion of the spins directly, taking the length of each spin to be unity. We first solve Eq. (\[Soft\]) for ${\bf S}(t)$, and then Eq. (\[S1oft\]) to obtain the ${\bf S}_i(t)$. We then evaluate the classical autocorrelation function ${\cal C}_{11}(t)=\langle {\bf S}_1(t)\cdot{\bf S}_1(0)\rangle$ by averaging over the length $S$ ($0\le S\le2$) of the total spin, the angle $\theta$ between ${\bf S}$ and ${\bf B}$, and the angle $\phi_0$ describing the initial relative configuration of ${\bf S}_1$ and ${\bf S}_2$, and requiring ${\cal C}_{11}(0)=1$. Since the procedure is analogous to that used for the four-spin ring,[@KL] the results are given in the Appendix. Here we only remark that the classical correlation functions also must satisfy a sum rule, $$\begin{aligned}
{1\over{6}}[1+2\cos(\tilde{B}t)]\langle S^2\rangle&=&{\cal C}_{11}(t)+{\cal C}_{12}(t).\end{aligned}$$
Fourier transforms
==================
Quantum Spins
-------------
The Fourier transform $\tilde{\cal C}_{11}(\omega)$ of the real part of the autocorrelation function ${\cal C}_{11}(t)$ for the quantum dimer of spin-$S$ spins is given by $$\begin{aligned}
\tilde{\cal C}_{11}(\omega)&=&\sum_{i=0}^{6S+1}f_{S,i}[\delta(\omega-\omega_{Si})+\delta(\omega+\omega_{Si})].\end{aligned}$$
The discrete mode frequencies $\omega_{S,i}$ and their amplitudes $f_{S,i}$ are given for $S=1/2$ and $S=5/2$ in Tables II and III in the Appendix.
Classical Spins
---------------
In order to compare the quantum and classical results, it is useful to evaluate the Fourier transform $\tilde{\cal
C}_{11}(\omega) = \int_{-\infty}^{\infty}dt \exp(i\omega t){\cal C}_{11}(t)$ of the classical autocorrelation function. Letting $\tilde{\omega}=\omega/|J|$ and $\overline{B}=\tilde{B}/|J|$, we find for positive $\omega$ that $$\begin{aligned}
\tilde{\cal
C}_{11}(\omega)&=&\delta(\tilde{\omega})C_{00}+
\delta(\tilde{\omega}-\overline{B})C_{01}+\delta\tilde{\cal C}_{11}(\tilde{\omega}).\label{cft}\end{aligned}$$ Exact expressions for the discrete amplitudes $C_{00}$ and $C_{01}$, and for the continuous part $\delta\tilde{\cal C}_{11}(\tilde{\omega})$ are given in the Appendix. Then, in order to check the numerical evaluation of the above quantities, we employ the frequency sum rule, $1=\int_0^{\infty}d\omega\tilde{\cal C}_{11}(\omega)/\pi$.
We note that $C_{00}$ also appears in the expression for $\tilde{\cal
C}_{11}(\omega)$ with $\omega\le0$. The coefficients $C_{00}$ and $C_{01}$ of the delta functions at $\omega=0$ and $\tilde{\omega}=\overline{B}$, respectively, are functions of both $\alpha$ and $\overline{B}$. These quantities arise from the long-time rms limit of ${\cal C}_{11}(t)$ and for the long-time oscillatory behavior at the driving frequency, respectively. This second delta function occurs at the resonant frequency in an EPR experiment.
Numerical Results
=================
Autocorrelation Function Spectra
--------------------------------
In Fig. 5, we plot the amplitudes $f_{5/2,i}$ of the delta-function modes of the zero-field quantum $S=5/2$ AFM $\tilde{\cal C}_{11}(\omega)$ for positive $\omega/|J|$, at the temperatures $\overline{T}=k_BT/[|J|S(S+1)]=0.1, 1, \infty$. The amplitudes of the modes at these three ${\overline T}$ values are indicated by the symbols +, $\times$, \*. For comparison, we also show the results of the zero-field classical calculation $\tilde{\cal C}_{11}(\omega)$. For $\overline{B}=0$, both of the delta-function amplitudes $C_{00}$ and $C_{01}$ appear at $\omega=0$, and their combined weight at these $\overline{T}$ values is indicated by the circle, triangle, and square, respectively. In addition, the continuous part $\delta\tilde{\cal C}_{11}(\omega)$ is plotted both by scaling its amplitude by $1/[S(S+1)]^{1/2}$ and the frequency by $[S(S+1)]^{1/2}$. This combined scaling allows us to compare with the quantum results, while preserving the area under the curves. For $S=5/2$, this scaling changes the maximum $\tilde{\omega}=\omega/|J|$ from 2 to $\sqrt{35}\approx5.91$. We note that as $T\rightarrow\infty$, the quantum and classical delta functions at $\omega=0$ agree exactly, both having the weight $\pi/2$. Also as $T\rightarrow\infty$, the classical $\delta\tilde{\cal C}_{11}(\omega)$ forms an envelope for the amplitudes of the quantum delta functions, except for a slight deviation at the larger frequencies. At $\overline{T}=1$, the agreement between the quantum and classical results is also pretty good, although there are deviations at nearly every quantum mode value. These deviations are more pronounced at $\overline{T}=0.1$, but the overall agreement is still rather good.
![Plots of the AFM quantum $S=5/2$ modes of $\tilde{\cal C}_{11}$ versus $\omega/|J|$ at $\overline{B}=0$ at $\overline{T}=k_BT/[|J|S(S+1)]=0.1, 1, \infty$, and for the combined discrete classical modes at $\omega=0$. The curves represent $\delta\tilde{\cal
C}_{11}(\omega)/[S(S+1)]^{1/2}$ versus $\omega[S(S+1)]^{1/2}/|J|$.[]{data-label="fig5"}](ekfig5.eps){width="45.00000%"}
![Plots of the AFM quantum $S=5/2$ modes of $\tilde{\cal C}_{11}$ versus $\omega/|J|$ at $\overline{B}=0.4$ at $\overline{T}=0.1,1,\infty$, and for the discrete classical modes at $\omega/|J|=0,\overline{B}$. The curves represent $\delta\tilde{\cal C}_{11}/3[S(S+1)]^{1/2}$ versus $\omega[S(S+1)]^{1/2}/|J|$.[]{data-label="fig6"}](ekfig6.eps){width="45.00000%"}
In Fig. 6, we show the corresponding curves for the $S=5/2$ AFM dimer at the field $\overline{B} = 0.4$. For $\overline{B}\ne0$, each of the quantum zero-field modes splits into three modes, and therefore to compare the continuous part $\delta\tilde{\cal
C}_{11}(\omega)$ with the split quantum modes, we divide the scaled continuous part of $\tilde{\cal C}_{11}$ by 3, plotting $\delta\tilde{\cal C}_{11}/3[S(S+1)]^{1/2}$ versus $\omega[S(S+1)]^{1/2}/|J|$. As in Fig. 5, the classical delta functions modes at $\tilde{\omega}=0,\overline{B}$ are not scaled, but in this case, they are distinct. For this rather strong field value, $\overline{B}=0.4$, the agreement between the classical and quantum results is very good as $T\rightarrow\infty$, and pretty good at the intermediate $\overline{T}=1$. However, at $\overline{T}=0.1$, these results can differ by more than a factor of two. We note that the classical curves are piecewise continuous functions of $\omega$, with several discontinuities in slope evident. These discontinuities in slope generally appear at the the frequencies corresponding to rather large jumps in the quantum mode amplitude values. At $\overline{T}$ values lower than those pictured, the classical curves develop into three distinct modes with finite widths.
![Plots of the FM quantum $S=5/2$ modes of $\tilde{\cal C}_{11}$ versus $\omega/|J|$ at $\overline{B}=0$ at $\overline{T}=0.1,1,\infty$, and for the combined discrete classical modes at $\omega=0$. The curves represent $\delta\tilde{\cal
C}_{11}(\omega)/[S(S+1)]^{1/2}$ versus $\omega[S(S+1)]^{1/2}/|J|$.[]{data-label="fig7"}](ekfig7.eps){width="45.00000%"}
![Plots of the FM quantum $S=5/2$ modes of $\tilde{\cal C}_{11}$ versus $\omega/|J|$ at $\overline{B}=0.1$ at $\overline{T}=0.1, 1, \infty$, and for the discrete classical modes at $\omega/|J|=0, \overline{B}$. The curves represent $\delta\tilde{\cal C}_{11}/3[S(S+1)]^{1/2}$ versus $\omega[S(S+1)]^{1/2}/|J|$.[]{data-label="fig8"}](ekfig8.eps){width="45.00000%"}
In Figs. 7 and 8, we show the results for the FM $S=5/2$ case at $\overline{B}=0, 0.1$, respectively. Choosing $\overline{B}=0.1$ in Fig. 8 allows us to display all of the results clearly on the same figure. Except for the different $\overline{T}$ values in Figs. 8 and 6, Figs. 7 and 8 correspond precisely to Figs. 5 and 6, with the same scaling of the continuous part of the Fourier transform of the classical correlation function, and the same symbols and linestyles. For the FM case at $\overline{B}=0, 0.1$, Figs. 7 and 8 show that the agreement between the classical and quantum modes is excellent as $T\rightarrow\infty$, pretty good at $\overline{T}=1$, but only qualitative at $\overline{T}=0.1$. In this case, the classical delta function at $\omega=0$ differs significantly from the quantum one, and the continuous classical curve cannot be scaled in this way at all. As shown for the equivalent neighbor model, the FM classical curves obey a different low-$T$ scaling relation, as the low $T$ peak center approaches $NJ$ linearly in $T$ from below as $T\rightarrow0$.[@KA] In Fig. 8, the development of the continuous classical curve into three low $T$ modes is evident, but the fit to the quantum case at $\overline{T}=0.1$ is not good at all. Not only should the low-$T$ classical curves require a different scaling relation, similar to those for $B=0$, the order of the relative heights of the three largest modes at the largest frequencies pictured for the classical curves is inverted relative to that for the quantum case.
![Plots of the AFM quantum $S=1/2$ modes of $\tilde{\cal C}_{11}$ versus $\omega/|J|$ at $\overline{B}=0.4$ at $\overline{T}=0.1,1,\infty$, and for the discrete classical modes at $\omega/|J|=0,\overline{B}$. The curves represent $\delta\tilde{\cal C}_{11}/3[S(S+1)]^{1/2}$ versus $\omega[S(S+1)]^{1/2}/|J|$.[]{data-label="fig9"}](ekfig9.eps){width="45.00000%"}
![Plots of the FM quantum $S=1/2$ modes of $\tilde{\cal C}_{11}$ versus $\omega/|J|$ at $\overline{B}=0.4$ at $\overline{T}=0.1,1,\infty$, and for the discrete classical modes at $\omega/|J|=0,\overline{B}$. The curves represent $\delta\tilde{\cal C}_{11}/3[S(S+1)]^{1/2}$ versus $\omega[S(S+1)]^{1/2}/|J|$.[]{data-label="fig10"}](ekfig10.eps){width="45.00000%"}
We now consider the case $S=1/2$. For the $B=0$ AFM and FM cases, there are two quantum modes for $0\le\omega$, at $\omega/|J|=0, 1$. As $T\rightarrow\infty$, these modes have equal intensity ($\pi/2$), for both FM and AFM cases. For the AFM case as $T$ decreases to 0, the amplitudes of the modes at $\omega/|J|=0 (1)$ decrease (increase) continuously to 0 ($\pi$), respectively. For the FM case, however, these mode amplitude values increase (decrease) with decreasing $T$ to $2\pi/3$ ($\pi/3$), respectively, at $T=0$. Somewhat surprisingly, for the AFM case, the amplitude of the combined classical delta functions at $\omega=0$ closely tracks that of the quantum $\omega=0$ mode. For the FM case, however, this tracking is not so good. However, the best that can be said for the continuous part is that as $T\rightarrow\infty$, the second quantum mode nearly falls upon the classical curve scaled as in Figs. 5 and 7, for both AFM and FM cases. At $\overline{T}=k_BT/[|J|S(S+1)]=1, 0.1$, this agreement becomes increasingly much worse, respectively. For the FM case, the classical treatment fails miserably as $T\rightarrow0$, as the amplitude of the classical delta function mode at $\omega=0$ approaches $\pi$, and the integrated intensity of the continuous classical mode peak vanishes as $T\rightarrow0$. Hence, the classical treatment does not describe the $S=1/2$ dimer at $B=0$ correctly, and the correct quantum treatment leads to just two modes for $\omega\ge0$.
For $B\ne0$, however, there are five modes for $\omega\ge0$ with $S=1/2$, so a classical treatment can approximate the quantum behavior a bit better than for $B=0$. In Figs. 9 and 10, the AFM and FM $S=1/2$ results for $\overline{B}=0.4$ at $\overline{T}=0.1, 1, \infty$ are shown. In both cases, the two quantum modes for $\omega/|J|=0,\overline{B}=0.4$ are compared with the two classical delta functions, and the remaining three quantum modes for $0\le\omega$ are compared with the continuous classical curves. In both cases, as $T\rightarrow\infty$, the five quantum modes have the amplitudes $\pi/6, \pi/3, \pi/6, \pi/6, \pi/6$, respectively. The classical results agree precisely with the first two, and form a qualitative envelope similar to the remaining three quantum modes. This agreement is qualitatively preserved at $\overline{T}=1$. However, as $T\rightarrow0$, the classical and quantum cases differ dramatically. For the AFM case pictured in Fig. 9, the five quantum modes for $0\le\omega$ approach the amplitudes 0, 0, $\pi/3,\pi/3,\pi/3$ as $T\rightarrow0$, respectively, and the continuous classical curve develops a strong peak at $\omega=0$, which is qualitatively different. The qualitative behavior of the two discrete classical AFM modes is not too bad, however. For the FM case pictured in Fig. 10, as $T\rightarrow0$, the quantum FM modes approach the amplitudes $\pi/3,\pi/3,0,0,\pi/3$, respectively. That is, the modes at $\omega=0,
|J+\tilde{B}|$ increase, the modes at $\omega=|J|, |J-\tilde{B}|$ decrease, and the mode at $\omega=\tilde{B}$ has a non-monotonic $T$ dependence. The classical treatment preserves these FM features only qualitatively, and is inaccurate for $\overline{T}\le0.1$.
Field dependencies of the quantum modes
---------------------------------------
In Fig. 11, we plotted the amplitudes $f_{1/2,i}$ of the five FM quantum modes for $S=1/2$ at $\overline{T}=0.1$ versus $\overline{B}$. The modes $\omega_{1/2,i}$ for $i=0,4$ correspond to $\omega=0$, $|\tilde{B}|$, $|J|$, $|J-\tilde{B}|$, and $|J+\tilde{B}|$, respectively. Note that at $B=0$, the $\omega_{1/2,i}$ for $i=0,1$ are degenerate but unequal in intensity, and the $\omega_{1/2,i}$ for $i=2,3,4$ are both degenerate and equal in intensity. At each field, the sum of the five intensities sums to $\pi$. From this figure, it is evident that the crossover from the low-field regime to the high field regime occurs at the rather low field value, $\overline{B}\approx0.1$. The high field regime is clearly consistent with the mode amplitudes at $\overline{B}=0.4$ pictured in Fig. 10, with the amplitudes of the modes at $\omega/|J|= 0, 0.4, 0.6, 1, 1.4$ approximately equal to $\pi/3,
\pi/3,0,0,\pi/3$, respectively.
![Plots of the FM quantum $S=1/2$ mode amplitudes $f_{1/2,i}$ versus $\overline{B}$ at $\overline{T}=0.1$ The numerical labels correspond to the $i$ values in Table I of the Appendix.[]{data-label="fig11"}](ekfig11.eps){width="45.00000%"}
![Plots of the AFM quantum $S=1/2$ mode amplitudes $f_{1/2,i}$ versus $\overline{B}$ at $\overline{T}=0.1$. The mode $i$ values correspond to Table I in the Appendix.[]{data-label="fig12"}](ekfig12.eps){width="45.00000%"}
In Fig. 12, we plotted the amplitudes $f_{1/2,i}$ of the five AFM quantum modes for $S=1/2$ at the $\overline{T}=0.1$. From Fig. 12, the five modes are difficult to discern clearly, due to the strong degeneracies involved. Clearly, the crossover from the low-field to the high-field regime occurs at $\overline{B}=1$. In the low-field regime, the modes at $\omega=0,\tilde{B}$ have nearly zero amplitudes, and the modes with $\omega=|J|,|J\pm\tilde{B}|$ have nearly equal amplitude $\pi/3$. This is the situation pictured for $\overline{B}=0.4$ in Fig. 9. However, Fig. 12 indicates that dramatic changes in the mode amplitudes at low $T$ should take place as $\overline{B}$ is increased from $\approx 0.8$ to $\approx 1.2$. Over this field range, the amplitudes of the modes at $\omega=|J|,|J+\tilde{B}|$ decrease from nearly $\pi/3$ to nearly 0, and the amplitudes of the modes at $\omega=0,\tilde{B}$ increase from nearly zero to nearly $\pi/3$. Meanwhile, the amplitude of the remaining mode at $\omega=|J|$ remains constant at $\pi/3$. Although not pictured for brevity, at $\overline{T}=1$, all five AFM modes for $S=1/2$ are clearly evident, and the crossover from the weak field regime to the strong field regime occurs at $\overline{B}\approx1.5$.
![Plots of the dominant FM quantum $S=5/2$ mode amplitudes $f_{5/2,i}$ versus $\overline{B}$ at $\overline{T}=0.1$. The mode $i$ values correspond to table II in the Appendix.[]{data-label="fig13"}](ekfig13.eps){width="45.00000%"}
![Plots of the dominant AFM quantum $S=5/2$ mode amplitudes $f_{5/2,i}$ versus $\overline{B}$ at $\overline{T}=0.1$. The mode $i$ values correspond to table II in the Appendix.[]{data-label="fig14"}](ekfig14.eps){width="45.00000%"}
The field dependencies of the most important modes for the FM and AFM $S=5/2$ dimers are shown at the rather low temperature $\overline{T}=0.1$ in Figs. 13 and 14, respectively. In Fig. 13, the FM modes shown are for $\omega=0$, $\tilde{B}$, $|5J\pm\tilde{B}|$, and $5|J|$. The crossover from the weak to the strong field limits occurs at about $\overline{B}\approx0.5$. In the weak field limit, this corresponds to the dominant modes at $\overline{B}=0.1$ shown in Fig. 8, for which the ranking of the five strongest low-$T$ modes is at $\omega/|J|=0.1, 0, 5.1, 5, 4.9$, respectively. At high fields, $\overline{B}>2$, there are only three important modes, at $\omega=0,\tilde{B}$, and $5J+\tilde{B}$, which have amplitudes that approach $5\pi/7$, $\pi/7$, $\pi/7$, at $B\rightarrow\infty$ respectively.
For the AFM case with $S=5/2$, the behavior of the relative mode amplitudes at $\overline{T}=0.1$ is more complicated, as shown in Fig. 14. There we plotted the field dependencies of the eight most important modes, which occur at $\omega=0$, $|2J\pm\tilde{B}|$, $2|J|$, $|\tilde{B}|$, $|J\pm\tilde{B}|$, $|J|$, respectively. From Fig. 14, it is evident that there are essentially three field regimes. The low field regime occurs for $\overline{B}<0.5$, the intermediate regime is for $1.5<\overline{B}<2.2$, and the high field regime for $\overline{B}>3$, approximately. An example of the low-field regime at the same temperature was shown for $\overline{B}=0.4$ in Fig. 6. In Figs. 6 and 14, the low-field rank-ordering of the six most important modes are those at $\omega=|\tilde{B}+J|$, $|J|$, $|\tilde{B}-J|$, $|\tilde{B}+2J|$, $2|J|$, and $|\tilde{B}-2J|$, where $J<0$. From the intermediate regime pictured in Fig. 14, however, the two most important modes are those at $\omega=|\tilde{B}+2J|$ and $|\tilde{B}+J|$, with the six other modes in the list contributing lesser, but comparable amounts. Finally, in the high-field regime, the two dominant modes are at $\omega=0$ and $|\tilde{B}|$, respectively.
Discussion
==========
We remark that there have been some experiments on one or more of the Fe2 single molecule magnet dimers. Le Gall [*et al.*]{} measured the four Fe2 dimers they made, and found that the zero-field magnetic susceptibility could be fit with the Heisenberg model, with exchange constants ranging from 15 to 19 cm$^{-1}$ (22 to 27 K, which corresponds to $\overline{B}=1$ at $B\approx 16-22$ T). A slightly different fit was made by Lascialfari [*et al.*]{}, obtaining $|J|\approx22K$ for the Fe2 dimer, \[Fe(OMe)(dpm)$_2$\]$_2$. [@Fe2] Those authors also refer to EPR measurements that give rise to a zero-field splitting anisotropy of about 7 K in the first excited state, which would complicate the analysis performed here. Such a zero-field splitting would still allow quantum level splitting, but the magnetization steps and the specific heat double peaks would not all be regularly spaced. We remark that these level crossing effects are purely quantum in nature, as the analogous classical calculations do not give rise to such effects, and hence are only approximate for $|\alpha|\agt1$. Nevertheless, the quantum effects predicted for the magnetization and specific heat ought to be observable with presently available facilities, as one should be able to see one or two magnetization steps, and one or two specific heat double peaks.
In addition, inelastic neutron scattering would see slightly different results from those presented here. In this situation, the classical envelope curves might be quite useful, as they can serve as a guide to the behavior that might be expected with specific zero-field splitting values. More important, inelastic neutron scattering can be used to probe the details of the magnetic interactions within an individual dimer. Provided that a single crystal of sufficient size for such studies can be obtained, one would perform the experiments at the wave vectors specific to the crystal structure. More precisely, the unpolarized inelastic neutron magnetic dynamical structure factor $S({\bf q},\omega)$ for a single crystal of equal-spin dimers is given by [@Fe8; @Cu2; @white; @fg] $$\begin{aligned}
S({\bf
q},\omega)&=&\sum_{\alpha,\beta=1}^3(\delta_{\alpha\beta}-\hat{q}_{\alpha}\hat{q}_{\beta})\nonumber\\
& &\times\int_{-\infty}^{\infty}{{dt}\over{2\pi}}e^{i\omega t}\langle
Q^{\dag}_{\alpha}({\bf q},t)Q_{\beta}({\bf q},0)\rangle,\label{Sgeneral}\end{aligned}$$ where the $\hat{q}_{\alpha}=\sin\theta_q\cos\phi_q,\sin\theta_q\sin\phi_q,\cos\theta_q$ for $\alpha=1,2,3$, respectively, and $\theta_q$, $\phi_q$ are the angles the scattering wave vector ${\bf q}$ makes with the spin quantization coordinates, and $\langle\ldots\rangle$ represents a thermal average, as in Eq. (\[c11oft\]). The dimer structure operator $$\begin{aligned}
Q_{\alpha}({\bf q},t)&=&f({\bf q})[e^{i{\bf q}\cdot{\bf d}}S_{1\alpha}(t)+e^{-i{\bf
q}\cdot{\bf d}}S_{2\alpha}(t)],\label{Q}\end{aligned}$$ where $2{\bf d}$ is the displacement vector between the dimer spins, and $f({\bf q})$ is the atomic magnetic form factor. [@Cu2; @fg] The low $T$, $B=0$ transitions in $S=1/2$ Cu$^{2+}$ and Yb$^{3+}$ and in $S=3/2$ Cr$^{3+}$ dimer powders were treated previously.[@Cu2; @Yb2; @Cr2] For ${\bf B}=0$, the quantization axis is arbitrary, so $\theta_q=\pi/2$ was chosen.[@Cu2] For ${\bf B}\ne0$, the quantization axis is parallel to ${\bf B}$, so $\theta_q$ and $\phi_q$ are the angles ${\bf q}$ makes with the coordinates $\hat{\bf z}||{\bf B}$ and $\hat{\bf x}\times\hat{\bf
y}=\hat{\bf z}$. For scattering wavevectors ${\bf q}_c$ directed along the special angle $$\begin{aligned}
\theta_q^c&=&\sin^{-1}(2/3)^{1/2},\end{aligned}$$ the components of each $\tilde{\cal C}_{ij}(\omega)$ are sampled equally, and $$\begin{aligned}
S({\bf q}_c,\omega)&=&{2\over{3\pi}}f^2({\bf q}_c)\bigl[\tilde{\cal
C}_{11}(\omega)+\tilde{\cal C}_{12}(\omega)\cos(2{\bf q}_c\cdot{\bf d})\bigr]
\nonumber\\
& &\\
&=&{2\over{3}}f^2({\bf q}_c)\Bigl[\tilde{\cal C}_{11}(\omega)[1-\cos(2{\bf
q}_c\cdot{\bf d})]/\pi+\nonumber\\
& &+{{\cos(2{\bf q}_c\cdot{\bf d})}\over{S(S+1)}}\Bigl(\langle
s_z^2\rangle\delta(\omega)+{1\over{2}}\langle {\bf s}^2-s_z^2\rangle\nonumber\\
& &\times[\delta(\omega-\tilde{B})+\delta(\omega+\tilde{B})]\Bigr)\Bigr],\end{aligned}$$ where we have employed the Fourier transform of the sum rule in Eq. (\[sumrule\]). For powder samples, one can still use the special angle technique with a field to obtain $\tilde{\cal C}_{11}(\omega)$, but since the direction of ${\bf d}$ is random, one obtains $$\begin{aligned}
\overline{\cos(2{\bf q}_c\cdot{\bf d})}&=&{{\sin(2q_cd)}\over{2q_cd}},\end{aligned}$$ where $\overline{\cdot\cdot\cdot}$ is a spatial average. [@fg].
For the general case of $\theta_q\ne\theta_q^c$, however, $S({\bf
q},\omega)$ cannot be written simply in terms of the $\tilde{\cal
C}_{ij}(\omega)$. There are four factors $h_i$ for $i=0,\ldots,3$, listed in the Appendix, that depend upon ${\bf q}\cdot{\bf d}$ and $\sin\theta_q$. We then find $$\begin{aligned}
S({\bf q},\omega)&=&{{f^2({\bf
q})}\over{2\pi}}\sum_{i=1}^{6S+1}f_{S,i}h_{S,i}\nonumber\\
& &\times[\delta(\omega-\omega_{Si})+\delta(\omega+\omega_{Si})],
\label{Sqomega}\end{aligned}$$ where $h_{S,i}$ is the appropriate $h_i$ for the modes $\pm\omega_{Si}$, as indicated for $S=1/2$ and 5/2 in the Appendix. The factors $h_0$ and $h_1$ are $\propto\cos^2({\bf q}\cdot{\bf d})$, and correspond respectively to the $\omega=0$ and $\pm\tilde{B}$ modes. The factors $h_2$ and $h_3$ are $\propto\sin^2({\bf q}\cdot{\bf d})$, and correspond respectively to the modes at $\pm nJ$ and $\pm|nJ\pm\tilde{B}|$, for $n=1,\ldots,2S+1$. Since $h_0$ and $h_2$ are also $\propto\sin^2\theta_q$, whereas $h_1$ and $h_3$ are $\propto(1-{1\over{2}}\sin^2\theta_q)$, the experimenter can fine tune the single crystal data by rotating ${\bf
B}$ and ${\bf q}$ relative to ${\bf d}$.
Neutron powder data on the deuterated $S=1/2$ dimer single molecule magnet VODPO$_4\cdot{1\over{2}}$D$_2$O (V2) were taken, resulting in a fit to the AFM Heisenberg model of $|J|=7.81(4)$ meV, [@V2neutron] close to the value 7.6 meV found in the susceptibility fit. [@Johnson] This corresponds to $\overline{B}=1$ at $B\approx 66$ T, which is too large for thermodynamic studies. However, inelastic neutron scattering at 0.1-0.2$\overline{B}$ ought to be possible for this material. Single crystal data could be particularly interesting. For the Fe2 dimers, inelastic neutron scattering in a field of $\overline{B}\alt0.4-0.5$ should be possible, which would not see any level crossing effects, but could prove interesting, as indicated in Fig. 6.
Acknowledgments
===============
We thank M. Ameduri and S. E. Nagler for useful discussions. DE gratefully acknowledges support from Project Number SFB463 with the MPI-PKS, the MPI-CPfS, and IFW, and the TU Dresden.
Appendix
========
Classical Time Correlation Function
-----------------------------------
From the classical equations of motion, $${\bf S}(t)=\hat{\bf B}S_{||}+S_{\perp}[\hat{\bf x}\cos(\tilde{B}t)-\hat{\bf
y}\sin(\tilde{B}t)],$$ where $\tilde{B}=\gamma B$, $\hat{\bf B}$ is a unit vector parallel to ${\bf B}$, and $\hat{\bf x}$, and $\hat{\bf y}$ are orthogonal unit vectors satisfying $\hat{\bf
x}\times\hat{\bf y}=\hat{\bf B}$. Since $S_{||}={\bf S}\cdot\hat{\bf B}$, we have $S_{||}=S\cos\theta$ and $S_{\perp}=S\sin\theta$, and hence $S^2=S_{||}^2+S_{\perp}^2$. ${\bf S}_1(t)$ is then found to be $$\begin{aligned}
S_{1,||}(t)&=&S_{1,||}+S_{1,\perp}\cos(JSt-\phi_0),\\
S_{1\pm}(t)&=&e^{\pm i\gamma Bt}\Bigl[{{S_{1,||}S_{\perp}}\over{S_{||}}}+{{
S_{1,\perp}S_{\perp}}\over{2}}\times\nonumber\\
& &\times\Bigl({{\exp[\mp i(JSt-\phi_0)]}\over{S_{||}-S}}+{{\exp[\pm
i(JSt-\phi_0)]}\over{S_{||}+S}}\Bigr)\Bigr],\nonumber\\\end{aligned}$$ where $S_{1\pm}=S_{1x}\pm iS_{1y}$, and $\phi_0$ fixes the initial relative configuration of the two spins. Since ${\bf S}_2={\bf S}-{\bf S}_1$, we have $S_{1,||}=S_{2,||}=S_{||}/2$ and $S_{1,\perp}=-S_{2,\perp}=(S_{\perp}/S)[1-S^2/4]^{1/2}$. After averaging over $\phi_0$, $$\begin{aligned}
{\cal C}_{11}(t)&=&{1\over{4}}\langle S_{||}^2+2{{S_{\perp}^2}\over{S^2}}(1-S^2/4)\cos(SJt)\nonumber\\
& &+S_{\perp}^2\cos(\tilde{B}t)+{{(1-S^2/4)}\over{S^2}}\Bigl((S+S_{||})^2\nonumber\\
& &\qquad\times\cos[(JS+\tilde{B})t]+
(S-S_{||})^2\nonumber\\
& &\qquad\times\cos[(JS-\tilde{B})t]\Bigr)\rangle.\label{phiaverage}\end{aligned}$$ Replacing $S$ with $s$ for elegance, these classical averages are evaluated from $$\begin{aligned}
\langle\ldots\rangle&=&Z^{-1}\int_0^2sds\int_0^{\pi}{{\sin\theta
d\theta}\over{2}}e^{\alpha s^2+bs\cos\theta}\ldots,\\
Z&=&\int_0^2sds\int_0^{\pi}{{\sin\theta
d\theta}\over{2}}e^{\alpha s^2+bs\cos\theta}.\end{aligned}$$ The integrals over $\theta$ can then be written in terms of $F_0(x)=\sinh(x)/x$ and its first and second derivatives, $F_1(x)=F_0'(x)=[\cosh(x)-\sinh(x)/x]/x$ and $F_2(x)=F_0''(x)=F_0(x)-2F_1(x)/x$, respectively. We note that $F_1(x)/F_0(x)=L(x)=\coth(x)-1/x$ is the Langevin function. We find
$$\begin{aligned}
{\cal C}_{11}(t)&=&{1\over{8Z}}\int_0^2sdse^{\alpha
s^2}\Bigl[2s^2G_1(bs,\tilde{B}t)\nonumber\\
& &+(4-s^2)\Bigl(\cos(sJt)G_2(bs,\tilde{B}t)\nonumber\\
& &+G_3(bs,\tilde{B}t)\sin(sJt)\Bigr)\Bigr],\label{C11oft}\\
G_1(x,y)&=&F_2(x)+[F_0(x)-F_2(x)]\cos(y),\\
G_2(x,y)&=&F_0(x)-F_2(x)+[F_0(x)+F_2(x)]\cos(y),\nonumber\\
& &\\
G_3(x,y)&=&2F_1(x)\sin(y).\end{aligned}$$
As $T\rightarrow\infty$. we set $t^*=|J|t$, and obtain, $$\begin{aligned}
\lim_{T\rightarrow\infty}{\cal C}_{11}(t)
&=&{{[1+2\cos(\tilde{B}t)]}\over{6}}f(t^*),\\
f(t^*)&=&1-{{[1+2\cos(2t^*)]}\over{t^{*2}}}\nonumber\\
& &+{{3\sin(2t^*)}
\over{t^{*3}}}-{{3[1-\cos(2t^*)]}\over{2t^{*4}}}.\label{foft}\end{aligned}$$ We note that $f(t^*)$ was obtained previously for the zero field case.[@Mueller] We also have $\lim_{T\rightarrow\infty}{\cal C}_{12}(t)=1-\lim_{T\rightarrow\infty}{\cal
C}_{11}(t)$. These forms clearly satisfy the requirement $\lim_{T\rightarrow\infty}{\cal C}_{11}(0)=1$.
Quantum Frequency Spectrum
--------------------------
The Fourier transform $\tilde{\cal C}_{11}(\omega)$ of the real part of the autocorrelation function ${\cal C}_{11}(t)$ for the quantum dimer of spin-$S$ spins is given by $$\begin{aligned}
\tilde{{\cal C}}_{11}(\omega)&=&\sum_{i=0}^{6S+1}f_{S,i}[\delta(\omega-\omega_{Si})+\delta(\omega+\omega_{Si})],\\
f_{S,i}&=&\pi a_{S,i}/{\cal C}^S(b,\alpha).\end{aligned}$$ For both $S=1/2$ and $S=5/2$, ${\cal C}^S(b,\alpha)$ is given by $$\begin{aligned}
{\cal
C}^S(b,\alpha)&=&4S(S+1)e^{2S[b-(2S+1)\alpha]}Z.\end{aligned}$$ We note that $D=S(S+1)Z$ and that ${\cal C}^S(0,0)=4S(S+1)(2S+1)^2$.
The factors $h_i$ that weight the modes in $S({\bf q},\omega)$ given by Eq. (\[Sqomega\]) can be derived from Eqs. (\[Sgeneral\]) and (\[Q\]). In Eq. (\[Sgeneral\]), the off-diagonal terms in $S({\bf q},\omega)$ with $\alpha,\beta=1,2$ sum to zero, and the remaining off-diagonal terms all vanish. Hence, we only require the matrix elements $$\begin{aligned}
{\cal
M}_{sm,\alpha}^{s'm'}&=&(1-\hat{q}_{\alpha}^2)\Bigl|\langle
sm|Q_{\alpha}^{\dag}({\bf q},0)|s'm'\rangle\Bigr|^2/f^2({\bf q}).\end{aligned}$$ From Eq. (\[m\]) and $s_{\pm}|sm\rangle=A^{\pm}_{sm}|s,m\pm1\rangle$, where $A^{\pm}_{sm}=[s(s+1)-m(m\pm1)]^{1/2}$, we write [@CS] $$\begin{aligned}
\langle
sm|S_{nz}|s'm'\rangle&=&\delta_{m',m}[m\delta_{s',s}/2+(-1)^{n-1}\nonumber\\
& &\times(B_{sm}\delta_{s',s+1}+C_{sm}\delta_{s',s-1})],\label{s1z}\\
\langle
sm|S_{n\pm}|s'm'\rangle&=&\delta_{m',m\pm1}[A_{sm}^{\pm}\delta_{s',s}/2+(-1)^{n-1}\nonumber\\
& &\times(B_{sm}^{\pm}\delta_{s',s+1}+C_{sm}^{\pm}\delta_{s',s-1})],\label{s1pm}\end{aligned}$$ for $n=1,2$, respectively. We note that $s_z=S_{1z}+S_{2z}$, $s_{\pm}=S_{1\pm}+S_{2\pm}$. We find $$\begin{aligned}
{\cal
M}_{sm,\alpha}^{s'm'}&=&\delta_{m',m}\bigl[\delta_{s',s}m^2h_0\nonumber\\
& &+4h_2\bigl(B_{sm}^2
\delta_{s',s+1}+C_{sm}^2\delta_{s',s-1}\bigr)\bigr]\nonumber\\
& &+\delta_{m',m\pm1}\Bigl[h_1\bigl(A_{sm}^{\pm}\bigr)^2\delta_{s',s}/2+2h_3
\nonumber\\
& &\times\Bigl(\bigl(B_{sm}^{\pm}\bigr)^2\delta_{s',s+1}+\bigl(C_{sm}^{\pm}
\bigr)^2\delta_{s',s-1}\Bigr)\Bigr],\end{aligned}$$ where the $h_i$ and the changes $\Delta s=s'-s$ and $\Delta m=m'-m$ in the matrix elements for which they occur are listed in Table I. For arbitrary $S$, each of the $h_{Si}$ is equal to one of the four $h_i$ listed in Table I. For $S=1/2$ and 5/2, the appropriate choices of the $h_{Si}$ are listed in Tables II and III, respectively. We remark that by setting $f({\bf q})=1$ and each of the $h_i=1$, $S({\bf q},\omega)\rightarrow\tilde{C}_{11}(\omega)/(2\pi)$.
$i$ $\Delta s$ $\Delta m$ $h_i$
----- ------------ ------------ ------------------------------------------------------------
0 0 0 $\sin^2\theta_q\cos^2({\bf q}\cdot{\bf d})$
1 $0$ $\pm1$ $(1-{1\over{2}}\sin^2\theta_q)\cos^2({\bf q}\cdot{\bf d})$
2 $\pm1$ 0 $\sin^2\theta_q\sin^2({\bf q}\cdot{\bf d})$
3 $\pm1$ $\pm1$ $(1-{1\over{2}}\sin^2\theta_q)\sin^2({\bf q}\cdot{\bf d})$
: Factors $h_i$ that appear in $S({\bf q},\omega)$, Eq. (31), and their associated transition quantum number changes.[]{data-label="tab1"}
### Quantum Frequencies for $S=1/2$
For $S=1/2$, we have $$\begin{aligned}
{\cal C}^{1/2}(b,\alpha)&=& 3(1+e^b+e^{2b}+e^{b-2\alpha}),\end{aligned}$$
The mode frequencies $\omega_{1/2,i}$, their relative amplitudes $a_{1/2,i}$, and the factors $h_{1/2,i}$ appearing in $S({\bf q},\omega)$ are given in Table II.
$i$ $\omega_{1/2,i}$ $h_{1/2,i}$ $a_{1/2,i}$
----- ------------------ ------------- -------------------------
0 0 $h_0$ $1+e^{2b}$
1 $|\tilde{B}|$ $h_1$ $(1+e^b)^2$
2 $|J|$ $h_2$ $e^b(1+e^{-2\alpha})$
3 $|J-\tilde{B}|$ $h_3$ $e^b(e^b+e^{-2\alpha})$
4 $|J+\tilde{B}|$ $h_3$ $1+e^{b-2\alpha}$
: Frequency $\omega_{S,i}$ and weighting factor $h_{S,i}$ spectra of the $S=1/2$ Heisenberg dimer in a magnetic field.[]{data-label="tab2"}
### Quantum Frequencies for $S=5/2$
For simplicity, we set $$\begin{aligned}
A_s(b)&=&{{(2s+1)\sinh[(2s+1)b/2]}\over{\sinh(b/2)}},\\
X_s(b)&=&{{e^{-b/2}}\over{4\sinh^3(b/2)}}\Bigl(e^{sb}\sinh(sb)\nonumber\\
& &-s\sinh(b)-4s^2\sinh^2(b/2)\Bigr),\\
Y_s(b)&=&X_s(-b)={{e^{b/2}}\over{4\sinh^3(b/2)}}\Bigl(e^{-sb}\sinh(sb)\nonumber\\
& &-s\sinh(b)+4s^2\sinh^2(b/2)\Bigr)\end{aligned}$$ We note that $X_1(b)=Y_1(b)=1$ and that $X_s(0)=Y_s(0)=s(4s^2-1)/3$. The mode frequencies $\omega_{5/2,i}$, their relative amplitudes $a_{5/2,i}$, and the $h_{5/2,i}$ factors are given in Table III. In evaluating the coefficients $a_{5/2,i}$ at $\alpha=b=0$, it is useful to employ the relations $$\begin{aligned}
\sum_{n=1}^{2S}\sum_{s=1}^ns^2&=&{{S(S+1)(2S+1)^2}\over{3}},\\
\sum_{s=0}^n(2s+1)^2&=&{{(n+1)(2n+1)(2n+3)}\over{3}}\end{aligned}$$
$i$ $\omega_{5/2,i}$ $h_{5/2,i}$ $a_{5/2,i}$
----- ------------------ ------------- ----------------------------------------------------------------------
0 0 $h_0$ $2e^{5(b-6\alpha)}\sum_{n=1}^{2S}e^{\alpha n(n+1)}\times$
$\times\sum_{s=1}^{n}s^2\cosh(sb)$
1 $|J|$ $h_2$ ${{35}\over{3}}(1+e^{-2\alpha})e^{5b-28\alpha}$
2 $2|J|$ $h_2$ ${{32}\over{15}}e^{5b-24\alpha}(1+e^{-4\alpha})\sum_{s=0}^1A_s(b)$
3 $3|J|$ $h_2$ ${{27}\over{35}}e^{5b-18\alpha}(1+e^{-6\alpha})\sum_{s=0}^2A_s(b)$
4 $4|J|$ $h_2$ ${{140}\over{441}}e^{5b-10\alpha}(1+e^{-8\alpha})\sum_{s=0}^3A_s(b)$
5 $5|J|$ $h_2$ ${1\over{9}}(1+e^{-10\alpha})e^{5b}\sum_{s=0}^4A_s(b)$
6 $|\tilde{B}|$ $h_1$ $2\coth(b/2)e^{5(b-6\alpha)}\sum_{n=1}^{2S}e^{\alpha n(n+1)}\times$
$\times\sum_{s=1}^{n}s\sinh(sb)$
7 $|\tilde{B}+J|$ $h_3$ ${{35}\over{3}}(1+e^{b-2\alpha})e^{4b-28\alpha}X_1(b)$
8 $|\tilde{B}+2J|$ $h_3$ ${{32}\over{15}}e^{3b-24\alpha}(1+e^{b-4\alpha})X_2(b)$
9 $|\tilde{B}+3J|$ $h_3$ ${{27}\over{35}}e^{2b-18\alpha}(1+e^{b-6\alpha})X_3(b)$
10 $|\tilde{B}+4J|$ $h_3$ ${{140}\over{441}}e^{b-10\alpha}(1+e^{b-8\alpha})X_4(b)$
11 $|\tilde{B}+5J|$ $h_3$ ${1\over{9}}(1+e^{b-10\alpha})X_5(b)$
12 $|\tilde{B}-J|$ $h_3$ ${{35}\over{3}}e^{5b-28\alpha}(e^{b}+e^{-2\alpha})Y_1(b)$
13 $|\tilde{B}-2J|$ $h_3$ ${{32}\over{15}}e^{6b-24\alpha}(e^b+e^{-4\alpha})Y_2(b)$
14 $|\tilde{B}-3J|$ $h_3$ ${{27}\over{35}}e^{7b-18\alpha}(e^b+e^{-6\alpha})Y_3(b)$
15 $|\tilde{B}-4J|$ $h_3$ ${{140}\over{441}}e^{8b-10\alpha}(e^b+e^{-8\alpha})Y_4(b)$
16 $|\tilde{B}-5J|$ $h_3$ ${1\over{9}}e^{9b}(e^b+e^{-10\alpha})Y_5(b)$.
: Frequency $\omega_{S,i}$ and weighting factor $h_{S,i}$ spectra of the $S=5/2$ Heisenberg dimer in a magnetic field.[]{data-label="tab3"}
Classical Frequency Spectrum
----------------------------
From Eq. (\[cft\]) and letting $\tilde{\omega}=\omega/J$ and $\overline{B}=\tilde{B}/J$, the classical spin Fourier transform $\tilde{\cal
C}_{11}(\omega)$ has the following discrete and continuous contributions,
$$\begin{aligned}
\tilde{\cal
C}_{11}(\omega)&=&\delta(\tilde{\omega})C_{00}+\delta(\tilde{\omega}-
\overline{B})C_{01}+\delta\tilde{\cal C}_{11}(\tilde{\omega}),\\
\delta\tilde{\cal C}_{11}(\tilde{\omega})&= &
\sum_{i=1}^4C_i(\tilde{\omega}),\\
C_{00}&=&{{\pi}\over{4Z|J|}}\int_0^2s^3dse^{\alpha
s^2}F_2(bs),\\
C_{01}&=&{{\pi}\over{4Z|J|}}\int_0^2s^3dse^{\alpha
s^2}[F_0(bs)-F_2(bs)],\\
C_1(\tilde{\omega})&=&{{\pi}\over{4Z|J|}}\Theta(2-\tilde{\omega})2\tilde{\omega}(1-\tilde{\omega}^2/4)e^{\alpha \tilde{\omega}^2}\times\nonumber\\
& &\times[F_0(b\tilde{\omega})
-F_2(b\tilde{\omega})],\\
C_2(\tilde{\omega})&=&{{\pi}\over{4Z|J|}}\Theta(2-\overline{B}-\tilde{\omega})
(\tilde{\omega}+\overline{B})[1-(\tilde{\omega}+\overline{B})^2/4]\times\nonumber\\
& &\times e^{\alpha(\tilde{\omega}+\overline{B})^2}\Bigl(F_0[b(\tilde{\omega}+\overline{B})]+\nonumber\\
& &
F_2[b(\tilde{\omega}+\overline{B})]+2F_1[|b|(\tilde{\omega}+\overline{B})]\Bigr),\\
C_3(\tilde{\omega})&=&{{\pi}\over{4Z|J|}}\Theta(\overline{B}-\tilde{\omega})\Theta(\tilde{\omega}+2-\overline{B})(\overline{B}-\tilde{\omega})
\times\nonumber\\
& &\times [1-(\tilde{\omega}-\overline{B})^2/4]e^{\alpha(\overline{B}-\tilde{\omega})^2}\Bigl(F_0[b(\overline{B}-\tilde{\omega})]\nonumber\\
& &+F_2[b(\overline{B}-\tilde{\omega})]+2F_1[|b|(\overline{B}-\tilde{\omega})]
\Bigr),\\
C_4(\tilde{\omega})&=&{{\pi}\over{4Z|J|}}\Theta(\tilde{\omega}-\overline{B})\Theta(2+\overline{B}-\tilde{\omega})(\tilde{\omega}-\overline{B})\times\nonumber\\
& &\times e^{\alpha(\tilde{\omega}-\overline{B})^2}\Bigl(F_0[b(\tilde{\omega}-\overline{B})]
\nonumber\\
& &+F_2[b(\tilde{\omega}-\overline{B})]-2F_1[|b|(\tilde{\omega}-\overline{B})]
\Bigr).\end{aligned}$$
Low Temperature Classical Modes
-------------------------------
We now investigate the low $T$ behavior of the various contributions $C_i({\omega})$ to the classical $\tilde{\cal C}_{11}(\omega)$. We follow the procedure used for the isosceles triangle and equivalent neighbor models in zero field. [@AK; @KA]
The FM modal spectrum as $T\rightarrow0$ is given by $$\begin{aligned}
\Omega_1(\overline{B})&=&|\overline{B}-2|,\\
\Omega_2(\overline{B})&=&2,\\
\Omega_3(\overline{B})&=&\overline{B}+2.\end{aligned}$$ We note that $C_1$ leads to $\Omega_2$, $C_2$ and $C_3$ combine to create $\Omega_1$, and $C_4$ leads to $\Omega_3$.
As $T\rightarrow0$, the AFM mode frequencies satisfy $$\begin{aligned}
\Omega_1(\overline{B})&=&0\Theta(2-\overline{B})+(\overline{B}-2)\Theta(\overline{B}-2),\label{zero}\\
\Omega_2(\overline{B})&=&\overline{B}\Theta(2-\overline{B})+2\Theta(\overline{B}-2),\\
\Omega_3(\overline{B})&=&2\overline{B}\Theta(2-\overline{B})+(\overline{B}+2)\Theta(\overline{B}-2).\end{aligned}$$ In $\Omega_1(\overline{B})$, the 0 indicates that the maximum of the mode is at $\tilde{\omega}=0$, the same position as for $C_{00}$.
D. Gatteschi, A. Caneschi, L. Pardi, and R. Sessoli, Science [**265**]{}, 1054 (1994). A. Lascialfari [*et al.*]{}, Phys. Rev. B [**55**]{}, 14341 (1997). M.-H.Julien, Z. H. Jang, A. Lascialfari, F. Borsa, M. Horvati[’c]{}, A. Caneschi, and D. Gatteschi, Phys. Rev. Lett. [**83**]{}, 227 (1999). R. Caciuffo, G. Amoretti, A. Murani, R. Sessoli, A. Caneschi, and D. Gatteschi, Phys. Rev. Lett. [**81**]{}, 4744 (1998). Y. Furukawa, A. Iwai, K. Kumagai, and A. Yakubovsky, J. Phys. Soc. Jpn. [**65**]{}, 2393 (1996). D. A. Tennant, S. E. Nagler, A. W. Garrett, T. Barnes, and C. C. Torardi, Phys. Rev. Lett. [**78**]{}, 4998 (1997). A. W. Garrett, S. E. Nagler, D. A. Tennant, B. C. Sales, and T. Barnes, Phys. Rev. Lett. [**79**]{}, 745 (1997). A. Zheludev, G. Shirane, Y. Sasago, M. Hase, and K. Uchinokura, Phys. Rev. B [**53**]{}, 11642 (1996). H. U. G[ü]{}del, A. Furrer, and H. Blank, Inorg. Chem. [**29**]{}, 4081 (1990); H. U. G[ü]{}del, Neutron News [**7**]{} (3) 24 (1996). H. U. G[ü]{}del, A. Furrer, W. B[ü]{}hrer, and B. H[ä]{}lg, Surface sci. [**106**]{}, 432 (1981). A. Lascialfari, F. Tabak, G. L. Abbati, F. Borsa, M. Corti, and D. Gatteschi, J. Appl. Phys. [**85**]{}, 4539 (1999). F. Le Gall, F. Fabrizi de Biani, A. Caneschi, P. Cinelli, A. Cornia, A. C. Fabretti, and D. Gatteschi, Inorg. Chim. Acta [**262**]{}, 123 (1997). G. M[ü]{}ller, J. Phys. (Paris) [**C8**]{}, 1403 (1988). M. Luban and J. Luscombe, Am. J. Phys. [**67**]{}, 1161 (1999). R. Klemm and M. Luban, Phys. Rev. B [**64**]{}, 104424 (2001); cond-mat/0105050. M. Ameduri and R. Klemm, cond-mat/0108213. C. Kittel and H. Shore, Phys. Rev. [**138**]{}, A1165 (1965). R. A. Klemm and M. Ameduri, Phys. Rev. B [**66**]{}, 012403 (2002); cond-mat/0112236. D. Mentrup, J. Schnack, and M. Luban, Physica A [**272**]{}, 153 (1999). J. W. Johnson, D. C. Johnston, A. J. Jacobson, and J. F. Brody, J. Am. Chem. Soc. [**106**]{}, 8123 (1984). R. M. White, [*Quantum Theory of Magnetism*]{}, (McGraw-Hill, New York, 1970), pp. 207-225. A. Furrer and H. U. G[ü]{}del, Phys. Rev. Lett. [**39**]{}, 657 (1977). E. U. Condon and G. H. Shortley, [*The Theory of Atomic Spectra*]{}, (Cambridge University Press, Cambridge, 1964).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The thermonuclear rate of the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction has been reevaluated based on a recent precise proton separation energy measurement of $S_p$($^{43}$V)=83$\pm$43 keV. The astrophysical impact of our new rates has been investigated through one-zone postprocessing type I x-ray burst calculations. It shows that the new experimental value of $S_p$ significantly affects the yields of species between A$\approx$40–45. As well, the precision of the recent experimental $S_p$ value constrains these yields to better than a factor of three.'
author:
- 'J.J. He$^1$'
- 'A. Parikh$^{2,3}$'
- 'B.A. Brown$^{4}$'
- 'T. Rauscher$^{5,6}$'
- 'S.Q. Hou$^{1,7}$'
- 'Y.H. Zhang$^{1}$'
- 'X.H. Zhou$^{1}$'
- 'H.S. Xu$^{1}$'
title: 'Thermonuclear $^{42}$Ti($p$,$\gamma$)$^{43}$V rate in type I X-ray bursts'
---
Introduction
============
Type I X-ray bursts (XRBs) arise from thermonuclear runaways within the accreted envelopes of neutron stars in close binary systems [@bib:woo76; @bib:jos77]. About one hundred bursting systems have been identified in the Galaxy, with light curves of about 10–100 s in duration, recurrence periods of $\sim$ hours to days, and peak luminosity L$_\mathrm{peak}$$\approx$10$^{4}$–10$^{5}$ L$_\mathrm{\odot}$ (similar, [*e.g.*]{}, to L$_\mathrm{peak}$ of classical novae). During the thermonuclear runaway, an accreted envelope enriched in H and He may be transformed to matter strongly enriched in heavier species (up to A$\approx$100 [@bib:sch01; @bib:elo09]) via the $\alpha$p-process and the rapid proton capture process (rp-process) [@bib:wal81; @bib:sch98; @bib:woo04]. Current XRB models do not predict the ejection of any appreciable amounts of synthesized material during the burst. Nonetheless, calculations indicate that radiative winds generated during some bursts may eject material. Studies are ongoing to examine the viability of detecting any associated absorption features. For reviews on aspects of type I X-ray bursts, see, [*e.g.*]{}, Refs. [@bib:lew93; @bib:str06; @bib:par13].
The rp-process is largely characterized by localized $({\rm p},\gamma)$-$(\gamma,{\rm p})$ equilibrium within particular isotonic chains near the proton drip-line. Slower $\beta$-decays (followed by fast ($p$,$\gamma$) reactions) connect these isotonic chains and set the timescale for processing towards heavier nuclei. In such an equilibrium situation the abundance distribution within an isotonic chain depends exponentially on nuclear mass differences as the abundance ratio between two neighboring isotones is proportional to $\exp[S_p/kT]$, where $S_p$ is the proton separation energy and $T$ the temperature of the stellar environment. In particular, those isotonic chains with sufficiently small $S_p$ values (relative to XRB temperatures - at 1 GK, $kT$$\approx$100 keV) need to be known with a precision of at least 50–100 keV [@bib:sch98; @bib:par09]. These include, among others, $S_p$($^{26}$P), $S_p$($^{43}$V), $S_p$($^{46,47}$Mn), $S_p$($^{61}$Ga), and $S_p$($^{65}$As) [@bib:par09]. As well, reliable nuclear physics input (including precise mass values and nuclear structure information) is needed for those nuclei along the rp-process path to calculate the thermonuclear reaction rates required for XRB models. Model predictions can then be compared with [*e.g.*]{}, observations of XRB light curves to extract quantitative information about the stellar environments [@bib:sch06].
The level structure of $^{43}$V is not experimentally known. The thermonuclear rate of the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction was first estimated by Wormer [*et al.*]{} [@bib:wor94] based entirely on the properties of four states in the mirror nucleus $^{43}$Ca [@bib:end78; @bib:end90]. Later, this rate was recalculated by Herndl [*et al.*]{} [@bib:her95] using two states determined through a shell model calculation of $^{43}$V. Theoretical rates calculated using statistical models are available [@bib:jina]; however, due to the low density of excited states expected in $^{43}$V near the proton threshold, such calculations are not ideal for this reaction [@bib:rau97; @bib:rau00; @bib:rau01]. A theoretical value of $S_p$=90$\pm$200 keV from the atomic mass evaluation (AME85) [@bib:aud85] was utilized in the above rate calculations. Another theoretical value of $S_p$=190$\pm$230 keV was adopted in later AME95 [@bib:aud95] and AME03 [@bib:aud03] compilations.
Recently, precise mass measurements of nuclei along the rp-process path have become available. These measurements were made at the HIRFL-CSR (Cooler-Storage Ring at the Heavy Ion Research Facility in Lanzhou) [@bib:xia02] in an IMS (Isochronous Mass Spectrometry) mode. Masses measured include those of a series of $T_z$=-1/2 nuclei ($^{63}$Ge, $^{65}$As, $^{67}$Se, and $^{71}$Kr) [@bib:tu11; @bib:zhang12] and $T_z$=-3/2 nuclei ($^{41}$Ti, $^{43}$V, $^{45}$Cr, $^{47}$Mn, $^{49}$Fe, $^{53}$Ni, and $^{55}$Cu) [@bib:yan13]. The proton separation energy of $^{43}$V has been experimentally determined to be $S_p$=83$\pm$43 keV for the first time [@bib:yan13]. Although the predicted values in the previous compilations (AME85, AME95 and AME03) agree with the experimental value within 1 $\sigma$ uncertainties, the latter is significantly more precise. This allows the uncertainty in the rate of the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction to be dramatically reduced. In this work, the thermonuclear rate of $^{42}$Ti($p$,$\gamma$)$^{43}$V has been reevaluated using the recent experimental $S_p$ ($^{43}$V) value and new calculated resonant and direct capture (DC) rates. The astrophysical impact of our new rates has been investigated through one-zone postprocessing x-ray burst calculations.
Reaction rate calculation
=========================
Resonant rate
-------------
We begin by estimating the $^{42}$Ti($p$,$\gamma$)$^{43}$V resonant rate using exactly the level energies, half-lives and single-particle spectroscopic factors from the mirror nucleus $^{43}$Ca [@bib:end78]. A similar approach was used in Ref. [@bib:wor94]. The resonant rate is calculated by the well-known narrow resonance formalism [@bib:wor94; @bib:her95; @bib:rol88],
$$\begin{aligned}
N_A\langle \sigma v \rangle_\mathrm{res}=1.54 \times 10^{11} (AT_9)^{-3/2} \omega\gamma \mathrm{[MeV]} \mathrm{exp} \left (-\frac{11.605E_r \mathrm{[MeV]}}{T_9} \right) [\mathrm{cm^3s^{-1}mol^{-1}}].
\label{eq1}\end{aligned}$$
Here, the resonant energy $E_r$ and strength $\omega\gamma$ are in units of MeV. For the proton capture reaction, the reduced mass $A$ is defined by $A_T$/(1+$A_T$) where $A_T$ is the target mass. The resonant strength $\omega\gamma$ is defined by $$\begin{aligned}
\omega\gamma=\frac{2J+1}{2(2J+1)}\frac{\Gamma_p\times\Gamma_\gamma}{\Gamma_\mathrm{tot}}.
\label{eq2}\end{aligned}$$ Here, $J_T$ and $J$ are the spins of the target and resonant state, respectively. $\Gamma_p$ is the partial width for the entrance channel, and $\Gamma_\gamma$ is that for the exit channel. In the excitation energy range considered in this work, other decay channels are closed [@bib:aud03], and hence the total width $\Gamma_\mathrm{tot}$$\approx$$\Gamma_p$+$\Gamma_\gamma$. Similar to the approach used by Wormer [*et al.*]{}, the gamma partial widths of the unbound states in $^{43}$V were estimated by the life-times ($\tau$) of the corresponding bound states in the mirror $^{43}$Ca via $\Gamma_\gamma$=$\hbar$/$\tau$; the proton partial widths were calculated by the following equation, $$\begin{aligned}
\Gamma_{p}=\frac{3\hbar^2}{AR^2}P_{\ell}(E)C^2S_p.
\label{eq3}\end{aligned}$$ Here, $R$=1.26$\times$(1+42$^{\frac{1}{3}}$) fm is the nuclear channel radius [@bib:wor94], $P_\ell$ the Coulomb penetrability factor, and $C^2S_p$ the proton spectroscopic factor of the resonance.
For this reaction, a temperature of 2 GK corresponds to a Gamow peak $E_{x}$($^{43}$V)$\approx$1.5 MeV with a width of $\Delta$$\approx$1.2 MeV [@bib:rol88]. Therefore, its resonant rate is determined by the excited states of $^{43}$V up to $\sim$2.1 MeV. This first estimate of the resonant rate shows that the first excited state ($E_x$=0.373 MeV) dominates the resonant contribution below 0.2 GK, the second excited state ($E_x$=0.593 MeV) dominates around 0.2–1.7 GK, and the high-lying 2.067 MeV state (with much shorter life-time $\tau$=30 fs) dominates at even higher temperature. It shows that the contribution owing to those high-lying states above 2.067 MeV is negligible at temperatures of interest in XRBs.
We then improved upon this first estimate of the resonant rate. The simplest model for calculating the isobaric-multiplet-mass-equation (IMME) is the $0f_{7/2}$ shell model used in [@bib:bro79] where the displacement energies in the mass region A=41-55 were used to deduce the effective isovector and isotensor two-body matrix elements. The root-mean-square difference between experiment and theory for 60 $\Delta Z$=1 displacement energies was 12 keV. With this model the $\Delta Z$=3 displacement energy difference between $^{43}$Ca and $^{43}$V (7/2$^-$) state is predicted to be 22.854(36) MeV compared to the new experimental value of 22.857(43) MeV. The agreement is impressive. In the framework of an OXBASH [@bib:oxbash] shell model, the resonant parameters of the three states discussed above have been recalculated and summarized in Table \[table1\]. These calculations are discussed in detail in Appendix A.
$E_x$($^{43}V$) (MeV) $E_r$ (MeV) $\tau$ (ps) $J^{\pi}$ $\ell$ $C^2S_p$ $\Gamma_\gamma$ (eV) $\Gamma_p$ (eV) $\omega\gamma$ (eV)
----------------------- -------------- ------------- ----------- -------- ---------- ----------------------- ----------------------- ------------------------------
0.436(0.050) 0.353(0.066) 22(2) 5/2$^-$ 3 0.15 3.04$\times$10$^{-5}$ 5.10$\times$10$^{-9}$ 1.5$\times$10$^{-8}$
lower: 1.4$\times$10$^{-10}$
upper: 6.3$\times$10$^{-7}$
0.537(0.050) 0.454(0.066) 117(6) 3/2$^-$ 1 0.046 3.42$\times$10$^{-6}$ 6.27$\times$10$^{-5}$ 6.5$\times$10$^{-6}$
lower: 2.2$\times$10$^{-6}$
upper: 1.1$\times$10$^{-5}$
2.067(0.100) 1.984(0.109) 0.03(0.01) 7/2$^-$ 3 0.0003 2.19$\times$10$^{-2}$ 3.45$\times$10$^{-2}$ 5.4$\times$10$^{-2}$
lower: 2.2$\times$10$^{-2}$
upper: 1.1$\times$10$^{-1}$
Resonance energies calculated using $E_r$=$E_x$($^{43}V$)-$S_p^{\mathrm{exp}}$, where $S_p^{\mathrm{exp}}$=83$\pm$43 keV. Estimated theoretical uncertainties in the parenthesis. Value from the previous ($p$,$d$) [@bib:sam68a] and ($d$,$t$) [@bib:do76] experiments. Averaged value from the ($d$,$p$) experiments [@bib:do66b; @bib:br74b]. Value calculated by the OXBASH code with same model-space and interactions as in Ref.[@bib:her95].
Direct capture rate
-------------------
The nonresonant direct capture (DC) rate can be estimated using methods presented in Refs. [@bib:rol88; @bib:her95],
$$\begin{aligned}
N_A\langle \sigma v \rangle = N_A\left (\frac{8}{\pi A}\right)^{1/2}\frac{1}{(kT)^{3/2}}\int^{\infty}_{0}S_\mathrm{dc}(E) \mathrm{exp} \left[-\frac{E}{kT}-\frac{b}{E^{1/2}} \right]dE
\label{eq5}\end{aligned}$$
If $S_\mathrm{dc}(E)$ factor is nearly a constant over the Gamow window, the nonresonant reaction rate can be approximated in a form of [@bib:rol88; @bib:her95]
$$\begin{aligned}
N_A\langle \sigma v \rangle_\mathrm{dc}=7.83 \times 10^{9} \left( \frac{Z}{A} \right)^{1/3}T_9^{-2/3}S_\mathrm{dc}(E_0) \mathrm{[MeV~b]}\times \mathrm{exp} \left[ -4.249 \left (\frac{Z^2 A}{T_9} \right)^{1/3} \right] [\mathrm{cm^3s^{-1}mol^{-1}}].
\label{eq6}\end{aligned}$$
The critical parameter is $S_\mathrm{dc}(E_0)$, the astrophysical $S$-factor at the Gamow energy $E_0$. Herndl *et al.* listed an effective $S_\mathrm{dc}(E_0)$ factor of 4.91$\times$10$^{-20}$ \[MeV b\] in their Table XIII. We have recalculated this factor and found that the above number is actually 4.91$\times$10$^{-2}$ \[MeV b\].
In this work, the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction rate from direct capture into ground state of $^{43}$V has been calculated with a RADCAP code [@bib:ber03; @bib:hua10] by using a Woods-Saxon nuclear potential (central + spin orbit) and a Coulomb potential of a uniform charge distribution. The nuclear potential parameters were determined by matching the bound-state energy ($E_b$=83 keV). A spectroscopic factor of $C^2S$=0.75 [@bib:her95], which agrees with the ($d$,$p$) experimental values of 0.68 [@bib:do66b] and 0.55 [@bib:br74b], was adopted in the present calculations. The DC rate contributes to the total rate only by 10–20% in the temperature region of 2–3 GK, and dominates the rate below 0.07 GK. The RADCAP calculations are described in detail in Appendix B.
Total reaction rate
-------------------
The total reaction rate of $^{42}$Ti($p$,$\gamma$)$^{43}$V has been calculated by simply summing up the resonant and DC contributions. Our new rate is tabulated in Table \[table2\] and plotted in Fig. \[fig1\]. The uncertainty in the present rate arises from uncertainties in our adopted $E_r$ (which also lead to uncertainties in the strengths since the values of $\Gamma_p$ and $\Gamma_\gamma$ have been scaled using the values of $E_r$ - see Appendix A) and the uncertainty in the DC contribution ($\approx$40% - see Appendix B). In addition, we have assumed a factor of two uncertainty in the adopted spectroscopic factors. The uncertainty of the total rate is dominated by the uncertainty of the $S_p$ value due to the exponential dependence of the rate on $S_p$. The rate based upon calculations in Herndl [*et al.*]{}, where only two states (at $E_x$=0.36, 0.55 MeV) were assumed, is also shown in Fig. \[fig1\] for comparison. Because the uncertainty of the DC contribution was not determined in Herndl [*et al.*]{}, the uncertainty of Herndl [*et al.*]{} rate shown originates only from those of the calculated resonant rates ([*i.e.*]{}, the error of $S_p$ propagating into the strengths). It shows our new rate calculated with the precise experimental $S_p$ value has much smaller uncertainties than the previous ones. This clearly demonstrates the importance of precise mass measurements.
![\[fig1\] (Color online) Total reaction rate calculated for the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction (in units of cm$^3$ mole$^{-1}$ s$^{-1}$). The upper and lower limits of the present rate (with $S_p$=83$\pm$43 keV [@bib:yan13]) are shown by the (red) thicker lines, and those of the Herndl *et al.* rate (with $S_p$=90$\pm$200 keV [@bib:her95]) are shown by the black thin lines. See text for details.](figure1.eps){width="8.8cm"}
T \[GK\] DC Resonant Total
---------- ------------------------ ------------------------- ------------------------
0.01 3.69$\times$10$^{-57}$ 2.96$\times$10$^{-178}$ 3.69$\times$10$^{-57}$
0.02 1.36$\times$10$^{-43}$ 9.44$\times$10$^{-90}$ 1.36$\times$10$^{-43}$
0.03 5.18$\times$10$^{-37}$ 2.31$\times$10$^{-60}$ 5.18$\times$10$^{-37}$
0.04 7.32$\times$10$^{-33}$ 1.00$\times$10$^{-45}$ 7.32$\times$10$^{-33}$
0.05 6.56$\times$10$^{-30}$ 5.65$\times$10$^{-37}$ 6.56$\times$10$^{-30}$
0.06 1.17$\times$10$^{-27}$ 3.66$\times$10$^{-31}$ 1.17$\times$10$^{-27}$
0.08 2.25$\times$10$^{-24}$ 6.15$\times$10$^{-24}$ 8.40$\times$10$^{-24}$
0.09 4.04$\times$10$^{-23}$ 1.53$\times$10$^{-21}$ 1.57$\times$10$^{-21}$
0.10 4.87$\times$10$^{-22}$ 1.24$\times$10$^{-19}$ 1.24$\times$10$^{-19}$
0.20 8.20$\times$10$^{-16}$ 7.58$\times$10$^{-11}$ 7.58$\times$10$^{-11}$
0.30 8.64$\times$10$^{-13}$ 1.64$\times$10$^{-7}$ 1.64$\times$10$^{-7}$
0.40 7.01$\times$10$^{-11}$ 8.05$\times$10$^{-6}$ 8.05$\times$10$^{-6}$
0.50 1.60$\times$10$^{-9}$ 7.87$\times$10$^{-5}$ 7.87$\times$10$^{-5}$
0.60 1.75$\times$10$^{-8}$ 3.44$\times$10$^{-4}$ 3.44$\times$10$^{-4}$
0.70 1.18$\times$10$^{-7}$ 9.54$\times$10$^{-4}$ 9.54$\times$10$^{-4}$
0.80 5.75$\times$10$^{-7}$ 2.00$\times$10$^{-3}$ 2.00$\times$10$^{-3}$
0.90 2.19$\times$10$^{-6}$ 3.47$\times$10$^{-3}$ 3.47$\times$10$^{-3}$
1.00 2.19$\times$10$^{-6}$ 5.32$\times$10$^{-3}$ 5.32$\times$10$^{-3}$
1.10 1.90$\times$10$^{-5}$ 7.44$\times$10$^{-3}$ 7.46$\times$10$^{-3}$
1.20 4.66$\times$10$^{-5}$ 9.75$\times$10$^{-3}$ 9.79$\times$10$^{-3}$
1.30 1.04$\times$10$^{-4}$ 1.22$\times$10$^{-2}$ 1.23$\times$10$^{-2}$
1.40 2.15$\times$10$^{-4}$ 1.48$\times$10$^{-2}$ 1.50$\times$10$^{-2}$
1.50 4.15$\times$10$^{-4}$ 1.77$\times$10$^{-2}$ 1.81$\times$10$^{-2}$
2.00 5.61$\times$10$^{-3}$ 5.60$\times$10$^{-2}$ 6.16$\times$10$^{-2}$
2.50 3.65$\times$10$^{-2}$ 2.46$\times$10$^{-1}$ 2.82$\times$10$^{-1}$
3.00 1.55$\times$10$^{-1}$ 7.90$\times$10$^{-1}$ 9.45$\times$10$^{-1}$
: \[table2\] Reaction rates calculated for $^{42}$Ti($p$,$\gamma$)$^{43}$V. All the rates are in units of cm$^3$ mole$^{-1}$ s$^{-1}$.
Figure \[fig2\] compares five different rates for the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction: (a) present rate (Fig. \[fig1\](a)); (b) the rate from Herndl [*et al.*]{} [@bib:her95]; (c) the rate from Wormer [*et al.*]{} [@bib:wor94]; (d) the statistical model rate ths8\_v4 available in the JINA REACLIB (with $S_p$=-0.0189 MeV [@bib:cyb10]); (e) the statistical model rate rath\_v2 in the REACLIB [@bib:jina] (with $S_p$=-0.411 MeV based on the FRDM mass model [@bib:frdm]). Because of the rather similar $S_p$ value used, our new rate does not deviate significantly from those of Herndl [*et al.*]{} and Wormer [*et al.*]{} in the temperature region of interest in XRBs. Our new rate, however, is very well constrained with the precise mass measurement as shown in Fig. \[fig1\]. The statistical-model calculations deviate from our new rate considerably over the entire temperature region of interest. This demonstrates again that the statistical-model is not ideally applicable for this reaction mainly owing to the low density of low-lying excited states in $^{43}$V.
![\[fig2\] (Color online) Ratios between the present rate (see Table \[table2\]) and other available ones (Herndl 1995 [@bib:her95], Wormer 1994 [@bib:wor94], rath\_v2 [@bib:jina] and ths8\_v4 [@bib:jina]).](figure2.eps){width="6.4cm"}
Astrophysical implications
==========================
The impact of our new $^{42}$Ti($p$,$\gamma$)$^{43}$V rate was examined in the framework of one-zone XRB models. Using the representative K04 thermodynamic history ($T_\mathrm{peak}$=1.4 GK [@bib:par08]), we performed a series of postprocessing calculations to explore the role of different $^{42}$Ti($p$,$\gamma$)$^{43}$V rates and $S_p$ values on the nuclear energy generation rate ($E_\mathrm{nuc}$) and XRB yields. Rates of all other reactions in the network were left unchanged during these calculations. To be clear, in the discussion below we will refer explicitly to $^{42}$Ti($p$,$\gamma$)$^{43}$V forward rates ([*e.g.*]{}, as shown in Fig. \[fig1\]) and to the $S_p$ value used to determine the corresponding reverse rates through the principle of detailed balance (see, [*e.g.*]{}, [@bib:par09]).
No significant differences in the respective nuclear energy generation rates were found by comparing XRB calculations with the (a) present forward rate ($S_p$=83 keV for the reverse rate); (b) Herndl *et al.* forward rate ($S_p$=88 keV); (c) Wormer *et al.* forward rate ($S_p$=88 keV); and (d) ths8\_v4 forward rate ($S_p$=-19 keV). $E_\mathrm{nuc}$ determined using the rath\_v2 forward rate ((e), $S_p$=-411 keV), however, was up to 10% lower than that from the above cases (a–d) during the burst. This (minor) difference is attributed to the very different $S_p$ value used in the rath\_v2 reverse rate: the $E_\mathrm{nuc}$ from an additional XRB calculation (f) performed with a reverse rate recalculated using the rath\_v2 forward rate and $S_p$=83 keV agreed well with the $E_\mathrm{nuc}$ from cases (a–d) above. This is because an equilibrium between the forward $^{42}$Ti($p$,$\gamma$)$^{43}$V and reverse $^{43}$V($\gamma$,$p$)$^{42}$Ti processes is quickly established owing to the relatively small $S_p$(=83 keV) of $^{43}$V relative to XRB temperatures ([*e.g.*]{}, at 1 GK, $kT$$\approx$100 keV). As a result, the actual rate of the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction is only of secondary importance; instead, it is the reaction $Q$-value (or $S_p$ value) that characterizes the equilibrium abundances of $^{42}$Ti and $^{43}$V and the energy release through subsequent reactions on these species.
![\[fig3\] Abundances following one-zone XRB calculations using the K04 thermodynamic history [@bib:par08]. Abundance variations determined using the present $^{42}$Ti($p$,$\gamma$)$^{43}$V forward rate with reverse rates calculated using $\Delta S_p$=43 keV (IMP [@bib:yan13], solid black line), and $\Delta S_p$=233 keV (AME03 [@bib:aud03], dotted grey line) are indicated. As well, abundances determined using the rath\_v2 forward rate [@bib:jina] along with reverse rates calculated with $S_p$=-411 keV (FRDM [@bib:frdm], open squares) and $S_p$=83 keV (IMP [@bib:yan13], open circles) are shown. Abundances determined with the ths8\_v4 rate [@bib:jina] ($S_p$=-19 keV [@bib:cyb10], open triangles) are also shown.](figure3.eps){width="8.5cm"}
----------- ------------------------------- --------------------------------- ----------------------------------- ---------------------- -----------------------
Species IMP $\Delta S_p$ [@bib:yan13] AME03 $\Delta S_p$ [@bib:aud03] rath\_v2 (FRDM $S_p$ [@bib:frdm]) rath\_v2 (IMP $S_p$) ths8\_v4 [@bib:cyb10]
$^{42}$Ti (2.3–5.3)$\times$10$^{-5}$ (1.6–75)$\times$10$^{-6}$ 8.7$\times$10$^{-5}$ 3.9$\times$10$^{-5}$ 7.1$\times$10$^{-5}$
$^{42}$Sc (1.0–2.3)$\times$10$^{-9}$ (6.9–33)$\times$10$^{-10}$ 3.9$\times$10$^{-9}$ 1.7$\times$10$^{-9}$ 3.1$\times$10$^{-9}$
$^{43}$Ti (1.1–2.2)$\times$10$^{-8}$ (3.2–31)$\times$10$^{-9}$ 3.5$\times$10$^{-8}$ 1.7$\times$10$^{-8}$ 2.9$\times$10$^{-8}$
$^{43}$V (2.3–4.2)$\times$10$^{-8}$ (8.0–55)$\times$10$^{-9}$ 1.0$\times$10$^{-9}$ 3.2$\times$10$^{-8}$ 1.1$\times$10$^{-8}$
$^{44}$Cr (1.0–1.9)$\times$10$^{-5}$ (3.6–25)$\times$10$^{-6}$ 1.6$\times$10$^{-6}$ 1.4$\times$10$^{-5}$ 5.0$\times$10$^{-6}$
----------- ------------------------------- --------------------------------- ----------------------------------- ---------------------- -----------------------
The effects on XRB yields by using different $^{42}$Ti($p$,$\gamma$)$^{43}$V forward rates and $S_p$ values have been investigated. Fig. \[fig3\] shows representative yields in this mass range for the different cases discussed above, as determined immediately following the respective XRB calculations. No significant differences in yields were observed for cases (a–c,f) above. The two cases (d,e) with reverse rates determined using negative $S_p$ values gave somewhat different yields for species with A=42–44. For example, the negative $S_p$ values produce relatively more $^{42}$Ti but less $^{43}$V.
The dominant role of the $S_p$ value used in the reverse rate in determining the yields is clearly seen in Fig. \[fig3\] from the comparison of cases (a) (labeled as “Present, IMP $\Delta S_p$"), (d) (labeled as “ths8\_v4"), (e) (labeled as “rath\_v2, FRDM $S_p$"), and (f) (labeled as “rath\_v2, IMP $S_p$"). It shows the yields calculated with the new experimental $S_p$ value (for the reverse rate) significantly differ from those yields with other theoretical $S_p$ values. In addition, to demonstrate the impact of the uncertainty in $S_p$, we performed additional XRB calculations using the present forward rate, along with reverse rates that reflect the one sigma uncertainties in $S_p$ from AME03 ($\Delta S_p$=233 keV) and the IMP mass measurement ($\Delta S_p$=43 keV). As shown in Fig. \[fig3\], the reduced uncertainty in $S_p$ directly influences the possible ranges of mass fractions for the affected species. Indeed, the uncertainty from the IMP mass measurement leads to variations, by less than a factor of three, in the yields of the most produced isotopes in this mass region, such as $^{42,43}$Ti, $^{42}$Sc, $^{43}$V and $^{43,44}$Cr.
Summary
=======
The thermonuclear rate of the $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction has been recalculated using the recent precise proton separation energy of $S_p$=83$\pm$43 keV measured at the HIRFL-CSR facility in Lanzhou, China. We have also used new, updated calculations of the direct capture and resonant contributions to the rate. Our new rate deviates significantly from other rates found in the literature. We confirm that statistical model calculations are not ideally applicable for this reaction primarily because of the low density of low-lying excited states in $^{43}$V. We recommend that out new rate be incorporated in future astrophysical network calculations.
The astrophysical impact of our new rate has been investigated through one-zone postprocessing Type I x-ray burst calculations. Even when using dramatically different rates, we find no significant changes to the calculated nuclear energy generation rate during a representative burst. This is because equilibrium between the forward $^{42}$Ti($p$,$\gamma$)$^{43}$V and reverse $^{43}$V($\gamma$,$p$)$^{42}$Ti processes rapidly develops at XRB temperatures. As such it is the reaction $Q$-value (or $S_p$) that mainly characterizes the equilibrium abundances of $^{42}$Ti and $^{43}$V. In this respect, the present $^{42}$Ti($p$,$\gamma$)$^{43}$V rate and $S_p$($^{43}$V) value are sufficiently well known to determine the nuclear energy generation rate within the framework of the adopted XRB model. In addition, we find that the new experimental value of $S_p$ affects significantly the yields of a limited number of species with A=42–44, such as $^{42,43}$Ti, $^{42}$Sc, $^{43}$V and $^{43,44}$Cr. The precision in $S_p$ achieved from the IMP mass measurement restricts the variation of these yields to better than a factor of three. It demonstrates clearly the importance of precise mass measurements for those key nuclei (especially those waiting-point nuclei) along the rp-process occurring in x-ray bursts.
**Acknowledgments**
This work was financially supported by the National Natural Science Foundation of China (Nos. 11135005, U1232208), the Major State Basic Research Development Program of China (2013CB834406, 2013CB834401). AP was supported by the Spanish MICINN (Nos. AYA2010-15685, EUI2009-04167), by the E.U. FEDER funds as well as by the ESF EUROCORES Program EuroGENESIS, BAB was supported by the NSF grant (PHY-1068217), and TR was supported by the Swiss NSF, EuroGENESIS and the ENSAR/THEXO collaboration within the 7th Framework Programme of the EU.
Calculation of resonant parameters
==================================
Below we summarize our calculations of the resonant parameters of the three states in $^{43}$V around 0.373 MeV (Resonance 1), 0.593 MeV (Resonance 2) and 2.067 MeV (Resonance 3).
Resonance 1
-----------
In the $0f_{7/2}$ model, the energy of the first excited 5/2$^-$ state is predicted to be 63 keV higher in $^{43}$V compared to $^{43}$Ca. Thus, with the experimental energy of 0.370 keV for $^{43}$Ca we obtain a predicted excitation energy of 0.436 MeV ($E_r$=0.353 MeV) for $^{43}$V.
The 5/2$^-$ to 7/2$^-$ transition in $^{43}$Ca has an experimental $B(M1)$ value of 0.023(2) $\mu_N^2$. In the $pf$ model space with the FPD6 interaction [@bib:fpd6] the $B(M1)$ values are predicted to be 0.018, 0.025 $\mu_N^2$ for the excited states in $^{43}$Ca and $^{43}$V, respectively. Therefore, a value of $B(M1)$=0.032 $\mu_N^2$ is derived for the predicted 0.436 MeV state, and $\Gamma_\gamma$ is thus calculated to be about 3.04$\times$10$^{-5}$ eV. This implies that the resonance strength for this state is determined by the much smaller $\Gamma_p$.
Wormer *et al.* estimated a resonant strength value of $\omega\gamma=$1.0$\times$10$^{-10}$ eV for the $E_r$=0.28 MeV state. A spectroscopic factor of 0.014 would reproduce their proton width of $\Gamma_p$$\approx$3.3$\times$10$^{-11}$ eV based on Eq. \[eq3\]. Later, Herndl [*et al.*]{} estimated a value of $\Gamma_p$=1.1$\times$10$^{-12}$ eV with a spectroscopic factor of 0.008, by the equation of $$\begin{aligned}
\Gamma_{p}=C^2S_p \times \Gamma_\mathrm{sp},
\label{eq4}\end{aligned}$$ where the single-particle width $\Gamma_\mathrm{sp}$ was calculated from the scattering phase shifts in a Woods-Saxon potential [@bib:bro93; @bib:cha93] whose depth was determined by matching the resonant energy. Based on the method introduced in Ref. [@bib:ili97], we have recalculated this proton width using using Equ. \[eq4\] with potential parameters of $E_r$=0.27 MeV, $r_0$=1.17 fm, $a$=0.69 fm and $r_c$=1.28 fm. The justification of the choice of the parameters can be found in [@bib:ili97]. We have obtained a single-particle width of $\Gamma_\mathrm{sp}$=1.76$\times$10$^{-10}$ eV, which roughly agrees with the value of 1.38$\times$10$^{-10}$ eV calculated by Herndl *et al.*
Neutron spectroscopic factor measurements imply values of $S_n$$\approx$0.15 [@bib:sam68a; @bib:do76], in disagreement with the values assumed by Wormer [*et al.*]{} and Herndl [*et al.*]{} We assumed $S_p$=$S_n$ in the following proton width calculations. With Equ. \[eq4\], a width of $\Gamma_p$=5.10$\times$10$^{-9}$ eV was obtained with the above parameters of $r_0$=1.17 fm, $a$=0.69 fm and $r_c$=1.28 fm (parameter Set 3 in Table \[table4\]). The proton width calculated by Eq. \[eq3\] is always larger than that by Eq. \[eq4\], because the former equation does not take the dimensionless single-particle reduced width $\theta_\mathrm{sp}^2$ [@bib:ili97] into account. $\theta_\mathrm{sp}^2$ is usually assumed to be unity, although this is not appropriate for many cases [@bib:ili97]. Here, $\theta_\mathrm{sp}^2$ is calculated to be 0.24.
Resonance 2
-----------
A description of the 3/2$^-$ excited state requires the full $pf$ shell-model basis. With the empirically determined isospin-nonconserving interactions for the $pf$ shell [@bib:orm89], the second excited 3/2$^-$ state in $^{43}$V is estimated to be located at $E_x$=0.537 MeV ($E_r$=0.454 MeV).
A spectroscopic factor of $C^2S$=0.046 averaged from the ($d$,$p$) experiments [@bib:do66b; @bib:br74b] was used for this state, as adopted in Ref. [@bib:wor94]. The proton width is calculated to be 6.27$\times$10$^{-5}$ eV ($\theta_\mathrm{sp}^2$=0.56) with the same parameter Set 3 (Table \[table4\]). Since it is difficult to make a reliable life-time calculation for this state, we estimated this $\Gamma_\gamma$ based on the mirror life-time. In the mirror $^{43}$Ca, this state decays either to the ground state ($J^{\pi}$=7/2$^-$) or to the first excited state ($J^{\pi}$=5/2$^-$) with branching ratios [@bib:end90] of 70.2% and 29.8%, respectively. The ground-state transition is a pure $E2$, whose width can be estimated by the relation of $\Gamma_\gamma(E2)$=$S$$\times$$\Gamma^W_\gamma(E2)$$\times$$BR$ [@bib:end79]. Here, $S$ is the strength of the transition in Wiesskopf units, and $BR$ is the branching ratio (70.2%). The Weisskopf-unit gamma width (in eV) for an $E2$ transition is $\Gamma^W_\gamma(E2)$=4.9$\times$10$^{-8}$A$^{4/3}$$E_\gamma^5$ [@bib:end79; @bib:wil60] with A=43. This results in a ground-state-transition width of $\Gamma_\gamma(E2)$$\approx$1.7$\times$10$^{-6}$ eV with $S$=7.2 [@bib:end79]. The first-excited-state transition is a mixture of $M1$ and $E2$, where the dominant $M1$ width can be calculated by the relation of $\Gamma_\gamma(M1)$=$S$$\times$$\Gamma^W_\gamma(M1)$$\times$$BR$. Here, a value of $S$=7.6$\times$10$^{-3}$ [@bib:end79] was adopted in the calculation. The Weisskopf-unit gamma width (in eV) for an $M1$ transition is $\Gamma^W_\gamma(M1)$=2.1$\times$10$^{-2}$$E_\gamma^3$ [@bib:end79; @bib:wil60]. $\Gamma_\gamma(M1)$ is estimated to be about 4.9$\times$10$^{-7}$ eV with a branching ratio of 29.8%. Therefore, only the ground-state-transition dominates the actual total $\Gamma_\gamma$ width, and the energy dependence of $\Gamma_\gamma$ can be accounted for by using the scale factor $E_\gamma^5$. For the 0.593-MeV state in $^{43}$Ca, $\Gamma_\gamma$ is about 5.62$\times$10$^{-6}$ eV (as estimated from the lifetime of 117 ps) In this work, we have adopted a value of $\Gamma_\gamma$=3.42$\times$10$^{-6}$ eV for the 0.537-MeV state in $^{43}$V by correcting for the energy difference between $^{43}$V ($E_x$=0.537 MeV) and $^{43}$Ca ($E_x$=0.593 MeV).
Resonance 3
-----------
The higher-lying 2.067-MeV 7/2$^-$ state in $^{43}$Ca is not described in the $pf$ model space, and requires nucleons to be excited from the $sd$ shell for its description. We do not have a good model for its displacement energy and simply use the same value for its excitation energy in $^{43}$V with an estimated error of 100 keV.
The $\Gamma_\gamma$ for this state was calculated to be 2.19$\times$10$^{-2}$ eV with a mirror life-time of $\tau$=0.03 ps. In the mirror $^{43}$Ca, this state mainly decays to the ground state ($J^{\pi}$=7/2$^-$) and to the first excited state ($J^{\pi}$=5/2$^-$) with branching ratios [@bib:end90] of 78% and 22%, respectively; both $\gamma$ transitions have $M1$($E2$) characters. By using the same strength $S$ value for the above 0.537 MeV state with respect to $E2$ and $M1$ transitions, $\gamma$ widths of the ground-state and first-excited transitions were calculated. It is found that the ground-state $E2$ transition dominates the total $\Gamma_\gamma$ for this state. Therefore, the factor $E_\gamma^5$ was again used to account for the energy dependence of $\Gamma_\gamma$. The proton width $\Gamma_p$ was calculated to be 3.45$\times$10$^{-2}$ eV with parameter Set 3 (Table \[table4\]). We have used a spectroscopic factor of 0.0003 as determined with the OXBASH code (using the same model-space and interactions as in Ref. [@bib:her95]). This factor may be larger in nature, and should be determined experimentally.
Calculation of direct capture rate
==================================
Parameters Set 1 [@bib:hua10] Set 2 [@bib:per63] Set 3 [@bib:ili97]
---------------------------- ------------------------------------------------------------ -------------------------------------------------------- ----------------------------------------------------------------------------------------------
$R_0$=$R_{so}$ (fm) 1.25$\times$(1+42)$^{\frac{1}{3}}$ \[$r_0$=$r_{so}$=1.26\] 1.25$\times$42$^{\frac{1}{3}}$ \[$r_0$=$r_{so}$=1.25\] 1.25$\times$42$^{\frac{1}{3}}$-0.23 \[=1.17$\times$42$^{\frac{1}{3}}$, $r_0$=$r_{so}$=1.17\]
$R_c$ (fm) 1.25$\times$(1+42)$^{\frac{1}{3}}$ \[$r_c$=1.26\] 1.25$\times$42$^{\frac{1}{3}}$ \[$r_c$=1.25\] 1.24$\times$42$^{\frac{1}{3}}$+0.12 \[=1.28$\times$42$^{\frac{1}{3}}$, $r_c$=1.28\]
$a_0$=$a_{so}$ (fm) 0.65 0.65 0.69
$V_{so}$ (MeV) -10.0 -10.0 -10.0
$V_{0}$ (MeV) -100.48 -101.73 -111.22
$S_\mathrm{dc}(0)$ (MeV b) 3.98$\times$10$^{-2}$ 3.84$\times$10$^{-2}$ 3.48$\times$10$^{-2}$
$r_0$, $r_{so}$ and $r_c$ are commonly defined as $r$=$R$/$A^{\frac{1}{3}}$ for comparison. The choice of parameters can be found in Ref. [@bib:ili97]. $V_0$ is varied to match the bound-state energy $E_b$=83 keV.
The astrophysical $S$ factor of the direct-capture $^{42}$Ti($p$,$\gamma$)$^{43}$V reaction has been calculated by the RADCAP code. The calculated $S_\mathrm{dc}$ factors are shown in Fig. \[fig4\] with three parameter sets listed in Table \[table4\]. With a spin-orbit potential of $V_{so}$=-10 MeV, the $S_\mathrm{dc}(E)$ factors calculated using three sets of parameters (Table \[table4\]) vary by no more than 15% over the energy range of 0–3 MeV. This energy range covers the Gamow window for a temperature up to 3 GK. The above changes can not be regarded as substantial. Since Huang *et al.* [@bib:hua10] reproduced successfully the $S$ factors for a series of radiative capture reactions, we have adopted their potential parameters (Set 1 in Table \[table4\]) in the final DC rate calculation. The present $S_\mathrm{dc}$ factors can be well parameterized in a Taylor-series form [@bib:rol88] of $S_\mathrm{dc}(E)$=$\sum\limits_{k=1}^k$$\frac{S^{(k)}(0)}{k!}$$E^k$, where $S$ factors are in units of \[MeV b\] and $E$ in MeV. The fitted parameters are $S(0)$=3.97$\times$10$^{-2}$ \[MeV b\] for the $S$ factor at zero energy, and the derivatives with respect to energy are $S^{(1)}(0)$=3.37$\times$10$^{-2}$, $S^{(2)}(0)$=1.31$\times$10$^{-2}$, $S^{(3)}(0)$=9.72$\times$10$^{-3}$ and $S^{(4)}(0)$=1.18$\times$10$^{-2}$, respectively.
![\[fig4\] (Color online) Direct-capture $S_\mathrm{dc}$ factors calculated with three parameter sets listed in Table \[table4\]. A previous constant value of $S_\mathrm{dc}(E_0)$=4.91$\times$10$^{-2}$ \[MeV b\] (Herndl 1995 [@bib:her95]) is shown for comparison.](figure4.eps){width="8.8cm"}
{width="12cm"}
In addition, the parameter dependence on $S_\mathrm{dc}(E)$ has been studied and the results are shown in Fig. \[fig5\]. It shows that $S_\mathrm{dc}$ factor is insensitive to the parameters $V_{so}$ and $R_c$ (or $r_c$), but rather sensitive to the parameters $R_0$ (or $r_0$) and $a$. The choice of parameter ranges is based on the literature values [@bib:per63; @bib:var91; @bib:ili97]. The error of the present DC rate is estimated simply by adding in quadrature the uncertainties originating from the potential parameters discussed above; it is about $\sim$40% in the energy range of 0–3 MeV. The DC rate as a function of temperature is calculated by numerical integration of our calculated $S$ factors using an EXP2RATE code [@bib:exp2].
[99]{} S.E. Woosley and R.E. Taam, Nature **263**, 101 (1976). P.C. Joss, Nature **270**, 310 (1977). H. Schatz [*et al.*]{}, Phys. Rev. Lett. **86**, 3471 (2001). V.-V. Elomaa [*et al.*]{}, Phys. Rev. Lett. **102**, 252501 (2009). R.K. Wallace and S.E. Woosley, Astrophys. J. Suppl. **45**, 389 (1981). H. Schatz [*et al.*]{}, Phys. Rep. **294**, 167 (1998). S.E. Woosley [*et al.*]{}, Astrophys. J. Suppl. **151**, 75 (2004). W. Lewin [*et al.*]{}, Space Sci. Rev. **62**, 223 (1993). T. Strohmayer, L. Bildsten, in: W. Lewin, M. van der Klis (Eds.), Compact Stellar X-Ray Sources, (Cambridge Univ. Press, Cambridge, 2006). A. Parikh [*et al.*]{}, Prog. Part. Nucl. Phys. **69**, 225 (2013). A. Parikh [*et al.*]{}, Phys. Rev. C [**79**]{}, 045802 (2009). H. Schatz and K. E. Rehm, Nucl. Phys. [**A777**]{}, 601 (2006). L. Van Wormer [*et al.*]{}, Astrophy. J. [**432**]{}, 326 (1994). P.M. Endt and C. Van Der Leun, Nucl. Phys. [**A310**]{}, 1 (1978). P.M. Endt, Nucl. Phys. [**A521**]{}, 1 (1990). H. Herndl [*et al.*]{}, Phys. Rev. C [**52**]{}, 1078 (1995). JINA Reaclib Database, please see, https://groups.nscl.msu.edu/jina/reaclib/db/. T. Rauscher [*et al.*]{}, Phys. Rev. C [**56**]{}, 1613 (1997). T. Rauscher and F.-K. Thielemann, At. Data Nucl. Data Tables **75**, 1 (2000). T. Rauscher and F.-K. Thielemann, At. Data Nucl. Data Tables **79**, 47 (2001). G. Audi and A.H. Wapstra, Nucl. Phys. [**A432**]{}, 1 (1985). G. Audi and A.H. Wapstra, Nucl. Phys. [**A595**]{}, 409 (1995). G. Audi [*et al.*]{}, Nucl. Phys. [**A729**]{}, 337 (2003). J. W. Xia [*et al.*]{}, Nucl. Instr. Meth. [**A 488**]{}, 11 (2002). X.L. Tu [*et al.*]{}, Phys. Rev. Lett. [**106**]{}, 112501 (2011). Y.H. Zhang [*et al.*]{}, Phys. Rev. Lett. [**109**]{}, 102501 (2012). X.L. Yan [*et al.*]{}, Astrophy. J. Lett. [**766**]{}, L8 (2013). C.E. Rolfs and W.S. Rodney, *Cauldrons in the Cosmos*, (University of Chicago Press, Chicago, 1988) B.A. Brown and R. Sherr, Nucl. Phys. **A322**, 61 (1979). B.A. Brown, E. Etchegoyen, W.D.M. Rae, and N.S. Godwin, OXBASH, 1984 (unpublished). S.M. Smith [*et al.*]{}, Nucl. Phys. [**A113**]{}, 303 (1968). P. Doll [*et al.*]{}, Nucl. Phys. [**A263**]{}, 210 (1976). W.E. Dorenbusch, T.A. Belote, and O. Hansen, Phys. Rev [**146**]{}, 734 (1966). G. Brown, A. Denning, J.G.B. Haigh, Nucl. Phys. [**A225**]{}, 267 (1974). C.A. Bertulani, Comput. Phys. Commun. **156**, 123 (2003). J.T. Huang [*et al.*]{}, Atom. Data Nucl. Data Tables **96**, 824 (2010). R.H. Cyburt [*et al.*]{}, Astrophys. J. Suppl. **189**, 240 (2010). P. Möller [*et al.*]{}, Atom. Data Nucl. Data Tables **59**, 185 (1995). A. Parikh [*et al.*]{}, Astrophys. J. Suppl. **178**, 110 (2008). W.A. Richter, M.G. van der Merwe, R.E. Julies, and B.A. Brown, Nucl. Phys. [**A523**]{}, 325 (1991). B.A. Brown [*et al.*]{}, Phys. Rev. C **48**, 1456 (1993). A.E. Champagne [*et al.*]{}, Nucl. Phys. **A556**, 123 (1993). C. Iliadis, Nucl. Phys. **A618**, 166 (1997). W.E. Ormand and B.A. Brown, Nucl. Phys. **A491**, 1 (1989). P.M. Endt, Atom. Data Nucl. Data Tables **23**, 3 (1979). D.H. Wilkinson, in *Nuclear Spectroscopy*, edited by F. Ajzenberg-Selove (Academic Press, New York, 1960), Vol. B. F. G. Perey, Phys. Rev. **131**, 745 (1963). R.L. Varner [*et al.*]{}, Phys. Rep. **201**, 57 (1991). T. Rauscher, *EXP2RATE v2.1*, http://nucastro.org/codes.html.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Resonances are associated with the trapping of an intermolecular complex, and are characterized by a series of quantum numbers such as the total angular momentum and the parity, representative of a specific partial wave. Here we show how at cold temperatures the rotational quenching of HF($j$=1,2) with H is strongly influenced by the presence of manifolds of resonances arising from the combination of a single value of the orbital angular momentum with different total angular momentum values. These resonances give rise up to a two-fold increase in the thermal rate coefficient at the low temperatures characteristic of the interstellar medium. Our results show that by selecting the relative geometry of the reactants by alignment of the HF rotational angular momentum, it is possible to decompose the resonance peak, disentangling the contribution of different total angular momenta to the resonance.'
author:
- 'P. G. Jambrina'
- 'L. González-Sánchez'
- 'M. Lara'
- 'M. Menéndez'
- 'F. J. Aoiz'
title: 'Unveiling shape resonances in H + HF collisions at cold energies.'
---
\#1 \#1
Scattering resonances are pure quantum mechanical effects that appear whenever the collision energy, $E_{\rm coll}$, matches the energy of a quasi-bound state of the intermolecular complex. [@L:ACP12] Unlike other quantum effects, they can be detected straight from the experiment, for example using molecular beams [@neumark_lee:jcp85; @VOCAGM:S15; @JBSKAGM:S20; @PDRKN:A20], where they manifest as local, sharp maxima in either the cross section or angular distribution of the products.
From a conceptual point of view, resonances are pictured as the result of the trapping of the intermolecular complex in potential wells after tunneling through the barrier (shape resonance) due to the presence of quasi-bound states, or as the excitation to a state of asymptotic higher energy, which is stabilized by the potential well (Feschbach resonance). Indeed, a dense resonance structure is observed for complex-forming reactions, (see for example Refs. [@BCJKRAB:JPCA15; @JABSBH:PCCP10; @LJAL:JCP15]) where the deep potential well can stabilize a myriad of quasi-bound states. [@BL:JCP20]
Very recently, Perreault *et al.* measured the role that the initial alignment of HD plays in H$_2$+ HD collisions, which in the cold energy regime, is governed by a resonance at around 1 K. [@PMZ:S17; @PMZ:NC18] Using quantum mechanical scattering calculations, it was possible to assign this resonance to a single partial wave ($L$=2) [@CBHG:PRL18; @CB:JCP19] and, to elucidate that, for a particular combination of initial and final states, this resonance could be controlled, vanishing for a suitable alignment of the HD internuclear axis.[@JCGBBA:PRL19]
In this manuscript, we turn our attention to the FH$_2$ system, one of the most widely studied systems in reaction dynamics both experimentally [@neumark_lee:jcp85; @skodje_liu:prl2000; @qiu:science2006; @dong:science2010; @wang:science2013; @kim:science2015; @YHXCWDLASZYN:NC19] and computationally (see for example refs. [@manolo:jcsft1997; @alexander:jcp2000; @lique:jcp2008; @jalde:pccp2011; @tizniti:natchem2014; @SVH:PCCP19; @FAC:JPCA20]). In particular, we will focus on H+HF inelastic collisions. HF is ubiquitous in the universe, [@ACWDENTD:AA2011; @Indriolo_2013] and is a key tracer of molecular hydrogen in diffuse interstellar medium. [@DL:MNRAS18] Here we will show that, as far as our calculations are concerned, collisions between H and HF($j$=1,2) in the cold energy regime are dominated by a manifold of resonances, which have a strong influence in the thermal coefficients for the range of temperatures relevant to the study of the chemistry in diffuse interstellar medium. We will also study the origin of these resonances and show the extent of control that can be achieved by preparing HF with a given internuclear axis distribution resulting from the alignment of its rotational angular momentum. Strikingly, our calculations predict that, for a certain alignment of the HF molecule, it is possible not only to enhance or to diminish the intensity of the resonance, but also to split the resonance peak, allowing us to disentangle its various contributions.
{width="1.0\linewidth"}
Since the aim of this work is i) the characterization of the observed resonances, and ii) to elucidate the extent of control that can be achieved, quantum mechanical (QM) scattering calculations were performed in a dense grid of collision energies starting at very low energies, $E_{\rm coll}/k_{\rm B}<$1 mK, and up to 100 K. To get an accurate description of the dynamics, it was necessary to propagate the wave-function up to very large distances ($6\cdot 10^4$ a$_0$), which made convenient to use the atom-rigid rotor approximation. Calculations were performed using the ASPIN code [@LOPEZDURAN2008821; @GonzalezSanchez-etal:15] on the LWA-78 Potential Energy Surface, [@LWLA:JCP07] which has been recently used to study H+HF collisions at higher energies.[@DL:MNRAS18] To check the validity of the atom rigid-rotor approximation, full-dimensional QM scattering calculations were carried out for a few energies using the ABC code. [@SCM:CPC00] The agreement between the two sets of calculations is good (see Fig. S1), although the atom-rigid rotor calculations underestimates the full-dimensional calculations at the highest energies by 20%. Both sets of calculations show the same overall behavior, and the main features of the full-dimensional results are well accounted for by the rigid-rotor calculations. Figure \[Fig1\] displays the energy dependence of the rotational quenching cross sections for HF($v$=0,$j$=1,2)+H collisions in the cold energy regime as function of the collision energy, $\sigma(E_{\rm coll})$, in the 1mK–100K $E_{\rm coll}$ range. Energies in this range are not sufficient to promote transitions to higher rotational states, and, besides, the probability for the H exchange channel is negligible.[@DL:MNRAS18; @JGASA:JPCA19] Hence, the only possible transitions are $j=1\rightarrow j'=0$, and $j=2\rightarrow j'=1$, 0. For a given $E_{\rm coll}$, $\sigma(E_{\rm coll})$ for $j=1\rightarrow j'=0$ is 3-5 times higher than for $j=2\rightarrow j'=1$ due to the wider gap between adjacent rotational quantum states with increasing $j$.[@JGASA:JPCA19] The Wigner regime for both channels is attained at energies below 5mK, where $\sigma(E_{\rm coll}) \propto
E_{\rm coll}^{-1/2}$. [@SBCEFMR:JPB00; @ISCJ:NJP11] The respective $L$ partial cross sections, $\sigma^L(E_{\rm coll})$, and $J$-partial cross section, $\sigma^J(E_{\rm
coll})$, are also shown in Fig. \[Fig1\]. As expected, at the Wigner regime only the s-wave ($L$=0, $J$=$j$) partial wave contributes significantly to the cross section. The corresponding Wigner limit, $\sigma^L \propto E_{\rm
coll}^{L-1/2}$, is also found for $L > 0$.[@LJLA:PRA15]
Most notably is that, at energies above the Wigner regime, $\sigma(E_{\rm coll})$ is dominated by a narrow peak located at around 5K for the $j=2 \rightarrow j'=1$ transition and around 5.5K for the $j=1 \rightarrow j'=0$ transition. At significantly higher energies, $E_{\rm coll} \sim$ 17K, a broader albeit smaller peak shows up. Results for $j=2 \rightarrow j'=0$ are shown in Fig. S2, and although $\sigma(E_{\rm coll})$ is smaller by at least one order of magnitude, it features the same resonance peaks as $j=2 \rightarrow j'=1$. From inspection of $\sigma^L(E_{\rm coll})$, it is clear that the sharp resonance peak at 5mK is exclusively due to $L$=3 and that the second broader peak can be mainly attributed to $L$=4. Nevertheless, the analysis of $\sigma^J(E_{\rm coll})$ shows that each peak can be decomposed in a series of maxima corresponding to different $J$s and the same $L$.
A further analysis can be carried out by plotting the contributions from the different $J$ to a given $L$, the double partial cross sections $\sigma^{L,J}(E_{\rm coll})$, as shown in Fig. \[Fig2\] for $L$=3 near the resonance. For $j=2 \rightarrow j'=1$, $J=1-5$ contribute to $L$=3 ($L - j \leq J \leq L+j$), and their respective $\sigma^{L=3,J}(E_{\rm coll})$ show maxima that are shifted in a relatively broad range of $E_{\rm coll}$, as shown in Table \[t1\]. From these peaks, $J$=1 and 5 contribute the most to the overall intensity, while the contributions of $J$=2 and 3 are almost negligible. The shifting of the maxima for $J$=1 and 5 leads to the small shoulder observed in the overall resonance peak. For the $j=1 \rightarrow j'=0$ transition, only $J=$2 and mainly $J$=4 contribute to $L$=3 (due to the parity conservation). In this case, the position of the maxima is very similar and hence there are no shoulders in the overall $\sigma(E_{\rm coll})$.
![ $L$=3 partial cross sections for $j=2 \rightarrow j'=1$ (top panel) and $j=1 \rightarrow j'=0$ (bottom panel). The white area correspond to the total cross section for $L$=3, $\sigma^{L=3}(E_{\rm coll})$, while the solid lines indicate the contribution of each $J$, $\sigma^{L,J}(E_{\rm coll})$. The positions of the resonance peaks predicted by the 1D-model are shown as solid vertical lines. []{data-label="Fig2"}](Figs/Fig2.pdf){width="1.0\linewidth"}
To characterize the nature of the aforementioned resonances, 1D adiabatic effective potentials for different rovibrational states have been calculated as a function of the atom-diatom distance (see supplementary information for further details). Two of them are shown in Fig \[Fig3\]. The binding character at short distances and the centrifugal barrier are evident in the figure. By tunneling, the trapping region is accessible, supporting quasibound states that give rise to shape resonances. The energies at which the 1D model predicts the peaks of each ($L$,$J$) resonance are shown in Table \[t1\], and in Fig. \[Fig2\] as vertical ticks. As can be observed, the agreement between the energies of the resonances and the maxima of the peaks is almost perfect, allowing us to attribute these peaks to shape resonances arising from different combinations of $J$ and $L$. The lifetimes and line shapes associated with each resonance were also calculated using the 1D model (Fig. \[Fig4\]), and the scattering probabilities. The results are compared in Table \[t1\] showing that the lifetimes obtained with the two methods are in good agreement, with the 1D model predicting slightly longer lifetimes. The lifetimes for the $L$=3 resonances are significantly longer, revealing that $L$=3 and $L$=4 resonances have a different character.
![ One-dimensional adiabatic effective intermolecular potential as a function of the distance between H and the center-of-mass of HF for two combinations of $J$ and $L$ of the $j=2 \rightarrow j'=1$ transition. The energies corresponding to the maximum time delay associated with these potentials are shown as dashed lines, along with the continuum wavefunctions. The shaded area is proportional to the square of the wavefunction. []{data-label="Fig3"}](Figs/Fig3-1.pdf){width="1.0\linewidth"}
[ |c|c|c|c|c| ]{}\
$J$ & $E$ (K) & $\tau$ (ps) & $E$ (K) 1D & $\tau$ (ps) 1D\
1 & 3.9 & 11.3 & 4.0 & 14.0\
2 & 5.5 & 4.4 & 5.6 & 5.2\
3 & 7.1 & 2.4 & 7.3 & 2.6\
4 & 7.0 & 2.6 & 7.0 & 2.8\
5 & 5.1 & 4.9 & 5.1 & 6.8\
\
6 & 16.2 & 0.90 & 16.3 & 1.0\
To further clarify the origin of the differences in the lifetimes of the various $J, L$ resonances, the effective adiabatic 1D potentials for the $L$=3–$J$=5 and $L$=4–$J$=6 particular cases are shown in Fig. \[Fig3\]. Along with the potentials, we show the energy of the quasibound states supported by these potentials (dashed lines) and the corresponding squares of the wavefunctions. For the $L$=3 peak, the energy of the quasi-bound state lies below the maximum of the centrifugal barrier, and it can be properly considered as a shape resonance. However, for the $L$=4 peak, the resonance energy lies slightly above the maximum of the barrier. As a consequence, the probability is more evenly spread over all radial distances, and the resonance exhibits a smaller lifetime. This kind of resonance can thus be characterized as an over-barrier (or “above-the-barrier”) resonance, the quantum equivalent to classical-orbiting [@PDRKN:A20].
![ $t_{delay}(E)$ as a function of the collision energy for $L$=3 manifold of resonances for $j=2\rightarrow j'=1$ transition. $t_{delay}(E)$ for $L$=4 and $J$=6 is also shown for the sake of comparison. The insert shows a blow-up of the energy range where the $L=3$ resonances show up.[]{data-label="Fig4"}](Figs/Fig4.pdf){width="1.0\linewidth"}
In Ref. it was found that the strength of the resonance peak can be tuned by alignment of rotational angular momentum and hence by changing the internuclear axis distribution. To check if it is also the case for H+HF collisions, we have used the procedure outlined in Refs. to investigate how the integral cross sections change by varying the angle $\beta$ between the polarization vector of the radiation field, used to prepare the HF molecule in specific rovibrational states, and the initial relative velocity vector. Each $\beta$ value entails a distribution of internuclear axis: if $\beta$=0$^{\circ}$ collisions are preferentially head-on, while $\beta$=90$^{\circ}$ implies a side-on geometry. The respective cross sections will be denoted by $\sigma^{\beta}(E_{\rm coll})$. These preparations can be contrasted with the “isotropic” distribution, with no external alignment.
{width="0.80\linewidth"}
The left panels in Fig \[Fig5\] display the cross sections for $j=2\rightarrow j'=1$ (Fig. \[Fig5\]-a) and $j=1 \rightarrow j'=0$ (Fig. \[Fig5\]-b) in the vicinity of the $L$=3 resonance for isotropic distribution, $\beta$=0$^{\circ}$ and $\beta$=90$^{\circ}$. For $j=2\rightarrow j'=1$ (a) the $L$=3 resonance is significantly enhanced for head-on ($\beta$=0$^{\circ}$) encounters while its intensity decreases for side-on encounters. Outside the resonance region, the effect of a preferential alignment is unimportant. The most interesting feature is that the $\beta$=0 alignment is able to disentangle the peaks for $J$=1 and $J$=5; the contribution of $J$=1, which in the isotropic case manifests as a shoulder in the $L$=3 resonance, is enhanced to the point of splitting the original peak in two. Since $\beta$=0 implies collisions with $\Omega$=0 exclusively, where $\Omega$ is the projection of the total angular momentum vector onto the relative velocity, this implies that the $J$=1 partial wave has a strong component of the perpendicular projection. For $j=1\rightarrow j'=0$ (Fig. \[Fig5\]-b) the cross-section is also enhanced for $\beta$=0, and this effect is particularly prominent at the resonance. In addition, the preference for head-on collisions is observed in a broad range of $E_{\rm coll}$. In this case, the resonance peak does not split for any HF preparation, as the energies corresponding to the two shape resonances contributing to this peak are very similar. These results for both transitions evince that the trapping of the collision complex is more efficient for head-on collisions.
Figures \[Fig5\] c-d show the thermal rate coefficients, $k(T)$, calculated for the two transitions averaged over the Maxwell-Boltzmann distribution. In general, it is not guaranteed that a resonance may influence the rate coefficients significantly. However, in the present case, the resonance has a strong effect on $k(T)$ in the 1-50K temperature range, leading to more than a two-fold increase of $k(T)$ for $j$=2 $\rightarrow$ $j'$=1 at 5 K. The different intermolecular axis preparations also have a strong effect on the $k(T)$ in the same temperature range, with $\beta=0$ encounters leading to the largest $k(T)$. It also worth noticing that, while for $j$=2 $\rightarrow$ $j'$=1 the $k(T)$ calculated with the resonance artificially removed rises monotonically with $T$, for $j$=1 $\rightarrow$ $j'$=0 it starts decreasing at 40 K, effect that is even more clear for $\beta$=0.
In summary, in this study we have demonstrated that H + HF inelastic collisions in the 1-10K ($E_{\rm coll}/k_{\rm B}$) energy range are dominated by shape resonances, which are themselves formed by a cluster of resonances, each of them characterized by orbital and total angular momentum values. We have shown that a 1-D model, based on the adiabatic effective potentials, can predict the position of the each $L$-$J$ resonance very accurately. Lifetimes and line-shapes of each of the resonances were determined using the phase-shift of the 1D continuum wavefunctions. In particular, the $L$=4 resonances exhibit shorter lifetimes and can be considered orbiting (over-the-barrier) resonances where the quasi-bound states lie slightly above the centrifugal barrier. In spite of the relatively large time delays associated to the resonances, alignment of HF prior the collision changes significantly both the intensity and shape of the excitation functions, which also manifest in the thermal rate coefficients. Remarkably, for $j=2\rightarrow j'=1$, when head-on collisions are promoted, the main resonance peak splits in two, each of them associated to collisions with a particular value of the total angular momentum. These results are in contrast to those found for HD + H$_2$ collisions in the cold energy regime, for which the resonance vanishes for head-on encounters, showing that the degree of control associated to the resonance is very sensitive of the topology of the system. The influence of the resonance persists after the energy averaging and it leads to up to two-fold increase of the thermal rate coefficient at relevant temperatures of the interstellar medium where HF is ubiquitous.
The authors thank Prof. Enrique Verdasco for his support and help with the calculations. Funding by the Spanish Ministry of Science and Innovation (grant and PGC2018-096444-B-I00) is also acknowledged. P.G.J. acknowledges funding by Fundación Salamanca City of Culture and Knowledge (programme for attracting scientific talent to Salamanca).
[48]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1126/science.aad2356) @noop [****, ()]{} @noop [ ()]{}, @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1051/0004-6361/201117578) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase https://doi.org/10.1016/j.cpc.2008.07.017) [****, ()](\doibase 10.1088/1367-2630/17/12/123003) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{}, (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.109.133201) @noop [**]{} (, ) @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{}
Supplementary Material
======================
Scattering calculations:
------------------------
The usual time-independent formulation of the coupled channel (CC) method is used to solve the Schrödinger equation, in the quantum scattering calculations of an atom with a diatomic molecule, as implemented in ASPIN code.[@LOPEZDURAN2008821] The HF molecule in its singlet ground state is treated as a rigid rotor, while the H atom is considered structureless. Details of the method have been given before and discussed recently in detail for the case of the Rb + OD$^-$/OH$^-$ system [@GonzalezSanchez-etal:15] and will not be reported here. The parameters used for applying the CC method are chosen for achieving numerical convergence of the final S-matrix elements. A maximum number of rotational channels up to $j_{max}$ = 10 has been included in each CC calculation, where at least five channels were included as closed channels for each collision energy, ensuring overall convergence of the inelastic cross sections. At the lowest considered collision energies the radial integration was extended out to $R_{max}$ =$5\cdot 10^4$ Å.
To check the validity of the rigid rotor approximations, full-dimensional QM calculations were also obtained using the ABC code. [@SCM:CPC00] Calculations were carried out for 80 energies, up to 150 K, including all partial waves to convergence. The propagation was carried out in 2000 log-derivative steps up to a hyperradius of 40 a$_0$. The basis included all the accessible states up to $E$=3.0 eV.
Partial cross sections
----------------------
The expression of the probability for the $j \to j'$ transition and for given total, $J$, and orbital, $L$, angular momentum values can be written as: $$\label{eq3}
P_{j \to j'}(J,L)=\sum_{L'} \sum_{\epsilon}
|S_{v',j'L'\, \, v,j,L}^{J{\rm \epsilon}}|^2,$$ where $S_{v',j'L'\,\,v,j,L}^{J{\rm \epsilon}}$ is the scattering matrix element for the process between the reactant channel $(v,j,L)$ and the product channel $(v',j',L')$, for $J$ and the parity $\rm \epsilon=(-1)^{(j+L)}=(-1)^{(j'+L')}$. In Eq. , the sums run over the possible values of the final orbital angular momentum $L'$, and the parity $\rm \epsilon$ .
for a given value of the collision energy, the double partial cross section is given by $$\label{eq10}
\sigma^{J,L}_{j \to j'}=\frac{\pi}{k^2}\, \frac{2J+1}{2j+1} P_{j \to j'}(J,L)$$ If summed over $J$ one gets the $J$ partial cross section: $$\label{eq11}
\sum_{J=|L-j|}^{L+j} \, \sigma^{J,L}_{j \to j'}=\frac{\pi}{k^2}\,
\sum_{J=|L-j|}^{L+j} \frac{2J+1}{2j+1} P_{j \to j'}(J,L) =\sigma^L_{j \to j'}$$ Alternatively, by summing over $L$, the $J$ partial cross section is retrieved: $$\label{eq12}
\sum_{L=|J-j|}^{J+j} \, \sigma^{J,L}_{j \to j'}=\frac{\pi}{k^2}\,
\frac{2J+1}{2j+1} \sum_{L=|J-j|}^{J+j} P_{j \to j'}(J,L) =\sigma^J_{j \to j'}$$
1D-model:
---------
To understand the nature of the peaks observed in the cross-sections, and confirm their resonant nature, we have used a simple, essentially elastic, one-dimensional model. It is based on the calculation of 1D adiabatic potentials, to describe the effective interaction felt by the diatom when it gets close to the H atom. It is convenient to use $(R,r,\gamma)$ Jacobi coordinates to describe the scattering process: $R$ is the distance between H and the centre-of-mass of HF, $r$ the HF internuclear distance (kept fixed in the rigid rotor approximation) and $\gamma$ the angle between $R$ and $r$. The potentials, which depend on the initial quantum numbers of the colliding partners, have been calculated by adiabatically separating $\gamma$ from $R$. For a given value of $R$, the matrix elements of all the terms in the rigid-rotor Hamiltonian, except for the radial collision kinetic energy, are calculated in a basis of reactant channels, labeled by the set of quantum numbers $(J,M,j,L)$ and the parity $\epsilon$. Only states with the same values of $J,M$ and $\epsilon$ are coupled. This way, we diagonalize the block matrix corresponding to each pair $(J,\epsilon)$ ($M$ being irrelevant) including as many different $j$s and $L$s as needed for convergence.[@PhysRevLett.109.133201] The resulting potentials,each of them correlating with a particular set of initial quantum numbers, support quasibound states, whose energies lie very close to the observed peaks.
To predict the approximate position, $E_0$, and the width, $\Gamma$, of the resonances, we calculate the time-delay function, $t_{\rm delay}(E)$. The time-delay is (in our context) a measurement of the time that the internuclear complex is trapped in the potential well. It can be calculated from the phase-shift of the 1D continuum wavefunction using the expression: [@child:book; @Goldberger:book] $$t_{\rm delay}(E)=2 \hbar (d\phi/dE),$$ The time-delay functions provided by the 1D model are depicted in Fig.\[Fig4\]. It is interesting to note their Lorentzian character. Indeed, in the absence of background, the time-delay has a pure Lorentzian line-shape centered precisely at $E_0$, [@Smith:PR1960; @Smith:PR1963] $$\label{tdel}
t_{\rm delay}(E)=\hbar \Gamma / [(E-E_0)^2 + (\Gamma/2)^2 ]$$ where $\Gamma$ is the FWHM of the line-shape. Accordingly, $\Gamma$ and $t_{\rm
delay}(E_0)$ are related as: $$t_{\rm delay}(E_0)=\frac{4 \hbar}{\Gamma},$$ Finally, the lifetime of the resonance (average time delay) is given by: [@Smith:PR1960] $$\tau=\frac{\hbar}{\Gamma}=\frac{t_{\rm delay}(E_0)}{4}.$$
These Eqs. would allow to calculate the features of a resonance starting from the time-delay functions in the absence of background scattering. However, in the presence of background scattering (as it is the case), we need to modify somewhat the fitting function. Assuming a background phase-shift which changes linearly with energy around $E_0$ ($\delta_{back}(E)=b+a(E-E_0)$), the maximum of the total time-delay will still provide $E_0$. In turn, once $E_0$ is known, the width can be determined fitting the time-delay to the analytical expression: $$\label{tdel}
t_{\rm delay}(E)=a+\hbar \Gamma / [(E-E_0)^2 + (\Gamma/2)^2 ]$$ The resulting positions and widths predicted by the 1D model are shown in Table I of the main text. They have been compared with those extracted from the full scattering calculation. Assuming a contribution up to first order in $(E-E_0)$ from the background scattering to the S-matrix elements,[@JBSKAGM:S20] we have used the following function to fit the $L$-$J$ inelastic probabilities: $$\label{tdel}
P(E)=P_3[(E-E_0)]/ [(E-E_0)^2 + (\Gamma/2)^2 ]$$ where $P_3[(E-E_0)]$ is a third degree polynomial in the variable $(E-E_0)$, whose unknown coefficients are also given by the fitting process.
Let us finally note that, as can be easily conclude from the data in Table I, the lifetimes provided by the 1D-model are an upper bound to the full-scattering ones. Indeed, the adiabatic Hamiltonian suppresses the kinetic couplings between different adiabatic curves. A resonance supported by one of this curves, can only decay by tunneling through the centrifugal barrier. However, under the exact Hamiltonian, resonances are coupled to other adiabatic states; hence there is an alternative mechanism of decay, which is expected to disminish the lifetime.
Alignment-dependent cross sections:
-----------------------------------
Let us define a scattering frame with the $z$ axis along the initial relative velocity and the $xz$ plane as the one determined by the initial and final relative velocities. Let us note with $\beta$ the polar angle that specifies the direction of the polarization vector in the scattering frame. Following Ref. , the cross sections for a given $\beta$, is given by: $$\sigma^{\beta}=\sigma_{\rm iso} \sum_{k}^{2 j} (2 k + 1 ) s^{(k)}_0 P_{k}(\cos \beta) A^{(k)}_0$$ where $\sigma_{\rm iso}$ is the unpolarized cross section, $P_{k}(\cos \beta)$ are the Legendre polynomials, and $s^{(k)}_0$ and $A^{(k)}_0$ are the intrinsic and extrinsic polarization parameters. The latter is a geometrical factor that, for optical pumping, is given by the $\langle j 0 , k 0 | j 0
\rangle$ Clebsch-Gordan coefficient. The $s^{(k)}_0$ polarization parameters can be calculated from the S-matrix as: [@AHJAJZ:JPCL12] $$s^{(k)}_0 = \frac{\pi}{\sigma_{\rm iso}\, k^2}\,\,\sum_J \sum_{\Omega', \Omega}
(2 J + 1) |S^{J}_{j',\Omega', j, \Omega}|^2
\langle j \Omega , k 0 | j \Omega \rangle$$ where $\Omega$ and $\Omega'$ are the helicities, [*i.e.*]{}, the projections of $J$ on the directions of the reactant’s approach and the product’s recoil, respectively. $\beta$=0$^{\circ}$ (which is equivalent to $\Omega=0$) collisions are preferentially head-on, while $\beta$=90$^{\circ}$ implies side-on encounters.
The values of $\sigma^{\beta}$ ($\beta=0^{\circ},\beta=90^{\circ}$) relative to $\sigma_{\rm iso}$ are shown in Fig. S3 for the three transitions considered in this work, and in a wide range of collision energies. At the lowest energies, close to the Wigner limit, $\sigma^{\beta}$ is independent on the preparation. [@AAMSA:JCP2005; @JCGBBA:PRL19] In the vicinity of the resonance, head-on collisions prevail while at energies above 200K side-on arrangements lead to higher cross sections. The preponderance of the $\beta=0^{\circ}$ preparation in the cold regime for the $j=1,2 \rightarrow j'=0$ transitions (upper panels of Fig. S3) stems from the fact that for $j'=$0 there is only one possible parity, $(-1)^J$, which includes $\Omega=0$, and hence the weight of this projection is higher than that when the two parities contribute to scattering (the $(-1)^{J+1}$ parity does not include $\Omega$=0), as is the case for the $j=2
\rightarrow j'=1$ transition. At energies above the 100-300K side-on collisions are preferred, as expected from classical arguments.
{width="0.90\linewidth"}
{width="1.0\linewidth"}
{width="0.8\linewidth"}
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Kaoru Hagiwara,'
- Kai Ma
- and Satyanarayan Mukhopadhyay
title: 'Closing in on the chargino contribution to the muon $g-2$ in the MSSM: current LHC constraints'
---
Introduction
============
The anomalous magnetic moment of the muon, $a_\mu = (g-2)_\mu/2$, is an accurately measured quantity, which, at the same time, is precisely predicted within the standard model (SM) [@Review]. Consequently, it is an important testing ground for the SM, since new states at the electroweak scale with coupling to muons can potentially contribute to $a_\mu$ via quantum corrections. Due to the chirality-flip nature of the magnetic moment operator, the sensitivity of $a_\mu$ to new particles coupling to leptons is parametrically higher by a factor of $(m_\mu/m_e)^2$, than that of the electron anomalous magnetic moment ($a_e$), despite the latter’s higher precision measurements.
The BNL E821 experiment [@Bennett; @Roberts:2010cj] provides the current best measurement for $a_\mu$, which reads $$a_{\mu}^{\rm EXP} = (11~659~208.9 \pm 6.3) \times 10^{-10}.$$ In the SM, $a_\mu$ receives relevant QED, electroweak and hadronic contributions, with the hadronic contribution being the dominant source of the uncertainty in the theoretical prediction. We refer the reader to Ref. [@Hagiwara_gm2SM] for a detailed recent discussion of the different contributions and their associated uncertainty estimates, while for the current study, we adopt the following SM prediction for $a_\mu$ [@Hagiwara_gm2SM]: $$a_{\mu}^{\rm SM} = (11~659~182.8 \pm 4.9) \times 10^{-10}.$$ Thus, the measured value of $a_\mu$ is larger than the SM prediction by $$a_{\mu}^{\rm EXP} - a_{\mu}^{\rm SM} = (26.1 \pm 8.0) \times 10^{-10},$$ which corresponds to a $3.3 \sigma$ discrepancy.
The measurement uncertainty on $a_\mu$ is expected to be reduced further by two upcoming experiments. The FermiLab FNAL E989 experiment [@Grange:2015fou], due to start data taking in 2017, is projected to achieve a factor of four reduction in the current measurement uncertainty. Using a completely different technology with an ultra-cold muon beam which does not share the same systematic uncertainties associated with the BNL and FNAL experiments, the J-PARC E34 experiment also has a competitive potential [@JPARC]. Therefore, if the current $3.3\sigma$ deviation is truly a sign of physics beyond the SM, these follow-up measurements would lead to an enhanced statistical significance for the present discrepancy. Simultaneously, the error in the theoretical prediction within the SM also needs better control, with the recent progress in estimating the hadronic light-by-light contribution from lattice QCD computations [@lattice] being promising in this regard.
The minimal supersymmetric standard model (MSSM) contains the necessary ingredients to accommodate the above discrepancy through contributions from scalar muons, the muon sneutrino, charginos and neutralinos (the latter two will be referred to as electroweak-inos in the following). A number of studies have been devoted to the computation of the MSSM contribution to the $\gm2$ [@Moroi:1995yh; @Chattopadhyay:1995ae; @Lopez:1993vi; @Cho:2000sf; @Cho:2001nfa; @Heinemeyer:2003dq; @Stockinger:2006zn; @Marchetti:2008hw; @vonWeitershausen:2010zr; @Cho_gm2; @Fargnoli:2013zia], the constraints on the relevant parameter space from a subset of the 8 TeV LHC data [@gm2_LHC], as well as the role of future $e^+e^-$ colliders (such as the ILC) in probing the light electroweak MSSM sector that can contribute to the $\gm2$ [@gm2_ILC]. The goal of the present study is to revisit the current constraints on the dominant contribution to $\gm2$ in the MSSM, namely that of the chargino and muon-sneutrino loop, in the light of recent data from both the 8 and 13 TeV runs of the LHC.
The ATLAS and CMS collaborations have carried out multiple sets of analyses in their search for the electroweak sector of the MSSM, and have practically covered all possible decay modes of the lighter chargino ($\C1$) and the second lightest neutralino ($\N2$). The search for electroweak-inos at the LHC is complex due to the multitude of decay modes available, determined by the mass hierarchies and mixing angles in the chargino, neutralino and slepton sectors. Due to the complexity in interpreting the searches in a multi-dimensional parameter space, so far, most of the ATLAS and CMS results assume one particular decay mode of the $\C1$ and $\N2$ to have a $100\%$ branching ratio (BR), and interpret the results within a simplified model setting. Implications of the LHC search results for the electroweak MSSM sector have also been explored in a number of phenomenological studies [@ewino_LHC].
In the most general scenario, the MSSM parameters relevant for the electroweak searches of our interest in this study are the mass parameters for the bino ($M_1$), wino ($M_2$) and Higgsino ($\mu$), the masses of the lighter ($M_{\tilde{\tau}_1}$) and heavier stau ($M_{\tilde{\tau}_2}$), the mixing angle in the stau sector ($\theta_{\tilde{\tau}}$), the soft masses of the left-smuon ($M_{\smuL}$) and left-selectron ($M_{\tilde{e}_L}$), and the ratio of the vacuum expectation values of the two Higgs doublets ($\tan \beta$). All of these parameters enter the search for electroweak-inos and the left-smuon, and all except the stau and selectron sector parameters also determine the MSSM contribution to the $\gm2$ at the one-loop level. Therefore, in order to perform a completely general analysis of the current status of $\gm2$ within the MSSM in the light of LHC searches, we need to perform a nine parameter global analysis, including, at the same time, several relevant LHC search channels. The computational resources required for this analysis is beyond the scope of our present study. However, as we shall argue in the subsequent sections, a relatively small set of computations is sufficient to obtain a broad understanding of the current constraints.
The rest of the sections and the general strategy adopted in this study are as follows. In Sec. \[sec:gm2\], we provide a brief overview of the different contributions to the $\gm2$ in the MSSM. In Sec. \[smuon\] we determine, to what extent the current LHC bound on the left-smuon mass parameter can be modified in the presence of $\C1$ and $\N2$ states lighter than the $\smuL$. Such a lower bound, in turn, leads to an estimate of the maximum possible chargino mass that can explain the $\gm2$ anomaly within $1\sigma$, depending upon the relevant mixing matrices in the chargino and neutralino sectors.
As far as the $\C1$ and $\N2$ searches are concerned, we first enumerate in Sec. \[sec:chargino\] the different possible mass hierarchies between the $\C1/\N2$ and the sleptons that determine the most relevant search channel(s). We then proceed to obtain the current LHC constraints on the stau NLSP scenario in Sec. \[stau\], using the ATLAS search results in the $\geq 2\tau_h+\met$ channel. Translating the LHC simplified model-based search constraints in this case is somewhat involved, due to the required modelling of the tau jet identification and reconstruction efficiencies. Therefore, we perform a detailed Monte Carlo (MC) analysis, including detector resolution effects, to recast the 8 and 13 TeV LHC search limits in the $\geq 2\tau_h+\met$ channel to the constraints in the $\mu-M_2$ plane. Variation in the LHC bounds due to changes in the LSP mass and the stau mixing angle are also illustrated here.
In Sec. \[trilepton\], we consider the smuon/selectron NLSP scenario, and utilize the LHC searches in the trilepton and $\met$ channel to estimate the current LHC limits on the electroweak-ino sector parameters for this mass hierarchy. In Sec. \[sec:sc3\], we then combine the limits from the multi-tau search channel and the trilpeton channel to obtain the LHC constraints for the scenario in which all three generation sleptons are lighter than the chargino. In Sec. \[combi\], we revisit the $1\sigma$ favoured region for the $\gm2$, in the light of the above constraints, paying particular attention to the mass hierarchies. We provide a summary of our results in Sec. \[sum\]. Our MC simulation setup for the $\geq 2\tau_h+\met$ search and its validation against the ATLAS 8 and 13 TeV search results are discussed in the Appendix.
Muon $(g-2)$ in the MSSM – a brief overview {#sec:gm2}
===========================================
In the MSSM, $a_{\mu}$ receives contributions at the one loop level from the chargino $(\tilde{\chi}^{\pm})-$muon sneutrino $(\tilde{\nu}_{\mu})$ loop and the neutralino $(\tilde{\chi}^0)-$smuon $(\tilde{\mu})$ loop [@Moroi:1995yh; @Chattopadhyay:1995ae; @Lopez:1993vi; @Cho:2000sf; @Cho:2001nfa; @Heinemeyer:2003dq; @Stockinger:2006zn; @Marchetti:2008hw; @vonWeitershausen:2010zr; @Cho_gm2; @Fargnoli:2013zia]. In order to understand the dependence of the supersymmetric contribution to the $\gm2$ on different MSSM parameters, we follow the discussion in Ref. [@Cho_gm2], and classify the contributions in the weak eigenstate basis. In this basis, the leading terms in the $m_{\mathrm{EW}}/m_{\mathrm{SUSY}}$ expansion (here, $m_{\mathrm{EW}}$ stands for $m_{\mu}$, $m_W$ or $m_Z$ and $m_{\mathrm{SUSY}}$ stands for $m_{\tilde{\mu}}$, $m_{\tilde{\nu}}$, $M_1$, $M_2$ or $\mu$), are given by five one-loop diagrams, see Fig. 1 and Eq. (2.6) in Ref. [@Cho_gm2]. The dominant contributions proportional to the wino-Higgsino mixing are given by the charged wino – charged Higgsino – muon sneutrino loop and the neutral wino – neutral Higgsino – left smuon loop as follows:
\[eq:gm21\] $$\begin{aligned}
\Delta a_{\mu}(\tilde{W}-\tilde{H}, \tilde{\nu}_{\mu})&=\frac{g^2m_{\mu}^2}{8\pi^2}\frac{M_2\mu\tan\beta}{m_{\tilde{\nu}_{\mu}}^4}
F_a(M_2^2/m_{\tilde{\nu}_{\mu}}^2,\mu^2/m_{\tilde{\nu}_{\mu}}^2),\label{a}\\
\Delta a_{\mu}(\tilde{W}-\tilde{H}, \tilde{\mu}_L)&=-\frac{g^2m_{\mu}^2}{16\pi^2}\frac{M_2\mu\tan\beta}{m_{\tilde{\mu}_L}^4}
F_b(M_2^2/m_{\tilde{\mu}_L}^2,\mu^2/m_{\tilde{\mu}_L}^2)\label{d},\end{aligned}$$
where, the loop functions in the above expressions are given as
$$\begin{aligned}
F_a(x,y) &= -\frac{G_3(x) - G_3(y)}{x-y},\\
F_b(x,y) &= -\frac{G_4(x) - G_4(y)}{x-y},\\
G_3(x) &= \frac{1}{2(x-1)^3}[(x-1)(x-3)+2\ln{x}],\\
G_4(x) &=\frac{1}{2(x-1)^3}[(x-1)(x+1)-2x\ln{x}].\end{aligned}$$
$F_a(x,y)$ and $F_b(x,y)$ are defined to be positive for all positive $x$ and $y$, and $F_a(x,y)$ is always larger than $F_b(x,y)$ for the same arguments [@Cho_gm2]. These two contributions are enhanced when the wino and Higgsino are maximally mixed, which requires $|M_2/\mu| \sim \mathcal{O}(1)$. In the approximation of equal masses for the left-smuon and the muon sneutrino, the arguments of $F_{a,b}(x,y)$ in Eqs. (\[a\], \[d\]) are the same. Taking into account the difference of a factor of two, and the fact that $F_a(x,y) > F_b(x,y)$, we can infer that the positive discrepancy between the data and the SM prediction can be explained with $M_2\mu > 0$. We note that the right-smuon is not relevant for these two contributions.
The terms proportional to the bino-Higgsino mixing are given by the bino – neutral Higgsino – left-smuon loop and the neutral Higgsino – bino – right-smuon loop as follows:
\[eq:gm22\] $$\begin{aligned}
\Delta a_{\mu}(\tilde{B}-\tilde{H},\tilde{\mu}_L)&=\frac{g_Y^2 m_{\mu}^2}{16\pi^2}\frac{M_1\mu\tan\beta}{m_{\tilde{\mu}_L}^4}
F_b(M_1^2/m_{\tilde{\mu}_L}^2,\mu^2/m_{\tilde{\mu}_L}^2),\label{c}\\
\Delta a_{\mu}(\tilde{B}-\tilde{H}, \tilde{\mu}_R)&=-\frac{g_Y^2 m_{\mu}^2}{8\pi^2}\frac{M_1\mu\tan\beta}{m_{\tilde{\mu}_R}^4}
F_b(M_1^2/m_{\tilde{\mu}_R}^2,\mu^2/m_{\tilde{\mu}_R}^2).\label{e}\end{aligned}$$
In a scenario with a light bino-like LSP as considered here, these terms are generally much smaller than the ones in Eq. \[eq:gm21\], due to the smaller hypercharge coupling (with $g_Y^2=g^2 \tan^2 \theta_W$, $g$ being the $\text{SU}(2)_L$ gauge coupling, and $\sin^2 \theta_W = 0.23$), as well as due to small bino-Higgsino mixing. Since the focus of our study is the dominant chargino-muon sneutrino loop contribution in Eq. \[a\], which does not depend upon the right-smuon mass, for our discussion, we shall assume the right-smuon to be decoupled. This implies that the contribution from Eq. \[e\] will be negligible. The neutralino contribution in Eq. \[c\] is determined by the same set of parameters that enter Eq. \[eq:gm21\] as well as $M_1$, and it is positive for $M_1 \mu > 0$.
Finally, the contribution proportional to the left-smuon and right-smuon mixing is given by the bino – left-smuon – right-smuon loop and reads: $$\begin{aligned}
\Delta a_{\mu}(\tilde{B},\tilde{\mu}_L-\tilde{\mu}_R)&=\frac{g_Y^2 m_{\mu}^2}{8\pi^2}\frac{\mu\tan\beta}{M_{1}^3}
F_b(m_{\tilde{\mu}_L}^2/M_1^2,m_{\tilde{\mu}_R}^2/M_1^2).\label{b}\end{aligned}$$ This contribution can be enhanced for a large value of $\mu \tan \beta$, with a non-trivial dependence on $M_1$. The LHC analyses considered by us do not probe this term, which can be significant when both the left-smuon and the right-smuon are light, while the charginos can be much heavier. Once again, as for Eq. \[e\], in the limit of a decoupled right-smuon, the contribution of Eq. \[b\] also becomes negligible. In our analysis, we include all five contributions given by Eqs. \[eq:gm21\]–\[b\]. Although we have used the above expressions for $\Delta a_{\mu}$ in the weak eigenstate basis in the subsequent sections, we have checked that using the mass eigenstate basis expressions do not lead to any significant difference for the decoupled right-smuon scenario.
Left-smuon mass limits from the LHC {#smuon}
===================================
The ATLAS and CMS collaborations have searched for scalar muon pair production in the dilepton and missing transverse momentum ($\met$) channel, and with $20.3 \fb^{-1}$ of data from the 8 TeV LHC, ATLAS obtained the following $95\%$ C.L. lower bounds [@ATLAS_Slepton]: $$\begin{aligned}
M_{\tilde{\mu}^\pm_L} & > 300 \gev, \text{~with}~ \tilde{\mu}^\pm_R \text{~decoupled} \nonumber \\
M_{\tilde{\mu}^\pm_R} & > 230 \gev, \text{~with}~ \tilde{\mu}^\pm_L \text{~decoupled} \nonumber \\
M_{\tilde{\mu}^\pm_L,\tilde{\mu}^\pm_R} & > 320 \gev, \text{~for a common} ~\tilde{\mu}^\pm_{L,R} ~\text{mass}.
\label{smuon_bound}\end{aligned}$$ The quoted lower bounds correspond to the case where ${\rm BR}({\tilde{\mu}^\pm_{L,R}} \rightarrow \mu^\pm \neu) = 1$, and $M_{\neu}=0$ GeV. These limits are not very sensitive to the choice of $M_{\neu}$, unless ${\tilde{\mu}^\pm_{L,R}}$ and ${\neu}$ are nearly degenerate, and we find that the $M_{\tilde{\mu}^\pm_L} > 300 \gev$ bound essentially remains unchanged for $M_{\neu}$ values of upto $150$ GeV [@ATLAS_Slepton]. We note that although smuon masses lower than around $94 \gev$ are not probed by the ATLAS search, such lower mass regions are excluded by the LEP2 experiment [@Abbiendi:2003ji].
![Parameter space in the $\mu-M_2$ plane where ${\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm + X)$ deviates from one, for $M_{\smuL}=300 \gev$. The green solid (red dashed) area encloses the region where ${\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm + X)<0.9 ~(0.5)$. In the region above the green solid line (the white space) ${\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm + X)$ is greater than 0.9.[]{data-label="fig:smuon_BR"}](smuon_BR.pdf)
As long as the $\C1$ and $\N2$ are heavier than $\smuL$, the above assumption of ${\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm \neu) = 1$ remains valid. However, in the MSSM parameter space of our interest, there are parameter regions where $\C1$ and $\N2$ can be lighter than $\smuL^\pm$, and it is important to determine to what extent the above lower bound of $300\gev$ can be modified. Taking for example a scenario with $M_1=50\gev$ and $\tan \beta=50$, we show by the green (red) shaded area in Fig. \[fig:smuon\_BR\] the parameter region in the $\mu-M_2$ plane in which the ${\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm + X) < 0.9 ~(0.5)$, where $X$ stands for anything other than the $\mu^\pm$ originating from the decay of $\smuL^{\pm}$. We find only a very small region of parameter space where the ${\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm + X)$ is smaller than $50\%$. The smuon decay BR to muons here includes contributions from the processes $\smuL^\pm \rightarrow \mu^\pm + \neu$, as well as from $\smuL^\pm \rightarrow \mu^\pm + \N2$ (and subdominantly from $\smuL^{\pm} \rightarrow \nu_{\mu} + \C1$, where subsequently $\C1 \rightarrow W^\pm (\rightarrow \mu^\pm \nu_\mu) \neu$). As we can see from Fig. \[fig:smuon\_BR\], the branching to $\C1$ is enhanced only when the $\C1$ is wino-like and light. In order to obtain the same event yield that can be excluded by the smuon search under consideration, the reduction in BR to final states containing muons can be compensated by an enhanced cross-section for smuon pair production. We can thus translate the $95\%$ C.L. lower bound on the left-smuon mass, 300 GeV in Eq. \[smuon\_bound\], as follows: $$\begin{aligned}
M_{\smuL} &> 287 \gev, \text{~~~for~} {\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm + X) = 0.9, \nonumber \\
M_{\smuL} &> 220 \gev, \text{~~~for~} {\rm BR}({\tilde{\mu}^\pm_{L}} \rightarrow \mu^\pm + X) = 0.5~.\end{aligned}$$
With the above lower bound on the left-scalar muon mass, the requirement of explaining the $\gm2$ anomaly by the chargino contribution would imply an upper bound on the chargino mass. The goal of the subsequent sections is thus to determine whether the current LHC searches allow for a chargino lighter than that upper bound. As mentioned in Sec. \[sec:gm2\], the MSSM contribution to $\gm2$ from diagrams involving the right-smuon is not discussed in this study, and we therefore take the right-smuon to be decoupled from the spectrum.
Chargino mass limits from the LHC {#sec:chargino}
=================================
As discussed in the Introduction, in a completely general MSSM scenario, all the electroweak sector mass and mixing parameters are relevant for interpreting the different LHC searches for the electroweak-inos. Thus, performing a global analysis of the MSSM parameter space, including several different LHC search channels, requires very large computational resources, and is beyond the scope of the present study. However, as we shall demonstrate in the following, to draw certain broad conclusions on the $\gm2-$compatible parameter space allowed by current constraints, performing a smaller subset of analyses is sufficient.
In the interpretation of different searches for $\tilde{\chi}_1^+ \tilde{\chi}_1^-$ and $\C1 \N2$ production carried out by the LHC collaborations, the hierarchies between the $\C1/\N2$ and slepton masses play a crucial role. Assuming for simplicity $M_{\smuL}=M_{\tilde{e}_L}$, and for fixed values of $M_{\tilde{\tau}_2}$ and $\theta_{\tilde{\tau}}$ (which, as we shall see later, are less relevant parameters), we then have the following possible mass hierarchies for a bino-like LSP ($\neu$) scenario:
1. $M_{\smuL}>M_{\C1},M_{\N2}>M_{\tilde{\tau}_1}$
2. $M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\smuL}$
3. $M_{\C1},M_{\N2}>M_{\tilde{\tau}_1},M_{\smuL}$ (with either hierarchy between $\tla$ and $\smuL$)
4. $M_{\smuL},M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}$.
In the following subsections, we take up each of the above hierarchies in turn, and discuss the constraints on the MSSM parameters of interest from the relevant LHC searches.
Scenario-1: stau NLSP {#stau}
---------------------
We first focus on the mass hierarchy $M_{\smuL}>M_{\C1},M_{\N2}>M_{\tilde{\tau}_1}>M_{\neu}$. In this scenario, the $\C1$ and $\N2$ would decay dominantly via the intermediate $\tla$, the tau-sneutrino, and $\tlb$ (when kinematically accessible), and subdominantly through $W/Z/h$, thereby leading to a multi-tau final state. Considering hadronic decays of the tau leptons ($\tau_h$), the final state of interest is then $\geq 2\tau_h+\met$. We follow the corresponding ATLAS search strategies [@ATLAS_tau_8; @ATLAS_tau_13] in this regard, on which our constraints are based. Utilizing both the $20.3 \fb^{-1}$ data from the 8 TeV LHC run [@ATLAS_tau_8] and the $14.8 \fb^{-1}$ of data from 13 TeV [@ATLAS_tau_13], the ATLAS collaboration has looked for $\C1 \N2$ production, and interpreted the results within a simplified model setup, assuming $M_{\tla,\tilde{\nu_\tau}}=(M_{\neu}+M_{\N2})/2$. The LHC8 and LHC13 searches lead to the following limit from ATLAS: $$M_{\C1} = M_{\N2} > 700 \gev, \text{~for~} M_{\neu}=0 \gev, \text{~at~}95\% \text{~C.L.}$$
Interpreting the search with at least two hadronically decaying tau leptons and $\met$ is considerably involved, as the crucial tau jet identification and reconstruction efficiencies are sensitive to the kinematics of the tau jets, and thus in turn to the mass splittings between the sparticles. We therefore study in detail how the LHC bounds translate to constraints in the corresponding MSSM parameter space, taking into account both modifications to the relevant production rates and BRs, as well as by performing a Monte Carlo study with a simple detector simulation to capture the possible changes in the detection efficiencies. The details of our MC simulation framework, the kinematic selection criteria employed, as well as the validation of our simulation framework are discussed in the Appendix. We note here that due to our modelling of the tau reconstruction and identification efficiencies based on a simple detector simulation, our event yields are in general larger than that reported by ATLAS. We therefore rescale our event yields to match the ATLAS numbers by a constant fudge factor each for the 8 and 13 TeV searches, as detailed in the Appendix.
In our analysis, we have incorporated certain simple but important effects not captured by the above simplified model study by ATLAS. To begin with, we include all possible electroweak-ino production modes, namely, $$\begin{aligned}
p p &\rightarrow \tilde{\chi}_i^+\tilde{\chi}_j^-, \text{~with~} i, j =1,2,\nonumber \\
&\rightarrow \tilde{\chi}_i^\pm\tilde{\chi}_j^0, \text{~with~} i =1,2, \text{~and~} j=1,2,3,4,\nonumber \\
&\rightarrow \tilde{\tau}_k^+\tilde{\tau}_\ell^-, \text{~with~} k,\ell=1,2. \end{aligned}$$ Clearly, not all of the production processes have appreciable cross-sections for a given set of parameters, but there are regions of parameter space (for example, for Higgsino-like $\tilde{\chi}_2^0$ and $\tilde{\chi}_3^0$), where including the production of heavier charginos and neutralinos is important. We do not make any assumptions about the elctroweak-ino decay branchings, and take into account their decays via staus, tau-sneutrino as well as via the SM gauge and Higgs bosons, according to the relevant BRs. Finally, the impact of variation in the stau mixing, the LSP mass, as well as $\tan \beta$ are also studied.
![*$95\%$ C.L. exclusion in the $\mu-M_2$ plane following the ATLAS search in the $\geq 2\tau_h+\met$ channel, using the 8 TeV, $20.3 \fb^{-1}$ data (green) and 13 TeV, $14.8 \fb^{-1}$ data (red) from the LHC. The results are shown for a scenario where the left-smuon is decoupled ($M_{\smuL}=3$ TeV), with $M_1=50 \gev, \tan \beta=50, M_{\tla}=100 \gev, M_{\tlb}=300 \gev$ and $\theta_{\tilde{\tau}}=\pi/4$. We set $M_{\tilde{e}_L}=M_{\smuL}$ for simplicity, and the right selectron and right smuon, which are not relevant for this study, are taken to be decoupled.*[]{data-label="fig:tau_search"}](tau_search_LHC.pdf)
For the dominant chargino contribution to the $\gm2$, apart from the left-smuon mass and $\tan\beta$, the two other important parameters are $\mu$ and $M_2$, which also determine the chargino and heavier neutralino production and decay rates. Therefore, we demonstrate the LHC constraints in the $\mu-M_2$ parameter plane. In Fig. \[fig:tau\_search\], we show the $95\%$ C.L. excluded region in this plane using the 8 TeV, $20.3 \fb^{-1}$ data (green) and 13 TeV, $20.3 \fb^{-1}$ data (red) from the LHC following the ATLAS search in the $\geq 2\tau_h+\met$ channel. The results are shown for the left-smuon decoupled scenario ($M_{\smuL}=3$ TeV), with $M_1=50 \gev, \tan \beta=50, M_{\tla}=100 \gev, M_{\tlb}=300 \gev$ and $\theta_{\tilde{\tau}}=\pi/4$. While the 8 TeV search covers $\C1$ masses of upto around 400 GeV, this reach is significantly extended with the currently analyzed 13 TeV dataset, namely $\C1$ masses of upto around 600 GeV are now excluded across the $\mu-M_2$ plane. Since for the same mass values, the wino-like $\C1 \N2$ production has a higher cross-section than the Higgsino-like one (by a factor of four), the exclusion is slightly stronger for high $\mu$ and smaller $M_2$ values. The cross-section reduction in the Higgsino-like region is, however, partially compensated by the associated production of additional Higgsino-like heavier neutralino states of similar mass with a $\C1$. We find that apart from the very light $\C1/\N2$ mass regions with low $\mu$ and high $M_2$ values, the 13 TeV search also covers the region excluded by the 8 TeV analysis. With this in mind, in subsequent figures, we shall only show the 13 TeV exclusion contours.
![*$95\%$ C.L. exclusion in the $\mu-M_2$ plane for different values of the LSP mass, $M_1=10\gev$ (blue-dashed line) and $M_1=50 \gev$ (red solid line). The exclusion contours are based on the 13 TeV, $14.8 \fb^{-1}$ data from the LHC, following the ATLAS search in the $\geq 2\tau_h+\met$ channel. The results are shown for a scenario where the left-smuon is decoupled ($M_{\smuL}=3$ TeV), with $\tan \beta=50, M_{\tla}=100 \gev, M_{\tlb}=300 \gev$ and $\theta_{\tilde{\tau}}=\pi/4$. We set $M_{\tilde{e}_L}=M_{\smuL}$ for simplicity, and the right selectron and right smuon, which are not relevant for this study, are taken to be decoupled.*[]{data-label="fig:M1_vary"}](M1_vary.pdf)
In order to demonstrate the impact of change in kinematic cut efficiencies due to different values of the $\neu$ LSP mass, we show in Fig. \[fig:M1\_vary\] the comparison of the current LHC 13 TeV exclusions for $M_1=10 \gev$ (blue dashed) and $M_1=50 \gev$ (red solid). The rest of the parameters are fixed as in Fig. \[fig:tau\_search\]. As we can see from this figure, the impact of varying $M_1$ is minimal, and as long as the bino-like $\neu$ is not degenerate with the $\tla$ NLSP, the kinematic cut efficiencies do not vary significantly. Therefore, in what follows, we shall show all of our results for the $M_1=50\gev$ scenario.
![*$95\%$ C.L. LHC13 exclusion contours for different values of the mixing angle in the stau sector. The red (solid), green (dashed) and blue (dot-dashed) lines correspond to the lighter stau being maximally mixed, right-stau and left-stau respectively. The other relevant parameters are fixed as in Fig. \[fig:tau\_search\].*[]{data-label="fig:stau_mix"}](stau_mix_effect.pdf)
We now consider the impact of varying the mixing angle in the stau sector on the LHC exclusions. In Fig. \[fig:stau\_mix\], we show results for three different choices for $\theta_{\tilde{\tau}}$: the lighter stau being a left-stau ($\theta_{\tilde{\tau}}=0$), a right-stau ($\theta_{\tilde{\tau}}=\pi/2$) or an equal admixture of both ($\theta_{\tilde{\tau}}=\pi/4$). In the first (second) scenario, the right (left) stau mass is set to be $3\tev$. For the first two cases, though we set the trilinear parameter $A_\tau = 0$, there is always a small mixing between the left and right stau due to non-zero values of $\mu \tan \beta$. When the lighter stau is a $\tilde{\tau}_L$, the tau sneutrino is also light and of similar mass, and therefore decays through the tau sneutrino leads to a reduction in ${\rm BR}(\N2 \rightarrow \tau^+ \tau^- \neu)$. This reduction in the BR to tau-rich final states is also applicable to the maximally mixed scenario, as decays through the sneutrino state are still accessible, albeit with a slightly smaller branching (due to a somewhat higher mass of the sneutrino compared to the $\tla$). When the lighter stau is a right-stau with the left-stau mass set at 3 TeV, the sneutrino is not kinematically accessible, thereby enhancing the BR of $\N2$ to tau’s, and hence increasing the exclusion reach in the wino-like region ($M_2<\mu$). In the Higgsino like region ($\mu<M_2$) there is no difference observed, as a Higgsino-like $\N2$ does not decay via sneutrinos. Keeping in mind that the $\theta_{\tilde{\tau}}=45^\circ$ scenario leads to slightly weaker bounds compared to the right-stau NLSP case, for subsequent figures we shall show results for the maximally mixed scenario only.
We note in passing that a left-right mixed $\tla$ NLSP with a bino-like LSP is well-motivated from the point of view of obtaining a DM candidate with the required thermal relic abundance (for $M_1 \geq \mathcal{O}(30 \gev)$), or for accommodating the Galactic Centre gamma ray excess through $\neu \neu \rightarrow \tau^+ \tau^-$ annihilation in the present epoch (for $M_1 \sim 10 \gev$) [@SM].
Scenario-2: selectron/smuon NLSP {#trilepton}
--------------------------------
The second mass hierarchy we consider is $M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\smuL}, M_{\tilde{e}_L}>M_{\neu}$. In this case, the $\C1$ and $\N2$ would decay dominantly through the $\smuL$ and $\tilde{e}_L$, their corresponding sneutrinos, and subdominantly through $W/Z/h$. Assuming a $100\%$ decay through $\smuL,\tilde{e}_L$ and their associated sneutrinos, the current ATLAS bounds from $\C1 \N2$ search in the $3\ell+\met$ channel (with $\ell=e^\pm,\mu^\pm$) using $13.3 \fb^{-1}$ of data from the 13 TeV LHC is as follows [@ATLAS_C1N2_13]: $$M_{\C1} = M_{\N2} > 1000 \gev, \text{~for~} M_{\neu}=0 \gev, \text{~at~}95\% \text{~C.L.}$$ For this ATLAS analysis, the $\C1$ and $\N2$ are assumed to be wino-like, while the $\neu$ is taken to be bino-like. These limits are expected to become slightly weaker in the Higgsino-like or wino-Higgsino mixed region for $\C1/\N2$.
![*Estimated $95\%$ C.L. exclusion region in the $\mu-M_2$ plane using the 13 TeV, $13.3 \fb^{-1}$ data (blue dashed line) from the LHC, following the ATLAS search in the $3\ell+\met$ channel. The results are shown for the scenario with both the staus decoupled ($M_{\tilde{\tau}_L}=M_{\tilde{\tau}_R}=3$ TeV), with $M_1=50 \gev, \tan \beta=50$ and $M_{\smuL}=300 \gev$. We set $M_{\tilde{e}_L}=M_{\smuL}$ for simplicity, and the right selectron and right smuon, which are not relevant for this study, are taken to be decoupled.*[]{data-label="fig:trilepton_all"}](trilepton_exclu.pdf)
Unlike in the stau-NLSP scenario discussed in the previous sub-section, we do not perform a detailed MC analysis of the kinematic selection efficiencies and detector effects for this mass hierarchy. This should not lead to significant modifications to the derived constraints, since in the very clean trilepton and $\met$ channel, the final state reconstruction and detector resolution effects are expected to be small, as long as we do not encounter degeneracies between the electroweak-ino, slepton and bino mass parameters. For example, though the above lower bound is quoted for $M_{\neu}=0 \gev$, it remains essentially unchanged in the range $0 \gev \leq M_{\neu} \lesssim 500 \gev$ [@ATLAS_C1N2_13]. Similarly, though the smuon/selectron mass for the simplified model based search is assumed to be given by $M_{\smuL,\tilde{e}_L}=(M_{\neu}+M_{\N2})/2$, it has been shown in Ref. [@ATLAS_EW_Summary] that away from the degenerate mass region, this mass gap does not affect the limits significantly. Therefore, for our discussion in this sub-section, we shall only consider the production cross-section and decay BR modifications of $\C1$ and $\N2$ in the $\mu-M_2$ plane.
In Fig. \[fig:trilepton\_all\], we show our estimate of the current 13 TeV exclusions in the $\mu-M_2$ plane using the trilepton channel (blue dashed curve), for $M_{\smuL,\tilde{e}_L}=300 \gev$, with the $\tilde{\tau}_L, \tilde{\tau}_R, \tilde{e}_R$ and $\tilde{\mu}_R$ decoupled (their masses are fixed at 3 TeV). The results are shown for $M_1=50\gev$ and $\tan\beta=50$. The smuon mass is set at its lower bound from 8 TeV LHC searches, as discussed in Sec. \[smuon\], which is valid in almost the entire parameter space except for very small $M_2$ values, where it can be slightly lower.
We observe two branches in the blue-dashed exclusion contour in Fig. \[fig:trilepton\_all\]. For the region where the $\C1/\N2$ is lighter than $\smuL/\tilde{e}_L$, they decay via the SM gauge and Higgs bosons, thereby reducing the BR to the trilepton final state. Therefore, only smaller $\C1/\N2$ mass values are accessible to the trilepton search when only the gauge/Higgs boson decay modes are open. For $M_{\C1}(=M_{\N2})$ larger than 300 GeV, decays through intermediate sleptons open up, and higher electroweak-ino masses can be probed as well. However, for the latter case, the bounds are weaker for Higgsino-like $\C1/\N2$ region compared to the wino-like one, as only the wino component of the $\C1/\N2$ has significant coupling to the left-smuon or the muon-sneutrino.
Scenario-3: all three generation sleptons lighter than chargino {#sec:sc3}
---------------------------------------------------------------
![*$95\%$ C.L. LHC13 exclusion contours, using the $\geq 2\tau_h+\met$ search only, for different $\smuL$ mass values. The red (solid) and blue (dashed) lines correspond to $M_{\smuL}=3000 \gev$ and $M_{\smuL}=300 \gev$ respectively. We set $M_{\tilde{e}_L}=M_{\smuL}$ for simplicity.*[]{data-label="fig:smuon_mass"}](light_smuon_effect.pdf)
We next study the mass hierarchy: $M_{\C1},M_{\N2}>M_{\tilde{\tau}_1},M_{\smuL,\tilde{e}_L}>M_{\neu}$ (with either hierarchy between $\tla$ and $\smuL$), which presents us with a scenario where both the $\geq 2\tau_h+\met$ and the $3\ell+\met$ searches described in the previous two sub-sections become important. Let us first consider how the bounds shown in Fig. \[fig:tau\_search\] get modified in the presence of a light left-smuon kinematically accessible in $\C1/\N2$ decays. In Fig. \[fig:smuon\_mass\], we show the LHC13 exclusion contours, using the $\geq 2\tau_h+\met$ search only, for two different $\smuL$ mass values. The red (solid) and blue (dashed) lines correspond to $M_{\smuL}=3000 \gev$ and $M_{\smuL}=300 \gev$ respectively, while the other relevant parameters are the same as in Fig. \[fig:tau\_search\]. It is observed that for the wino-like $\C1$ and $\N2$ region, there is a decrease in the reach of the tau search channel, since the wino-like $\C1/\N2$ will also have a significant decay branching to smuon and muon sneutrino states, thereby reducing the BR to the multi-tau final state.
The region in which the $\geq 2\tau_h+\met$ search loses power, the $3\ell+\met$ search becomes sensitive. In general, a statistical combination of both these search channels should lead to the strongest constraints in the $\mu-M_2$ plane. There is, however, a degree of complementarity to the power of these two searches. We demonstrate this fact in Fig. \[fig:stau\_smuon\], which shows that we can separate different areas in the $\mu-M_2$ plane in which one of the two searches has a higher exclusion power. For reasons discussed above, in the lower $\mu$ and higher $M_2$ region, the multi-tau channel is stronger, while in the lower $M_2$ and higher $\mu$ region, the multilepton channel dominates. Along the diagonal, with $\mu \sim M_2$, both searches become less powerful, especially for $M_{\C1}>500 \gev$.
![*$95\%$ C.L. LHC13 exclusion contours, using the $\geq 2\tau_h+\met$ search (red solid) and the $3\ell+\met$ search (blue dashed), in a scenario where both $\tla$ and $\smuL$ can be lighter than the $\C1/\N2$. The results are shown for $\tan \beta=50, M_1=50 \gev, M_{\smuL}=300 \gev, M_{\tla}=100 \gev, M_{\tlb}=300 \gev$ and $\theta_{\tilde{\tau}}=\pi/4$. We set $M_{\tilde{e}_L}=M_{\smuL}$ for simplicity. For most regions of the parameter space, the two searches are found to be complementary.*[]{data-label="fig:stau_smuon"}](stau_smuon_light.pdf)
Scenario-4: all sleptons heavier than chargino {#no_slep}
----------------------------------------------
Finally, we comment on the mass hierarchy in which $M_{\smuL},M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\neu}$. Since in this case all the sleptons are heavier than $\C1$ and $\N2$, the latter can only decay via $W/Z/h$ bosons. Based on $20.3 \fb^{-1}$ of data from the 8 TeV LHC run (the 13 TeV analysis for these $\C1/\N2$ decay modes are not yet published by the LHC collaborations), the ATLAS collaboration obtains the following limit using a combination of dilepton and trilepton searches [@ATLAS_Slepton]: $$M_{\C1} = M_{\N2} > 415 \gev, \text{~for~} M_{\neu}=0 \gev, \text{~at~}95\% \text{~C.L.}~~\text{(WZ~mode)}.$$ In deriving the above limit, it is assumed that $\text{BR}(\C1 \rightarrow W^\pm \neu)=1$ and $\text{BR}(\N2 \rightarrow Z \neu)=1$. If on the other hand, one adopts a simplified model in which $\text{BR}(\C1 \rightarrow W^\pm \neu)=1$ and $\text{BR}(\N2 \rightarrow h \neu)=1$, the above bound becomes considerably weaker, and a combination of search results targeting different Higgs decay modes yields the following bound from ATLAS (using $20.3 \fb^{-1}$ of data from the 8 TeV LHC run) [@ATLAS_Wh]: $$M_{\C1} = M_{\N2} > 250 \gev, \text{~for~} M_{\neu}=0 \gev, \text{~at~}95\% \text{~C.L.}~~\text{(Wh~mode)}.$$ The assumption of either $\text{BR}(\N2 \rightarrow Z \neu)=1$ or $\text{BR}(\N2 \rightarrow h \neu)=1$ does not hold in most regions of the $\mu-M_2$ parameter space, and a combination of $\N2 \rightarrow Z \neu$ and $\N2 \rightarrow h \neu$ decays take place depending upon the Higgsino component of $\N2$ and $\neu$. We do not study this mass hierarchy in this paper, and ideally a statistical combination of the searches targeting the $Z\neu$ and $h\neu$ decay modes of $\N2$ should be performed in order to determine the current best constraints in the $\mu-M_2$ plane for this scenario. We note that although this scenario leads to the weakest bounds on the electroweak-inos, making the sleptons, especially the left-smuon heavier, is not favourable to the chargino contribution to the $\gm2$. However, it is expected that a part of the $\gm2$ favoured parameter space would be allowed for this mass hierarchy, even after the 8 TeV LHC constraints are taken into account.
Impact of LHC constraints on the ${\gm2}$ favoured parameter space {#combi}
==================================================================
We are now in a position to discuss the impact of the LHC constraints on the MSSM parameter space in which the $\gm2$ anomaly can be explained within $1\sigma$. As emphasized in Sec. \[sec:gm2\], we focus on the right-smuon decoupled scenario, where the dominant contribution to $\Delta a_\mu$ comes from the chargino-muon sneutrino loop. For our numerical analysis, we have included all the relevant chargino and neutralino contributions to $\Delta a_\mu$, with $M_{\tilde{\mu}_R}$ fixed at 3 TeV.
Scenario-1: stau NLSP {#scenario-1-stau-nlsp}
---------------------
We start with the first mass hierarchy : $M_{\smuL}>M_{\C1},M_{\N2}>M_{\tilde{\tau}_1}>M_{\neu}$. In Fig. \[fig:gm21\], we show two representative scenarios for the $\gm2$ favoured region at $1\sigma$, with $M_{\smuL}=500 \gev$ (blue shaded) and 800 GeV (violet shaded). As we can see in Fig. \[fig:gm21\], for $M_{\smuL}=500 \gev$, the above mass hierarchy assumption is not valid in a small region of the parameter space, in which $M_{\smuL}<M_{\C1},M_{\N2}$ (the blue shaded region above the grey dashed line for $M_{\C1}=500 \gev$). However, $\C1/\N2$ decays via the left-smuon (and/or left-selectron) in this region are kinematically suppressed, and therefore, the $\geq 2\tau_h+\met$ search results shown remain effectively unchanged, even though they were obtained in Sec. \[stau\] with the assumption of $M_{\smuL}=3$ TeV. We see from Fig. \[fig:gm21\] that the $\gm2$ favoured parameter space at $1\sigma$ in the stau-NLSP scenario is severely constrained by the latest 13 TeV LHC results in the $\geq 2\tau_h+\met$ channel (red solid line). This conclusion is not modified under the variation of the relevant MSSM parameters, namely, $\tan \beta$, the mass of $\tla$ and the mixing angle in the stau sector, as we discuss below.
In Fig. \[fig:gm21\], we have shown the results for $\tan\beta=50$. For lower values of $\tan\beta$, we have checked that the LHC search constraints are not modified significantly. However, as shown in Eqs. \[eq:gm21\]–\[b\], the MSSM contribution to $\Delta a_\mu$ is proportional to $\tan \beta$. Therefore, for lower values of $\tan\beta$, the $\gm2$ favoured region shifts towards lower values of the $\C1$ mass (for a fixed $\smuL$ mass), which translates to smaller values of $\mu$ and $M_2$. It is thus clear that for $\tan \beta$ values smaller than $50$, the current LHC constraints on the $\gm2$ favoured region are even stronger.
As discussed in Sec. \[stau\], the variation in $M_1$ does not affect the LHC search constraints, and its impact on $\Delta a_\mu$ through the neutralino-smuon loop is also small in the $\tilde{\mu}_R$ decoupled scenario. Furthermore, as explained in Sec. \[stau\], with a maximally mixed $\tla$, the LHC constraints are only slightly weaker compared to the right-stau NLSP case, while the constraints are very similar with a left-stau NLSP. Therefore, changing the stau mixing would not relax the constraints compared to the red solid line in Fig. \[fig:gm21\]. As far as the $\tla$ mass is concerned, within the assumptions of the above mass hierarchy, its impact should be minimal as well, unless we encounter degeneracies between the $\C1/\N2$ and $\tla$ mass values.
Therefore, we can conclude that in the stau NLSP scenario, with the mass hierarchy, $M_{\smuL}>M_{\C1},M_{\N2}>M_{\tilde{\tau}_1}>M_{\neu}$, the current 13 TeV LHC constraints already disfavour most of the parameter space in which the $\gm2$ anomaly can be accommodated within $1\sigma$ through the chargino contribution.
![*Current $95\%$ C.L. LHC constraints on the $\gm2$ favoured parameter region in the $\mu-M_2$ plane, for the mass hierarchy $M_{\smuL},M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\neu}$. In the blue (violet) shaded region, the $\gm2$ anomaly can be explained at $1\sigma$, with the choice $M_{\smuL}=500 \gev ~(800 \gev)$. The values of other relevant MSSM parameters are fixed as in Fig. \[fig:tau\_search\].*[]{data-label="fig:gm21"}](gm21.pdf)
Scenario-2: selectron/smuon NLSP {#sec:five_two}
--------------------------------
We now consider the second mass hierarchy: $M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\smuL}, M_{\tilde{e}_L}>M_{\neu}$. Since the $\smuL$ is the NLSP in this scenario, for our discussion, we choose the lowest value of $M_{\smuL}$ allowed by the LHC8 search for left-smuons described in Sec. \[smuon\]. In Fig. \[fig:gm22\], we show the $\gm2$ favoured region at $1\sigma$ with $M_{\smuL}=300 \gev$ (brown shaded area), with $\tan \beta=50$. As we can see from this figure, in a substantial region of the $\mu-M_2$ plane, $M_{\C1}$ is lower than 300 GeV, which does not satisfy the hierarchy assumption of $M_{\C1},M_{\N2}>M_{\smuL}, M_{\tilde{e}_L}$. However, the $3\ell+\met$ search results shown in this figure (blue dashed line) were obtained with $M_{\smuL}=300 \gev$ in Sec. \[trilepton\], and therefore the constraints shown remain valid in the entire region of the parameter space. In the region with $M_{\C1},M_{\N2}<M_{\smuL}, M_{\tilde{e}_L}$, the trilepton events originate from $\C1/\N2$ decays to $W^\pm/Z$ bosons. As we see from Fig. \[fig:gm22\], even though the trilepton search is very powerful and strongly excludes most of the $1\sigma$ favoured parameter space for $\gm2$ with this mass hierarchy, in the Higgsino-like $\C1/\N2$ region, we can find a small window allowed by the current LHC search constraints. As discussed earlier, this is because of the reduced coupling of a left-smuon or left-selectron to a $\C1/\N2$, when the latter is mostly a Higgsino-like state. In the currently allowed region, the Higgsino mass parameter ($\mu$) is found to be around $500\gev$, with the wino-mass parameter ($M_2$) taking values higher than about 1 TeV, and the chargino mass falls in the range $300 \gev < M_{\C1}, M_{\N2} < 500 \gev$.
![*Current $95\%$ C.L. LHC constraints on the $\gm2$ favoured parameter region in the $\mu-M_2$ plane, for the mass hierarchy $M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\smuL}, M_{\tilde{e}_L}>M_{\neu}$. In the brown shaded region, the $\gm2$ anomaly can be explained at $1\sigma$, with the choice $M_{\smuL}=300 \gev$. The values of other relevant MSSM parameters are fixed as in Fig. \[fig:trilepton\_all\].*[]{data-label="fig:gm22"}](gm22.pdf)
Scenario-3: all three generation sleptons lighter than chargino {#sec:five_three}
---------------------------------------------------------------
The third mass hierarchy considered in this study is $M_{\C1},M_{\N2}>M_{\tilde{\tau}_1},M_{\smuL,\tilde{e}_L}>M_{\neu}$. In Fig. \[fig:gm23\], as in the previous case above, we show the $\gm2$ favoured region at $1\sigma$ for the left-smuon mass value of $M_{\smuL}=300 \gev$ (brown shaded), with $\tan\beta=50$. For discussing the LHC constraints, we choose, as in Fig. \[fig:stau\_smuon\], $M_{\tla}=100 \gev, M_{\tlb}=300 \gev$ and $\theta_{\tilde{\tau}}=\pi/4$, with $\tan \beta=50$ and $M_1=50 \gev$ (this choice of parameter values is motivated by the abundance for relic $\neu$ DM). Thus, both the $\smuL$ and the $\tla$ masses are fixed close to their current lower limits (for the case of $\tla$, the strongest bound from the LEP experiments is about $90 \gev$ [@Abbiendi:2003ji; @SM]). Once again, as in the previous scenario, though taking $M_{\smuL}=300 \gev$ may violate the hierarchy of $M_{\C1},M_{\N2}>M_{\smuL}$, the constraints shown in Fig. \[fig:gm23\], namely that from the $\geq 2\tau_h+\met$ (red solid line) and $3\ell+\met$ (blue dashed line) searches were derived in Sec. \[sec:sc3\] for the same value of $M_{\smuL}$, and therefore they remain valid in the entire parameter space.
As discussed in Sec. \[sec:sc3\], both the $\geq 2\tau_h+\met$ and the $3\ell+\met$ searches provide important complementary constraints for this mass hierarchy, and taken together, they exclude bulk of the $1\sigma$ favoured parameter region for $\gm2$, with $M_{\smuL}=300 \gev$. A small window remains allowed when $M_2 \sim \mu$ and $500 \gev < M_{\C1}, M_{\N2} < 700 \gev$, where both the searches lose their power. If $\smuL$ is made heavier, we will approach the stau NLSP scenario discussed above, and the overall constraints on the parameter space of interest will continue to be strong.
![*Current $95\%$ C.L. LHC constraints on the $\gm2$ favoured parameter region in the $\mu-M_2$ plane, for the mass hierarchy $M_{\C1},M_{\N2}>M_{\tilde{\tau}_1},M_{\smuL,\tilde{e}_L}>M_{\neu}$. In the brown shaded region, the $\gm2$ anomaly can be explained at $1\sigma$, with the choice $M_{\smuL}=300 \gev$. The values of other relevant MSSM parameters are fixed as in Fig. \[fig:stau\_smuon\].*[]{data-label="fig:gm23"}](gm23.pdf)
Comments on the lower bound on left-smuon mass and its impact on $\gm2$
-----------------------------------------------------------------------
In this subsection, we comment on the impact of having a weaker bound on $\smuL$ when the chargino is lighter than the left-smuon. As discussed in Sec. \[smuon\], relaxing the 8 TeV LHC lower bound on the left-smuon mass from 300 GeV to around 220 GeV is possible only for a very light wino-like chargino, with $M_{\C1}$ smaller than $M_{\smuL}$. We show in Fig. \[fig:gm24\] the difference in the $1\sigma$ allowed regions for $\gm2$ for the two choices of $\smuL$ mass, $M_{\smuL}=300 \gev$ (brown shaded) and 220 GeV (green shaded). Focussing on the region below the grey dashed line along which $M_{\C1}<300 \gev$, we can see from this figure that the additional $\gm2$ favoured parameter space gained by lowering $M_{\smuL}$ is for $\mu>1500 \gev$ and $M_2<300 \gev$, where the lighter chargino is wino-like. However, from Figs. \[fig:gm22\] and \[fig:gm23\], we find that this additional parameter space is excluded by the current LHC search results on electroweak-inos. Therefore, the possible lowering of the $\smuL$ mass limit does not have any impact on our conclusions.
We also note in passing that for very high $M_2$ and low $\mu$ values, the smuon mass bound cannot be relaxed, as the $\C1$ is Higgsino-like here, and thus the $\smuL$ cannot decay to a $\C1$. Therefore, the additional $\gm2$ allowed parameter space with $M_2>1700$ GeV and $\mu<300$ GeV in Fig. \[fig:gm24\] is not viable either.
![*Parameter region in the $\mu-M_2$ plane in which the $\gm2$ anomaly can be explained within $1\sigma$: for $M_{\smuL}=300 \gev$ (brown shaded) and $M_{\smuL}=220 \gev$ (green shaded).*[]{data-label="fig:gm24"}](gm24.pdf)
Summary {#sum}
=======
In this study, we revisit the current LHC constraints on the $\gm2$ favoured parameter space in the MSSM, focussing on the scenario where the chargino-muon sneutrino loop contribution largely accounts for the discrepancy. Since the chargino-muon sneutrino loop leads to the dominant contribution to $\gm2$ in most of the MSSM parameter space (except in the region in which the left- and right-scalar muons are light and there is a large mixing between them), it is crucial to thoroughly probe this scenario at the LHC.
The most relevant LHC searches in this context are that of the left-smuon and electroweak-inos, for which we include constraints from both the 8 TeV (with $20.3\fb^{-1}$ data) and 13 TeV (with upto $14.8\fb^{-1}$ data) LHC runs. In interpreting the LHC constraints from the above searches in the context of the $\gm2$, we organize our study by paying particular attention to the different possible mass hierarchies between the electroweak-inos and the three generations of sleptons.
Since the LHC bounds on electroweak-inos become weaker if they decay via an intermediate stau or a tau sneutrino instead of the first two generation sleptons, we begin our study with the stau NLSP scenario, with the mass hierarchy $M_{\smuL}>M_{\C1},M_{\N2}>M_{\tilde{\tau}_1}>M_{\neu}$. For this mass hierarchy, we have carried out a detailed Monte Carlo analysis of the search for electroweak-inos in the $\geq 2\tau_h+\met$ channel. This is necessary because tau-jet identification and reconstruction efficiencies are sensitive to the event kinematics, and therefore to the mass differences among the supersymmetric particles involved. After carefully validating our MC simulation framework against the ATLAS results (as detailed in the Appendix), we translate the LHC8 and LHC13 bounds to the relevant Higgsino-wino mass ($\mu-M_2$) plane. Particular attention is paid to the possible dependence of the LHC constraints on the LSP mass and the mixing angle in the stau sector. It is observed that while the LSP mass does not play any significant role (unless we encounter degeneracies between the sparticle masses), the bounds are weaker for a left-stau or maximally mixed stau NLSP scenario, compared to a right-stau NLSP case. We note that the strongly mixed stau NLSP with a bino-like LSP is motivated by the DM relic density requirement (for $M_{\tilde{B}} \geq \mathcal{O}(30 \gev)$), or for accommodating the Galactic Centre gamma ray excess (for $M_{\tilde{B}} \sim 10 \gev$).
While the LHC constraints are found to be slightly weaker in a region where the $\C1/\N2$ are Higgsino-like, due to lower values of $\C1\N2$ production rates, this difference is reduced by the opening of new substantial production modes in the Higgsino-like region, such as that of $\C1 \tilde{\chi}_3^0$. With the left-smuon heavier than the $\C1/\N2$ for this mass hierarchy, we find that the $\gm2$ favoured region at $1\sigma$ is severely constrained even in the stau NLSP scenario, especially in the light of the recent 13 TeV data. Such strong constraints are obtained on setting $\tan \beta$ values as high as 50, and the LHC constraints on the $\gm2$ favoured parameter region become stronger for lower values of $\tan\beta$.
When the left-smuon (and/or left-selectron) is the NLSP, we have the mass hierarchy $M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\smuL}, M_{\tilde{e}_L}>M_{\neu}$. In such a case, the LHC constraints come from the trilepton and $\met$ search channel, for which the limits on the electroweak-ino sector are generically stronger, and using the recently analyzed 13 TeV LHC data, ATLAS excludes $\C1/\N2$ masses of upto 1 TeV. We estimate the current constraints in the $\mu-M_2$ plane for this mass hierarchy as well, taking into account the modification to the $\C1\N2$ production cross-section and the corresponding decay BRs, in comparison to the simplified model based ATLAS study. With a left-smuon mass close to its current lower bound of around 300 GeV, we find that a small region of the $\gm2$ favoured parameter space at $1\sigma$ is still viable in view of the current LHC13 exclusion estimate for this mass hierarchy. In the allowed region, the Higgsino mass parameter ($\mu$) is found to be around $500\gev$, with the wino-mass parameter ($M_2$) taking values higher than about 1 TeV, and the chargino mass falls in the range $300 \gev < M_{\C1}, M_{\N2} < 500 \gev$. For such parameter values, the $\C1$ and $\N2$ are Higgsino-like, thereby reducing their BR to left-smuons, which leads to a weaker exclusion limit from the $3\ell+\met$ search.
The third possible mass hierarchy considered in this study is a scenario where all three generation sleptons are lighter than the chargino, i.e., $M_{\C1},M_{\N2}>M_{\tilde{\tau}_1},M_{\smuL,\tilde{e}_L}>M_{\neu}$. In this scenario, we find an interplay of the $\geq 2\tau_h+\met$ and the $3\ell+\met$ search channels, and the constraints ensuing from them in the $\mu-M_2$ plane are found to be largely complementary. For the Higgsino-like $\C1/\N2$ region ($\mu < < M_2$) the $\geq 2\tau_h+\met$ channel leads to stronger limits due to the higher decay rate of $\C1/\N2$ via an intermediate stau, whereas for the wino-like region ($\mu>>M_2$) the $3\ell+\met$ search leads to more stringent constraints. In the intermediate region, where the $\C1$ and the $\N2$ are nearly a maximal mixture of wino and Higgsino states ($\mu \sim M_2$), a small window of parameter space is found to be allowed based on the current constraints, in which the $\gm2$ anomaly can still be explained within $1\sigma$, with $500 \gev < M_{\C1}, M_{\N2} < 700 \gev$. Though individually both the multi-tau and multi-lepton searches lose power for such values of $\mu$ and $M_2$, it is plausible that a statistical combination of these two search channels would lead to stronger constraints.
To sum up, on performing a thorough study of the constraints on the electroweak-ino sector parameters relevant for explaining the $\gm2$ anomaly via the dominant chargino-muon sneutrino loop in the MSSM, we find that with a bino LSP, and a slepton NLSP, the region of parameter space allowed by current LHC constraints is rather narrow. While for the stau NLSP scenario most of the viable parameter regions are already excluded, with a smuon (and/or selectron) NLSP, $300 \gev < M_{\C1}, M_{\N2} < 500 \gev$ are still allowed, when the lighter chargino is Higgsino-like. With all three generations of sleptons lighter than the chargino, the surviving region is found to be $500 \gev < M_{\C1}, M_{\N2} < 700 \gev$, with the lighter chargino being a maximal mixture of wino and Higgsino states.
We have not performed a detailed analysis for the fourth possible mass hierarchy of $M_{\smuL},M_{\tilde{\tau}_1}>M_{\C1},M_{\N2}>M_{\neu}$, in which all three generations of sleptons are heavier than the $\C1$ and $\N2$. In this scenario, the electroweak-inos would decay via SM gauge and Higgs bosons, and the LHC constraints become somewhat weaker compared to the above three scenarios. Therefore, it is expected that a part of the $\gm2$ favoured parameter space would be allowed, though the viable region shrinks compared to the above three cases due to the left-smuon being required to be heavier than the chargino. Apart from this mass hierarchy, the $\gm2$ anomaly can also be accommodated by the contribution from the bino-left smuon-right smuon loop, with a large mixing between the left and right smuons, which in turn requires a large value of $\mu \tan \beta$; in addition, both the scalar muon states should be light.
In view of the fact that the upcoming FermiLab and J-PARC $\gm2$ experiments have the potential to confirm and possibly lead to an enhanced statistical significance of the $\gm2$ anomaly currently based on the BNL result, it is important at this point to determine the viability of well-motivated beyond standard model scenarios like the MSSM in explaining the current $3.3\sigma$ discrepancy. In this regard, as found in our study, the current LHC search results are already powerful enough to constrain large regions of the favoured parameter space. Future large statistics data sets from the LHC would not only consolidate the current constraints, but also probe regions of parameter space yet unexplored, which still remain promising from the point of view of explaining the $\gm2$ anomaly within the MSSM framework.
Note added {#note-added .unnumbered}
==========
After we submitted this paper for publication, new results from the 13 TeV LHC have been reported by the ATLAS and CMS collaborations which are relevant to the analyses presented here. The most important update comes from slepton search in the two leptons and missing transverse momentum final state, for which 13 TeV LHC results including $36.1{~\rm fb}^{-1}$ data is now available [@multilepton_latest]. This improves the lower bound on $M_{\tilde{\mu}^\pm_L,\tilde{\mu}^\pm_R}$ from around $300$ GeV to $500$ GeV (for a common $\tilde{\mu}^\pm_{L,R}$ mass), with ${\rm BR}({\tilde{\mu}^\pm_{L,R}} \rightarrow \mu^\pm \neu) = 1$, and $M_{\neu}=0$ GeV. Furthermore, the update from the previously reported $13.3{~\rm fb}^{-1}$ to the current $36.1{~\rm fb}^{-1}$ results from 13 TeV LHC also modifies the multilepton search limits on chargino-neutralino pair production [@multilepton_latest].
On including these new LHC results, we have obtained updated constraints in the $\gm2$ favoured parameter space in scenarios 2 and 3, which were studied in Sec \[trilepton\], \[sec:sc3\], \[sec:five\_two\] and \[sec:five\_three\]. In particular, we show the updated versions of Figs. \[fig:gm22\] and \[fig:gm23\] in Fig. \[fig:update\_slepton\] (left and right columns respectively). To begin with, the $\gm2$ favoured region with $M_{\tilde{\mu}_L}=500$ GeV is now pushed to lower values of the $\C1$ mass. In the scenario with the selectron and/or smuon NLSP, in the currently allowed region, the Higgsino mass parameter ($\mu$) continues to be around $500\gev$, with the wino mass parameter ($M_2$) taking values between $500$ GeV and 1 TeV. We note that this region in the $\mu-M_2$ plane was covered by the multilepton search (blue dashed line in Fig. \[fig:gm22\]) with the previous bound of $300$ GeV on $M_{\tilde{\mu}_L}$. Since in this region $300 \gev < M_{\C1}, M_{\N2} < 500 \gev$, taking into account the updated constraint of $M_{\tilde{\mu}_L}>500$ GeV, the appropriate mass hierarchy to consider now is either scenario 1 with $M_{\neu} < M_{\tilde{\tau}_1} < M_{\C1},M_{\N2} < M_{\smuL}$ (in which case it is already excluded, see Fig. \[fig:gm21\]), or scenario 4 with $M_{\neu} < M_{\C1},M_{\N2} < M_{\smuL},M_{\tilde{\tau}_1}$ (in which case it should still be allowed). With the stau sector parameter choices made in Fig. \[fig:update\_slepton\], left column, the currently allowed region belongs to scenario 4.
In the scenario with all three generation sleptons lighter than $\C1$ and $\N2$, we find that with $M_{\tilde{\mu}_L}=500$ GeV, the $\gm2$ favoured parameter region is now covered almost entirely by the 13 TeV LHC search in the $\geq 2\tau_h+\met$ channel, as seen in Fig. \[fig:update\_slepton\], right column. We note that even though the $\geq 2\tau_h+\met$ channel search results have also been updated by the ATLAS collaboration including upto $36.1{~\rm fb}^{-1}$ data [@tau_update], the small improvements in the $\C1$ and $\N2$ mass limits do not impact our conclusions on the $\gm2$ favoured parameter space, which are already disfavoured in our analyses using the $14.8{~\rm fb}^{-1}$ dataset, following Ref. [@ATLAS_tau_13].
![*Updated LHC constraints on the $\gm2$ favoured parameter space using $36.1{~\rm fb}^{-1}$ of data in the multilepton$+\met$ search channel at the 13 TeV LHC [@multilepton_latest]: for the selectron and/or smuon NLSP scenario (left column) and when all three generation sleptons are lighter than $\C1$ and $\N2$ (right column). See text in the Note added section for details.*[]{data-label="fig:update_slepton"}](smuon_NLSP_500GeV "fig:") ![*Updated LHC constraints on the $\gm2$ favoured parameter space using $36.1{~\rm fb}^{-1}$ of data in the multilepton$+\met$ search channel at the 13 TeV LHC [@multilepton_latest]: for the selectron and/or smuon NLSP scenario (left column) and when all three generation sleptons are lighter than $\C1$ and $\N2$ (right column). See text in the Note added section for details.*[]{data-label="fig:update_slepton"}](mixed_NLSP_smuon500 "fig:")
Acknowledgments {#acknowledgments .unnumbered}
===============
S.M. would like to thank Yasuhito Sakaki for help with [Mathematica]{} graphics. K.H. and S.M. are supported in part by the U.S. Department of Energy under grant No. DE-FG02-95ER40896 and in part by the PITT PACC. K.M. is supported by the China Scholarship Council, and the National Natural Science Foundation of China under Grant No. 11647018, and partially by the Project of Science and Technology Department of Shaanxi Province under Grant No. 15JK1150.
Appendix {#appendix .unnumbered}
========
In this Appendix, we provide the details of our Monte Carlo (MC) simulation framework for the $\C1$ and $\N2$ searches in the $\geq 2\tau_h+\met$ channel. We follow the ATLAS analyses in this regard, details of which can be found in Ref. [@ATLAS_tau_8] for the search using $20.3 \fb^{-1}$ of data from the 8 TeV LHC, and in Ref. [@ATLAS_tau_13] for the search with $14.8\fb^{-1}$ of data from the 13 TeV run.
For our MC simulation of a general MSSM scenario, we obtain the MSSM mass spectra from weak scale inputs of the soft SUSY breaking parameters using [SUSPECT2]{} [@Suspect]. We then compute the decay branching ratios (BR) of the supersymmetric particles using [SDECAY]{} [@Sdecay], with the help of the [SUSY-HIT]{} package [@Susyhit]. The resulting mass spectra and decay BRs are used as inputs to [MadGraph5]{} [@MG5] for generating parton level events for electroweak-ino and stau pair production, which are then passed onto [PYTHIA6]{} [@Pythia] for including the effects of parton shower, hadronization and underlying events. We use [DELPHES2]{} [@Delphes] for including the effects of object reconstruction and detector resolution. Jets are reconstructed using the anti-$k_T$ algorithm [@antikt] with radius parameter $R=0.4$ using [FastJet]{} [@Fastjet]. We employ the [CTEQ6L1]{} [@cteq; @LHAPDF] parton distribution functions, and the factorization and renormalization scales have been fixed at the default event-by-event choice of [MadGraph5]{}.
![*Comparison of the $95\%$ C.L. exclusion contours obtained by our MC simulation (blue solid contour) with the corresponding ATLAS pMSSM exclusion (green dashed contour) in the $\mu-M_2$ plane, using $20.3\fb^{-1}$ of data from the 8 TeV LHC, for the parameter choices $M_1=50 \gev, M_{\tla}=90 \gev$ and $\tan \beta=50$. For this analysis, the $\tla$ is taken to be a right-stau. See text for details.*[]{data-label="fig:valid_8TeV"}](ATLAS_vs_MC_8TeV.pdf)
The 8 TeV ATLAS search is divided into four different signal regions, primarily targeting the $\C1\N2, \C1\C1$ and direct stau pair production processes. The 8 TeV results, however, are not sensitive to direct stau pair production exclusively [@ATLAS_tau_8], and most of the current sensitivity results from electroweak-ino pair production. The kinematic selection criteria used are as described in Ref. [@ATLAS_tau_8], see in particular the discussion in Sections 6 and 7, with the definition of different signal regions in Table 1. For the three different MSSM reference points and the phenomenological MSSM (pMSSM) parameter space in the Higgsino-wino mass ($\mu-M_2$) plane studied by ATLAS, we find that our event yields for the signal process are somewhat larger than the numbers obtained by ATLAS. A very likely origin of this difference is the simple modelling of tau-jet identification and reconstruction efficiencies in our MC simulation. It is, however, encouraging that we can approximately match the ATLAS results by multiplying our event yields with a constant fudge factor. After including an average K-factor in the range of $1.2-1.3$ to take into account the effects of higher order corrections to our leading order cross-section computation (we compute the next-to-leading order cross-sections using [Prospino]{} [@Prospino]), for the 8 TeV analysis, this fudge factor is found to be around $0.4$. In Fig. \[fig:valid\_8TeV\], we compare our $95\%$ C.L. exclusion (blue solid contour) with the ATLAS pMSSM exclusion (green dashed contour) in the $\mu-M_2$ plane, for the parameter choices $M_1=50 \gev, M_{\tla}=90 \gev$ and $\tan \beta=50$. For this analysis, the $\tla$ is taken to be a right-stau (see Figure 10 in Ref. [@ATLAS_tau_8] for the corresponding ATLAS result). As we can see in this figure, our results are broadly in agreement with the ATLAS limits, after the above fudge factor is taken into account. In deriving the 8 TeV exclusion contours discussed in other sections of this study, we have adopted the same methodology.
![*Comparison of the $95\%$ C.L. exclusion contours obtained by our MC simulation (blue solid contour) with the corresponding ATLAS exclusion (green dashed contour) in a simplified model study of $\C1\N2$ pair production, using $14.8\fb^{-1}$ of data from the 13 TeV LHC. The tau-slepton and tau-sneutrino masses are set to be $(M_{\C1}+M_{\neu})/2$, with $M_{\C1}=M_{\N2}$. See text for details.*[]{data-label="fig:valid_13TeV"}](ATLAS_vs_MC_13TeV.pdf)
For the 13 TeV LHC study, we follow the recent ATLAS note [@ATLAS_tau_13], see in particular Sections 4 and 5 as well as Table 1 in Ref. [@ATLAS_tau_13] for details on the object reconstruction, event selection and the definition of the different signal regions. For the 13 TeV analysis, so far, ATLAS has reported their results in terms of simplified models, and the corresponding interpretations for the pMSSM parameter space is not yet available. We therefore compare our results with the ATLAS limits on $\C1\N2$ pair production, where the $\C1$ and $\N2$ decay via an intermediate tau-slepton and tau-sneutrino of mass $(M_{\C1}+M_{\neu})/2$, with $M_{\C1}=M_{\N2}$. We compare the $95\%$ C.L. exclusion contours based on our MC simulation (blue solid) with the corresponding ATLAS results (green dashed) in Fig. \[fig:valid\_13TeV\]. As in the 8 TeV analysis, we found our event yields to be larger than the ATLAS numbers, and it was necessary to normalize our signal yields by a fudge factor of around $0.14$, after taking the K-factor correction into account. This fudge factor is somewhat smaller than the one required for the 8 TeV analysis. This is possibly due to the dependence of the tau reconstruction and identification efficiencies on the tau-jet kinematics, which can be quite different depending upon the centre of mass energy, and the difference in the kinematic selection criteria adopted. Once again, after taking the fudge factor into account, the overall agreement of our MC simulation with the ATLAS results is found to be reasonably good, especially in the intermediate $\C1/\N2$ mass region, with our bounds being slightly weaker in the high mass region, and slightly stronger in the lower mass region [^1].
[99]{} For a review, see, for example, F. Jegerlehner and A. Nyffeler, “The Muon $g-2$,” Phys. Rept. [**477**]{} (2009) 1. G. W. Bennett [*et al.*]{} \[Muon g-2 Collaboration\], “Final Report of the Muon E821 Anomalous Magnetic Moment Measurement at BNL,” Phys. Rev. D [**73**]{} (2006) 072003. B. L. Roberts, “Status of the Fermilab Muon $(g-2)$ Experiment,” Chin. Phys. C [**34**]{} (2010) 741. K. Hagiwara, R. Liao, A. D. Martin, D. Nomura and T. Teubner, “$(g-2)_\mu$ and $\alpha(M_Z^2)$ re-evaluated using new precise data,” J. Phys. G [**38**]{} (2011) 085003. J. Grange [*et al.*]{} \[Muon g-2 Collaboration\], “Muon (g-2) Technical Design Report,” arXiv:1501.06858 \[physics.ins-det\]. N. Saito \[J-PARC g-2/EDM Collaboration\], “A novel precision measurement of muon g-2 and EDM at J-PARC,” AIP Conf. Proc. [**1467**]{} (2012) 45; [http://g-2.kek.jp/portal/documents.html]{}
T. Blum, N. Christ, M. Hayakawa, T. Izubuchi, L. Jin, C. Jung and C. Lehner, “Connected and leading disconnected hadronic light-by-light contribution to the muon anomalous magnetic moment with physical pion mass,” arXiv:1610.04603 \[hep-lat\]. T. Moroi, “The Muon anomalous magnetic dipole moment in the minimal supersymmetric standard model,” Phys. Rev. D [**53**]{} (1996) 6565 Erratum: \[Phys. Rev. D [**56**]{} (1997) 4424\]. U. Chattopadhyay and P. Nath, “Probing supergravity grand unification in the Brookhaven g-2 experiment,” Phys. Rev. D [**53**]{} (1996) 1648. J. L. Lopez, D. V. Nanopoulos and X. Wang, “Large (g-2)-mu in SU(5) x U(1) supergravity models,” Phys. Rev. D [**49**]{} (1994) 366. G. C. Cho, K. Hagiwara and M. Hayakawa, “Muon g-2 and precision electroweak physics in the MSSM,” Phys. Lett. B [**478**]{} (2000) 231 G. C. Cho and K. Hagiwara, “Supersymmetric contributions to muon g-2 and the electroweak precision measurements,” Phys. Lett. B [**514**]{} (2001) 123
S. Heinemeyer, D. Stockinger and G. Weiglein, “Two loop SUSY corrections to the anomalous magnetic moment of the muon,” Nucl. Phys. B [**690**]{} (2004) 62. D. Stockinger, “The Muon Magnetic Moment and Supersymmetry,” J. Phys. G [**34**]{} (2007) R45. S. Marchetti, S. Mertens, U. Nierste and D. Stockinger, “Tan(beta)-enhanced supersymmetric corrections to the anomalous magnetic moment of the muon,” Phys. Rev. D [**79**]{} (2009) 013010. P. von Weitershausen, M. Schafer, H. Stockinger-Kim and D. Stockinger, “Photonic SUSY Two-Loop Corrections to the Muon Magnetic Moment,” Phys. Rev. D [**81**]{} (2010) 093004.
G. C. Cho, K. Hagiwara, Y. Matsumoto and D. Nomura, “The MSSM confronts the precision electroweak data and the muon $g-2$,” JHEP [**1111**]{} (2011) 068. H. Fargnoli, C. Gnendiger, S. Pa§ehr, D. Stockinger and H. Stockinger-Kim, “Two-loop corrections to the muon magnetic moment from fermion/sfermion loops in the MSSM: detailed results,” JHEP [**1402**]{} (2014) 070 M. Endo, K. Hamaguchi, S. Iwamoto and T. Yoshinaga, “Muon g-2 vs LHC in Supersymmetric Models,” JHEP [**1401**]{} (2014) 123; S. P. Das, M. Guchait and D. P. Roy, “Testing SUSY models for the muon g-2 anomaly via chargino-neutralino pair production at the LHC,” Phys. Rev. D [**90**]{} (2014) no.5, 055011; M. A. Ajaib, B. Dutta, T. Ghosh, I. Gogoladze and Q. Shafi, “Neutralinos and sleptons at the LHC in light of muon $(g-2)_{\mu}$,” Phys. Rev. D [**92**]{} (2015) no.7, 075033; K. Kowalska, L. Roszkowski, E. M. Sessolo and A. J. Williams, “GUT-inspired SUSY and the muon $g-2$ anomaly: prospects for LHC 14 TeV,” JHEP [**1506**]{}, 020 (2015); T. T. Yanagida and N. Yokozaki, “Muon $g-2$ in MSSM gauge mediation revisited,” Phys. Lett. B [**772**]{}, 409 (2017); For a comprehensive analysis of the LHC limits on generic weak scale particles contributing to the $(g-2)_\mu$, see, A. Freitas, J. Lykken, S. Kell and S. Westhoff, “Testing the Muon g-2 Anomaly at the LHC,” JHEP [**1405**]{} (2014) 145 Erratum: \[JHEP [**1409**]{} (2014) 155\].
M. Endo, K. Hamaguchi, T. Kitahara and T. Yoshinaga, “Probing Bino contribution to muon $g - 2$,” JHEP [**1311**]{} (2013) 013; M. Endo, K. Hamaguchi, S. Iwamoto, T. Kitahara and T. Moroi, “Reconstructing Supersymmetric Contribution to Muon Anomalous Magnetic Dipole Moment at ILC,” Phys. Lett. B [**728**]{} (2014) 274.
S. Gori, S. Jung and L. T. Wang, “Cornering electroweakinos at the LHC,” JHEP [**1310**]{} (2013) 191; T. Han, S. Padhi and S. Su, “Electroweakinos in the Light of the Higgs Boson,” Phys. Rev. D [**88**]{} (2013) no.11, 115010; M. Berggren, T. Han, J. List, S. Padhi, S. Su and T. Tanabe, “Electroweakino Searches: A Comparative Study for LHC and ILC (A Snowmass White Paper),” arXiv:1309.7342 \[hep-ph\]; M. Chakraborti, U. Chattopadhyay, A. Choudhury, A. Datta and S. Poddar, “The Electroweak Sector of the pMSSM in the Light of LHC - 8 TeV and Other Data,” JHEP [**1407**]{} (2014) 019; T. A. W. Martin and D. Morrissey, “Electroweakino constraints from LHC data,” JHEP [**1412**]{} (2014) 168; M. Chakraborti, U. Chattopadhyay, A. Choudhury, A. Datta and S. Poddar, “Reduced LHC constraints for higgsino-like heavier electroweakinos,” JHEP [**1511**]{} (2015) 050; A. Choudhury and S. Mondal, “Revisiting the Exclusion Limits from Direct Chargino-Neutralino Production at the LHC,” Phys. Rev. D [**94**]{} (2016) no.5, 055024; A. Datta, N. Ganguly and S. Poddar, “New Limits on Heavier Electroweakinos and their LHC Signatures,” Phys. Lett. B [**763**]{} (2016) 213; C. Arina, M. Chala, V. Martin-Lozano and G. Nardini, “Confronting SUSY models with LHC data via electroweakino production,” arXiv:1610.03822 \[hep-ph\].
G. Aad [*et al.*]{} \[ATLAS Collaboration\], “Search for direct production of charginos, neutralinos and sleptons in final states with two leptons and missing transverse momentum in $pp$ collisions at $\sqrt{s} =$ 8 TeV with the ATLAS detector,” JHEP [**1405**]{} (2014) 071.
G. Abbiendi [*et al.*]{} \[OPAL Collaboration\], Search for anomalous production of dilepton events with missing transverse momentum in e+ e- collisions at $s^{(1/2)}$ = 183-Gev to 209-GeV,” Eur. Phys. J. C [**32**]{} (2004) 453. G. Aad [*et al.*]{} \[ATLAS Collaboration\], “Search for the direct production of charginos, neutralinos and staus in final states with at least two hadronically decaying taus and missing transverse momentum in $pp$ collisions at $\sqrt{s}$ = 8 TeV with the ATLAS detector,” JHEP [**1410**]{} (2014) 096 The ATLAS collaboration \[ATLAS Collaboration\], “Search for electroweak production of supersymmetric particles in final states with tau leptons in $\sqrt{s} =$ 13TeV pp collisions with the ATLAS detector,” ATLAS-CONF-2016-093. K. Hagiwara, S. Mukhopadhyay and J. Nakamura, “10 GeV neutralino dark matter and light stau in the MSSM,” Phys. Rev. D [**89**]{} (2014) no.1, 015023.
The ATLAS collaboration \[ATLAS Collaboration\], “Search for supersymmetry with two and three leptons and missing transverse momentum in the final state at $\sqrt{s}=13$ TeV with the ATLAS detector,” ATLAS-CONF-2016-096. G. Aad [*et al.*]{} \[ATLAS Collaboration\], “Search for the electroweak production of supersymmetric particles in $\sqrt{s}$=8 TeV $pp$ collisions with the ATLAS detector,” Phys. Rev. D [**93**]{} (2016) no.5, 052002.
G. Aad [*et al.*]{} \[ATLAS Collaboration\], “Search for direct pair production of a chargino and a neutralino decaying to the 125 GeV Higgs boson in $\sqrt{s} = 8$ TeV ${pp}$ collisions with the ATLAS detector,” Eur. Phys. J. C [**75**]{} (2015) no.5, 208. The ATLAS collaboration \[ATLAS Collaboration\], “Search for electroweak production of supersymmetric particles in the two and three lepton final state at $\boldmath{\sqrt{s}=13\,}$TeV with the ATLAS detector,” ATLAS-CONF-2017-039. M. Aaboud [*et al.*]{} \[ATLAS Collaboration\], “Search for the direct production of charginos and neutralinos in final states with tau leptons in $\sqrt{s} = $ 13 TeV $pp$ collisions with the ATLAS detector,” Eur. Phys. J C [**78**]{}, 154 (2018).
A. Djouadi, J. L. Kneur and G. Moultaka, “SuSpect: A Fortran code for the supersymmetric and Higgs particle spectrum in the MSSM,” Comput. Phys. Commun. [**176**]{} (2007) 426 doi:10.1016/j.cpc.2006.11.009 \[hep-ph/0211331\].
M. Muhlleitner, A. Djouadi and Y. Mambrini, “SDECAY: A Fortran code for the decays of the supersymmetric particles in the MSSM,” Comput. Phys. Commun. [**168**]{} (2005) 46.
A. Djouadi, M. M. Muhlleitner and M. Spira, “Decays of supersymmetric particles: The Program SUSY-HIT (SUspect-SdecaY-Hdecay-InTerface),” Acta Phys. Polon. B [**38**]{} (2007) 635. J. Alwall, M. Herquet, F. Maltoni, O. Mattelaer and T. Stelzer, “MadGraph 5 : Going Beyond,” JHEP [**1106**]{}, 128 (2011); J. Alwall [*et al.*]{}, “The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations,” JHEP [**1407**]{} (2014) 079. T. Sjostrand, S. Mrenna and P. Z. Skands, “PYTHIA 6.4 Physics and Manual,” JHEP [**0605**]{}, 026 (2006). S. Ovyn, X. Rouby and V. Lemaitre, “DELPHES, a framework for fast simulation of a generic collider experiment,” arXiv:0903.2225 \[hep-ph\]; J. de Favereau [*et al.*]{} \[DELPHES 3 Collaboration\], “DELPHES 3, A modular framework for fast simulation of a generic collider experiment,” JHEP [**1402**]{} (2014) 057.
M. Cacciari, G. P. Salam and G. Soyez, “The Anti-k(t) jet clustering algorithm,” JHEP [**0804**]{} (2008) 063.
M. Cacciari, G. P. Salam and G. Soyez, “FastJet user manual,” Eur. Phys. J. C [**72**]{} (2012) 1896; M. Cacciari and G. P. Salam, “Dispelling the $N^{3}$ myth for the $k_t$ jet-finder,” Phys. Lett. B [**641**]{} (2006) 57. J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. M. Nadolsky and W. K. Tung, “New generation of parton distributions with uncertainties from global QCD analysis,” JHEP [**0207**]{} (2002) 012.
M. R. Whalley, D. Bourilkov and R. C. Group, “The Les Houches accord PDFs (LHAPDF) and LHAGLUE,” hep-ph/0508110.
W. Beenakker, M. Klasen, M. Kramer, T. Plehn, M. Spira and P. M. Zerwas, “The Production of charginos / neutralinos and sleptons at hadron colliders,” Phys. Rev. Lett. [**83**]{} (1999) 3780 Erratum: \[Phys. Rev. Lett. [**100**]{} (2008) 029901\].
[^1]: Without taking into account the fudge factor, the obtained constraints from the multi-tau and missing transverse momentum channel would be stronger. However, since the $\gm2$ favoured parameter region where this search channel is important is already disfavoured in our analyses even with the fudge factor, our conclusions would remain the same.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Experiments have been performed using a spherical superconducting microwave resonator that simulates the geometric structure of the C$_{60}$ fullerene molecule. The objective was to study with very high resolution the exceptional spectral properties emerging from the symmetries of the icosahedral structure of the carbon lattice. In particular, the number of zero modes has been determined to test the predictions of the Atiyah-Singer index theorem, which relates it to the topology of the curved carbon lattice. This is, to the best of our knowledge, the first experimental verification of the index theorem.'
author:
- 'B. Dietz'
- 'T. Klaus'
- 'M. Miski-Oglu'
- 'A. Richter'
- 'M. Bischoff'
- 'L. von Smekal'
- 'J. Wambach'
title: 'Test of the Atiyah-Singer Index Theorem for Fullerene with a Superconducting Microwave Resonator'
---
— The spectrum of graphene, a monolayer of carbon (C) atoms arranged on a hexagonal lattice, has been the focus of extensive theoretical [@Beenakker2008; @Castro2009] and experimental studies [@Ponomarenko2008]. Its universal properties were often also investigated experimentally in analog systems, so-called ’artificial graphene’ [@Polini2013a], e.g., in our group in photonic crystals [@Bittner2010; @Kuhl2010; @Bellec2013; @Rechtsman2013; @Khanikaev2013; @Dietz2015]. Moreover, theoretically much attention has been devoted to curved graphene structures like fullerene molecules [@Kroto1985; @Manousakis1991; @Gonzalez1992; @Gonzalez1993; @Kolesnikov2006; @Vozmediano2010] and the connection between their spatial symmetries and electronic properties. Here, the most famous example is the C$_{60}$ molecule. It consists of 60 carbon atoms at the vertices of a truncated icosahedron and has the shape of a soccer ball. Concerning the spectral properties of fullerenes the number of near-zero modes, i.e., of electronic states with excitation energies close to zero, have been of particular interest since they determine the electrical conductivity. In [@Pachos2007; @Pachos2007a] an index theorem has been derived that allows the computation of the number of such near-zero modes from the topology of the surface. It was deduced from the renowned Atiyah-Singer index theorem [@Atiyah1963; @Atiyah1968; @Atiyah1968a; @Atiyah1968b] which states that the analytic index of an elliptic differential operator on a compact manifold equals the topological one, in other words, that there is a connection between the number of zero modes of the operator and the topology of the manifold on which it is defined. The aim of the high-resolution experiments presented in this letter was to test these predictions in experiments with a superconducting microwave resonator of the same topology as the C$_{60}$ molecule.
First we briefly review the salient features of graphene and fullerenes and outline the derivation of the index theorem from [@Pachos2007; @Pachos2007a] for deformed graphene sheets. We then describe the experimental setup and compare the results of the measurements to the predictions from the index theorem and to tight-binding model (TBM) calculations. These allow us to study the approach to the thermodynamic limit of an infinite number of carbon atoms.
[*Graphene, fullerenes and the Atiyah-Singer theorem.*]{}— The honeycomb structure of graphene is formed by two interpenetrating triangular sublattices. As a consequence, at half filling the Fermi surface in graphene reduces to two independent points in the first Brillouin zone, the so-called ’Dirac points’, denoted by ${\bf K_+}$ and ${\bf K_-}$ [@Beenakker2008; @Castro2009] that are conical intersections of the valence and the conduction band. Low energy excitations within the cone regions around ${\bf K_\pm}$ have a linear dispersion with a slope given by the Fermi velocity $v_F$. On an infinte graphene sheet they are therefore described by a Dirac Hamiltonian for massless spin-1/2 quasiparticles consisting of partner Hamiltonians $$H_\pm=\pm v_F \, \sigma^\alpha q_\alpha$$ which describe excitations with momentum ${\bf q}=(q_x,q_y)$ in each of the two Dirac cones around ${\bf K_\pm}$. The Pauli matrices $\sigma^\alpha $ with $\alpha = x,y$ act on the two sublattice components of the excitations, combined in two-dimensional spinors and hence referred to as quasi-spin. Both cones together then yield a four-component Dirac equation [@Beenakker2008].
Fullerene molecules can be constructed by introducing positive curvature into an initially flat graphene sheet [@Lammert2004]. The bending is realized by replacing hexagons by pentagons, ensuring at the same time that the lattice is not stretched and each C atom keeps three neighbors. To determine the number of pentagons $n_5$ necessary to generate a spherical fullerene molecule with $n_6$ hexagons one uses the Euler formula [@Millman1977] which relates the number of vertices $V$, of edges $E$, faces $F$ and open ends $N_\mathrm{open}$ of an arbitrary two-dimensional lattice to the genus $g$ of the surface formed by it, via the Euler characteristic $$\chi = V-E+F = 2 ( 1-g) - N_\mathrm{open} \, .$$ For a lattice of pentagons and hexagons, $V=(5n_5+6n_6)/3$, $E=(5n_5+6n_6)/2$ and $F=n_5+n_6$, this gives $\chi =n_5/6$. Without open ends $\chi$ must be an even integer on a closed orientable surface due to the Gauß-Bonnet theorem [@Millman1977]. Hence, for a flat graphene sheet with periodic boundary conditions one has $g =1$ for the torus and $n_5=0$, while a sphere with $g=0$ needs $n_5 = 12$ pentagons to avoid open ends. Consequently, fullerenes are grown from the C$_{60}$ molecule by increasing the number of hexagons, i.e., always have twelve pentagons at the same relative positions. This also applies to the thermodynamic limit, and one expects that their low-energy electronic excitations are described by a Dirac equation on a sphere.
To introduce a pentagon into the honycomb lattice, a $\pi /3$ sector is cut out and then the edges are glued together [@Gonzalez1992; @Gonzalez1993; @Pachos2007; @Pachos2007a] as illustrated in Fig. \[fig1\].
![(Color online) Left panel: A two-dimensional graphene sheet. The red and the blue dots mark the two independent triangular sublattices. In order to form a curved sheet which contains one pentagon, a $\pi/3$ segment is cut out from the sheet. Right panel: The conically deformed graphene sheet with the pentagon at the apex.[]{data-label="fig1"}](fig1.eps){width="\linewidth"}
Thereby a pentagon is created at the apex of the emerging cone. Along the seam, two C atoms from the same triangular sublattice, e.g., the red ones in Fig. \[fig1\], are connected. This results in a coupling of the Dirac operators associated with the $\bf K_\pm$ points. Indeed, when the four-dimensional spinor associated with the Dirac equation of the flat graphene sheet is transported around the apex by an angle $2\pi$, it is forced to jump at the seam from a red site to another red one instead of to a blue one. It thus acquires a non-trivial phase, which can be accounted for by introducing a non-Abelian gauge field $A_\mu$ in the Hamiltonian which yields a flux of $ (\pi/2) \, \tau^y $ when integrated along a closed loop around the apex. The Pauli matrix $\tau^y$ thereby couples the ${\bf K_+}$ and ${\bf K_-}$ spinor components [@Beenakker2008]. This description entails the existence of a ficitious magnetic monopole inside the surface. In the case of fullerenes it is located at the center of the spherical molecule, yielding a flux of $1/8$ through each of the twelve pentagons. Thus the total magnetic monopole charge inside the sphere equals 3/2. In addition to that, analogous to the daily rotation of Foucault’s pendulum, a deficit angle of $\pi/3$ arises when moving a frame along a loop around the apex. It is here described by a quasi-spin connection $Q_\mu$ with circulation $ -({\pi}/{6})\, \sigma^z $ around the apex.
The coupling of the ${\bf K_\pm}$ spinor components in the resulting four-dimensional Dirac equation can be removed by a rotation, which leads to two independent two-dimensional Dirac equations denoted by $l=1,2$ [@Pachos2007a], $$\slashed D^l\psi^l= v_F \, \sigma^\alpha e_\alpha^\mu\left(q_\mu -iQ_\mu-iA_\mu^l\right)\psi^l= E\psi^l\, ,
\label{DE}$$ where $e_\alpha^\mu$ is the Zweibein in the tangent plane of the surface, and $A^l_\mu$ are the components of $A_\mu$ in the rotated basis with circulation $\pm \pi/2 $ for $l= 1$ and $2$, respectively, where $\vec A^l$ is now an Abelian gauge field.
The four-dimensional Dirac equation obtained from Eq. (\[DE\]) provides a good description of the low-energy excitations of the C molecules [@Gonzalez1992; @Gonzalez1993; @Kolesnikov2006]. It yields the long-wavelength excitations of the deformed graphene sheet in the vicinity of the Dirac points, and thus also the zero modes that we are interested in. For the fullerenes the Dirac operators $\slashed D^l$ are elliptic and defined on a compact surface. Hence the Atiyah-Singer index theorem [@Atiyah1963; @Atiyah1968; @Atiyah1968a; @Atiyah1968b] applies. Ten years after its first formulation a new proof was provided based on the heat equation [@Atiyah1973] which was later employed for the derivation of an index theorem for graphene sheets deformed by pentagons and heptagons [@Pachos2007; @Pachos2007a] as briefly reviewed in the following.
Each Dirac operator in Eq. (\[DE\]) can be written in terms of off-diagonal partner operators $P$ and $ P^\dagger$ [@Stone1984], so $(\slashed D^l)^2$ contains only diagonal operators $PP^\dagger$ and $P^\dagger P$ that have the same number of zero modes as $P^\dagger$ and $P$, respectively. Furthermore, the non-zero eigenvalues of $PP^\dagger$ and $P^\dagger P$ are identical. The analytic index of $\slashed D^l$ is given by the difference of the numbers of zero modes of $P$ and $P^\dagger$ denoted by $\nu_\pm$, respectively, i.e., $\mathrm{index}(\slashed D^l)=\nu_+-\nu_-$ [@Pachos2007; @Pachos2007a]. More importantly, however, this index is related to the total flux of the effective gauge field via the Atiyah-Singer index theorem [@Pachos2007], $$\mathrm{index}(\slashed D^l)=\frac{1}{2\pi}\iint_\Omega {\mathcal{F}}^l {\rm d}\Omega\, .
\label{AS}$$ The integral is taken over the compact surface $\Omega$ and ${\mathcal{F}}^l=\partial\wedge A^l$ are the field strengths associated with the now Abelian gauge potentials $A^l$. Stokes’ theorem then implies from the closed loops around each apex that $$\frac{1}{2\pi}\iint_{\Omega}{\mathcal{F}}^l {\rm d}\Omega =\frac{1}{2\pi}n_5\oint A^l_\mu \, {\rm d}s^\mu \equiv\pm\frac{3}{2}\chi\, .$$ The Euler formula thus leads to the Atiyah-Singer index theorem for fullerenes in the form, $\mathrm{index}(\slashed D^l)=\pm 3(1-g)$. In two dimensions either $\nu_+$ or $\nu_-$ vanish. Hence, the index theorem provides the number of zero modes [@Pachos2007; @Pachos2007a].
The total number of zero modes is the sum of those of the subsystems corresponding to $l=1,2$. Consequently, according to the index theorem, the zero modes of the four-dimensional Dirac operator for spherical fullerenes correspond to two triplets. The same result has been obtained in a continuum model for the low-energy electronic states of icosahedral fullerenes [@Gonzalez1992; @Gonzalez1993; @Kolesnikov2006; @Vozmediano2010]. We emphasize, however, that the eigenvalues of the near-zero modes tend to zero, i.e., coincide with the energy at the Dirac points, only in the thermodynamic limit of an infinity number of $C$ atoms. In a sufficiently large but finite fullerene molecule they are expected to lie much closer to the Dirac energy than all the other ones.
[*Experimental setup and resonance spectra.*]{}— Hitherto, experiments have been performed with flat, superconducting microwave resonators, so-called ’microwave billiards’ [@Richter1999] to address problems from the fields of quantum chaos [@Haake2001; @StoeckmannBuch2000] and compound nucleus reactions [@Mitchell2010]. In this context, the equivalence of the Helmholtz equation and the non-relativistic Schrödinger equation of the corresponding quantum billiard is exploited which holds below a maximum microwave frequency $f_{\rm max}=c/(2d)$ with $c$ the velocity of light and $d$ the height of the billiard. Consequently, the eigenvalues of a quantum billiard can be obtained experimentally from the eigenfreqencies of the microwave billiard of corresponding shape. Recently, we realized experiments with superconducting microwave Dirac billiards and studied universal spectral properties of graphene sheets [@Wurm2011] with unprecedented accuracy [@Bittner2012; @Dietz2015; @Dietz2015a].
The aim of the experiments presented here was the investigation of the universal spectral properties of the fullerene C$_{60}$ molecule attributed to its lattice structure and to determine the number of zero modes which, according to the Atiyah Singer index theorem solely depends on the number of pentagons. For this we use a system exhibiting the same topological properties, namely a quantum fullerene billiard on a sphere, consisting of a network of 60 circular billiards at the positions of the C atoms connected by three of the altogether 90 straight leads with three adjacent ones. We studied them experimentally by using instead of a planar, superconducting microwave (Dirac) billiard a cavity, which is imprinted on a sphere. The microwave fullerene billiard displayed in Fig. \[fig2\] was constructed by milling a total of 60 circular cavities (vertices) and 90 rectangular channels (edges) out of a brass sphere and then closing them with small triangular brass plates of 5 mm thickness and 3 mm thick rectangular ones, respectively. Before the parts were screwed together, they were covered with lead, which is superconducting below $T_c=7.2$ K.
![(Color online) Lead plated fullerene billiard used in the experiments. In the left part the small plates that cover the circular cavities and the rectangular channels were removed. The red caps protect the antenna ports. The billiard is superconducting below T$_c$=7.2 K.[]{data-label="fig2"}](fig2.eps){width="0.6\linewidth"}
The diameter of the sphere of 160 mm was limited by the size of the liquid Helium cryostat in which the resonator was cooled down to 4.2 K in order to attain superconductivity. The radius of the circular cavities was 12 mm, the widths of the waveguides 14 mm, before lead coating them. Thus the cutoff frequency for the first propagating mode in the latter is $f_c^1\gtrsim 10.714$ GHz. In total, 8 antennas were attached to the triangular plates. Two, covered with red caps, are visible in Fig. \[fig1\]. The height of the resonator was 3 mm corresponding to $f_{\rm max}=50$ GHz.
For the measurement of the transmission spectrum shown in the upper panel of Fig. \[fig3\] microwave power was coupled into the microwave billiard via antenna a and the output signal was received at antenna b, with a and b denoting two of the 8 antennas. A vectorial network analyzer determined the relative phase and amplitude of the output and input signals, thus yielding the scattering matrix element $S_{\rm ba}$. The smallest resonance frequency equals $f=8.254$ GHz, so we show the spectrum from 8 - 40 GHz. Due to the high-quality factor $Q>10^5$ of the resonator, all resonances could be resolved in that frequency range. We concentrate our discussion here on the region between 8.254 and 18.801 GHz.
![Transmission spectrum of the fullerene billiard (upper panel) up to 40 GHz. The first band ranges from 8.254 to 8.779 GHz. It contains $60$ resonances, that are separated into 15 groups with the number of resonances indicated in the middle panel, which shows a zoom into it. The Dirac frequency $f_D=8.504$ GHz is marked by an arrow. The two triplets of interest (the zero modes) are clearly resolved and shown in the lower panel.[]{data-label="fig3"}](fig3){width="\linewidth"}
The spectrum exhibits three distinct bands, containing 60, 210 and 90 resonances, respectively, in the frequency intervals \[8.254, 8.779\] GHz, \[11.492, 16.657\] GHz and \[18.312, 18.801\] GHz. They are located around the eigenfrequencies $\tilde f_1\simeq 8.4$ GHz, $\tilde f_2\simeq 13.5$ GHz and $\tilde f_3\simeq 18.5$ GHz of the first three quasibound states in an open circular billiard of the same size as the cavities in the resonator with openings at the positions of the waveguides. The modes excited inside the circular cavities resemble within a given band the corresponding mode in the open circular billiard, and are described by $J_0$, $J_1$ and $J_2$ Bessel functions in the first, second and third band, respectively. In the latter two cases they are twofold degenerate due to the mirror symmetry. Note, that the circular cavities exhibit no threefold symmetry, because each of them is part of one pentagon and two hexagons and the internal angles differ.
The first band is located well below the cutoff frequency of the waveguide. Consequently, the electric field modes excited inside the circular cavities are only weakly coupled to those in the neighboring ones. The resonance frequencies in the second band are above the cutoff frequency. Accordingly, the modes in the cavities are coupled via the modes inside the waveguides, and thus mimick a situation where the C atoms are coupled to the neighboring ones via an extra atom, thus explaining the number of resonances in this band. The third band is still below the frequency $f_c^2\gtrsim 20.143$ GHz of the second propagating mode in the waveguides. As a result, the number of possible mode configurations is restricted due to the symmetry properties of the modes excited inside the cavities, that prefereably couple to the second excited mode inside the waveguide. Above 20.232 GHz, i.e., beyond $f_c^2$, several bands are intertwined.
In summary, only the first band can be used to model the situation in the fullerene C$_{60}$ molecule. The middle panel of Fig. \[fig3\] shows a magnification of it. Fifteen groups of nearly degenerate resonances are clearly visible. The number of resonances identified in each of them is indicated and coincides with the degrees of degeneracy predicted on the basis of group theoretical considerations for the eigenfrequencies because of its truncated icosahedral structure [@Manousakis1991; @Dresselhaus1996]. In the group with degeneracy degree 9, in fact, the energy values of 5 and 4 degenerate eigenfrequencies, respectively, are accidentally the same. We emphasize, that we were only able to identify *all* 60 resonance frequencies because the degeneracies were lifted. The reason is that the symmetries of the resonator structure were slightly perturbed due to the presence of the antennas and unavoidable marginal inhomogeneities in the lead coating. The influence of the former turned out to be negligible for sufficiently short antennas. The effect of the latter on the size of the splittings of the nearly degenerate resonance frequencies was tested by smoothing the surface of the resonator which indeed induced a reduction of the splittings in each group. The pair of triplets visible in the middle panel of Fig. \[fig3\] and shown in a further magnification of the spectrum in the lowest panel, is closest to the Dirac frequency at $f_D=8.504$ GHz which was determined as described below. The zoom demonstrates the high resolution necessary to resolve the 6 resonances. These are the 6 modes conjectured by the Atiyah-Singer index theorem that we were looking for. As is discussed next they are corroborated also by TBM calculations.
[*Tight-binding model description of the spectra.*]{}— The eigenvalues of the C$_{60}$ molecule have been computed previously using the TBM [@Manousakis1991], however, a stringent test of its applicability was missing. Given the experimental results on the eigenfrequencies in the first band of the fullerene resonator we are now in a position to check the validity of the TBM in detail. As stated above, the modes excited in the 60 cavities are weakly coupled, which is an essential prerequisit for the applicability of the TBM. Detailed calculations showed a quantitative agreement between the computed and the measured frequencies only when including next-nearest, and second and third-nearest neighbor couplings with strengths $t_1,\, t_2$ and $t_3$, respectively. This yielded for the frequency of the isolated cavities $f_0=8.515$ GHz and the coupling parameters $t_1=-0.0929$ GHz, $t_2=0.0035$ GHz and $t_3=0.0005$ GHz. The eigenvalues deduced from the TBM appear as 15 groups of degenerate eigenvalues with the same multiplicities as the resonances in the spectrum shown in the middle panel of Fig. \[fig3\]. An even better agreement was achieved by taking into account the fact that, due to the inhomogeneities in the lead coating, the radii of the cavities are slightly different. In order to estimate the deviations thus induced in $f_0$, we used the fact, that $f_0$ is given by the first zero of the $J_0$ Bessel function, $J_0(kR)=0$, for a circular cavity of radius $R$. We inserted for each of the 60 cavities the measured radius and replaced $f_0$ by the individual values. Thereby, the degeneracies were removed. In panel a) of Fig. \[fig4\] we compare the resonance density $\rho (f)=\sum_i\delta(f-f_i)$ obtained for the frequencies $f_i$ in the first band (black full line) with the TBM result (red dashed line). For display purposes we have replaced the $\delta$ functions by Lorentzians of finite width of $\Gamma=2$ MHz. The good agreement reassures the applicability of the TBM and thus justifies its use for further numerical studies with larger fullerene molecules with the parameters determined from the experiment.
![(Color online) Panel a): Resonance density determined from the resonance frequencies of the first band (middle panel in Fig. \[fig3\]) of the fullerene billiard (black solid line) compared to the TBM calculations (red dashed line). Panel b): Resonance densities computed within the TBM for C$_{60}$, C$_{240}$, C$_{540}$ and C$_{720}$ (lowermost to upmost curve). Panel c): Experimentally determined resonance density of a rectangular Dirac billiard [@Dietz2013; @Dietz2015].[]{data-label="fig4"}](fig4){width="\linewidth"}
Panel b) of Fig. \[fig4\] shows a comparison between the resonance densities of the C$_{60}$, C$_{240}$, C$_{540}$ and C$_{720}$ molecules in ascending order. Here, we used Lorentzians of width $\Gamma=5$ MHz for all cases. As stated above, all molecules contain the same number of pentagons whereas the number of hexagons increases. The resonance densities should thus resemble more and more that of a graphene sheet. They exhibit a minimum bounded by two increasingly sharp peaks, that evolve into van Hove singularities in the limit of an infinite number of atoms [@VanHove1953; @Dietz2013]. A comparison of the resonance densities with that of a rectangular graphene sheet with periodic boundary conditions (panel c) of Fig. \[fig4\]), which has been obtained in measurements with a microwave Dirac billiard [@Dietz2015], shows that they resemble for large fullerenes. In contradistinction to the latter, however, the resonance densities of the fullerenes all exhibit a peak of similar size located at the minimum which is due to the two triplets of zero modes. This remains true in the limit of an infinite number of atoms [@Gonzalez1992; @Gonzalez1993; @Kolesnikov2006; @Vozmediano2010] and is thus a distinct feature of the spatial curvature and topology. According to the Atiyah-Singer index theorem a plane graphene sheet with periodic boundary conditions should not exhibit any zero modes. This is observed in panel c). Zero modes are expected only, if part of the graphene sheet is terminated with zigzag edges [@Wurm2011]. The associated states are called edge states, because their wave functions vanish everywhere except at these edges [@Kuhl2010; @Bittner2012; @Dietz2015]. We have also computed the wave functions of the fullerene molecules under consideration using the TBM and found, that those of the 6 zero modes are localized at the pentagons, which may be considered to be equivalent to zigzag edges within the hexagon network.
The central frequencies of the two triplets, marked by squares for the one close to the Dirac frequency (dotted line) and by a circle for the one further away are displayed in Fig. \[fig5\] as a function of the number $n$ of C atoms. As is clearly visible, the distance between the triplets decreases with increasing size of the fullerene molecule and both approach the Dirac frequency. This behavior is well fitted by a function $f(n)= f_D+a/n^b$ yielding the parameter values given in the caption of Fig. \[fig5\], and for the Dirac frequency finally a value of $f_D=8.504$ GHz. Note that this is essentially the only way to determine the Dirac frequency of a C$_{60}$ molecule.
![(Color online) Calculated frequencies of the zero modes of the fullerenes C$_{60}$, C$_{180}$, C$_{240}$, C$_{540}$ and C$_{720}$ vs. the number $n$ of C atoms. All pairs of triplets (squares and circles) are located slightly above the Dirac frequency (dotted line). The experimental values are marked by a red filled circle and square. The dashed lines correspond to fits of the function $f(n)=f_D+a/n^b$ to the data points yielding for the Dirac frequency $f_D=8.504$ GHz, $a=1.188$ GHz, $b=0.661$ for the lower zero modes and $a=1.211$ GHz, $b=0.869$ for the upper ones.[]{data-label="fig5"}](fig5){width="\linewidth"}
[*Conclusions.*]{}— The lowest 60 eigenvalues of a C$_{60}$ fullerene were determined in high-precision experiments using a superconducting microwave billiard of corresponding shape. They appear in 15 groups of nearly degenerate ones, where the multiplicity coincides with that determined based on the group theory of the truncated icosahedral structure of C$_{60}$. We have demonstrated in TBM calculations for spherical fullerene molecules of increasing size that the two triplets of resonances detected close to the Dirac frequency correspond to the triplets of zero modes predicted by the Atiyah-Singer index theorem, and thus provided to the best of our knowledge the first experimental test of it. The exact value of the Dirac frequency was obtained as the asymptotic value attained by the frequencies of the triplets in the limit of an infinite number of atoms.
This work was supported by the DFG within the Collaborative Research Center 634.
[36]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , , , ****, ().
, , , , , , , ****, ().
, , , , , ****, ().
, , , , , , ****, ().
, , , , , , , ****, ().
, , , , ****, ().
, , , , , , ****, ().
, , , , , , ****, ().
, , , , ****, ().
, , , , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ** (, ).
, , , ****, ().
, ****, ().
, in **, edited by , , , (, , ), vol. , p. .
, ** (, , ).
, ** (, , ).
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , .
, , , ** (, , ).
, , , , , , , ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'I. Brott'
- '[S. E. de Mink]{}'
- '[M. Cantiello]{}'
- '[N. Langer]{}'
- '[A. de Koter]{}'
- '[C. J. Evans]{}'
- '[I. Hunter]{}'
- '[C. Trundle]{}'
- '[J.S. Vink]{}'
bibliography:
- '../../literatur.bib'
date: 'Received: Accepted: '
title: 'Rotating Massive Main-Sequence Stars I: Grids of Evolutionary Models and Isochrones'
---
Introduction
============
Massive stars can be considered as cosmic engines. With their high luminosities, strong stellar winds and violent deaths they drive the evolution of galaxies throughout the history of the universe. Even before galaxies formed, massive stars are believed to have played an essential role in re-ionizing the universe [@Haiman97], with important consequences for its subsequent evolution. Massive stars are visible out to large distances in the nearby universe [@Kudritzki08], and as ensembles even in star-forming galaxies at high redshift [e.g., @Douglas09]. Their explosions as supernovae or gamma-ray bursts shine through a major fraction of the universe, probing intervening structures as well as massive star evolution at the lowest metallicities [@Savaglio06; @Ohkubo09]. It is therefore of paramount importance for many areas within astrophysics to obtain accurate models of massive star evolution. In this paper we present an extensive grid of evolutionary models of rotating massive stars focusing on the main-sequence stage, which is at the root of understanding their subsequent evolution.
Basic concepts about massive main-sequence stars are long established and provide a good guide to their expected properties, most importantly the mass–luminosity relation [@Mitalas84; @Gonzalez05]. However, two major issues have plagued evolutionary models of massive stars until today: mixing and mass loss [@Chiosi86]. We concentrate here on the role of mixing in massive stars as, on the main sequence, the effects of mass loss remain limited in the considered mass and metallicity range.
Concerning thermally driven mixing processes, the occurrence of semiconvection and the related question of the appropriate choice of the criterion for convection do not alter the main sequence evolution of massive stars significantly [@Langer85]. However, the efficiency of convective overshooting in massive main-sequence stars is still not well known. In particular, the cool edge of the main-sequence band is not well determined, as stars in observational samples are found in the gap predicted between main-sequence stars and blue core-helium burners [e.g. @Fitzpatrick90; @Evans06_lmc_smc]. Here we follow a promising approach to determine the cool edge of the main sequence band, and thus the overshooting efficiency, from the rotational properties of B-type stars [see @Hunter08_vrot; @Vink10].
Rotationally-induced mixing processes have been invoked in the past decade to explain the surface enrichment of some massive main-sequence stars with the products from hydrogen burning, in particular nitrogen [@Heger00; @Meynet00; @Maeder00]. Rotationally-induced mixing was further used to explain the ratio between O-stars and various types of Wolf-Rayet stars at different metallicities [@MeynetMaeder05_wolfrayetpopulations], the variety of core collapse supernovae [@Georgy09] and the evolution towards long gamma-ray bursts [@Yoon05]. However, in view of the fact that all the mentioned observed phenomena have alternative explanations, a direct test of rotational mixing in massive stars appears of paramount importance.
In view of the recent results from the FLAMES Survey of Massive Stars [@Evans05_gal], which provided spectroscopic data for large samples of massive main sequence stars, we pursued the strategy of computing dense model grids of rotating massive main sequence stars, and predict in detail and comprehensively their observable properties. This evolutionary model data [^1], which was calibrated using results from the FLAMES Survey, is tailored to provide the first direct and quantitative test of rotational mixing in rapidly rotating stars[@Hunter08_letter; @Hunter09_nitrogen; @Brott10_popsyn from here on Paper II].
Our new model grids deliver predictions for many observables, with a wide range of potential applications. Besides helium and the CNO elements, we show how the light elements lithium, beryllium and boron are affected by rotational mixing, as well as the elements fluorine and sodium. We also provide isochrones which predict complex structures near the turn-off point of young star clusters. In Sec. \[sec:bec\], we describe the set-up and calibration of our massive star models, Sec. \[sec:evolgrid\] and \[sec:abund\] provide the evolutionary tracks in the HR diagram and the corresponding isochrones, and the surface abundances of all considered elements. We end with a brief summary in Sec. \[sec:summary\].
Stellar evolution code {#sec:evolution_models}
======================
\[sec:bec\]
We use a one-dimensional hydrodynamic stellar evolution code that takes into account the physics of rotation, magnetic fields and mass-loss. This code has been described extensively by @Heger00. Recent improvements are presented in @Petrovic05_GRB and @Yoon06.
To allow for deviations from spherical symmetry due to rotation in this one-dimensional code, we consider the stellar properties and structure equations on mass shells that correspond to isobars. According to e.g. @Zahn92 turbulence efficiently erases gradients along isobaric surfaces and enforces shellular rotation [@MeynetMaeder97] allowing us to use the one-dimensional approximation. The effect of the centrifugal acceleration on the stellar structure equations is considered according to @KippenhahnThomas70 and @EndalSofia76 as described in @Heger00.
Transport of chemicals and angular momentum
-------------------------------------------
All mixing processes are treated as diffusive processes. Convection is modeled using the Ledoux criterion, adopting a mixing-length parameter of $\alpha_{\rm{MLT}}=1.5$ [@BohmVitense58; @Langer91]. Semi-convection is treated as in @Langer+83 adopting an efficiency parameter $\alpha_{\rm{SEM}}=1$ [@Langer91]. We take into account convective core-overshooting using an overshooting parameter of $0.335$ pressure scale heights i.e. the radius of the convective core is equal to the radius given by the Ledoux criterium plus an extension equal to 0.335 Hp, where Hp is the pressure scale height evaluated at the formal boundary of the convective core (see Sec. \[sec:overshooting\] for the calibration). Furthermore, we consider various instabilities induced by rotation that result in mixing: Eddington-Sweet circulation, dynamical and secular shear instability, and the Goldreich-Schubert-Fricke instability [@Heger00].
Transport of angular momentum is also treated as a diffusive process following @EndalSofia78 and @Pinsonneault89 as described in @Heger00. The turbulent viscosity is determined as the sum of the diffusion coefficients for convection, semi-convection, and those resulting from rotationally induced instabilities. In addition, we take into account the transport of angular momentum by magnetic fields due to the Spruit-Tayler dynamo [@Spruit02], implemented as described in @Petrovic05_GRB. We do not consider possible transport of chemical elements as a result of the Spruit-Tayler dynamo because its validity is still controversial [@Spruit06]. We note that, while the dynamo process itself was confirmed through numerical calculations by @Braithwaite06, it is still unclear under which conditions it can operate and whether it is active in massive stars [@Zahn07]. Heuristically, angular momentum transport through magnetic fields produced by the Spruit-Tayler dynamo, as the so far only mechanism, appears to reproduce the rotation rates of white dwarfs and neutron stars quite well [@Heger05; @Suijs08].
In agreement with @MaederMeynet05 we find that magnetic fields keep the star near rigid rotation throughout its main sequence evolution, suppressing the mixing induced by shear between neighboring layers. The dominant rotationally induced mixing process in our models is the Eddington-Sweet circulation, a large scale meridional current that is caused by thermal imbalance in rotating stars between pole and equator [e.g. @Tassoul78].
Some of the diffusion coefficients describing rotational mixing are based on order of magnitude estimates of the relevant time and length scales [@Heger00]. To consider the uncertainties efficiency factors are introduced, which need to be calibrated against observational data. The contribution of the rotationally induced instabilities to the total diffusion coefficient is reduced by a factor $f_c =0.0228$ (see Sec. \[sec:rotmix\] for the calibration) while their full value enters in the expression for the turbulent viscosity [see @Heger00 for details]. The inhibiting effect of chemical gradients on the efficiency of rotational mixing processes is regulated by the parameter $f_\mu$, see @Heger00. We adopt $f_\mu=0.1$ after @Yoon06 who calibrated this parameter to match observed surface helium abundances.
Mass loss
---------
We updated the treatment of mass loss by stellar winds by implementing the prescription of @Vink00 [@Vink01], based on the method by @deKoter97, for winds from early O- and B-type stars. This mass loss recipe predicts a fast increase of the mass-loss rate as one moves to lower temperatures near $22\,000\,$K. This increase is related to the recombination of Fe[iv]{} to Fe[iii]{} at the sonic point and is commonly referred to as the bi-stability jump. These rates are derived for $12.5~\mathrm{kK} \la T_\mathrm{eff} \la 50~\mathrm{kK}$. We note that the formulae on the hot and cool side of the jump have [*not*]{} been derived for the intermediate temperature range between 22.5 and 27.5 kK. In this range, we thus perform a linear interpolation. In order to accommodate for a strong mass-loss increase when approaching the HD limit, we switch to the empirical mass loss rate of @Nieuwenhuijzen90, when the Vink et al rate becomes smaller than that from @Nieuwenhuijzen90 at any temperature lower than the critical temperature for the bi-stability jump. This ensures a smooth transition between the two mass loss prescriptions. This strategy also naturally accounts for the increased mass loss at the [*second*]{} bi-stability jump at $\sim$12.5 KK.
To account for the effects of surface enrichment on the stellar winds, we follow the approach of @Yoon06. The mass-loss rate of @Vink00 [@Vink01] is employed for stars with a surface helium mass fraction, $Y_\mathrm{s}$, below 0.4. We interpolate between the mass-loss rate of @Vink00 [@Vink01] and the Wolf-Rayet mass-loss rate of @Hamann95 reduced by a factor of 10 [@Yoon06] for $0.4 \le Y_\mathrm{s} \le 0.7$. When the helium surface abundance exceeds $Y_\mathrm{s} > 0.7$, we adopt the Wolf-Rayet mass-loss rate.
Rotational mass loss enhancement is included according to @Yoon05.
The initial chemical composition {#sec:chem}
--------------------------------
We adopt three different initial compositions that are suitable for comparison with OB stars in the Small and Large Magellanic Cloud (SMC, LMC) and a mixture that is tailored to the Galactic sample of the FLAMES survey of massive stars [@Evans05_gal], to which we will refer in this paper as the Galactic (GAL) mixture for brevity. In contrast with several previous studies we do not simply adopt Solar-scaled abundance ratios. Even though such mixtures may be sufficient to study the overall effects of metallicity, they are not accurate enough for direct comparison with observed surface abundances. For example, @Kurt98 find that the carbon to nitrogen ratio in HII regions in the LMC and SMC is considerably larger than the solar ratio. In the stellar interior, carbon is converted into nitrogen, which can reach the surface due to mixing and/or mass loss. @Brott08 found that the surface nitrogen abundance can be enhanced by a factor of 11 with a realistic initial SMC mixture. A Solar-scaled mixture with the same iron abundance would lead to enhancement of only a factor of 4, for the same initial parameters. This demonstrates the need for tailored initial chemical mixtures.
[C]{} [N]{} [O]{} [Mg]{} [Si]{} [Fe]{}
----- ------- ------- ------- -------- -------- --------
LMC 7.75 6.90 8.35 7.05 7.20 7.05
SMC 7.37 6.50 7.98 6.72 6.80 6.78
GAL 8.13 7.64 8.55 7.32 7.41 7.40
: Initial abundances for C, N, O, Mg, Si, Fe adopted in our chemical compositions for the Magellanic Clouds and the Galaxy (see Sec. \[sec:chem\]).
\[tab:compositions\]
For the initial abundances of C, N and O we use data from HII regions by @Kurt98. Surface abundances for these elements in B-type stars may no longer reflect the initial composition due to nuclear processing and mixing. For Mg and Si we use measurements of B-type stars as these stars are not expected to significantly alter the abundances of these elements during their main-sequence evolution. The B-type stars in different fields and clusters in the LMC show only small variations in Mg and Si of less than 0.02 dex [@Trundle07; @Hunter07_chem_Bstars; @Hunter09_nitrogen]. Also for the SMC these authors find no evidence for any systematic differences. The Fe abundance for the SMC is taken from a study of A-supergiants by @Venn99. For all the other elements we adopt solar abundances by @Asplund05 scaled down by 0.4 dex for the LMC mixture and 0.7 dex for the SMC mixture. The initial abundances are summarized in Tab. \[tab:compositions\]. These mixtures have metallicities of $Z = 0.0047$ for the LMC and $Z=0.0021$ for the SMC, where $Z$ is defined as the mass fraction of all elements heavier than helium (see Table \[tab:Z\]).
Choosing a chemical composition suitable for comparison with stars in the Milky Way is not straight forward due to the chemical gradients within the disk of our galaxy. We adopt an initial Mg and Si abundance based on the analysis of the Galactic cluster NGC6611 by @Hunter07_chem_Bstars. For C we adopted the NLTE corrected value from @Hunter07_chem_Bstars. N and O abundances are based on the average of HII-regions, see @Hunter08_vrot [@Hunter09_nitrogen] and references therein. The Fe abundance is taken from measurements of A-supergiants by @Venn95. For all other elements we have adopted solar abundances by @Asplund05. The resulting metallicity of our Galactic mixture is $\rm{Z}=0.0088$, which is lower than the solar metallicity of $\rm{Z}=0.012$ found by @Asplund05 and the solar neighborhood metallicity of $\rm{Z}=0.014$ measured by @Przybilla08. We point out we assumed the stellar wind mass loss rates as well as the opacities in our models to depend on the iron abundance, for which the spread in measurements is much smaller than for the total metallicity. Our wind mass loss rate scales with $(\rm{Fe}_{\rm{Surf}}/\rm{Fe}_{\odot})^{0.85}$, where we use the iron abundance from @Grevesse96 as solar reference, for consistency reasons with older models.
For helium we assume that the mass fraction scales linearly with Z between the primordial helium mass fraction of $\rm{Y}=0.2477$ [@Peimbert07] at $\rm{Z}=0$ and $\rm{Y}=0.28$ at the solar value [@Grevesse96]. We adopt the OPAL opacity tables [@opal] using $(\rm{Fe}_{Surf}/\rm{Fe}_{\odot})\times \rm{Z}_{\odot}$ to interpolate between tables of different metallicities. The solar reference values are again taken from @Grevesse96.
X Y Z
----- -------- -------- --------
LMC 0.7391 0.2562 0.0047
SMC 0.7464 0.2515 0.0021
GAL 0.7274 0.2638 0.0088
: Resulting hydrogen (X), helium (Y) and metal (Z) mass fractions for the chemical mixtures used in our models.
\[tab:Z\]
Calibration of overshooting {#sec:overshooting}
---------------------------
Mixing beyond the convective stellar core can occur when convective cells penetrate into the radiative region, when they “overshoot” the boundary due to their non-zero velocity. This effect is accounted for through a parameter $\alpha$ which measures the extension of the affected region in units of the local pressure scale height. In general, overshooting leads to larger stellar cores resulting in higher luminosities and larger stellar radii, especially towards the end of the main sequence evolution.
Various attempts have been undertaken to constrain the value of this parameter. Using eclipsing binary stars, @Schroeder97 derived values between 0.25 and 0.32 for stars in the mass range of 2.5 to 7[$\rm{M}_\odot$]{}. @Ribas00 and @claret07 found $\alpha$ in the range of 0.1 to 0.6, with a systematic increase of the amount of overshooting with the stellar mass. @Briquet07 used asteroseismological measurements and obtained $\alpha = 0.44\pm 0.07$ for the $\beta$-Cephei $\theta$ Ophiuchi. Another way to constrain the overshooting parameter is to fit the width of the main-sequence band for clusters with turn-off masses below approximately 15[$\rm{M}_\odot$]{}[@Mermilliod86]. Using non-rotating evolutionary models, @Mermilliod86 and @MaederMeynet87 found an overshooting parameter between 0.25 and 0.3. This is supported by findings of @Napiwotzki91. In general, however, this method does not work well for masses above 5...10[$\rm{M}_\odot$]{}, since the red edge of the main sequence band remains mostly unidentifiable [@Vink10].
In this work we re-calibrate the overshooting parameter for our models using data of the FLAMES survey of massive stars. We use the fact that the properties of stars near the end of their main-sequence evolution depend on their core size. For larger overshooting parameters, the end of the main sequence shifts to lower effective temperatures and lower surface gravities. After the end of their main sequence evolution, the rapid stellar expansion results in a strong spin down of the envelope. This phase of evolution occurs on the Kelvin-Helmholtz timescale, which is very short compared to the nuclear timescale. Observing stars during this phase is unlikely.
In Fig. \[fig:logg-vsini\] we plot the projected rotational velocity against the surface gravity for stars in the LMC sample of the FLAMES survey [adapted from @Hunter08_vrot]. Stars with surface gravities of $\log g > 3.2$ show a wide range of projected rotational velocities, ${\varv \sin i }$, in clear contrast with the stars with lower surface gravities which have projected velocities of about 50[$\rm{km/s}$]{}or lower. We interpret this transition as the division between main-sequence stars that are born with a wide range of rotational velocities and a different, more evolved, population that has experienced significant spin down. The surface gravity at which this transition occurs seems to be independent of mass, at least within the range available in this sample. To show this we color-coded the masses of the observed stars in Fig. \[fig:logg-vsini\].
In an alternate scenario discussed by@Vink10, the population of slow rotators might be main-sequence stars, that have spun down due to increased mass loss (bi-stability breaking). However, this would require a huge overshooting, in order to extend the lifetime of main sequence stars at temperatures below the bi-stability temperature of about 22kK. With the current overshooting calibration, bi-stability breaking plays a role only well above 30[$\rm{M}_\odot$]{}. The nature of the evolved slow rotators in Fig. \[fig:logg-vsini\] remains puzzling. The fact that all of them have a strong nitrogen surface enhancement might propose that they are post-red supergiants which acquired the nitrogen enhancement through the first dredge-up [@Vink10], but they might as well be products of binary interaction or of so far unidentified physical processes.
For the overshooting calibration we use a stellar model with a mass and rotation rate that are representative for the entire sample. The typical mass is about 16[$\rm{M}_\odot$]{} and the mean projected velocity of the stars in the sample with ${\log g }\geq 3.2 {\rm dex}$ is $\langle {\varv \sin i }\rangle =110\,{$\rm{km/s}$\xspace}$. This corresponds to an average rotational velocity of $\langle v \rangle=142\,{$\rm{km/s}$\xspace}$ assuming a random orientation of the spin axes, i.e. $\langle \sin i \rangle=\pi/4$. Throughout the main-sequence evolution the equatorial velocity is expected to be nearly constant due to the reduced effect of stellar winds in the low metallicity environment of the LMC and due to coupling of the expanding envelope to the contracting core. Therefore, we use the typical velocity as an estimate of the average initial rotational velocity.
The full black line in Fig. \[fig:logg-vsini\] shows a model with the typical mass and rotation rate. The overshooting parameter has been adjusted to $\alpha=0.335$, such that the end of the main sequence coincides with the drop in rotational velocities at ${\log g }=3.2$. Fig. \[fig:logg-vsini\] demonstrates that the adopted amount of overshooting, derived for a 16[$\rm{M}_\odot$]{}star, appears to be valid for the mass range 10 to 20 [$\rm{M}_\odot$]{}. However, it is unconstrained outside this range, as for example the 30 [$\rm{M}_\odot$]{}tracks in Fig. \[fig:logg-vsini\] have no observed counterparts near the TAMS. We therefore use an overshooting parameter of $\alpha=0.335$ for our entire grid. Given the uncertainty in the surface gravities ($\sim$0.15 dex), we estimate the uncertainty of the overshooting parameter in the mass range 10 to 20 [$\rm{M}_\odot$]{}to be $~0.1$.
A priori we do not expect the overshooting parameter to depend on metallicity. To check this we show the projected rotational velocities against surface gravity for the SMC and Galactic sample of the FLAMES survey in Fig. \[fig:logg-vsini-mw-smc\]. The drop in rotation rate occurs at the same surface gravity within the measurement errors. The Galactic sample shows a second feature around ${\log g }=3.5$ caused mainly by stars below 5[$\rm{M}_\odot$]{}. @Dufton06_vrot already described this stars as over-luminous for their age and a general disappointing agreement of the clusters HR-diagram with isochrones. When we vary the initial composition from SMC to the Galactic mixture, we find that the end of the main sequence for our 16[$\rm{M}_\odot$]{}model shifts from ${\log g }\sim 3.3 $ to 3.1, adopting the same value for the overshooting parameter. This shift is also within the measurement uncertainties. Therefore, adopting one fixed value for the overshooting parameter in our entire model grid is a reasonable assumption.
Calibration of the rotational mixing efficiency {#sec:rotmix}
-----------------------------------------------
As mentioned in the beginning of Sec. \[sec:bec\] we treat all mixing processes as diffusive processes, where the diffusion coefficient is taken as the sum of the individual diffusion coefficients for the various mixing processes. The contributions of all rotationally induced mixing processes are reduced by a factor, $f_c$, before they are added to the total diffusion coefficient for transport of chemicals. @Pinsonneault89 was the first to introduce this parameter. In order to explain the solar lithium abundance they needed to reduce the efficiency of rotationally induced mixing processes by a factor $f_c=0.046$. Based on theoretical considerations @ChaboyerZahn92 proposed $ f_c = 0.033$, which was adopted in the models of @Heger00. We note that we cannot compare our efficiency parameter directly with these values due to differences in the implementation and the considered specific set of mixing and angular momentum transport processes.
We calibrate the mixing efficiency for our models using the B-stars in the LMC sample of the FLAMES survey. We aim to reproduce the main trend of the observed nitrogen surface abundances with the projected rotational velocity for the nitrogen enriched fast rotators. In Fig. \[fig:hunterplot\] we present the data, indicating different regions or boxes which have been numbered for reference ([@Hunter08_letter] and ). The majority of the sample, about 53% of the stars, is not significantly enriched in nitrogen, given the typical error of 0.2 dex for the nitrogen abundance measurements. This group is indicated as Box 5 in Fig. \[fig:hunterplot\]. Fast rotators with significantly enhanced surface abundances ($> 7.2$ dex; see Box 3) constitute 18% of the sample. A detailed discussion of the stars in Box 1 and 2, which deviate from the main trend, and of observational biases and other factors that might influence this numbers are provided in and @Hunter08_letter.
For the calibration we use stellar models with a mass that is typical for the stars for which we have nitrogen surface abundance measurements, i.e. 13[$\rm{M}_\odot$]{}. We adjust the mixing efficiency such that the nitrogen surface abundances reached at the end of the main-sequence evolution follow the trend of the stars in Box 3 while we make sure that the slow rotators do not enrich more than 0.2 dex to match the stars in Box 5. We find that this is obtained for a mixing efficiency of $f_c$=0.0228. In Fig. \[fig:hunterplot\] we plot evolutionary tracks for several 13[$\rm{M}_\odot$]{}models with different initial rotation rates. Adopting this mixing efficiency, a model with the typical mass and average rotational velocity (${\varv_{\mathrm{ZAMS}} }=142\,{$\rm{km/s}$\xspace}$) reaches a surface nitrogen enhancement of 7.2 dex at the end of its main sequence. To illustrate the sensitivity of the nitrogen surface abundances to the adopted mixing efficiency, we depict the surface abundances reached at the end of the main sequence for $f_c = 0.015$ and $f_c=0.03$ (dotted lines and grey band). As we have no reason to believe that the mixing efficiency $f_c$ depends on metallicity, we adopt the same value for the SMC and the Galactic grid.
{width=".31\textwidth"}{width=".31\textwidth"}{width=".31\textwidth"}{width=".06\textwidth"}
Grid of evolutionary tracks and isochrones: initial parameters and description of provided data {#sec:init}
-----------------------------------------------------------------------------------------------
Using the calibration and the three initial compositional mixtures described above we computed a grid of evolutionary models. We adopt initial masses between 5 and 60[$\rm{M}_\odot$]{}and initial equatorial rotational velocities (${\varv_{\mathrm{ZAMS}} }$) ranging from 0 to about $600\,{$\rm{km/s}$\xspace}$. For each initial composition we computed about 200 evolutionary tracks. The spacing between the initial parameters was chosen to provide a proper coverage of the parameter space and to provide extra resolution for massive fast rotating stars whose evolution depends sensitively on the initial rotation rate (see Fig. \[fig:parameterspace\]).
The results of our calculations are made available online through VizieR in table format. For each evolutionary track we provide the basic stellar parameters at different ages such as the mass, effective temperature, luminosity, radius and mass-loss rate as well as the surface abundances of various elements. We also list the central and surface mass fractions of the most important isotopes of these elements. In Table \[tab:trackdata\] we summarized all the provided data. The time steps are chosen to resolve the evolutionary changes of the main stellar parameters. Tracks with higher time resolution are available from the authors upon request.
Based on these evolutionary tracks we generated sets of isochrones, using our population synthesis code STARMAKER . The isochrones cover the evolution until central hydrogen exhaustion for stellar masses between 5 and 60[$\rm{M}_\odot$]{}. We provide them for ages between 0 and 35Myr at intervals of 0.2Myr, for initial rotational velocities between 0 and 540[$\rm{km/s}$]{}with steps of 10[$\rm{km/s}$]{}and for three different initial compositions: SMC, LMC and a Galactic mixture. We provide the basic stellar parameters and surface abundances as listed in Tab. \[tab:aiso\_data\].
------------------------------------------------------------- ------- ----------- ----------- -----------
stellar surf.
parameters abun
Age (t \[yrs\]) H $^1$H
Mass (m \[[$\rm{M}_\odot$]{}\]) He $^3$He $^4$He
Eff. temp. (${\rm{T}_{\rm{eff}}}$\[K\]) Li $^7$Li
Luminosity ($\log \rm(L/L_{\odot})$) Be $^9$Be
Radius (R \[$\rm{R_{\odot}}$\]) B $^{10}$B $^{11}$B
Mass loss rate (M \[[$\rm{M}_\odot$]{}/yr\]) C $^{12}$C $^{13}$C
Surface gravity (${\log g }$) N $^{14}$N $^{15}$N
Surface velocity ($\varv_{\rm{surf}}$ \[[$\rm{km/s}$]{}\]) O $^{16}$O $^{17}$O $^{18}$O
Rotation period (P \[days\]) F $^{19}$F
Critical velocity ($\varv_{\rm{crit}}$ \[[$\rm{km/s}$]{}\]) Ne $^{20}$Ne $^{21}$Ne $^{22}$Ne
Eddington factor ($\Gamma_{\rm e}$) Na $^{23}$Na
Mg $^{24}$Mg $^{25}$Mg $^{26}$Mg
Al $^{26}$Al $^{27}$Al
Si $^{28}$Si $^{29}$Si $^{30}$Si
Fe $^{56}$Fe
------------------------------------------------------------- ------- ----------- ----------- -----------
: Description of the online tables containing the results of our evolutionary calculations as a function of time, for given initial composition, initial mass and initial rotation rate.
\[tab:trackdata\]
------------------------------------------------------------- --------------------------
stellar surf.
parameters abun
Age (t \[yrs\]) H
Mass (m \[[$\rm{M}_\odot$]{}\]) He
Effective temperature (${\rm{T}_{\rm{eff}}}$\[K\]) B
Luminosity ($\log \rm(L/L_{\odot})$) C
Radius (R \[$\rm{R_{\odot}}$\]) N
Mass loss rate (M \[[$\rm{M}_\odot$]{}/yr\]) O
Surface gravity (${\log g }$) Ne
Surface velocity ($\varv_{\rm{surf}}$ \[[$\rm{km/s}$]{}\]) Na
Rotation period (P \[days\]) Mg
Critical velocity ($\varv_{\rm{crit}}$ \[[$\rm{km/s}$]{}\]) Si
Eddington factor ($\Gamma_{\rm e}$) $^{1}$H (mass fraction)
initial mass and velocity $^{4}$He (mass fraction)
------------------------------------------------------------- --------------------------
: Description of the online tables containing stellar parameters along isochrones computed for different initial rotational velocities as provided by our population synthesis code . See also Tab. \[tab:trackdata\].
\[tab:aiso\_data\]
Evolutionary tracks and isochrones {#sec:models_hrd}
==================================
\[sec:evolgrid\]
Evolutionary tracks
-------------------
Rotation affects evolutionary tracks in the Hertzsprung-Russell diagram [e.g. @MaederMeynet2000_Review and references therein]. The centrifugal acceleration reduces the effective gravity resulting in cooler and slightly less luminous stars. However, rotation also induces internal mixing processes which can have the opposite effect, leading to more luminous and hotter stars. Which of these effects dominates depends on the initial mass, rotation rate, metallicity and the evolutionary stage. In this section we describe the effects of rotation on our models.
In rotating stars, the radiative energy flux depends on the local effective gravity [@vonZ24], and temperature and luminosity become latitude dependent. The resulting thermal imbalance drives large scale meridional currents. In most stars, rotating at moderate rotation rates, the expected latitude dependence of temperature and luminosity is week [@Hunter09_nitrogen]. In models close to critical rotation, luminosity and temperature differences are expected to be more significant and may even give rise to polar winds and equatorial outflows caused by critical rotation [@Maeder99]. Temperatures, luminosities and gravities given for the models presented in this paper are surface averaged values.
#### Mass dependence of rotational mixing:
For intermediate mass main-sequence stars the main effect of rotation is a reduction of the effective gravity by centrifugal acceleration. Fig. \[fig:hrd\_tracks\_mvar\_vconst\] shows that, during the main-sequence evolution, the tracks of the fast rotating 5 $M_\odot$ stars are about 2000K cooler compared to the nonrotating counterparts. Rotational mixing does affect the surface abundances of those elements that are so fragile that the relatively low temperatures in the stellar envelope are sufficient to induce nuclear reactions on them. However, the mixing is not efficient enough to significantly alter the structure of the star. As a consequence, the effect of rotation on the corresponding evolutionary tracks in the Hertzsprung-Russell diagram remains limited.
In rapid rotators more massive than about 15[$\rm{M}_\odot$]{}, the effect of rotational mixing becomes more important than that of the reduced effective gravity: towards the cool edge of the main sequence band, the rotating stars become more luminous than the nonrotating ones. The main effect of rotation is an effective increase of the size of the stellar core, similar to the effect of overshooting (Sec. \[sec:overshooting\]). This can be seen in Fig. \[fig:hrd\_tracks\_mvar\_vconst\] when one compares the tracks of the 15 and 16[$\rm{M}_\odot$]{}models near the end of their main-sequence evolution. The larger core mass in the fast rotating models results in a higher luminosity. In the most massive stars at low metallicity ($ \gtrsim 16$[$\rm{M}_\odot$]{}for the SMC, $M \gtrsim 19$[$\rm{M}_\odot$]{}for the LMC), and for the extreme rotators shown in Fig. \[fig:hrd\_tracks\_mvar\_vconst\], mixing induced by rotation is so efficient that almost all the helium produced in the center is transported throughout the entire envelope of the star. During their main-sequence evolution they become brighter and hotter, evolving up- and bluewards in the Hertzsprung-Russell diagram. After all hydrogen in the center has been converted to helium, the star contracts, reaching effective temperatures of up to 100,000K. This type of evolution is referred to as “quasi-chemically homogeneous evolution” [@Maeder87; @Yoon05; @Woo06]. The blueward and redward evolution after the exhaustion of hydrogen in the center is analogous to the “hook” at the end of the main sequence seen in tracks of nonrotating stars. Note that all models have been computed beyond central hydrogen exhaustion, in most cases up to the ignition of helium. This allows us to later compare the main sequence surface abundances with those resulting from the first dredge-up in the red supergiant regime.
#### Metallicity dependence of rotational mixing:
Fig. \[fig:hrd\_tracks\_mconst\_vvar\] depicts evolutionary tracks of models with various initial rotation rates. The most striking feature in these diagrams is the bifurcation of the evolutionary tracks occurring at high masses and low metallicity, most clearly visible at SMC metallicity. Stars that rotate faster than a certain threshold are so efficiently mixed that they evolve almost chemically homogeneously. Stars that rotate slower than this threshold build a chemical gradient at the boundary between the convective core and the radiative envelope. This gradient itself has an inhibiting effect on the mixing processes, strongly reducing the transport of material from the core to the envelope. The minimum rotation rate required for chemically homogeneous evolution decreases with increasing mass [@Yoon06].
At high metallicity, rotational mixing is less efficient. In addition, mass and angular momentum loss due to stellar winds becomes important, slowing down the rotation rate and therefore the efficiency of the mixing processes. The fastest rotating stars at high metallicity initially evolve blue- and upward in the Hertzsprung-Russell diagram, see in the lower panel of Fig. \[fig:hrd\_tracks\_mconst\_vvar\]. However, the combined effects of spin down by a stellar wind and the build up of an internal chemical gradient reduces the efficiency of internal mixing processes. The star switches onto a redward evolutionary track, similar to that of a non rotating star. However, its larger core mass results in a higher luminosity compared to the slower rotating counterparts.
{width="33.00000%"}{width="33.00000%"}{width="33.00000%"}
Isochrones {#sec:aiso-hrd}
----------
While for a given metallicity, a classical isochrone can be represented by a single line in the Hertzsprung-Russell diagram, the isochrones of rotating stars span an area for a given age and initial composition. This is shown in Fig. \[fig:aiso\]. The isochrone constructed from nonrotating evolutionary models is plotted in red. The isochrone corresponding to the fastest initial rotational velocity for which the models do not yet follow a blueward evolution in the HRD is plotted in black. The isochrone with the slowest initial velocity that shows a clear bluewards evolution is shown in green. In blue and yellow we have selected two isochrones from the transition region that may help the reader to asses the sudden transition from classical to chemically homogeneous evolution.
For rotational velocities above 350[$\rm{km/s}$]{}, when the most massive stars undergo chemically homogeneous evolution, the isochrones deviate strongly from the nonrotating case. The maximum spread in effective temperature occurs around 4 Myr (second row in Fig. \[fig:aiso\]). At lower metallicity the isochrones split into more clearly separated branches. At SMC metallicity (e.g. the third panel in the first column in Fig. \[fig:aiso\]) the isochrone based on models rotating initially at 500[$\rm{km/s}$]{}moves straight to the blue. In contrast, the comparable LMC isochrone returns to the red for stars above $\sim$50[$\rm{M}_\odot$]{}. This behavior is directly related to the feature in the evolutionary LMC tracks in Fig. \[fig:hrd\_tracks\_mconst\_vvar\], see for example the 50[$\rm{M}_\odot$]{}track for 550[$\rm{km/s}$]{}. At Galactic metallicity the blueward evolution does not appear in the models. Nevertheless, the area spanned by the isochrones extends over a wide range of effective temperatures.
Between about 5 and 10 Myr the most massive nonrotating stars have evolved off the main sequence. However, at low metallicity, the most massive fast rotators, which undergo chemically homogeneous evolution are still in their main-sequence phase at this time, forming a blue straggler-like blue population (see the bottom panels of Figs.\[fig:aiso\]). If homogeneously evolving stars exist in nature they would most likely be found in low metallicity star clusters with ages between 5 and 10Myr.
Surface Abundances {#sec:tracks_surf_changes}
==================
\[sec:abund\]
Abundances as a function of time
--------------------------------
A direct observable consequence of rotationally induced mixing is the enrichment or depletion of certain elements in the atmospheres of main-sequence stars. In Fig. \[fig:track\_abundance\_changes1\_smc\]-\[fig:track\_abundance\_changes1\_mw\] we show the evolution of the surface abundance of various elements as a function time. The effects of rotational mixing are more pronounced at lower metallicity, at higher masses and for higher rotational velocity (see also Fig. \[fig:parameterspace\] and Sec. \[sec:evolgrid\]).
As is usual in observational work, we express the surface abundances relative to the abundance of hydrogen. When stars become significantly hydrogen depleted at the surface, using hydrogen as a reference element may not be the most logical choice. Changes in the abundance may partially reflect changes in the reference element hydrogen. We plot the abundances in red when these effects become important (i.e. when the helium mass fraction at the surface becomes larger than 40%) .
Most of our stellar models evolve to the red supergiant stage directly after the end of core hydrogen burning. This leads to a large vertical step in the surface abundances of many elements in Fig. \[fig:track\_abundance\_changes1\_smc\]-\[fig:track\_abundance\_changes1\_mw\], which is due to the convective dredge-up in the red supergiant stage. In the following, we discuss the evolution of the surface abundances of various groups of elements, focusing on the changes occurring over the course of the main-sequence evolution.
### Helium
Even though helium is the main product of hydrogen burning, the abundance of helium at the surface remains remarkably constant in most evolutionary tracks during the main-sequence phase. For the 12[$\rm{M}_\odot$]{}models the enhancement is less than 0.2 dex. Only the fast rotators of 30 and 60 [$\rm{M}_\odot$]{}show significant helium surface enrichments, especially at low metallicity (Fig. \[fig:track\_abundance\_changes1\_smc\]-\[fig:track\_abundance\_changes1\_mw\]). These behaviors can be understood as the combination of two effects. Firstly, the production of helium occurs on the nuclear timescale. This occurs slower than, for example, the production of nitrogen or destruction of Li. Secondly, with the production of helium a steep gradient in mean molecular weight is established at the boundary between the core and the envelope. Such gradients inhibit the efficiency of mixing processes and prevent the transport of helium to the surface. Significant amounts of helium can be transported to the surface in models were mixing processes are efficient enough to prevent the build-up of a chemical barrier between core and envelope.
### The fragile elements, Li, Be, B and F {#sec:LiBeB}
The elements Li, Be, B are destroyed by proton captures at temperatures higher than $2.5\times 10^{6}\,$K for lithium, $3.5\times 10^{6}\,$K for beryllium, $5 \times 10^{6}\,$K for boron [@McWilliam04 p. 123]. These elements can only survive in the outermost layers. In rotating stars the surface abundances of these elements decrease gradually as the outer layers are mixed with deeper layers in which these elements have been destroyed. The decrease is most rapid for the most fragile element, lithium. The surface abundance of fluorine behaves similarly. It can survive temperatures of up to about $20 \times 10^{6}$K. We note that the rates of the reactions in which fluorine is involved are quite uncertain [@Arnould99].
In addition, stellar winds can affect the surface abundances especially for the most massive stars at high metallicity. Due to mass loss deeper layers are revealed in which the fragile elements have been destroyed. For nonrotating stars a sudden drop in the surface abundances of the fragile elements is visible in Fig. \[fig:track\_abundance\_changes1\_smc\]-\[fig:track\_abundance\_changes1\_mw\], when mass loss has removed the layers in which these elements can survive. In rotating models the change is more gradual.
After the end of the main-sequence, dredge-up leads to a further decrease of the surface abundances of these elements. However, for some models with masses between 20 and 40[$\rm{M}_\odot$]{}we find that small amounts of Li and Be are produced in the hydrogen burning shell by the p-p chain. As a result of the interplay between a convective zone on top of the shell source and the convective envelope reaching into these layers after hydrogen exhaustion, we find that the surface abundances of these elements can be significantly increased in these models. Even though the possible production of these elements by massive stars has several interesting applications, the robustness of the predictions for these elements requires further investigation.
### Carbon, nitrogen, oxygen and sodium
Carbon, nitrogen and oxygen take part as catalysts in the conversion of hydrogen into helium. Although their sum remains roughly constant, nitrogen is produced when the cycle settles into equilibrium, at the expense of carbon and, somewhat later, oxygen. The CN-equilibrium is achieved very early in the evolution, before a strong mean molecular weight gradient has been established between the core and the envelope. Therefore, rotational mixing can transport nitrogen throughout the envelope to the stellar surface. While the surface abundance of nitrogen increases, carbon is depleted. Since full CNO-equilibrium is achieved only after a significant amount of hydrogen has been burnt in the core, changes in the oxygen surface abundance are only found in later stages in the more massive and faster rotating models.
Sodium is produced in the extension of the CNO-cycle, the NeNa-chain. The changes in the surface abundances of sodium resemble the changes is nitrogen, even though the enhancements are smaller and appear a little bit later at the stellar surface, see Fig. \[fig:track\_abundance\_changes1\_smc\]-\[fig:track\_abundance\_changes1\_mw\].
The relative increase of the nitrogen abundance depends on the initial amount of carbon available for conversion into nitrogen. These amounts are different for the different mixtures. The C/N ratio in the SMC and LMC composition are similar (7.4 and 7.1, respectively), while the ratio in our Galactic composition is smaller (3.1). In Fig. \[fig:nitrogen-time-z\] we show the evolution of the surface nitrogen abundance in models of 15 and 40[$\rm{M}_\odot$]{}with initial rotation rates of 0 and 270[$\rm{km/s}$]{}for SMC, LMC and Galactic composition. The initial abundance of nitrogen in the Magellanic clouds is much lower than in the Galactic mixture. Nevertheless, the rotating SMC and LMC models reach surface nitrogen abundances at the TAMS which can be higher than in the initial Galactic mixture. The steep rise at the end of the curves is due to the dredge-up when the star becomes a red supergiant. The TAMS is located at the base of the steep rise.
Surface abundances for selected ages {#sec:aiso-abun}
------------------------------------
Due to the effects of rotation, the age or mass of a star can no longer be determined uniquely from its position in the Hertzsprung-Russell diagram by fitting evolutionary tracks or isochrones. This holds especially for stars that are located near the zero-age main sequence. A young, unevolved, slowly rotating star may have the same effective temperature and luminosity as a less massive, fast rotating, evolved star. A determination of the surface composition and to some extent the projected rotational velocity may help to lift this degeneracy.
In the panels of Fig. \[fig:aiso\_abundance\_changes\_smc\]-\[fig:aiso\_abundance\_changes\_mw\] we show which abundance distributions are predicted at a given time, as function of the considered mass and rotational velocity. Models of more massive stars show more pronounced surfaces abundances changes for a given age. The kinks in the lines of 400[$\rm{km/s}$]{}at LMC and Galactic composition occur due to difficulties in the interpolation between stars that follow blueward and redward evolutionary tracks effects.
These plots also show the effect of rotation on the main sequence life time. For example, Fig. \[fig:aiso\_abundance\_changes\_mw\] shows that the lines representing models with initial rotation rates of 400 and 500[$\rm{km/s}$]{}predict that stars of 50 and 60[$\rm{M}_\odot$]{}are still present at 4Myr, whereas nonrotating models predict that they have left the main sequence. On a much smaller scale this effect is also visible if one compares lines for rotation rates between 0 and 300[$\rm{km/s}$]{}based on models that follow normal evolutionary tracks.
The steep drop in the boron abundance along the isochrones based on nonrotating models is related to the stellar wind mass loss. For example, at LMC composition, the drop occurs around 35[$\rm{M}_\odot$]{}, indicating that the winds efficiently remove the outer layers of stars of 35 [$\rm{M}_\odot$]{}and higher (see also Sec. \[sec:LiBeB\]).
Summary {#sec:summary}
=======
In this paper we have presented an extensive grid of models for rotating massive main sequence stars. We provide three sets of initial compositions that are suitable for comparison with early OB stars in the SMC, the LMC and the Galaxy. The models cover the main-sequence evolution and in most cases these have been computed up to helium ignition. In terms of overshooting and rotational mixing efficiency, our models have been calibrated against the FLAMES survey of massive stars. We are using a new method to calibrate the amount of overshooting that makes use of the observed drop in projected rotation rates for stars with surface gravities lower than $\log g = 3.2$ dex. Interpreting this drop as the end of the main sequence, we find an overshooting parameter of $0.34\pm 0.1$ pressure scale heights.
We have also presented a detailed set of isochrones based on models of rotating massive main sequence stars. Whereas classical isochrones can be represented by a single line in the Hertzsprung-Russell diagram, the isochrones of rotating stars span a wide range of effective temperatures at a given luminosity. Therefore, the mass and age of an observed star can no longer be determined uniquely from its location in the Hertzsprung-Russell diagram. We also provided detailed predictions for the changing surface abundances of rotating massive main sequence stars. While we believe that the data provided here can be useful to many future studies of massive stars, we make use of it in for undertaking a quantitative test of rotational mixing of massive stars in the LMC.
[^1]: Our model data is made available through VizieR at <http://cdsarc.u-strasbg.fr/cgi-bin/...>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that elastic currents that take into account variations of the tunnel transmitivity with voltage and a large ratio of majority to minority spin densities of states of the $s$ band, can account for the low voltage current anomalies observed in magnet-oxide-magnet junctions. The anomalies can be positive, negative or have a mixed form, depending of the position of the Fermi level in the $s$ band, in agreement with observations. Magnon contribution is negligible small to account for the sharp drop of the magnetoresistance with the voltage bias.'
address:
- |
Instituto de Física ‘Gleb Wataghin’,\
Universidade Estadual de Campinas (UNICAMP),\
C. P. 6165, Campinas 13.083-970 SP, Brazil\
and
- |
Laboratorio de Física de Sistemas Pequeños y Nanotecnología,\
Consejo Superior de Investigaciones Científicas (CSIC),\
Serrano 144, E-28006 Madrid, Spain
author:
- 'G. G. Cabrera[@address]'
- 'N. García'
date: 'November, 2000'
title: 'Low Voltage I-V Characteristics in Magnetic Tunnel Junctions'
---
Tunneling of electrons in metal-insulator-metal junctions is an old phenomenon studied from a long time ago[@old; @simmons]. However, it is only quite recently that spin-dependent tunneling between two ferromagnetic metals has been shown to produce the magnetoresistance effect observed in those systems[@mr; @zhang]. In $3d$ ferromagnets, most of the spin polarization comes from the $d$ bands, while tunneling currents are dominated by $s$ band contributions. This is so, because $d$ wave functions are more localized and their effective tunneling barrier is higher[@nico1]. For $Ni$, it has been estimated that the tunneling probability of the $s$ electrons is of the order of $100-1000$ times that of the $d$ electrons, thus leading to a positive spin polarization in $Ni$ field emission experiments[@meservey]. In the context of tunneling experiments, the large magnetoresistance effect (25-30 %) found in [@mr; @zhang] is puzzling, since it points to a large polarization of the $s$ band, with a ratio of the densities of states for majority $\left(
N^{(M)}(E)\right) $ and minority $\left( N^{(m)}(E)\right) $ electrons at the Fermi level $\left( E_{F}\right) $ of the order of $$N^{(M)}(E_{F})/N^{(m)}(E_{F})\approx 2.0-2.5\quad , \label{ratio}$$ in apparent contradiction with energy band calculations for ferromagnetic metals[@moruzzi].
In addition, a remarkable dependence of the junction conductance with the voltage bias $\left( V\right) $ has been observed at low voltages (of the order of a few hundred millivolts). As usual in magnetoresistance experiments, one compares the resistances for the cases where the magnetizations at the electrodes are anti-parallel (AP) and parallel (P). In several experiments reported in Ref.[@mr; @zhang], the junction resistance drops significantly with the applied voltage, with a peak at zero bias (called [*zero-bias anomaly)* ]{}that is more pronounced for the AP alignment. The effect is also temperature dependent, the peak being less sharp at room temperature. Finally, it is found that the junction magnetoresistance (JMR) has a large decrease with voltages, up to 60% at 0.5 V in some cases[@mr]. It has been argued that this effect can be attributed to the excitation of internal degrees of freedom by hot electrons (even at liquid He temperature). Scattering from surface magnons has been proposed as a mechanism to randomize the tunneling process and open the spin-flip channels that leads eventually to a sharp drop of the MR[@zhang]. However, this explanation is controversial, since magnon scattering cross sections are negligibly small to account for such a big drop of resistance and no spin-flip events have been observed in experiments with polarized injected electrons in tunnelling phenomena[@spinflip]. Also, the theory given in Ref. [@zhang] , and uses a perturbation scheme only valid for voltages smaller than $\sim 40$ $mV$, while the data extend to $\sim 400\ mV$.
In the present Letter, we show that the variations of the conductance with the voltage bias can be simply accounted for by the lowering of the barrier height with voltages, as given by the Simmons’ tunneling theory[@simmons]. The structure at zero bias is obtained, when one properly takes into account variations of the density of states with the bias at both magnetic electrodes. Assuming that the tunneling current comes from the $s$ band, we formulate a simple model with a parabolic dispersion (free-electron[* *]{}like). We obtain different behaviors for the zero-bias anomaly, whether the Fermi level is located near the bottom ([*peak*]{}) or top of the band ([*dip*]{}). Fitting with the experiments[@mr; @zhang] can only be obtained if one assumes a large spin polarization corresponding to relation (\[ratio\]).
In order to develope our calculation, one has to rewrite Simmons’ formulae with the conductance current written in the form $$J^{\left( C\right) }(V)=A\sum_{\sigma ,\mu }%
%TCIMACRO{\dint }
%BeginExpansion
\displaystyle \int %
%EndExpansion
\limits_{-\infty }^{\infty }\ dE\ T\left( E,\Delta s,\phi ,V\right) \
N_{L}^{(\sigma )}(E)N_{R}^{(\mu )}(E+V)\left[ f_{L}\left( E\right)
-f_{R}(E+V)\right] \ , \label{current}$$ where $T\left( E,\Delta s,\phi ,V\right) $ is the transmitivity through the barrier for energy $E$, parametrized with the mean barrier height $\phi $ and width $\Delta s$[@simmons], the index $C=P,AP$ refers to the magnetic configuration (parallel or anti-parallel), and $\ N_{L,R}$ and $%
f_{L,R}$ are the densities of states and the Fermi distributions for the left and right electrodes, respectively. In ferromagnets, one has to distinguish between [*majority* ]{}$(M)$ and [*minority* ]{}$(m)$[* *]{}spin bands and the super-indices in the densities of states and in the sum in expression (\[current\]) label the allowed processes for spin conserving tunneling, for both magnetic configuration, $P$ and $AP$. For parallel alignment, the factor of the densities of states that enter in (\[current\]) is $$N_{L}^{(m)}(E)N_{R}^{(m)}(E+V)+N_{L}^{(M)}(E)N_{R}^{(M)}(E+V)\ , \label{pc}$$ while for the anti-parallel configuration, where majority and minority are interchanged for the left and right electrodes, one has to consider $$N_{L}^{(m)}(E)N_{R}^{(M)}(E+V)+N_{L}^{(M)}(E)N_{R}^{(m)}(E+V)\ . \label{apc}$$ Concerning equation (\[current\]), several remarks are in order.
1. In his original treatment of the tunneling problem[@simmons], Simmons considers the case of very flat conduction bands for the metal electrodes and takes the densities of states as constants. However, for $s$ bands the density of states varies as the square root of the energy, and for magnetic junctions this cannot be neglected, especially near the band edges, where the variation is bigger. Zero-bias anomalies in normal non-magnetic metals has been previously reported, in cases where the structure of the density of state is important[@wyatt].
2. Expression (\[current\]) involves an integral over all energies, but states that are deep in the band are cut off exponentially by the tunneling probability. As a net result, the conductance is dominated by electrons that are near the Fermi level, and (\[current\]) approximately factorizes in the form $$J^{\left( C\right) }(V)\approx \left( \sum_{\sigma ,\mu =m,M}^{C}N^{(\sigma
)}(E_{F})N^{(\mu )}(E_{F}+V)\right) J^{\left( S\right) }(V)\
=D^{(C)}(E_{F},V)\ J^{\left( S\right) }(V) \label{factor}$$ where $J^{\left( S\right) }(V)$ is the Simmons’ tunneling current as a function of the voltage bias and $$D^{(C)}(E_{F},V)=\sum_{\sigma ,\mu =m,M}^{C}N^{(\sigma )}(E_{F})N^{(\mu
)}(E_{F}+V)\ . \label{density}$$ In (\[factor\]), we are assuming that both electrodes are made from the same ferromagnetic metal. The term $J^{\left( S\right) }(V)$ is the Simmons’ contribution, is spin independent and carries all the information concerning the tunneling barrier. As shown in [@simmons], it has no quadratic term in the voltage for small bias, and no zero-bias anomaly.
In Fig. 1, we show the variation with voltage of the Simmons resistance for typical barriers, with the resistance normalized at zero bias. A large variation is observed in all the examples, but the resistance has no peak or dip at zero voltage. Except for the structure at zero bias, the overall variation of the Simmons’ resistance is of the order of what is observed in experiments (or even may vary faster with voltage in some cases). Some experimental results are also shown for comparison.
Next, we introduce the factor $D^{(C)}(E_{F},V)$, defined in (\[density\]), in the conductance calculation. We model the density of states of the $s$ bands with a parabolic dependence (free-electron like) in the form $$N^{(\sigma )}(E)=\frac{\Omega }{4\pi ^{2}}\left( \frac{2m_{e}}{\hbar ^{2}}%
\right) ^{3/2}\sqrt{\pm \left( E-E_{\sigma }\right) },\qquad \sigma =m,M,$$ where $\Omega $ is the volume of the sample (electrode), $m_{e}$ is the electron mass, and the $\pm $ sign refers to the cases where we are in the bottom or in the top of the conduction band, respectively. In formulating the Stoner model within a naive band theory, $\left| E_{m}-E_{M}\right| $ should yield the exchange of the $s$ band. But Fermi surfaces of transition metals are very intricate, with contributions from electron and hole-like carriers and with different shapes for majority and minority spin sheets. In this context, $E_{m}$ and $E_{M}$ come from the band structure and $\Delta
E=\left| E_{m}-E_{M}\right| $ may be very different from the true exchange of the band.
To parametrized our results, and denoting by $E_{F}$ the Fermi energy, we define $$\begin{array}{l}
E_{F}^{M}\equiv \left| E_{F}-E_{M}\right| , \\
E_{F}^{m}\equiv \left| E_{F}-E_{m}\right| , \\
E_{F}^{M}\equiv \lambda \ E_{F}^{m},\quad \lambda >1,
\end{array}$$ which includes both cases, bottom and top of the band. The ratio of the densities of states at the Fermi level is given by $%
N_{L}^{(M)}(E_{F})/N_{L}^{(m)}(E_{F})=\sqrt{\lambda }$. Several possibilities can be realized, wether majority and minority carriers are electrons or holes. When both are electrons or holes, the factors $%
D^{(C)}(E_{F},V)$ can be expanded in series in $V$, yielding a linear term in $V$ that is responsible for the zero-bias anomaly: $$\begin{array}{l}
D_{\pm }^{(P)}(V)\approx \left( \left[ N^{(m)}(E_{F})\right] ^{2}+\left[
N^{(M)}(E_{F})\right] ^{2}\right) \left( 1\pm p^{(P)}\left| V\right| \right)
, \\
\\
D_{\pm }^{(AP)}(V)\approx \left( 2N^{(m)}(E_{F})N^{(M)}(E_{F})\right) \left(
1\pm p^{(AP)}\left| V\right| \right) ,
\end{array}$$ where the $\pm $ sign labels the bottom and top cases respectively, with the slopes of the linear terms given by $$\begin{array}{l}
p^{(P)}=%
%TCIMACRO{\dfrac{1}{E_{F}^{m}\left( 1+\lambda \right) }}
%BeginExpansion
{\displaystyle {1 \over E_{F}^{m}\left( 1+\lambda \right) }}%
%EndExpansion
\ , \\
\\
p^{(AP)}=%
%TCIMACRO{\dfrac{\lambda +1}{4\lambda E_{F}^{m}}}
%BeginExpansion
{\displaystyle {\lambda +1 \over 4\lambda E_{F}^{m}}}%
%EndExpansion
\ .
\end{array}$$ When we have a mixed case, [*i.e.* ]{}one of the spin is electron-like and the other hole-like, no linear term appears in $D^{(P)}(V)$. On the other hand, for $D^{(AP)}(V)$, the slope of the linear term is given by $$p^{(AP)}=\mp \left(
%TCIMACRO{\dfrac{\lambda -1}{4\lambda E_{F}^{m}}}
%BeginExpansion
{\displaystyle {\lambda -1 \over 4\lambda E_{F}^{m}}}%
%EndExpansion
\right) \ ,$$ where the $-$ ($+$) sign applies when the majority carriers are electrons (holes). In Fig. 2, we display results of our calculation for examples of typical barriers. The value of the magnetoresistance at zero bias was taken from Ref. [@zhang], with $$N_{L}^{(M)}(E_{F})/N_{L}^{(m)}(E_{F})=\sqrt{\lambda }\approx 2.2\quad .$$ In Fig. 2 [*a)*]{}, we show the case when the Fermi level is in the bottom of the $s$ band, with a linear decrease of the resistance with the voltage bias for both magnetic configurations ($AP$ and $P$). If the Fermi level is in the top of both spin bands, we initially get a linear increase of the resistance which, after some voltage value, is dominated by the Simmons’ term. This case is displayed in part [*c)*]{} of Fig. 2. In Fig. 2 [*b)*]{}, we display the situation where the majority band ($\uparrow $) is almost filled (holes) and the minority ($\downarrow $) is almost empty (electrons). The resistance for the $P$ setup, exhibits no linear term. In Fig. 2 [*a)*]{}, we also show experimental results taken from Ref. [@zhang]. We have not tried an optimum fitting with experiments, but it is clear that experimental results can only be explained assuming a large polarization of the $s$ band. Note that the insets in Fig. 2 [*a)*]{}-2 [*c)*]{} sketch the band configurations for both spins.
The change in tunnel resistance or magnetoresistance (MR) is given by $$\frac{\Delta R}{R}=\frac{R_{AP}-R_{P}}{R_{AP}}\quad , \label{mr}$$ where again, $AP$ and $P$ refer to the magnetic configuration of the ferromagnetic electrodes. This ratio, as it is evident from relation (\[factor\]), is almost independent of the Simmons’ term, not depending on details of the tunneling process. In Fig. 3[* A)*]{}, we display results of $\Delta R/R$ corresponding to the examples of Fig. 2. In [*B)*]{}, we take different experimental results found in the literature[@mr]. Note that when the Fermi level lies near the top of the band, there is an increase of the MR. Eventually, we may reach the minority spin band edge, with a vanishing density of states, for which $$R_{AP}\rightarrow \infty \quad .$$
Temperature $(T)$ effects can also be taken into account through relation (\[current\]), with the broadening of the Fermi distributions, but a rough estimation shows that the effect should be similar to that of an applied voltage $V\approx 2T$, with an effective lowering of the barrier height, a smaller resistance, and the softening of the zero-bias anomaly, in agreement with experiments.
From our calculations presented above the following conclusions are pertinent:
1. The overall variation of the tunnel current with voltage [@mr; @zhang] can be explained by elastic tunneling using the well known Simmons’ formula[@simmons] and is due to the lowering of the barrier by the applied voltage. This is at variance with the calculations in Ref. [@zhang], where they argue that this effect is negligible. Therefore, magnons are not needed to explain the experiments;
2. The anomalies in the currents and the magnetoresistances can be explained within this simple framework, provided that the ratio of majority spin to minority spin electrons is of the order of $2.2-2.5$, for the data of Ref.[@mr; @zhang]. If one is allowed to choose the adequate configuration of the s bands (see Fig.2), a maximum, a minimum or a mix of both can appear at the anomaly (as it has been observed in Ref.[@sharma]);
3. From band structures calculations [@moruzzi], it is not clear to us that the above polarization of the $s$ band can be justified. There may be other oxidation states inside the metal, at the interface, and in the oxide layer, that contribute to the polarization of the current;
4. Alternatively, it may also happen, as it has been suggested in Ref.[@nico1; @berko; @ivan], that the current is dominated by conduction paths that provide large values of magnetoresistance[@nico2] due to domain wall scattering[@wall], and then there is also contribution of $d$-electrons. In this case, the density of states will have mixed contributions from $s$ and $d$-electrons, with a variety of topologies in the MR[@nico3];
5. The main conclusion is that the magnetoresistance is a mapping of the spin up and down densities of states in the metals and the barrier and cannot be assigned only to the bulk ferromagnetic metals, and many mixing possibilities exist for explaining the physical measurements.
[**Acknowledgments.**]{} GGC acknowledges partial support from Brazilian FAPESP [*(Fundação de Amparo à Pesquisa do Estado de São Paulo)* ]{}and CNPq [*(Conselho Nacional de Desenvolvimento Científico e Tecnológico)*]{}.
Visiting scientist at [*Laboratorio de Física de Sistemas Pequeños y Nanotecnología*]{}, Consejo Superior de Investigaciones Científicas (CSIC), Madrid, Spain
R. Holm, J. Appl. Phys. [**22**]{}, 569 (1951);W. A. Harrison, Phys. Rev. [**123**]{}, 85 (1961); J. C. Fisher and I. Giaever, J. Appl. Phys. [**32**]{}, 172 (1961); M. Julliere, Phys. Lett. [**54A**]{}, 225 (1975).
J. G. Simmons, J. Appl. Phys. [**34**]{}, 1793 (1963); J. Phys. D: Appl. Phys. [**4**]{}, 613 (1971).
J. S. Moodera, L. R. Kinder, T. M. Wong, and R. Meservey, Phys. Rev. Lett. [**74**]{}, 3273 (1995); J. S. Moodera, J. Nowak, and R. J. M. van de Veerdonk, Phys. Rev. Lett. [**80**]{}, 2941 (1998).
S. Zhang, P. M. Levy, A. C. Marley, and S. S. P. Parkin, Phys. Rev. Lett. [**79**]{}, 3744 (1997). This paper uses perturbation theory in $\left( eV/\hbar \omega _{c}\right) $, where $\hbar \omega _{c}\approx
100\ meV$, is roughly the maximum magnon energy, and the calculation should be valid when this ratio is much smaller than one. However, the theory is extrapolated to values of $V\sim 400\ meV$, where it is clearly not valid. Also, for this range, the assumption made that the values of the tunneling probability are voltage independent is not feasible, as can be easily deduced from Simmons’ tunneling theory[@simmons] and as shown in Fig. 1 of this paper. No justification whatsoever is given for the extremely high magnon scattering cross section. This is just another adjustable parameter of the theory.
N. García, Appl. Phys. Lett. [**77**]{}, 1351 (2000).
R. Meservey and P. M. Tedrow, Phys. Rep. [**238**]{}, 173 (1994).
V. L. Moruzzi, J. F. Janak, and A. R. Williams, [*Calculated Electronic Properties of Metals*]{} (Pergamon Press, New York, 1978).
H. C. Siegmann, private communication.
A. F. G. Wyatt, Phys. Rev. Lett. [**13**]{}, 401 (1964).
M. Sharma, S. X. Wang, and J. H. Nickel, Phys. Rev. Lett. [**82**]{}, 616 (1999).
C. L. Platt, A. S. Katz, R.C. Dynes, and A. E. Berkowitz, Appl. Phys. Lett. [**75**]{}, 127 (1999)
B.J. Jönsson-Akerman, R. Escudero, C. Leighton, S. Kim, I. K. Schuller, Appl. Phys. Lett. [**77**]{}, 1870 (2000).
N. García, M. Muñoz, and Y.-W. Zhao, Phys. Rev. Lett. [**82**]{}, 2923 (1999).
G. G. Cabrera and L. M. Falicov, Phys. Stat. Sol. (b) [**61**]{}, 539 (1974); G. Tatara, Y.-W. Zhao, M. Muñoz, and N. García, Phys. Rev. Lett. [**83**]{}, 2030 (1999).
N. García, H. Rohrer, I. G. Saveliev, and Y.-W. Zhao, Phys. Rev. Lett. [**85**]{}, 3053 (2000); to be published.
[**FIGURE CAPTIONS**]{}
[**Fig. 1** ]{}Variation of the Simmons’ resistance with voltage for several tunnel barriers. Data is normalized at zero bias. Experimental results from [@zhang] are also shown (solid triangles) as a reference.
[**Fig. 2** ]{}Resistance as a function of the voltage bias for the two configurations of the magnetic electrodes and for different $s$ band structures (they are shown in the insets). Parameters for the tunneling barriers are given in each figure. Spin $\uparrow $ is taken as the majority band in all cases. As a reference, experimental results take from [@zhang] are shown in part [*a)*]{}, where a good agreement with our calculation is obtained.
[**Fig. 3** ]{}Magnetoresistance, as defined in (\[mr\]), for all the cases depicted in Fig. 2. Densities of states are adjusted at the zero bias value. In [*A)*]{},[* *]{}we compare with results from [@zhang], while part [*B)*]{} compares with Ref.[@mr].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we study using Destination Artificial Noise (DAN) besides Source Artificial Noise (SAN) to enhance physical layer secrecy with an outage probability based approach. It is assumed that all nodes in the network (i.e. source, destination and eavesdropper) are equipped with multiple antennas. In addition, the eavesdropper is passive and its channel state and location are unknown at the source and destination. In our proposed scheme, by optimized allocation of power to the SAN, DAN and data signal, a minimum value for the outage probability is guaranteed at the eavesdropper, and at the same time a certain level of signal to noise ratio (SNR) at the destination is ensured. Our simulation results show that using DAN along with SAN brings a significant enhancement in power consumption compared to methods that merely adopt SAN to achieve the same outage probability at the eavesdropper.'
author:
- Ali Rahmanpour
- 'Vahid T. Vakili'
- 'S. Mohammad Razavizadeh'
bibliography:
- 'Paper-1\_Refs\_New.bib'
title: |
Enhancement of Physical Layer Security\
Using Destination Artificial Noise\
Based on Outage Probability
---
Introduction
============
Due to weakness of traditional security methods which are based on using cryptography algorithms in upper layers, physical layer security attracts many attentions in the recent years. Physical Layer Security was first addressed by Wyner’s celebrated paper [@Wiretap:1975] in which a wiretap channel was studied and the notion of secrecy capacity was introduced. Since then, several studies have extended the Wyner’s work, such as [@Gaussian:1978] for Gaussian wiretap channels, [@SecureFading:2008] for fading channels and [@SecrecyMIMO:2011] for Multiple-Input Multiple-Output (MIMO) channels.
One practical method in this area is using beamforming techniques in combination with emission of an Artificial Noise (AN) at the data source side (Alice) to corrupt an eavesdropper’s reception during transmission of data to the destination (Bob) [@Guaranteeing:2008]. During past years several works have developed this idea that most of them aim to maximize secrecy capacity. In a different way, the authors in [@Outage:2012] proposed a probability based approach for employing AN to guarantee a minimum level of outage probability at an unknown Eve while satisfying a certain Quality of service (QoS) requirement at the Bob. It should be noted that all these works are based on exploiting Source Artificial Noise (SAN) that means sending a noise-like signal at the data-source side of a wireless link.
Recently, thanks to the development of full-duplex communications, one node can transmit and receive data signals at the same time and the same frequency band [@Co-Channel:2010], [@Breaking:2013]. By full-duplex communications, it is possible to adopt Destination Artificial Noise (DAN) at the destination along with the SAN at the source. This idea was first proposed in [@SecureCommunication:2012] and then in [@Improving:2013] for Single-Input Multiple-Output (SIMO) communication systems. In [@Securing:2013] the MIMO communication systems when only one of the receiver antennas is used for data reception has been studied, but in [@Application:2014] destination can allocate more antennas for receiving the data. Similar to SAN, most works on the DAN are also based on maximizing secrecy capacity.
In this paper unlike to the previous works, we adopt an outage probability based approach for the characterization of the physical layer security that uses both DAN and SAN. In addition, we propose an optimal power allocation method to the DAN, SAN and source information signal to ensure the security requirements. We also investigate the effect of considering a power constraint at the destination for the situation in which Bob can only cancel a certain amount of self-interference. It is shown how this power constraint increases the total required power for a certain level of security.
Our simulation results show that the proposed method brings a significant reduction in the total power consumption compared to the previous works where only SAN is used to ensure a certain outage probability at the Eve. We also represent the effect of Eve’s location on the performance of the proposed method. It is shown that in contrast to previous works, we are able to ensure secrecy requirements even where Eve is very close to the Bob.
This paper is organized as follow. After presenting system model in Section 2, the power allocation problem is proposed in Section 3. In Section 4, simulation results are presented for both constrained and unconstrained power scenarios at the destination. Finally in Section 5, the paper has been concluded.
*Notation*: Bold symbols in small and capital letter denote vectors and matrices. In addition $\boldsymbol{(.)}^{H}$ denotes the conjugate transpose and $\left \| \boldsymbol{.} \right \|$ is the norm operator.
SYSTEM MODEL
============
In this paper, we consider a network consisting of a data source (Alice), a destination (Bob) and an eavesdropper (Eve). It is assumed that the source and the destination are equipped with $N_{A}\geq2$ and $N_{B}\geq2$ antennas, respectively, while Eve is equipped with $N_{E}\geq1$ antenna. The channel gains between Alice and Bob is represented by a matrix $\boldsymbol{H}_{AB}$ of size $N_{A}\times N_{B}$ and it is assumed to be known to all nodes. In the other hand the channel gains between Alice and Eve is represented by a matrix $\boldsymbol{H}_{AE}$ of size $N_{A}\times N_{E}$ and the channel gains between Bob and Eve by a matrix $\boldsymbol{H}_{BE}$ of size $N_{B}\times N_{E}$ which are unknown for the legal nodes (Alice and Bob). All channel gains are modelled by independent zero-mean complex Gaussian random variables. The variances of the above channels are $\sigma_{H_{AB}}^{2}$, $\sigma_{H_{AE}}^{2}$ and $\sigma_{H_{BE}}^{2}$ for the channel between Alice and Bob, Alice and Eve and Bob and Eve, respectively. If $\boldsymbol{H}$ be a channel gains matrix (small scale fading effects), by considering the *path loss* effect, we can define $\hat{\boldsymbol{H}}$ as [@Goldsmith],[@Application:2014] $$\label{eq:Hhat}
\hat{\boldsymbol{H}}= (\lambda_{0} r^{-\kappa})^{1/2} \boldsymbol{H}$$ where $\lambda_{0}$ is a constant that is used for showing the power at a reference signal and it determined by empirical measurements. In addition $r$ is the distance between two nodes and $\kappa$ is the path loss exponent and depends on the propagation environment. The values of this parameter is typically between 2 and 6.
We assume that the Eve’s location is unknown to the legal nodes and has a uniform distribution in a circle with radius $r_{ab}$, where $r_{ab}$ is the distance between Alice and Bob. Also it is assumed that the legitimate users are located at the center of the area (Fig. \[fig:model\]).
![System model[]{data-label="fig:model"}](paper1model){width="\columnwidth"}
To enhance the security of the system, we assume that only one of the destination antennas is employed for receiving the data signal and the other antennas are assigned to the AN propagation (i.e. DAN). On the other hand, all source antennas could be used to transmit both precoded data signal and AN (i.e. SAN). To decrease the power consumption and enhance the performance at the Bob, we select its best antenna (i.e. the *j*th antenna) for receiving data from Alice. One approach for selecting this antenna is as follows: $$\label{i-arg}
j=\arg \max_{i} \left \{ \left \| \boldsymbol{H}^{i}_{AB} \right \| \right \}$$ where $\boldsymbol{H}^{i}_{AB}$ is the *i*th column of $\boldsymbol{H}_{AB}$ which is related to the *i*th antenna at the destination and $i=1,2,...,N_{B}$. Therefore the channel assigned to the data transmission between Alice and Bob is denoted by a vector $\boldsymbol{h}_{AB}$ of size $N_{A}$. In addition, the channel matrix between the remaining Bob’s antennas and Eve is presented by a matrix $\boldsymbol{\tilde{H}}_{BE}$ of size $(N_{B}-1)\times N_{E}$.
In addition assuming that Eve uses the selection combining method, the channel vector between Eve’s *k*th antenna and Alice is denoted by $\boldsymbol{h}_{AE}$ of size $N_{A}$, while $\boldsymbol{h}_{BE}$ of size $(N_{B}-1)$ denotes the channel between Bob’s remained antennas and Eve’s *k*th antenna.
Now let $\boldsymbol{x}$, $y_{B}$ and $\boldsymbol{y}_{E}$ be the precoded data signal transmitted by Alice, the received signal at Bob and the received signal at Eve, respectively. Therefore $$\label{y:received}
y_{B}=\hat{\boldsymbol{h}}^{H}_{AB}\boldsymbol{x}+n_{B}$$ $$\label{z:received}
\boldsymbol{y}_{E}=\hat{\boldsymbol{H}}^{H}_{AE} \boldsymbol{x} + \hat{\boldsymbol{\tilde{H}}}^{H}_{BE} \boldsymbol{\nu} + \boldsymbol{n}_{E}$$ where $\boldsymbol{\nu}$ is the DAN emitting by Bob and we assume that its self-interference can be completely canceled at the destination. Assuming there is no knowledge about the Eve’s Channel State Information (CSI), DAN power must be equally distributed among $(N_{B}-1)$ Bob’s antennas that are assigned to DAN propagation. Hence we can write $$\label{DAN}
\boldsymbol{\nu} = \sqrt{\frac{P}{N_{B}-1}}\sum_{\substack{i=1 \\ i\neq j}}^{N_{B}}\eta \boldsymbol{\psi}_{i}$$ where $P\geq 0$ is the power allocated to DAN, $\eta$ is a complex scalar with unit magnitude and uniform random phase, $\boldsymbol{\psi}_{i}$ is a vector of size $N_{B}$ whose *i*th element is equal to one and all other elements are equal to zero. $n_{B}$ and $\boldsymbol{n}_{E}$ of size $N_{E}$ are the Gaussian noise terms with zero mean and variances $\sigma_{B}^{2}$ and $\sigma_{E}^{2} \boldsymbol{I}$, respectively. $\boldsymbol{x}$ is the beamformed signal transmitted by Alice and could be written as $$\label{x-signal}
\boldsymbol{x}=\sqrt{\varphi \acute{P}}d\boldsymbol{t}+\sqrt{(1-\varphi ) \acute{P}} \boldsymbol{\eta}$$ where $0\leq \varphi \leq 1$ is a parameter that determines the ratio of power assigned to the information signal and SAN, $\acute{P}>0$ is Alice’s total power, $d$ is the information symbol with $\mathbb{E} \left \{ \left | d \right |^{2} \right \}=1$ and $\boldsymbol{t}$ of size $(N_{A})$ is normalized $(\left \| \boldsymbol{t} \right \|=1)$ beamforming vector. $\boldsymbol{\eta}$ is SAN vector of size $(N_{A})$ that is orthogonal to $\boldsymbol{t}$ $(i.e., \boldsymbol{t}^{H} \boldsymbol{\eta} =0)$ and its covariance matrix is denoted by $\boldsymbol{C}_{\eta}$ as $$\label{covariance:SAN}
\boldsymbol{C}_{\eta}=\mathbb{E}\left \{ \boldsymbol{\eta} \boldsymbol{\eta}^{H} \right \}$$ Therefore we have $Tr\left \{ \boldsymbol{C}_{\eta} \right \}=1$. For the beamforming at the source node, as proposed in [@Physical-Layer:2011], if $\boldsymbol{t}_{i}$ is the *i*th eigenvector of $\boldsymbol{h}_{AB} \boldsymbol{h}_{AB}^{H}$ and $\boldsymbol{t}_{1}$ is assumed to be the principal eigenvector, it is assumed that $\boldsymbol{t}=\boldsymbol{t}_{1}$. Based on the orthogonality of the eigenvectors of $\boldsymbol{h}_{AB} \boldsymbol{h}_{AB}^{H}$, $\boldsymbol{\eta}$ is a linear combination of $N_{A}-1$ eigenvectors, and hence it lies in the nullspace of $\boldsymbol{h}_{AB}$. Since Eve’s CSI is not known to the legal nodes, the noise power equally is distributed to these eigenvectors as follows $$\label{SAN}
\boldsymbol{\eta} =\sqrt{\frac{1}{N_{A}-1}}\sum _{i=2}^{N_{A}}\eta \boldsymbol{t}_{i}.$$
According to the above beamforming model, the signal to noise ratios (SNRs) at the Bob and Eve’s *k*th antenna can be derived as $$\label{SNR:B}
\mathrm{SNR_{B}}=\frac{\varphi \acute{P} \Vert \hat{\boldsymbol{h}}_{AB} \Vert^{2}}{\sigma_{B}^{2}}$$ and $$\label{SNR:E}
\mathrm{SNR_{E}^{k}}=\frac{\varphi \acute{P} \hat{\boldsymbol{h}}_{AE}^{H} \boldsymbol{t}_{1} \boldsymbol{t}_{1}^{H} \hat{\boldsymbol{h}}_{AE} }{(1-\varphi ) \acute{P} \hat{\boldsymbol{h}}_{AE}^{H} \boldsymbol{C}_{\eta } \hat{\boldsymbol{h}}_{AE}+\frac{P}{N_{B}-1} \hat{\boldsymbol{h}}_{BE}^{H} \hat{\boldsymbol{h}}_{BE}}.$$
The Power Allocation Problem
============================
In this section, we discuss about the problem of power allocation to SAN and DAN. To ensure secrecy, $\gamma _{b}$ and $\gamma _{e}$ QoS constraints should be satisfied at Bob and Eve and hence the power optimization problem can be written as:
\[eq:opt1\] $$\begin{aligned}
\min_{P, \acute{P}, \varphi } \quad & P+\acute{P} \\
\label{eq:opt1-2}
\mathrm{s.t.} \quad & \mathrm{SNR_{B}} \ge \gamma _{b} \\
\label{eq:opt1-3}
& \mathbb{P} \lbrack \mathrm{SNR_{E}} \le \gamma _{e}\rbrack \ge \beta.\end{aligned}$$
In above optimization problem, (\[eq:opt1-2\]) is for satisfying a certain level of SNR at Bob and (\[eq:opt1-3\]) guarantees a minimum value of outage probability at Eve.
As it can be seen, our object is a joint power optimization for source and destination. Considering that Eve uses the selection combining method and the channel matrix coefficients are independent, we have $$\label{P:SNR_Ek}
\mathbb{P} \lbrack \mathrm{SNR_{E}} \le \gamma _{e}\rbrack = \prod_{k=1}^{N_{E}} \mathbb{P} \lbrack \mathrm{SNR_{E}^{k}} \le \gamma _{e}\rbrack \\
= \mathbb{P} \lbrack \mathrm{SNR_{E}^{k}} \le \gamma _{e}\rbrack^{N_{E}}$$ By substituting (\[SNR:B\]), (\[SNR:E\]) and (\[P:SNR\_Ek\]) into (\[eq:opt1\]) and considering (\[eq:Hhat\]), we have:
\[eq:opt2\] $$\begin{aligned}
\label{eq:opt2-1}
& \min _{P, \acute{P}, \varphi } \quad P+ \acute{P} \\
\label{eq:opt2-2}
& \mathrm{s.t.} \nonumber \\
& \quad \varphi \acute{P} \ge \gamma _{b}\sigma _{B}^{2} (\lambda_{0} r_{AB}^{-\kappa})^{-1} \Vert {\boldsymbol{h}}_{AB}\Vert^{-2} \\
\label{eq:opt2-3}
& \mathbb{P} \lbrack \frac{\varphi \acute{P} (\lambda_{0} r_{AE}^{-\kappa}) {\boldsymbol{h}}_{AE}^{H} \boldsymbol{t}_{1} \boldsymbol{t}_{1}^{H} {\boldsymbol{h}}_{AE} }{(1-\varphi ) \acute{P} (\lambda_{0} r_{AE}^{-\kappa}) {\boldsymbol{h}}_{AE}^{H} \boldsymbol{C}_{\eta } {\boldsymbol{h}}_{AE}+\frac {P} {N_{B}-1} (\lambda_{0} r_{BE}^{-\kappa}) {\boldsymbol{h}}_{BE}^{H} {\boldsymbol{h}}_{BE}} \nonumber \\
& \le \gamma _{e} \rbrack \ge \beta^{\frac{1}{N_{E}}}\end{aligned}$$
We can rewrite (\[eq:opt2-3\]) as $$\label{eq:opt3-3}
\mathbb{P} \lbrack \boldsymbol{h}^{H}_{AE} \boldsymbol{a} \boldsymbol{h}_{AE} + b \boldsymbol{h}_{BE}^{H} \boldsymbol{h}_{BE} \le \sigma _{E}^{2} \rbrack \ge \beta^{\frac{1}{N_{E}}}$$ where $\boldsymbol{a}$ and $b$ are defined as below
\[eq:AandB\] $$\begin{aligned}
\label{eq:a}
& \boldsymbol{a}= \acute{P} \lambda_{0} \bar{r}_{ae}^{-\kappa} (\varphi \gamma
_{e}^{-1} \boldsymbol{t}_{1} \boldsymbol{t}_{1}^{H} - (1-\varphi ) \boldsymbol{C}_{\eta }) \\
\label{eq:b}
& b= - \frac {P} {N_{B}-1} \lambda_{0} \bar{r}_{be}^{-\kappa} \end{aligned}$$
where $\bar{r}_{ae}=\mathbb{E}[r_{ae}]$ and $\bar{r}_{be}=\mathbb{E}[r_{be}]$. The left side in (\[eq:opt3-3\]) can be interpreted as a Cumulative Distribution Function (CDF). By defining $$\label{Xhah}
X=\boldsymbol{h}^{H}_{AE} \boldsymbol{a} \boldsymbol{h}_{AE}$$ $$\label{Yhh}
Y=\boldsymbol{h}_{BE}^{H} \boldsymbol{h}_{BE}$$ and $Z=X+bY$, we have \[eq:Fz\] $$\begin{aligned}
%\label{eq:Fz-1}
F_{Z}(z)&= \mathbb{P} \lbrack X+bY \le z \rbrack \nonumber \\
& = \mathbb{P} \lbrack \boldsymbol{h}^{H}_{AE} \boldsymbol{a} \boldsymbol{h}_{AE} + b \boldsymbol{h}_{BE}^{H} \boldsymbol{h}_{BE} \le z \rbrack \nonumber \\
\label{eq:Fz-3}
& =\int_{-\infty}^{+\infty}{f_{Y}(y)}dy \int_{-\infty}^{z-by} {f_{X}(x)}dx \end{aligned}$$ where $\int_{-\infty}^{z-by} {f_{X}(\tau)}d\tau=F_{X}(z-by)$ is the CDF of $X$. On the other hand $X$ is an indefinite Hermitian quadratic form for $x \ge 0$ [@distribution:2009; @Outage:2012], so $F_{X}(x)$ can be derived as [@distribution:2009]: $$\label{eq:Fx}
F_{X}(x)=u(x)+\frac{\alpha _{1}}{\vert \lambda _{1}\vert }e^{(\frac{-x}{\lambda _{1}})} u(\frac{x}{\lambda _{1}})$$ where $$\label{eq:alpha}
\alpha_{1}=-\frac{\lambda _{1}}{(1-\frac{\lambda _{2}}{\lambda_{1}})^{(N_{A}-1)}}$$ and
\[eq:lambda\] $$\begin{aligned}
\label{eq:lambda-1}
& \lambda _{1}=\varphi \acute{P}\lambda _{0} \bar{r}_{ae}^{-\kappa }\gamma_{e}^{-1}\sigma _{H_{AE}}^{2}\ge 0 \\
\label{eq:lambda-2}
& \lambda _{2}=-(1-\varphi )\acute{P}\lambda _{0} \bar{r}_{ae}^{-\kappa }\sigma_{H_{AE}}^{2}(N_{A}-1)^{-1}\le 0.\end{aligned}$$
In addition, $Y$ is sum of the squares of $2(N_{B}-1)$ independent normal random variables and has chi-squared ($\chi^{2}$) distribution as $$\label{eq:fY}
f_{Y}(y;2(N_{B}-1))=\frac{(\frac{y}{\sigma_{H_{BE}}^{2}})^{(N_{B}-2)}.e^{(\frac{-y}{2\sigma_{H_{BE}}^{2}})}}
{\sigma _{H_{BE}}^{2}2^{(N_{B}-1)}\Gamma (N_{B}-1)} u(y).$$
Therefore (\[eq:Fz-3\]) can be rewritten as $$\label{eq:Fz2}
F_{Z}(z)=\int_{0}^{+\infty }{f_{Y}(y) dy}\int_{-\infty}^{z-by}{f_{X}(x)}dx,$$ Considering $(z-by) \geq 0$, substituting (\[eq:Fx\]) and (\[eq:fY\]) into (\[eq:Fz2\]) and considering $\sigma_{H_{BE}}^{2}=1$, we have \[eq:Fz3\] $$\begin{aligned}
% \label{eq:Fz3-1}
F_{Z}(z)& =\int_{0}^{+\infty}{\frac{1}{2^{(N_{B}-1)}(N_{B}-2)!}y^{(N_{B}-2)}e^{-0.5y}dy} \nonumber \\
\label{eq:Fz3-2}
& +\int_{0}^{+\infty}{\frac{\alpha _{1}e^{(\frac{-z}{\lambda_{1}})}}{2^{(N_{B}-1)}(N_{B}-2)!\vert \lambda _{1}\vert}y^{(N_{B}-2)}e^{(\frac{b}{\lambda _{1}}-0.5)y}dy}\end{aligned}$$ After some manipulations, $F_{Z}(z)$ in (\[eq:Fz3-2\]) derived as $$\label{eq:Fz4}
F_{Z}(z)=(1+\frac{(-1)^{N_{B}}e^{(\frac{-z}{\lambda_{1}})}}{2^{(N_{B}-1)}(1-\frac{\lambda _{2}}{\lambda_{1}})^{N_{A}-1}(\frac{b}{\lambda _{1}}-0.5)^{(N_{B}-1)}}).$$
Substituting (\[eq:lambda\]) and (\[eq:Fz4\]) into (\[eq:opt2\]), the optimization problem in (\[eq:opt2\]) is simplified as
\[eq:opt4\] $$\begin{aligned}
\label{eq:opt4-1}
& \min _{P, \acute{P}, \varphi } \quad P+ \acute{P} \\
\label{eq:opt4-2}
& \mathrm{s.t.} \quad \varphi \acute{P} \ge \gamma _{b}\sigma _{B}^{2} (\lambda_{0} r_{AB}^{-\kappa})^{-1} \Vert {\boldsymbol{h}}_{AB}\Vert^{-2} \\
\label{eq:opt4-3}
& (1-\frac{2^{(1-N_{B})}e^{(\frac{-\gamma _{e}\sigma _{E}^{2}}{\varphi \acute{P}\lambda
_{0} \bar{r}_{ae}^{-\kappa }\sigma_{H_{AE}}^{2}})}}{(1+\frac{(1-\varphi )\gamma _{e}}{\varphi (N_{A}-1)})^{(N_{A}-1)}(\frac{P \bar{r}_{be}^{-\kappa }\gamma _{e}}{(N_{B}-1) \varphi \acute{P} \bar{r}_{ae}^{-\kappa }\sigma _{H_{AE}}^{2}}+0.5)^{(N_{B}-1)}}) \nonumber \ge \beta^{\frac{1}{N_{E}}}. \\\end{aligned}$$
According to (\[eq:opt4-3\]), it is obvious that the optimum value for (\[eq:opt4-2\]) is $\varphi \acute{P} = \gamma _{b}\sigma _{B}^{2} (\lambda_{0} r_{AB}^{-\kappa})^{-1} \Vert {\boldsymbol{h}}_{AB}\Vert^{-2}$, therefore the optimization method needs to only be applied to (\[eq:opt4-3\]). We use standard numerical methods to solve the problem.
Simulation and Numerical Results
================================
In this section, we evaluate the performance of our proposed method by computer simulations. In our simulations, it is assumed that $\sigma_{H_{AB}}^{2}=\sigma_{H_{AE}}^{2}=\sigma_{H_{BE}}^{2}=1$, $r_{AB}=2$ *Km*, $\gamma_{E}=0.5$ and $\sigma^{2}_{B}=\sigma^{2}_{N}=4 \times 10^{-14}$. For the path loss, it is assumed that $\lambda_{0}=10^{-1}$ and $\kappa=3$. Also the channel between Alice and Bob, $\boldsymbol{H}_{AB}$ is generated randomly with a distribution of $\boldsymbol{H}_{AB} \sim \mathcal{CN}(0,\sigma_{H_{AB}}^{2} \boldsymbol{I})$.
To represent how changes in the main parameters of our problem influence the security and power consumption, we first consider a scenario in which the destination has no power constraint for DAN. Then we also investigate the effect of a power constraint on DAN at the destination.
Case a) Destination has no power constraint for DAN {#case-a-destination-has-no-power-constraint-for-dan .unnumbered}
---------------------------------------------------
As mentioned before, we assume that the Eve’s location is distributed uniformly on the area and $\bar{r}_{ae}=\bar{r}_{be}=1000$. In addition we first assume that $N_{E}=N_{A}=N_{B}=4$. As it is seen in Fig. \[fig:beta\], in the case that both DAN and SAN are used, the required power for AN to achieve a given outage probability at the Eve is much less than the case that only SAN is used. In addition, in both cases, with increasing $\gamma_{B}$, more power is needed to guarantee the given outage probability at the Eve.
In Fig. \[fig:Ne\], we investigate the effect of different number of antennas at the Eve on the required power for the case $N_{A}=N_{B}=4$. It could be seen that when Eve has more capabilities, more power is needed to achieve a given outage probability. In Fig. \[fig:Nab\], the required power for AN is shown for different values of $N_{A}$ and $N_{B}$, when $N_{E}=4$.
In another scenario, we assume Alice and Bob are located at locations $(-1000,0)$ and $(1000,0)$, $N_{E}=N_{A}=N_{B}=4$, $\gamma_{B}=0.5$ and $\beta=0.6$. The total power should be allocated to AN to achieve a target outage probability at Eve $(\beta)$ while Eve moves from $(-1500,0)$ to $(15000,0)$ is shown in Fig. \[fig:distance\]. From this figure, it is seen that by using DAN we need less power to guarantee a similar $\beta$. Especially in the case that Eve is close to Bob, it is still possible to guarantee a given security.
![The required power for AN (total SAN and DAN) versus the target outage probability ($\beta$) at the Eve, for $N_{E}=4$, for $\gamma_{E}=0.4$ and $0.6$.[]{data-label="fig:beta"}](sim1){width="\columnwidth"}
![The required power for AN versus the number of antennas at Eve, for $\beta=0.6$ and $0.8$.[]{data-label="fig:Ne"}](sim1Ne){width="\columnwidth"}
![The required power for AN versus the target outage probability ($\beta$) at the Eve, for different sets of $N_{A}$ and $N_{B}$ when $N_{E}=4$.[]{data-label="fig:Nab"}](sim1Nab){width="\columnwidth"}
![The optimal power is needed for AN (total SAN and DAN) versus Eve’s location changes from $(-1500,0)$ to $(1500,0)$, for $\beta=0.6$[]{data-label="fig:distance"}](sim2){width="\columnwidth"}
Case b) Destination has a power constraint {#case-b-destination-has-a-power-constraint .unnumbered}
------------------------------------------
Although *Full Duplex* technique is advancing so fast and today it is possible to cancel even high levels of *Self-Interference* at the destination ([@Co-Channel:2010] and [@Breaking:2013]), but it is still useful to investigate how power consumption will be changed by limiting the maximum power allocated to the DAN. This is the case that is considered in Fig. \[fig:constraint\]. As we see in this figure, at the outage probability $\beta=0.9$ and for $N_{E}=N_{A}=N_{B}=4$, when DAN power limitation decreases from $2 mW$ to $0.5 mW$, total required Power for AN increases from $15 mW$ to $20 mW$.
![The optimal power is needed for AN (total SAN and DAN) versus target outage probability ($\beta$) at Eve, while there is a power constraint at the Destination[]{data-label="fig:constraint"}](sim3){width="\columnwidth"}
Conclusion
==========
In this paper, using destination artificial noise along with source artificial noise is introduced to guarantee an intended outage probability at Eve, while ensuring a certain SNR at Bob. For both constrained and unconstrained power allocation scenarios it has been shown how using the DAN, decreases the total power which is required to guarantee an intended outage probability at Eve. Using outage probability approach instead of secrecy capacity, makes our solution appropriate for the quasi-static channels.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We extend the characterization of the integrability of an almost complex structure $J$ on differentiable manifolds via the vanishing of the Frölicher-Nijenhuis bracket $[J, J]^{FN}$ to an analogous characterization of torsion-free $G_2$-structures and torsion-free ${\mbox{\rm Spin}(7)}$-structures. We also explain the Fernández-Gray classification of $G_2$-structures and the Fernández classification of ${\mbox{\rm Spin}(7)}$-structures in terms of the Frölicher-Nijenhuis bracket.'
address:
- 'Graduate School of Mathematical Sciences, University of Tokyo, 3-8-1, Komaba, Meguro, Tokyo 153-8914, Japan'
- 'Current address: Gakushuin University, 1-5-1, Mejiro, Toshima,Tokyo, 171-8588, Japan'
- 'Institute of Mathematics CAS , Zitna 25, 11567 Praha 1, Czech Republic'
- 'Fakultät für Mathematik, Technische Universität Dortmund, Vogelpothsweg 87, 44221 Dortmund, Germany'
author:
- Kotaro Kawai
- Hông Vân Lê
- Lorenz Schwachhöfer
title: 'The Frölicher-Nijenhuis bracket and the geometry of $G_2$-and ${\mbox{\rm Spin}(7)}$-manifolds'
---
[^1]
Introduction
============
A $G_2$-structure on a $7$-dimensional manifold $M^7$ is a $3$-form $\varphi \in {{\Omega}}^3(M^7)$ which at each point $p \in M^7$ is contained in a certain open subset of $\Lambda^3 T^\ast_p M^7$; similarly, a ${\mbox{\rm Spin}(7)}$-structure on an $8$-dimensional manifold $M^8$ is given by a $4$-form $\Phi \in {{\Omega}}^4(M^8)$ which at each point is contained in a certain subset of $\Lambda^4 T^\ast_p M^8$. Such structures induce both an orientation and a Riemannian metric on the underlying manifold, denoted by $g_\varphi$ and $g_\Phi$, respectively, and $\Phi$ is self-dual w.r.t. this metric.
Manifolds with $G_2$-structures have first been investigated by Fernández and Gray [@FG1982], and ${\mbox{\rm Spin}(7)}$-structures by Fernández [@Fernandez1986] who showed that the covariant derivatives $\nabla \varphi \in {{\Omega}}^1(M^7, \Lambda^3 T^\ast M^7)$ and $\nabla \Phi \in {{\Omega}}^1(M^8, \Lambda^4 T^\ast M^8)$, respectively, decompose into four irreducible components in case of $G_2$-structures and into two irreducible components in the case of ${\mbox{\rm Spin}(7)}$-structures. Thus, the conditions of the vanishing of some of these components yield $2^4 = 16$ classes of $G_2$-structure and $2^2 = 4$ classes of ${\mbox{\rm Spin}(7)}$-structures, respectively, and the underlying geometries were discussed in [@FG1982] and [@Fernandez1986]; see also Section \[subs:16classesG2\] below.
A $G_2$-structure (${\mbox{\rm Spin}(7)}$-structure, respectively) is called [*torsion-free*]{}, if $\varphi$ ($\Phi$, respectively) is parallel. As it turns out, the parallelity of $\varphi$ and $\Phi$, respectively, is equivalent to $\varphi$ and $\Phi$ being harmonic forms.
Alternatively, $G_2$- and ${\mbox{\rm Spin}(7)}$-structures may be characterized via certain [*($2$-fold or $3$-fold) cross products on the tangent bundle*]{}. These are given as the sections $$Cr_\varphi:= \delta_{g_\varphi} \varphi \in {{\Omega}}^2(M^7, TM^7), \quad
\chi_\varphi := -\delta_{g_\varphi} \ast \varphi \in {{\Omega}}^3(M^7, TM^7)$$ in case of $G_2$-structures, and as $$P_\Phi := -\delta_{g_\Phi} \Phi \in {{\Omega}}^3(M^8, TM^8),$$ where $\delta_g: {{\Omega}}^{k+1}(M) \to {{\Omega}}^k(M, TM)$ is the contraction of a differential form with the Riemannian metric $g$ and have natural interpretations via octonian multiplication. The triple cross product $\chi \in {{\Omega}}^3(M^7, TM^7)$ on a manifold with a $G_2$-structure has been introduced by Harvey-Lawson [@HL1982] and was used in many papers on deformation of associative submanifolds, see e.g.[@McLean1998], [@Kawai2014a], [@LV2016]. The $3$-fold cross product $P$ on ${{\mathbb R}}^8$ has been first explicitly constructed by Brown and Gray [@BG1967]. They also proved that (up to the $G_2$-action) there are exactly two non-equivalent $3$-fold cross products on ${{\mathbb R}}^8={{\mathbb O}}$. In [@HL1982] Harvey and Lawson intensively used the $3$-fold cross product on ${{\mathbb R}}^8$ which is related to the Cayley $4$-form and hence is invariant under the action of ${\mbox{\rm Spin}(7)}$. Fernandez showed the uniqueness of a ${\mbox{\rm Spin}(7)}$-invariant $4$-form on ${{\mathbb R}}^8$ (up to a multiplicative constant) and used the associated $3$-fold cross product to classify ${\mbox{\rm Spin}(7)}$-structures on $8$-manifolds [@Fernandez1986].
In this article, we view these cross products as elements of the [*Frölicher-Nijenhuis Lie algebra* ]{}${{\Omega}}^\ast(M, TM)$. Namely, it was shown by Frölicher-Nijenhuis in [@FN1956] that ${{\Omega}}^\ast(M, TM)$ can be given the structure of a graded Lie algebra using the [*Frölicher-Nijenhuis bracket $[\;,\; ]^{FN}$*]{} in a natural way. Thus, given a manifold with a $G_2$-structure $(M^7, \varphi)$, we may consider the Frölicher-Nijenhuis brackets $${}[Cr_\varphi, \chi_\varphi]^{FN} \in {{\Omega}}^5(M^7, TM^7), \quad [\chi_\varphi, \chi_\varphi]^{FN} \in {{\Omega}}^6(M^7, TM^7),$$ (observe that $[Cr, Cr]^{FN} = 0$ due to graded skew-symmetry), and analogously, for a manifold with a ${\mbox{\rm Spin}(7)}$-structure $(M^8, \Phi)$ we may consider $${} [P_\Phi, P_\Phi]^{FN} \in {{\Omega}}^6(M^8, TM^8).$$ These brackets may be regarded as a natural generalization of the Nijenhuis tensor of an almost complex structure $J$. Indeed, regarding such a structure as an element $J \in {{\Omega}}^1(M, TM)$, it turns out that $[J,J]^{FN} \in {{\Omega}}^2(M, TM)$ coincides – up to a constant multiple – with the Nijenhuis tensor of $J$, whence $J$ is integrable if and only if $[J,J]^{FN} = 0$ [@FN1956b].
Our main result is that the Frölicher-Nijenhuis bracket also characterizes the torsion-freeness of $G_2$- and ${\mbox{\rm Spin}(7)}$-structures, respectively. Namely, we show the following.
\[thm:Brackets-intro\] Let $(M^7, \varphi)$ be a manifold with a $G_2$-structure and the associated Riemannian metric $g = g_\varphi$, and let $\nabla$ be the Levi-Civita connection of $g$. Then for every $p \in M^7$ the following are equivalent.
1. The $G_2$-structure is torsion-free at $p$, i.e., $(\nabla \varphi)_p = 0$.
2. $[Cr_\varphi, \chi_\varphi]_p^{FN} = 0 \in \Lambda^5 T_p^*M^7 \otimes T_pM^7$.
3. $[\chi_\varphi, \chi_\varphi]_p^{FN} = 0 \in \Lambda^6 T_p^*M^7 \otimes T_pM^7$.
In fact, we show in Theorem \[thm:Brackets\] that $(\nabla \varphi)_p$ is characterized by either $[\chi_\varphi, \chi_\varphi]_p^{FN}$, or by the projection of $[Cr_\varphi, \chi_\varphi]_p^{FN}$ onto a subspace isomorphic to $T_pM^7 \otimes T_pM^7$.
\[thm:Brackets-Spin-intro\] Let $(M^8, \Phi)$ be a manifold with a ${\mbox{\rm Spin}(7)}$-structure, and let $\nabla$ be the Levi-Civita connection of the associated Riemannian metric $g = g_\Phi$. Then for every $p \in M^8$ the following are equivalent.
1. The ${\mbox{\rm Spin}(7)}$-structure is torsion-free at $p$, i.e., $(\nabla \Phi)_p = 0$.
2. $[P_\Phi, P_\Phi]_p^{FN} = 0 \in \Lambda^6 T_p^*M^8 \otimes T_pM^8$.
Namely, we show in Theorem \[thm:Brackets-Spin\] that $(\nabla \Phi)_p$ is characterized by the projection of $[P_\Phi, P_\Phi]_p^{FN}$ onto a subspace isomorphic to $W^7_p
\otimes T_pM^8$ for some rank-$7$ bundle $\Lambda^6 T^\ast M^8 \supset W^7 \to M^8$.
These explicit descriptions also allow us to give a complete characterization of the $16$ cases of $G_2$-structures in terms of $[Cr_\varphi, \chi_\varphi]_p^{FN}$ and $[\chi_\varphi, \chi_\varphi]_p^{FN}$, and of the $4$ classes of ${\mbox{\rm Spin}(7)}$-structures in terms of $[P_\Phi, P_\Phi]_p^{FN}$; cf. Section \[subs:16classesG2\].
Our paper is organized as follows. In Section \[sec:pre\] we recall the Frölicher-Nijenhuis bracket on ${{\Omega}}^\ast(M, TM)$. Then we turn to the case of $G_2$-structures in Section \[sec:g2\], characterizing the torsion endomorphism and showing the results that lead us to Theorem \[thm:Brackets-intro\]. In Section \[sec:spin7\], we repeat this discussion for the case of ${\mbox{\rm Spin}(7)}$-structures which leads to Theorem \[thm:Brackets-Spin-intro\]. Finally, the characterization of the 16 classes of $G_2$-structures and the $4$ classes of ${\mbox{\rm Spin}(7)}$-structures in terms of the Frölicher-Nijenhuis bracket is given in Section \[subs:16classesG2\]. The appendix then contains the proofs of some identities on representations of $G_2$ and ${\mbox{\rm Spin}(7)}$ which are used throughout the paper.
Preliminaries {#sec:pre}
=============
The Frölicher-Nijenhuis bracket {#subs:fnb}
-------------------------------
Let $M$ be a manifold and $({{\Omega}}^\ast(M), \wedge) = (\bigoplus_{k \geq 0} {{\Omega}}^k(M), \wedge)$ be the graded algebra of differential forms. We shall use superscripts to indicate the degree of a form, i.e., $\alpha^k$ denotes an element of ${{\Omega}}^k(M)$.
Evidently, contraction $\imath_X: {{\Omega}}^k(M) \to {{\Omega}}^{k-1}(M)$ with a vector field $X \in {{{\mathfrak {X}}}}(M)$ is a derivation of degree $-1$. More generally, for $K \in {{\Omega}}^k(M, TM)$ we define $\imath_K \alpha^l$ as the [*contraction of $K$ with $\alpha^l \in {{\Omega}}^l(M)$* ]{}pointwise by $$\imath_{\kappa^k \otimes X} \alpha^l := \kappa^k \wedge (\imath_X \alpha^l) \in {{\Omega}}^{k+l-1}(M),$$ where $\kappa^k \in {{\Omega}}^k(M)$ and $X \in {{{\mathfrak {X}}}}(M)$ is a vector field, and this is a derivation of ${{\Omega}}^*(M)$ of degree $l-1$. Thus, the [*Nijenhuis-Lie derivative along $K \in {{\Omega}}^k(M, TM)$* ]{}defined as $$\label{eq:LK-deriv}
{{\mathcal L}}_K (\alpha^l) := [\imath_K, d] (\alpha^l) = \imath_K (d\alpha^l) + (-1)^k d(\imath_K \alpha^l) \in {{\Omega}}^{k+l}(M)$$ is a derivation of ${{\Omega}}^*(M)$ of degree $k$. Observe that for $k = 0$ in which case $K \in {{\Omega}}^0(M, TM)$ is a vector field, both $\imath_K$ and ${{\mathcal L}}_K$ coincide with the standard notion of contraction with and Lie derivative along a vector field.
In [@FN1956] [@FN1956b], it was shown that ${{\Omega}}^*(M, TM)$ can be given a unique Lie algebra structure, called the [*Frölicher-Nijenhuis bracket*]{} and denoted by $[\cdot, \cdot]^{FN}$, such that ${{\mathcal L}}$ defines an action of ${{\Omega}}^*(M, TM)$ on ${{\Omega}}^*(M)$, that is, $$\label{eq:FN-homom}
{{\mathcal L}}_{[K_1, K_2]^{FN}} = [{{\mathcal L}}_{K_1}, {{\mathcal L}}_{K_2}] =: {{\mathcal L}}_{K_1} \circ {{\mathcal L}}_{K_2} - (-1)^{|K_1||K_2|} {{\mathcal L}}_{K_2} \circ {{\mathcal L}}_{K_1}.$$ It is given by the following formula for $\alpha^k \in {{\Omega}}^k(M)$, $\beta^l \in{{\Omega}}^l (M)$, $X_1, X_2 \in {{\mathfrak X}}(M)$ [@KMS1993 Theorem 8.7 (6), p. 70]: $$\begin{aligned}
\nonumber [\alpha^k \otimes X_1, &\beta^l \otimes X_2]^{FN} = \alpha^k \wedge \beta^l \otimes [ X_1, X_2]\nonumber \\
& + \alpha^k \wedge {{\mathcal L}}_{X_1} \beta^l \otimes X_2
- {{\mathcal L}}_{X_2} \alpha^k \wedge \beta^l \otimes X_1 \label{eq:kms}\\
&+ (-1)^{k} \left( d \alpha^k \wedge (\imath_{X_1} \beta^l) \otimes X_2
+ (\imath_{X_2} \alpha^k) \wedge d \beta^l \otimes X_1 \right).
\nonumber\end{aligned}$$ In particular, for a vector field $X \in {{\mathfrak X}}(M)$ and $K \in {{\Omega}}^*(M, TM)$ we have [@KMS1993 Theorem 8.16 (5), p. 75] $$\begin{aligned}
{{\mathcal L}}_X (K) = [X, K] ^{FN},\end{aligned}$$ that is, the Frölicher-Nijenhuis bracket with a vector field coincides with the Lie derivative of the tensor field $K \in {{\Omega}}^*(M, TM)$ which means that $\exp(tX): {{\Omega}}^*(M, TM) \to {{\Omega}}^*(M, TM)$ is the action induced by (local) diffeomorphisms of $M$.
\[ex:Lk-J\] Let $A \in {{\Omega}}^1(M, TM)$ be an endomorphism field on $M$. Then [@KMS1993 Remark 8.17, p. 75] $$[A, A]^{FN} = 2 [A, A]_N,$$ where $[A, A]_N$ is the Nijenhuis tensor of $A$. W.r.t. a local frame $(e_i)$ with dual frame $(e^i)$ we can write $A = e^i \otimes A e_i$, whence $$\begin{aligned}
\nonumber
{{\mathcal L}}_A \alpha^k &\stackrel{(\ref{eq:LK-deriv})}= e^i \wedge (\imath_{A e_i} d\alpha^k) - d(e^i \wedge (\imath_{A e_i} \alpha^k))\\
\label{eq:der-endo}
&= A \cdot d\alpha^k - d (A \cdot \alpha^k),\end{aligned}$$ where we denote by $\cdot$ the pointwise action of $A_p \in End(T_pM)$ on $\Lambda^k T_p^*M$. Observe that by (\[eq:FN-homom\]) we have ${{\mathcal L}}_{[A,A]^{FN}} = 2 ({{\mathcal L}}_A)^2$, so that the derivation ${{\mathcal L}}_A: {{\Omega}}^k(M) \to {{\Omega}}^{k+1}(M)$ is a differential iff $[A,A]_N = 0$.
For instance, if $A = Id$ then $I \cdot \alpha^k = e^i \wedge (\imath_{e_i} \alpha^k) = k \alpha^k$, so that $${{\mathcal L}}_I \alpha^k \stackrel{(\ref{eq:der-endo})}= I \cdot d\alpha^k - d (I \cdot \alpha^k) = (k+1) d\alpha^k - d(k\alpha^k) = d\alpha^k.$$ To see another example, let $A = J$ be an almost complex structure. Then $[J,J]^{FN} = 2 [J,J]_N = 0$ iff $J$ is integrable, and in this case one calculates from (\[eq:der-endo\]) that ${{\mathcal L}}_J = -d^c = i({{\partial}}- \bar{{{\partial}}})$ is the negative of the complex differential, where $d = {{\partial}}+ \bar{{{\partial}}}$ is the decomposition into the holomorphic and anti-holomorphic part of $d$. In particular, $H^*_J({{\Omega}}^*(M)) \cong H^*_{dR}(M)$ coincides with the deRham cohomology.
We end this section by providing a formula for the Frölicher-Nijenhuis bracket for those types of forms which we shall be concerned with. Recall from the introduction that on a Riemannian manifold $(M, g)$ we define the map $$\label{eq:def-partial}
\delta = \delta_g: \Lambda^k V^* \longrightarrow \Lambda^{k-1} V^* \otimes V, \qquad \delta_g(\alpha^k) := (\imath_{e_i} \alpha^k) \otimes (e^i)^\#,$$ taking the sum over some basis $(e_i)$ of $T_pM$ with dual basis $(e^i)$ of $T^\ast_pM$. This implies that to each $\Psi \in {{\Omega}}^{k+1}(M)$ we may associate a section $\delta_g(\Psi) \in {{\Omega}}^k(M, TM)$.
\[prop:FN\] Let $(M, g)$ be an $n$-dimensional Riemannian manifold of dimension $n$ and let $\Psi_l \in {{\Omega}}^{k_l+1}(M)$, $l = 1,2$. Moreover, let $$K_l := \delta_g(\Psi_l) \in {{\Omega}}^{k_l}(M, TM)$$ with the map $\delta_g$ from (\[eq:def-partial\]).
Then the Frölicher-Nijenhuis bracket at $p \in M$ is given as $$\begin{aligned}
{}[K_1, K_2&]^{FN}_p = \Big( (\imath_{e_i} \Psi_1) \wedge (\imath_{e_j} \nabla_{e_i} \Psi_2) - (-1)^{k_1} (\imath_{e_j} \imath_{e_i} \Psi_1) \wedge e^k \wedge \imath_{e_i} \nabla_{e_k} \Psi_2\\
- &(\imath_{e_j} \nabla_{e_i} \Psi_1) \wedge (\imath_{e_i} \Psi_2) - (-1)^{k_1} e^k \wedge \imath_{e_i} \nabla_{e_k} \Psi_1 \wedge (\imath_{e_j} \imath_{e_i} \Psi_2)\Big) \otimes (e^j)^\#,\end{aligned}$$ where $(e_i)$ is an arbitrary basis of $T_pM$ with dual basis $(e^i)$ of $T_p^*M$. In particular, if $K_1 = K_2 =: K$ and $k_1 = k_2$ is odd, then $$\begin{aligned}
{}[K, K]^{FN}_p = & 2 \Big( (\imath_{e_i} \Psi) \wedge (\imath_{e_j} \nabla_{e_i} \Psi) + (\imath_{e_j}\imath_{e_i} \Psi) \wedge e^k \wedge \imath_{e_i} \nabla_{e_k} \Psi \Big) \otimes (e^j)^\#.\end{aligned}$$
If $K_1 = K_2 = K$ and $k_1 = k_2$ is even, then $[K, K]^{FN} = 0$ due to the graded skew symmetry of the bracket. Furthermore, observe that $(e^j)^\# = e_j$ in case $(e_i)$ is an orthonormal basis.
Evidently, if this formula holds for [*some* ]{}basis $(e_j)$ with dual basis $(e^j)$, then it holds for [*any* ]{}basis. Therefore, it suffices to show the assertion for an orthonormal basis $(e_j)$ in which case $(e^j)^\# = e_j$.
Choose geodesic normal coordinates $(x^i)$ around $p \in M$ in such a way that $({{\partial}}_i)_p := ({{\partial}}/ {{\partial}}x^i)_p$ is an orthonormal basis of $T_pM$. The dual basis of ${{\partial}}_i$ is $dx^i$, whence $(dx^i)^\# = g^{ij} {{\partial}}_j$. Thus, $$K_l = (\imath_{{{\partial}}_i} \Psi_l) \otimes (dx^i)^\# = g^{ij} (\imath_{{{\partial}}_i} \Psi_l) \otimes {{\partial}}_j.$$ Thus, by (\[eq:kms\]) $$\begin{aligned}
{}[K_1, K_2]^{FN} =
&[g^{ij} (\imath_{{{\partial}}_i} \Psi_1) \otimes {{\partial}}_j, g^{rs} (\imath_{{{\partial}}_r} \Psi_2) \otimes {{\partial}}_s]^{FN}\\
=
&\left(
g^{ij} (\imath_{{{\partial}}_i} \Psi_1) \wedge {{\mathcal L}}_{{{\partial}}_j} (g^{rs} (\imath_{{{\partial}}_r} \Psi_2)) \otimes {{\partial}}_s \right.\\
&-
{{\mathcal L}}_{{{\partial}}_s} (g^{ij} (\imath_{{{\partial}}_i} \Psi_1)) \wedge g^{rs} (\imath_{{{\partial}}_r} \Psi_2) \otimes {{\partial}}_j \\
&+
(-1)^{k_1} d (g^{ij} (\imath_{{{\partial}}_i} \Psi_1)) \wedge \imath_{{{\partial}}_j} (g^{rs} (\imath_{{{\partial}}_r} \Psi_2)) \otimes {{\partial}}_s \\
&+(-1)^{k_1}(\imath_{{{\partial}}_s}(g^{ij} (\imath_{{{\partial}}_i} \Psi_1)) \wedge d(g^{rs} (\imath_{{{\partial}}_r} \Psi_2)) \otimes {{\partial}}_j.\end{aligned}$$ Since at $p$, $g_{ij} = g^{ij} = \delta_{ij}, {{\partial}}_r g_{ij} = 0$, ${{\mathcal L}}_{{{\partial}}_j} \Psi = \nabla_{e_j} \Psi$, $\nabla_{{{\partial}}_i} {{\partial}}_j = 0$, and ${{\partial}}_j = (e^j)^\#$, the asserted formula follows.
Cross products and $G_2$-structures {#sec:g2}
===================================
$G_2$-structures and associated cross products {#subs:g2str}
----------------------------------------------
In this section we collect some basic facts on $G_2$-structures, see e.g. [@Humphreys], [@Bryant1987], [@FG1982], [@HL1982] for references.
Let $M$ be an oriented 7-manifold. A [*$G_2$-structure on $M$* ]{}is a $3$-form $\varphi \in {{\Omega}}^3(M)$ such that at each $p \in M$ there is a positively oriented basis $(e_i)$ of $T_pM$ with dual basis $(e^i)$ such that $$\label{varphi}
\varphi_p = e^{123} + e^{145} + e^{167} + e^{246} - e^{257} - e^{347} - e^{356},$$ where $e^{i_1 \dots i_k}$ is short for $e^{i_1} \wedge \cdots \wedge e^{i_k}$. We call such a basis a [*$G_2$-frame*]{}. The stabilizer of $\varphi_p$ is isomorphic to the exceptional group $G_2$, and there is a unique $G_2$-invariant Riemannian metric $g_\varphi$ on $M$ such that each $G_2$-frame is orthonormal. In particular, the Hodge-dual of $\varphi$ w.r.t. $g_\varphi$ is given by $$\label{varphi*}
\ast_{g_\varphi} \varphi = e^{4567} + e^{2367} + e^{2345} + e^{1357} - e^{1346} - e^{1256} - e^{1247}.$$
The set of $G_2$-frames yields a principal $G_2$-bundle $$G_2(M) = G_2(M, \varphi) \longrightarrow M,$$ whence for each $G_2$-module $V$ we denote by
$$\label{eq:def-G2-bundles}
V(M) := G_2(M) \times_{G_2} V \longrightarrow M$$
the associated vector bundle over $M$. For instance, $$V_7(M) \cong TM \cong T^*M.$$
\[def:Crchi\] [@HL1982] Let $(M, \varphi)$ be an oriented manifold with a $G_2$-structure $\varphi$ and the induced Riemannian metric $g = g_\varphi$. Then the $TM$-valued forms $Cr_\varphi \in {{\Omega}}^2(M, TM)$ and $\chi_\varphi \in {{\Omega}}^3(M, TM)$ are defined by $$Cr_\varphi := \delta_{g_\varphi}(\varphi) \qquad \mbox{and} \qquad \chi_\varphi := -\delta_g(\ast \varphi),$$ and are called the [*$2$-fold and $3$-fold cross product on $M$*]{}, respectively. That is, for $x,y,z,w \in TM$ we have $$\begin{aligned}
g_\varphi(Cr_\varphi (x,y), z) = \varphi (x,y,z), \qquad
g_\varphi(\chi_\varphi (x,y,z), w) = * \varphi (x,y,z,w).\end{aligned}$$
We shall usually suppress the indices $\varphi$ for $g, Cr$ and $\chi$ if it is clear from the context which $G_2$-structure $\varphi$ is used.
If we use the $G_2$-structure to identify each $T_pM \cong Im {{\mathbb O}}$ with the imaginary octonians, then $Cr$ and $\chi$ can be interpreted w.r.t. the octonian product $\cdot: {{\mathbb O}}\times {{\mathbb O}}\to {{\mathbb O}}$ as $$Cr(x, y) := (x \cdot y)_{Im {{\mathbb O}}} \qquad \mbox{and}\qquad \chi(x,y,z) := ((x \cdot y) \cdot z - x \cdot (y \cdot z))_{Im {{\mathbb O}}}.$$
We summarize important known facts about the decomposition of tensor products of $G_2$-modules into irreducible summands which are well known, see e.g. [@Kar2005 Section 2]. We denote by $V_k$ the $k$-dimensional irreducible $G_2$-module if there is a unique such module. For instance, $V_7$ is the irreducible $7$-dimensional $G_2$-module from above, and $V_7^* \cong V_7$. For its exterior powers, we obtain the decompositions $$\label{eq:DiffForm-V7}
\begin{array}{rlrl}
\Lambda^0 V_7 \cong \Lambda^7 V_7 \cong V_1, \qquad
& \Lambda^2 V_7 \cong \Lambda^5 V_7 \cong V_7 \oplus V_{14},\\[2mm]
\Lambda^1 V_7 \cong \Lambda^6 V_7 \cong V_7, \qquad
& \Lambda^3 V_7 \cong \Lambda^4 V_7 \cong V_1 \oplus V_7 \oplus V_{27},
\end{array}$$ where $\Lambda^k V_7 \cong \Lambda^{7-k} V_7$ due to the Hodge isomorphism. We denote by $\Lambda^k_l V_7 \subset \Lambda^k V_7$ the subspace isomorphic to $V_l$ in the above notation. Evidently, $\Lambda^3_1 V_7$ and $\Lambda^4_1 V_7$ are spanned by $\varphi$ and $\ast \varphi$, respectively. For the decompositions of $\Lambda^2 V_7$ and $\Lambda^5 V_7$ the following descriptions are well known.
$$\begin{aligned}
\nonumber
\Lambda^2_7 V_7 &= \{ \imath_v \varphi \mid v \in V_7\},\\
\nonumber
\Lambda^2_{14} V_7 &= \{ \alpha^2 \in \Lambda^2 V_7 \mid \ast \varphi \wedge \alpha^2 = 0\},\\
\nonumber
\Lambda^3_1 V_7 &= {{\mathbb R}}\varphi,\\
\label{decom-L-V7}
\Lambda^3_7 V_7 &= \{ \imath_v \ast \varphi \mid v \in V_7\},\\
\nonumber
\Lambda^4_1 V_7 &= {{\mathbb R}}\ast \varphi,\\
\nonumber
\Lambda^4_7 V_7 &= \varphi \wedge V_7 = \{ \varphi \wedge v \mid v \in V_7\},\\
\nonumber
\Lambda^5_7 V_7 &= \ast \varphi \wedge V_7 = \{ \ast \varphi \wedge v \mid v \in V_7\},\\
\nonumber
\Lambda^5_{14} V_7 &= \{ \alpha^5 \in \Lambda^5 V_7 \mid \alpha^5 \wedge (\imath_v \varphi) = 0 \; \mbox{for all $v \in V_7$}\}.\end{aligned}$$
We also point out that all representations of $G_2$ are of real type, meaning that for any real irreducible representation $V$ of $G_2$ the complexified space $V^{{\mathbb C}}:= V \otimes {{\mathbb C}}$ is (complex) irreducible; equivalently, a real irreducible representation of $G_2$ does not admit a $G_2$-invariant complex structure.
These decompositions are used in the appendix to obtain many formulas which will be used in the sequel.
The torsion of manifolds with a $G_2$-structure
-----------------------------------------------
Let $(M, \varphi)$ be a manifold with a $G_2$-structure with the corresponding Riemannian metric $g = g_\varphi$, and let $\nabla$ be the Levi-Civita connection of $g$. In general, $\varphi$ and $\ast \varphi$ will not be parallel w.r.t. $\nabla$, and the failure of their parallelity can be described in the following way which is essentially a reformulation of the intrinsic torsion of a $G_2$-structure discussed in [@Bryant2005] and [@Kar2005].
\[prop:torsionG2\] Let $(M, \varphi)$ be a manifold with a $G_2$-structure with associated Riemannian metric $g = g_\varphi$ and Levi-Civita connection $\nabla$. Then there is a section $T \in {{\Omega}}^1(M, TM) = \Gamma(End(TM))$ such that for all $v \in TM$ we have $$\label{eq:nabla-G2}
\nabla_v \varphi = \imath_{T(v)} \ast \varphi \qquad \mbox{and} \qquad \nabla_v * \varphi = -(T(v))^\flat \wedge \varphi.$$
Thus, the section $T \in {{\Omega}}^1(M, TM)$ measures how $\varphi$ fails to be parallel, and this has been described in Fernández and Gray [@FG1982] by slightly different means. In fact, it contains the same information as the intrinsic torsion of the $G_2$-structure in the sense of [@Bryant2005], whence we use the following terminology.
\[def:torsionG2\] Let $(M, \varphi)$ be a manifold with a $G_2$-structure. The section $T \in {{\Omega}}^1(M, TM)$ for which (\[eq:nabla-G2\]) holds is called the [*torsion endomorphism* ]{}of the $G_2$-structure.
For an orthonormal frame $(e_i)$ of $T_pM$ we define the coefficients of $T$ by $$\label{def-tij}
t_{ij} := {\langle}T(e_i), e_j {\rangle}, \qquad \mbox{so that} \qquad T(e_i) = t_{ij} e_j.$$ Furthermore, we define the form $$\label{def:tau}
\tau := t_{ij} e^{ij} = \dfrac12 (t_{ij} - t_{ji}) e^{ij} = e^i \wedge T(e_i)^\flat \in \Lambda^2 V_7^*.$$ For the exterior derivatives of $\varphi$ and $\ast \varphi$, we have $$\label{d*-T}
d\varphi_p = T_p^\top(e_i)^\flat \wedge (\imath_{e_i} \ast \varphi_p) \qquad \mbox{and} \qquad
d\ast \varphi_p = -\tau_p \wedge \varphi_p,$$ where we sum over an orthonormal basis $(e_i)$ of $T_pM$ in the first equation and where $T_p^\top$ denotes the transpose matrix of $T_p$. In particular, it is now a straightforward calculation to show that $(M, \varphi)$ is torsion free at $p \in M$ (i.e., $T_p = 0$) iff $d\varphi_p = 0$ and $d\ast \varphi_p = 0$ (cf. [@FG1982]).
The Frölicher-Nijenhuis brackets on a manifold with a $G_2$-structure
---------------------------------------------------------------------
In this section, we shall compute part of their Frölicher-Nijenhuis brackets of the sections $Cr = \delta_g \varphi \in {{\Omega}}^2(M, TM)$ and $\chi = -\delta_g \ast \varphi \in {{\Omega}}^3(M, TM)$. from Definition \[def:Crchi\] on a manifold $M$ with a $G_2$-structure $\varphi$.
The Frölicher-Nijenhuis bracket $[Cr, Cr]^{FN}$ vanishes identically due to the graded skew-symmetry of the bracket. On the other hand, the Frölicher-Nijenhuis brackets $[Cr, \chi]^{FN}$ and $[\chi, \chi]^{FN}$ are elements of ${{\Omega}}^5(M, TM)$ and ${{\Omega}}^6(M, TM)$, respectively.
Due to the decomposition $\Lambda^5 V_7 = \Lambda^5_7 V_7 \oplus \Lambda^5_{14} V_7$ as a $G_2$-module, we may decompose $${{\Omega}}^5(M, TM) = \Gamma(M, \Lambda^5_7 T^*M \otimes TM) \oplus \Gamma(M, \Lambda^5_{14} T^*M \otimes TM),$$ and we denote the projections onto the two summands by $\pi_7$ and $\pi_{14}$, respectively. We now wish to show Theorem \[thm:Brackets-intro\] from the introduction. In order to work towards the proof, we first calculate $\pi_7([Cr, \chi]^{FN})$.
\[prop:Cr-chi\] Let $(M, g, \varphi)$ be a manifold with a $G_2$-structure and let $T \in {{\Omega}}^1(M, TM)$ be its torsion endomorphism. Then for each $p \in $M, $$\label{eq:pi-Crchi}
\pi_7([Cr, \chi]_p^{FN}) = 2 \ast \varphi \wedge \left( \left(T_p^\top - 2 T_p - tr(T_p)\right) e_i\right)^\flat \otimes e_i,$$ summing over an orthonormal basis $(e_i)$ of $T_pM$, where $T^\top$ denotes the transpose of $T$. In particular, $\pi_7([Cr, \chi]^{FN}_p) = 0$ if and only if $T_p = 0$ if and only if $[Cr, \chi]^{FN}_p = 0$.
We fix $p \in M$ and use normal coordinates around $p$. Then in order to calculate $[Cr, \chi]^{FN}_p$ we apply Proposition \[prop:FN\] to $\Psi_1 = \varphi$ and $\Psi_2 = \ast \varphi$ and obtain $${}-[Cr, \chi]^{FN}_p = [(\imath_{e_i} \varphi) \otimes e_i, (\imath_{e_j} \ast \varphi) \otimes e_j]^{FN}_p =: \beta_j \otimes e_j,$$ where $$\begin{aligned}
\nonumber
\beta_j =&\; (\imath_{e_k} \varphi) \wedge (\imath_{e_j} \nabla_{e_k} \ast \varphi) - (\imath_{e_j}\imath_{e_k} \varphi) \wedge e^l \wedge (\imath_{e_k} \nabla_{e_l} \ast \varphi)\\
\label{Cr-chiVer1}
&-(\imath_{e_j} \nabla_{e_k} \varphi) \wedge (\imath_{e_k} \ast \varphi) - e^l \wedge (\imath_{e_k} \nabla_{e_l} \varphi) \wedge (\imath_{e_j} \imath_{e_k} \ast \varphi).\end{aligned}$$
Decomposing $\Lambda^5 T_p^*M$ according to (\[decom-L-V7\]), we write $\beta_j = \ast \varphi \wedge v_j^\flat + \beta_j^{14}$ with $v_j \in T_pM$ and $\beta_j^{14} \in \Lambda^5_{14} T_p^*M$, so that $\pi_7([Cr, \chi]^{FN}) = \ast \varphi \wedge v_j^\flat \otimes e_j$. Let $$\label{def-bij}
b_{ij} = \ast ((\imath_{e_i} \varphi) \wedge \beta_j).$$ Then as $(\imath_{e_i} \varphi) \wedge \beta_j^{14} = 0$ by (\[decom-L-V7\]) and $(\imath_{e_i} \varphi) \wedge \ast \varphi \wedge v_j^\flat = 3 {\langle}e_i, v_j{\rangle}{\mbox{\rm vol}}$ by (\[eq:form-0\]), it follows that $$\label{def-pi7}
\pi_7(-[Cr, \chi]^{FN}_p) = \dfrac13 b_{ij} (\ast \varphi \wedge e^i) \otimes e_j.$$
In order to determine the coefficients $b_{ij}$, we decompose $\beta_j$ into the four summands from (\[Cr-chiVer1\]). Then from the first summand we get
$$\begin{aligned}
\nonumber
(\imath_{e_i} \varphi) \wedge (\imath_{e_k} \varphi) \wedge (\imath_{e_j} \nabla_{e_k} \ast \varphi)\stackrel{(\ref{eq:nabla-G2})} =&\; - (\imath_{e_i} \varphi) \wedge (\imath_{e_k} \varphi) \wedge (\imath_{e_j} (T(e_k)^\flat \wedge \varphi))\\
\nonumber
=&\; - (\imath_{e_i} \varphi) \wedge (\imath_{e_k} \varphi) \wedge (t_{kj} \varphi - T(e_k)^\flat \wedge (\imath_{e_j} \varphi))\\
\nonumber
\stackrel{(\ref{eq:form-1}),(\ref{eq:form-2})} =&\; - 6 t_{kj} \delta_{ik} {\mbox{\rm vol}}+ 2 (\delta_{ik} t_{kj} + t_{ki} \delta_{kj} + t_{kk} \delta_{ij}) {\mbox{\rm vol}}\\
\label{b_ij-1}
=&\; 2 (t_{ji} - 2 t_{ij} + tr(T) \delta_{ij}) {\mbox{\rm vol}}.\end{aligned}$$
From the second summand we obtain
$$\begin{aligned}
\nonumber
-(\imath_{e_i} \varphi) \wedge (\imath_{e_j} \imath_{e_k} \varphi) \wedge e^l \wedge &(\imath_{e_k} \nabla_{e_l} \ast \varphi)\\
\nonumber
\stackrel{(\ref{eq:nabla-G2})} =&\; (\imath_{e_i} \varphi) \wedge (\imath_{e_j}\imath_{e_k} \varphi) \wedge e^l \wedge (\imath_{e_k} (T(e_l)^\flat \wedge \varphi))\\
\nonumber
=&\; (\imath_{e_i} \varphi) \wedge (\imath_{e_j}\imath_{e_k} \varphi) \wedge e^l \wedge (t_{lk} \varphi - T(e_l)^\flat \wedge (\imath_{e_k} \varphi))\\
\nonumber
\stackrel{(\ref{eq:form-3})} =&\; t_{lk} \Big(2(\delta_{kl} \delta_{ji} - \delta_{jl} \delta_{ki}) {\mbox{\rm vol}}- 2 e^{jkli} \wedge \varphi\Big)\\
\nonumber
& - (\imath_{e_i} \varphi) \wedge (\imath_{e_j}\imath_{e_k} \varphi) \wedge e^l \wedge T(e_l)^\flat \wedge (\imath_{e_k} \varphi)\\
\nonumber
=&\; 2(tr(T) \delta_{ij} - t_{ji}) {\mbox{\rm vol}}- 2 e^{ij} \wedge \tau \wedge \varphi\\
\nonumber
& - \frac12 (\imath_{e_i} \varphi) \wedge \imath_{e_j}((\imath_{e_k} \varphi) \wedge (\imath_{e_k} \varphi)) \wedge \tau\\
\nonumber
\stackrel{(\ref{formulas2})}=&\; 2(tr(T) \delta_{ij} - t_{ji}) {\mbox{\rm vol}}- 2 e^{ij} \wedge \tau \wedge \varphi\\
\nonumber
& - 3 (\imath_{e_i} \varphi) \wedge (\imath_{e_j}\ast \varphi) \wedge \tau\\
\nonumber
\stackrel{(\ref{eq:form13})}=&\; 2(tr(T) \delta_{ij} - t_{ji}) {\mbox{\rm vol}}- 2 e^{ij} \wedge \tau \wedge \varphi\\
\nonumber
& -3 (-2 \ast e^{ji} + e^{ji} \wedge \varphi) \wedge \tau\\
\nonumber
=&\; 2(tr(T) \delta_{ij} - t_{ji}) {\mbox{\rm vol}}+ e^{ij} \wedge \tau \wedge \varphi\\
\nonumber
& + 6 t_{kl}(\delta_{jk} \delta_{il} - \delta_{jl} \delta_{ik}) {\mbox{\rm vol}}\\
\label{b_ij-2}
=&\; 2(tr(T) \delta_{ij} + 2 t_{ji} - 3 t_{ij}) {\mbox{\rm vol}}+ e^{ij} \wedge \tau \wedge \varphi.\end{aligned}$$
From the third term in (\[Cr-chiVer1\]) we get
$$\begin{aligned}
\nonumber
-(\imath_{e_i} \varphi) \wedge (\imath_{e_j} \nabla_{e_k} \varphi) \wedge& (\imath_{e_k} \ast \varphi)\\
\nonumber
\stackrel{(\ref{eq:nabla-G2})}=&\; -(\imath_{e_i} \varphi) \wedge (\imath_{e_j} \imath_{T(e_k)} \ast\varphi) \wedge (\imath_{e_k} \ast \varphi)\\
\nonumber
\stackrel{(\ref{eq:form-4})}=&\; -2 (t_{kk} \delta_{ji} - \delta_{jk} t_{ki}) {\mbox{\rm vol}}- e^j \wedge T(e_k)^\flat \wedge e^{ki} \wedge \varphi\\
\label{b_ij-3}
=&\; -2 (tr(T) \delta_{ij} - t_{ji}) {\mbox{\rm vol}}- e^{ij} \wedge \tau \wedge \varphi.\end{aligned}$$
Finally, from the last term in (\[Cr-chiVer1\]) we get
$$\begin{aligned}
\nonumber
-(\imath_{e_i} \varphi) \wedge e^l \wedge (\imath_{e_k} \nabla_{e_l} \varphi) \wedge& (\imath_{e_j} \imath_{e_k} \ast \varphi)\\
\nonumber
\stackrel{(\ref{eq:nabla-G2})}=&\; -(\imath_{e_i} \varphi) \wedge e^l \wedge (\imath_{e_k} \imath_{T(e_l)} \ast \varphi) \wedge (\imath_{e_j} \imath_{e_k} \ast \varphi)\\
\nonumber
\stackrel{(\ref{formulas-3})}=&\; 2 (\imath_{e_i} \varphi) \wedge e^l \wedge (\imath_{T(e_l)} \varphi) \wedge (\imath_{e_j} \varphi)\\
\nonumber
\stackrel{(\ref{eq:form-2})}=&\; 4 (t_{li} \delta_{lj} + \delta_{li} t_{lj} + t_{ll} \delta_{ij}) {\mbox{\rm vol}}\\
\label{b_ij-4}
=&\; 4 (t_{ji} + t_{ij} + tr(T) \delta_{ij}) {\mbox{\rm vol}}.\end{aligned}$$
Thus, adding (\[b\_ij-1\]) through (\[b\_ij-4\]), we get from (\[Cr-chiVer1\]) that
$$\begin{aligned}
b_{ij} {\mbox{\rm vol}}= (\imath_{e_i} \varphi) \wedge \beta_j =&\; 2 (t_{ji} - 2 t_{ij} + tr(T) \delta_{ij}) {\mbox{\rm vol}}\\
&+ 2(tr(T) \delta_{ij} + 2 t_{ji} - 3 t_{ij}) {\mbox{\rm vol}}+ e^{ij} \wedge \tau \wedge \varphi\\
&- 2 (tr(T) \delta_{ij} - t_{ji}) {\mbox{\rm vol}}- e^{ij} \wedge \tau \wedge \varphi\\
&+ 4 (t_{ji} + t_{ij} + tr(T) \delta_{ij}) {\mbox{\rm vol}}\\
=&\; 6 (2 t_{ji} - t_{ij} + tr(T) \delta_{ij}) {\mbox{\rm vol}},\end{aligned}$$
and hence, (\[def-pi7\]) implies (\[eq:pi-Crchi\]).
Thus, $\pi_7([Cr, \chi]^{FN}_p) = 0$ iff $T_p^\top - 2 T_p - tr(T_p)Id = 0$. Taking the trace, this implies that $tr(T_p) - 2 tr(T_p) - 7 tr(T_p) = 0$ and hence, $tr(T_p) = 0$, and $T_p^\top - 2 T_p = 0$ evidently implies that $T_p = 0$. That is, $\pi_7([Cr, \chi]^{FN}_p) = 0$ iff $T_p = 0$, showing the last statement.
Next, let us consider the bracket $[\chi, \chi]^{FN}$.
\[prop:chi-chi\] Let $(M, g, \varphi)$ be a manifold with a $G_2$-structure and let $T \in {{\Omega}}^1(M, TM)$ be its torsion endomorphism with the associated form $\tau \in {{\Omega}}^2(M)$ from (\[def:tau\]). Then for each $p \in $M, $$\label{eq:pi-chichi}
{}[\chi, \chi]_p^{FN} = -4 \ast (T_p + T_p^\top)(e_i) \otimes e_i + 6 e^i \wedge \tau_p \wedge \varphi \otimes e_i,$$ summing over an orthonormal basis $(e_i)$ of $T_pM$. In particular, $[\chi, \chi]_p^{FN} = 0$ if and only if $T_p = 0$.
According to Proposition \[prop:FN\] we have $${}[\chi, \chi]^{FN}_p = \gamma_j \otimes e_j,$$ where $$\label{chichiVer1}
\gamma_j = 2 ((\imath_{e_k} \ast \varphi) \wedge (\imath_{e_j} \nabla_{e_k} \ast \varphi) + (\imath_{e_j} \imath_{e_k} \ast \varphi) \wedge e^l \wedge (\imath_{e_k} \nabla_{e_l} \ast \varphi)).$$
Now let $c_{ij} := \ast ( e^i \wedge \gamma_j ) = {\langle}e_i, (\ast \gamma_j)^\flat{\rangle}$. Then $$\label{def-pi7chichi}
{}[\chi, \chi]^{FN}_p = c_{ij} \ast e^i \otimes e_j,$$
In order to evaluate the coefficients $c_{ij}$, we consider the two summands in (\[chichiVer1\]) separately, and obtain from the first one
$$\begin{aligned}
\nonumber
e^i \wedge (\imath_{e_k} \ast \varphi) &\wedge (\imath_{e_j} \nabla_{e_k} \ast \varphi) \stackrel{(\ref{eq:nabla-G2})}=\; - e^i \wedge (\imath_{e_k} \ast \varphi) \wedge (\imath_{e_j} (T(e_k)^\flat \wedge \varphi))\\
\nonumber
=&\; - t_{kj} e^i \wedge (\imath_{e_k} \ast \varphi) \wedge \varphi + e^i \wedge (\imath_{e_k} \ast \varphi) \wedge T(e_k)^\flat \wedge (\imath_{e_j} \varphi)\\
\nonumber
\stackrel{(\ref{eq:form-11}), (\ref{eq:form-5})}=&\; -4 t_{kj} \delta_{ik} {\mbox{\rm vol}}+ 2(\delta_{ik} t_{kj} - t_{kk} \delta_{ij}) {\mbox{\rm vol}}+ T(e_k)^\flat \wedge e^{ikj} \wedge \varphi\\
\label{c_ij-1}=&\; -2(t_{ij} + tr(T) \delta_{ij}) {\mbox{\rm vol}}+ e^{ij} \wedge \tau \wedge \varphi .\end{aligned}$$
From the second summand in (\[chichiVer1\]) we calculate
$$\begin{aligned}
\nonumber
e^i \wedge (\imath_{e_j} & \imath_{e_k} \ast \varphi) \wedge e^l \wedge (\imath_{e_k} \nabla_{e_l} \ast \varphi)\\
\nonumber
\stackrel{(\ref{eq:nabla-G2})}=&\; - e^{il} \wedge (\imath_{e_j} \imath_{e_k} \ast \varphi) \wedge (\imath_{e_k} (T(e_l)^\flat \wedge \varphi))\\
\nonumber
=&\; - t_{lk} e^{il} \wedge (\imath_{e_j} \imath_{e_k} \ast \varphi) \wedge \varphi + e^i \wedge (\imath_{e_j} \imath_{e_k} \ast \varphi) \wedge \tau \wedge (\imath_{e_k} \varphi)\\
\nonumber
\stackrel{(\ref{eq:form-6}), (\ref{formulas-4})}=&\; - t_{lk} \big(2 (\delta_{lj} \delta_{ik} - \delta_{ij} \delta_{lk}) {\mbox{\rm vol}}- e^{iljk} \wedge \varphi\big) + 3 e^{ij} \wedge \tau \wedge \varphi\\
\label{c_ij-2}=&\; - 2 (t_{ji} - tr(T) \delta_{ij}) {\mbox{\rm vol}}+ 2 e^{ij} \wedge \tau \wedge \varphi.\end{aligned}$$
Thus, adding (\[c\_ij-1\]) and (\[c\_ij-2\]), equation (\[chichiVer1\]) yields
$$\begin{aligned}
c_{ij} {\mbox{\rm vol}}= e^i \wedge \gamma_j =&\; 2 \Big( -2(t_{ij} + tr(T) \delta_{ij}) {\mbox{\rm vol}}+ e^{ij} \wedge \tau \wedge \varphi\\
&\qquad - 2 (t_{ji} - tr(T) \delta_{ij}) {\mbox{\rm vol}}+ 2 e^{ij} \wedge \tau \wedge \varphi\Big)\\
=&\; 2 (-2(t_{ij} + t_{ji}) + 3 {\langle}\ast e^i, e^j \wedge \tau \wedge \varphi {\rangle}) {\mbox{\rm vol}},\end{aligned}$$
and from this and (\[def-pi7chichi\]), the formula (\[eq:pi-chichi\]) follows.
In order to show the last statement, observe that $[\chi, \chi]^{FN}_p=0$ iff $c_{ij} = 0$ for all $i, j$. Since then $c_{ij} + c_{ji} = -8(t_{ij} + t_{ji})$, it follows that $t_{ij} + t_{ji} = 0$ and hence $0 = c_{ij} = 6 \ast ( e^{ij} \wedge \tau \wedge \varphi)$ for all $i, j$ which implies that $\tau = t_{kl}e^{kl} = 0$ and hence, $t_{kl}= t_{lk}$. All of this together implies that $t_{ij} = 0$ for all $i, j$, and hence, $T_p=0$ as asserted.
We are now ready to show the following result which immediately implies Theorem \[thm:Brackets-intro\] from the introduction.
\[thm:Brackets\] Let $(M^7, \varphi)$ be a manifold with a $G_2$-structure with associated metric $g = g_\varphi$, let $\nabla$ be the Levi-Civita connection of $g$, and let $T \in {{\Omega}}^1(M^7, TM^7)$ be its torsion endomorphism defined in Definition \[def:torsionG2\]. Then for every $p \in M^7$ the following are equivalent.
1. $T_p = 0 \in T_p^*M^7 \otimes T_pM^7$.
2. The $G_2$-structure is torsion-free at $p$, i.e., $(\nabla \varphi)_p = 0$.
3. $\pi_7([Cr, \chi]_p^{FN}) = 0 \in \Lambda^5_7 T_p^*M^7 \otimes T_pM^7$.
4. $[Cr, \chi]_p^{FN} = 0 \in \Lambda^5 T_p^*M^7 \otimes T_pM^7$.
5. $[\chi, \chi]_p^{FN} = 0 \in \Lambda^6 T_p^*M^7 \otimes T_pM^7$.
The equivalence of the first two statements is well known, see e.g. [@FG1982]. Proposition \[prop:Cr-chi\] shows the equivalence of the first and the third, whereas Proposition \[prop:chi-chi\] shows the equivalence of the first and the last statement. That $(\nabla \varphi)_p=0$ implies $[Cr, \chi]_p^{FN} = 0$ is immediate from the formula of the bracket in Proposition \[prop:FN\], and obviously, $[Cr, \chi]_p^{FN} = 0$ implies $\pi_7([Cr, \chi]_p^{FN}) = 0$.
Cross products and ${\mbox{\rm Spin}(7)}$-structures {#sec:spin7}
====================================================
${\mbox{\rm Spin}(7)}$-structures and associated cross products {#subs:spin7str}
---------------------------------------------------------------
The exposition in this section mainly follows the references [@Bryant1987], [@Fernandez1986], [@HL1982].
Let $M$ be an oriented $8$-manifold. A [*${\mbox{\rm Spin}(7)}$-structure on $M$* ]{}is a $4$-form $\Phi \in {{\Omega}}^4(M)$ such that at each $p \in M$ there is a positively oriented basis $(e_\mu)_{\mu=0}^7$ of $T_pM$ with dual basis $(e^\mu)_{\mu=0}^7$ such that $\Phi_p \in \Lambda^4 T_pM$ is of the form $$\begin{aligned}
\label{Phi4}
\Phi_p & := & e^{0123} + e^{0145} + e^{0167} + e^{0246} - e^{0257} - e^{0347} - e^{0356}\\
\nonumber & & + e^{4567} + e^{2367} + e^{2345} + e^{1357} - e^{1346} - e^{1256} - e^{1247}.\end{aligned}$$
Throughout this section, we shall use Greek indices $\mu, \nu, \ldots$ to run over $0, \ldots, 7$, whereas Latin indices $i,j, \ldots$ range over $1, \ldots, 7$.
A basis $(e_\mu)$ of $T_pM$ whose dual basis $(e^\mu)$ satisfies (\[Phi4\]) is called a [*${\mbox{\rm Spin}(7)}$-frame*]{}. Observe that if we define for a ${\mbox{\rm Spin}(7)}$-frame $(e_\mu)$ the forms $\varphi_p$ and $\ast_7 \varphi_p$ on $V_p := {\mbox{\rm span}}(e_i)_{i=1}^7 \subset T_pM$ as in (\[varphi\]) and (\[varphi\*\]), then $$\Phi_p = e^0 \wedge \varphi_p + \ast_7 \varphi_p.$$
The stabilizer of $\Phi_p$ is the group ${\mbox{\rm Spin}(7)}$ acting on $T_pM$ via the spinor representation, and there is a unique ${\mbox{\rm Spin}(7)}$-invariant Riemannian metric $g_\Phi$ on $M$ such that each ${\mbox{\rm Spin}(7)}$-frame is orthonormal. In particular, $\Phi$ is self-dual w.r.t. $g_\Phi$. The set of all ${\mbox{\rm Spin}(7)}$-frames forms a principal ${\mbox{\rm Spin}(7)}$-bundle $${\mbox{\rm Spin}(7)}_M = {\mbox{\rm Spin}(7)}_{(M, \Phi)} \longrightarrow M,$$ and again, for each ${\mbox{\rm Spin}(7)}$-module $W$ we obtain the associated vector bundle $$\label{eq:def-Spin7-bundles}
W(M) := {\mbox{\rm Spin}(7)}_M \times_{{\mbox{\rm Spin}(7)}} W \longrightarrow M.$$ For instance, if we denote the $k$-dimensional irreducible ${\mbox{\rm Spin}(7)}$-module by $W_k$ (in case the dimension uniquely specifies this module), then $$W_8(M) \cong TM \cong T^*M.$$
It is well known that the action of ${\mbox{\rm Spin}(7)}$ on $W_8$ is transitive on the unit sphere $S^7 \subset W_8$, and the stabilizer of an element is isomorphic to $G_2 \subset {\mbox{\rm Spin}(7)}$. In analogy of the products $Cr$ and $\chi$ on manifolds with a $G_2$-structure in Definition \[def:Crchi\], we define on a ${\mbox{\rm Spin}(7)}$-manifold $M$ a triple product as follows.
\[def:P\] Let $(M, \Phi)$ be manifold with a ${\mbox{\rm Spin}(7)}$-structure, and let $g = g_\Phi$ be the induced Riemannian metric. Then the $TM$-valued form $P =P_\Phi \in {{\Omega}}^3(M, TM)$ is defined by $$P_\Phi := -\delta_{g_\Phi}(\Phi),$$ and is called the [*$3$-fold cross product on $M$*]{}. That is, for $x,y,z,w \in TM$ we have $$\label{def-P}
g(P(x,y,z), w) = \Phi (x,y,z,w).$$
We shall usually suppress the indices $\Phi$ for $g$ and $P$ if it is clear from the context which ${\mbox{\rm Spin}(7)}$-structure $\Phi$ is used.
${\mbox{\rm Spin}(7)}$-representations {#subs:decomspin7}
--------------------------------------
In this section, we shall discuss the decomposition of symmetric and anti-symmetric powers of $W_8$ as ${\mbox{\rm Spin}(7)}$-modules. For its exterior powers, we obtain the decompositions $$\begin{aligned}
\nonumber
\Lambda^0 W_8 &\cong \Lambda^8 W_8 \cong W_1, \qquad
& \Lambda^2 W_8 \cong \Lambda^6 W_8 &\cong W_7 \oplus W_{21},\\
\label{decom1-L-W8}
\Lambda^1 W_8 &\cong \Lambda^7 W_8 \cong W_8, \qquad
& \Lambda^3 W_8 \cong \Lambda^5 W_8 &\cong W_8 \oplus W_{48},\\
\nonumber
\Lambda^4 W_8 &\cong W_1 \oplus W_7 \oplus W_{27} \oplus W_{35}\end{aligned}$$ where $\Lambda^k W_8 \cong \Lambda^{8-k} W_8$ via the Hodge-$\ast$. Again, we denote by $\Lambda^k_l W_8 \subset \Lambda^k W_8$ the subspace isomorphic to $W_l$ in the above notation.
Moreover, there are also irreducible decompositions of the symmetric powers of $W_7$ and $W_8$ as $$\label{decom-Sk-W7}
\odot^2 W_7 \cong W_1 \oplus W_{27}, \qquad \odot^2 W_8 \cong W_1 \oplus W_{35}$$ into the induced ${\mbox{\rm Spin}(7)}$-invariant metric and the trace free symmetric tensors; see [@Humphreys].
\[lem:lambdas\] Let $e^0 \in W_8$ be a unit vector, let $V_7 := e_0^\perp$ on which ${\mbox{\rm Spin}(7)}$ acts as the double cover of $SO(7)$, so that $V_7 \cong W_7$ as a ${\mbox{\rm Spin}(7)}$-module. Then the following maps are ${\mbox{\rm Spin}(7)}$-equivariant embeddings. $$\label{def:lambda}
\lambda^k: W_7 \longrightarrow \Lambda^k W_8, \qquad
\begin{array}{ll} \lambda^2(v) := e^0 \wedge v^\flat + (\imath_v \varphi)\\[2mm]
\lambda^4(v) := e^0 \wedge (\imath_v \ast_7 \varphi) - v^\flat \wedge \varphi\\[2mm]
\lambda^6(v) := \Phi \wedge \lambda^2(v) = 3 \ast \lambda^2(v)
\end{array}$$ Here $\ast$ and $\ast_7$ denote the Hodge-$\ast$ in $W_8$ and $V_7$, respectively.
The decompositions in (\[decom-Sk-W7\]) imply that there are ${\mbox{\rm Spin}(7)}$-equivariant maps $\lambda^k: W_7 \to \Lambda^k W_8$, and these are unique up to rescaling.
The equivariance of $\lambda^2$ follows from [@SW2017 p.68], and thus, $\Phi \wedge \lambda^2(v) \in \Lambda^6_7 W_8$, whence $\Phi \wedge \lambda^2(v) = 3 \ast \lambda^2(v)$ follows from [@SW2017 Theorem 9.8]. This shows the statement on $\lambda^6$.
By [@SW2017 Theorem 9.8], $\Lambda^4_7 W_8$ is the infinitesimal orbit of $\Phi$ under the action of ${{\mathfrak{so}}}(W_8) \cong \Lambda^2 W_8$. That is, $$\Lambda^4_7 W_8 = \{(u^\flat \wedge v^\flat) \cdot \Phi \mid u, v \in W_8\} = \{ u^\flat \wedge (\imath_v \Phi) - v^\flat \wedge (\imath_u \Phi) \mid u, v \in W_8\}.$$ Setting $u := e_0$ and picking $v \in e_0^\perp \cong W_7$ for a ${\mbox{\rm Spin}(7)}$-frame $(e_\mu)$, it follows that the image of $\lambda^4$ equals $\Lambda^4_7 W_8$, and since $\lambda^4$ is evidently $G_2$-equivariant, it must coincide with the ${\mbox{\rm Spin}(7)}$-equivariant map $W_7 \to \Lambda^4_7 W_8$.
From this lemma, we obtain the following descriptions of the decompositions, which essentially recapitulates [@SW2017 Theorem 9.8].
$$\begin{aligned}
\nonumber
\Lambda^k_7 W_8 &= \{\lambda^k(v) \mid v \in V_7\} \quad \mbox{for $k = 2,4,6$},\\
\nonumber
\Lambda^2_{21} W_8 &= \{ \alpha^2 \in \Lambda^2 W_8 \mid \alpha^2 \wedge \Phi \wedge \lambda^2(v) = 0\; \mbox{for all $v \in V_7$}\},\\
\nonumber
\Lambda^6_{21} W_8 &= \{ \alpha^6 \in \Lambda^6 W_8 \mid \alpha^6 \wedge \lambda^2(v) = 0\; \mbox{for all $v \in V_7$}\},\\
\nonumber
\Lambda^3_8 W_8 &= \{ \imath_a \Phi \mid a \in W_8\},\\
\nonumber
\Lambda^5_8 W_8 &= \{ a^\flat \wedge \Phi \mid a \in W_8\},\\
\label{decom-L-W8}
\Lambda^3_{48} W_8 &= \{ \alpha^3 \in \Lambda^3 W_8 \mid \Phi \wedge \alpha^3 = 0\},\\
\nonumber
\Lambda^5_{48} W_8 &= \{ \alpha^5 \in \Lambda^3 W_8 \mid \Phi \wedge \ast \alpha^5 = 0\},\\
\nonumber
\Lambda^4_1 W_8 &= {{\mathbb R}}\Phi,\\
\nonumber
\Lambda^4_{27} W_8 &= {\mbox{\rm span}}\{\lambda^2(v) \wedge \lambda^2(w) \mid v,w \in V_7, {\langle}v,w {\rangle}= 0\}\\
\nonumber
\Lambda^4_{35} W_8 &= \{ \alpha^4 \mid \ast \alpha^4 = - \alpha^4\}.\end{aligned}$$
We also recall the decomposition of the tensor product $$\label{eq:W8W7}
Lin(W_8, W_7) := W_8^* \otimes W_7 = W_8 \oplus W_{48}.$$ Here, the summand isomorphic to $W_8$ is given as $$\label{eq:W8W7-summand}
\{ (\imath_a \lambda^2(e_i)) \otimes e_i \mid a \in W_8\},$$ where the sum is taken over an orthonormal basis $(e_i)$ of $V_7 \cong W_7$. Finally, we define the ${\mbox{\rm Spin}(7)}$-invariant tensor $\sigma \in (W_8 \otimes W_7 \otimes W_8 \otimes W_7)^*$ by $$\label{eq:sigma}
\sigma(a, u, b, v) := \frac12 \ast (a^\flat \wedge b^\flat \wedge \lambda^4(u) \wedge \lambda^2(v)).$$
Contraction with the inner products on $W_7$ and $W_8$ induces a ${\mbox{\rm Spin}(7)}$-equivariant map $$\begin{aligned}
\label{eq:phi-sigma}
\phi_\sigma: Lin(W_8, W_7) &\longrightarrow Lin(W_8, W_7)\\
\nonumber \phi_\sigma(A)(a) &:= \sigma(a, A(e_\mu), e_\mu, e_i)\; e_i.\end{aligned}$$ We calculate $$\begin{aligned}
\nonumber
\lambda^4(u) \wedge \lambda^2(v) & =\; (e^0 \wedge (\imath_u \ast_7 \varphi) - u^\flat \wedge\varphi) \wedge (e^0 \wedge v^\flat + \imath_v \varphi)\\
\nonumber
& =\; e^0 \wedge (\imath_u \ast_7 \varphi) \wedge (\imath_v \varphi) + e^0 \wedge u^\flat \wedge v^\flat \wedge \varphi - u^\flat \wedge \varphi \wedge (\imath_v \varphi)\\
\label{eq:l4l2} & \stackrel{(\ref{eq:form13}), (\ref{eq:form14})}=\; 2 u^\flat \wedge v^\flat \wedge (e^0 \wedge \varphi - \ast_7 \varphi) - 2 e^0 \wedge (\ast_7 ( u^\flat \wedge v^\flat )).\end{aligned}$$
\[lem:phisigma\] The map $\phi_\sigma$ has eigenvalues $-1$ and $6$ with multiplicity $48$ and $8$, respectively.
Observe that the ${\mbox{\rm Spin}(7)}$-invariant inner products on $W_7$ and $W_8$ induce an inner product on $Lin(W_8, W_7) = W_8^* \otimes W_7$ for which $(e^\mu \otimes e_i)$ is an orthonormal basis whenever $(e_\mu)$ is an orthonormal basis of $W_8$ so that $V_7 = e_0^\perp$ is spanned by $(e_i)$. This induced inner product satisfies $${\langle}\phi_\sigma(e^\mu \otimes e_i), e^\nu \otimes e_j {\rangle}_{Lin(W_8, W_7)} = {\langle}\phi_\sigma(e^\mu \otimes e_i)(e_\nu), e_j{\rangle}_{W_7} = \sigma(e_\nu, e_i, e_\mu, e_j),$$ and since $\sigma(e_\mu, e_i, e_\mu, e_j) = 0$, it follows that the matrix representation of $\phi_\sigma$ w.r.t. the basis $(e^\mu \otimes e_i)$ has $0$’s on the diagonal, whence $tr(\phi_\sigma) = 0$. Furthermore, $\phi_\sigma$ is self-adjoint since $\sigma(a,u,b,v) = \sigma(b,v,a,u)$ by (\[eq:sigma\]) and (\[eq:l4l2\]), whence has real eigenvalues.
Decomposing $Lin(W_8, W_7) \stackrel{(\ref{eq:W8W7})}\cong W_8 \oplus W_{48}$, (\[eq:W8W7-summand\]) implies that the elements in the summand congruent to $W_8$ are given by the maps $$A_a: W_8 \longrightarrow W_7, \qquad A_a(b) := {\langle}(\imath_a \lambda^2(e_i))^\#, b{\rangle}e_i$$ for a fixed $a \in W_8$. In order to calculate $\phi_\sigma(A_a)$, observe that ${\mbox{\rm Spin}(7)}$ acts transitively on the unit sphere, whence we may assume w.l.o.g. that $a = e_0$, so that $$A_{e_0}(b) = {\langle}e_i, b{\rangle}e_i = pr_{e_0^\perp}(b),$$ where $pr_{e_0^\perp}: W_8 \rightarrow e_0^\perp = V_7$ is the orthogonal projection. Thus, $$\begin{aligned}
\phi_\sigma(A_{e_0})(b) &=\; \sigma(b, A_{e_0}(e_\mu), e_\mu, e_i)\; e_i = \sigma(b, e_j, e_j, e_i)\; e_i\\
\stackrel{(\ref{eq:l4l2})}=&\;
\ast \Big( b^\flat \wedge e^j \wedge \Big(e^j \wedge e^i \wedge (e^0 \wedge \varphi - \ast_7 \varphi) - e^0 \wedge (\ast_7 e^{ji})\Big) \Big) e_i\\
=&\; \ast \Big( e^0 \wedge b^\flat \wedge e^j \wedge (\ast_7 e^{ij}) \Big) e_i\\
=&\; (1 - \delta_{ij}) \ast (e^0 \wedge b^\flat \wedge \ast_7 e^i) e_i\\
=&\; 6 {\langle}b, e_i{\rangle}e_i = 6 A_{e_0}(b),\end{aligned}$$ so that $\phi_\sigma(A_a) = 6 A_a$ for all $A_a$. By Schur’s lemma and since $\phi_\sigma$ is self-adjoint, $\phi_\sigma|_{W_{48}} = c Id_{W_{48}}$ for some $c \in {{\mathbb R}}$, whence $$0 = tr(\phi_\sigma) = 6 \dim W_8 + c \dim W_{48},$$ and from this, $c = -1$ and the lemma follows.
For a manifold with a ${\mbox{\rm Spin}(7)}$-structure $(M, \Phi)$ and induced metric $g = g_\Phi$, the covariant derivative $g_\Phi$ $\nabla_v \Phi$ w.r.t. the Levi-Civita connection is contained in the infinitesimal orbit of ${{\mathfrak{so}}}(T_pM, g_p)$ [@Bryant1987] and hence in $\Lambda^4_7 T^\ast M$. That is, there is a section $T \in {{\Omega}}^1(M, W_7(M)) = \Gamma(Lin(TM, W_7 (M)))$ such that $$\label{eq:nabla-Spin7}
\nabla_v \Phi = \lambda^4(T(v)) = e^0 \wedge (\imath_{T(v)} \ast \varphi) - (T(v))^\flat \wedge \varphi$$ with the map $\lambda^4: W_7 \to \Lambda^4_7 T_pM$ from (\[def:lambda\]). In analogy to Definition \[def:torsionG2\], we use the following terminology.
\[def:torsionSpin7\] Let $(M, \Phi)$ be a manifold with a ${\mbox{\rm Spin}(7)}$-structure. The section $T \in {{\Omega}}^1(M, W_7(M))$ for which (\[eq:nabla-Spin7\]) holds is called the [*torsion endomorphism* ]{}of the ${\mbox{\rm Spin}(7)}$-structure.
The Frölicher-Nijenhuis brackets on a manifold with a ${\mbox{\rm Spin}(7)}$-structure
--------------------------------------------------------------------------------------
Recall the section $P = - \delta_g \Phi \in {{\Omega}}^3(M, TM)$ on a manifold with a ${\mbox{\rm Spin}(7)}$-structure $(M, \Phi)$ from Definition \[def:P\]. We wish to relate its Frölicher-Nijenhuis bracket to its torsion. In order to do this, recall that $[P,P]^{FN} \in {{\Omega}}^6(M, TM)$.
Due to the decomposition $\Lambda^6 W_8 = \Lambda^6_7 W_8 \oplus \Lambda^6_{21} W_8$ as a $G_2$-module, we may decompose $${{\Omega}}^6(M, TM) = \Gamma(M, \Lambda^6_7 T^*M \otimes TM) \oplus \Gamma(M, \Lambda^6_{21} T^*M \otimes TM),$$ and we denote the projections onto the two summands by $\pi_7$ and $\pi_{21}$, respectively.
\[prop:PP\] Let $(M, \Phi)$ be a manifold with a ${\mbox{\rm Spin}(7)}$-structure with the torsion endomorphism $T \in {{\Omega}}^1(M, W_7(M))$ from (\[eq:nabla-Spin7\]), and let $P = - \delta_g \Phi \in {{\Omega}}^3(M, TM)$ be as before. Then for $p \in M$, $$\label{eq:pi7-Phi}
\pi_7([P, P]_p^{FN}) = -\dfrac23 \Phi \wedge \lambda^2\Big( (4 T_p + \phi_\sigma(T_p))(e_\mu) \Big) \otimes e_\mu.$$ In particular, $\pi_7([P, P]_p^{FN}) = 0$ iff $T_p = 0$.
By Proposition \[prop:FN\], $[P,P]^{FN}_p = \gamma_\mu \otimes e_\mu$, where $\gamma_\mu \in \Lambda^6 T_p^*M$ is given by $$\gamma_\mu = 2 ((\imath_{e_\nu} \Phi) \wedge \imath_{e_\mu} (\nabla_{e_\nu} \Phi) + (\imath_{e_\mu} \imath_{e_\nu} \Phi) \wedge e^\rho \wedge (\imath_{e_\nu} \nabla_{e_\rho}\Phi)).$$
If we decompose $\gamma_\mu = \Phi \wedge \lambda^2(v_\mu) + \gamma_\mu^{21}$ with $\gamma_\mu^{21} \in \Lambda^6_{21} T_p^*M$, then for any $v \in V_7 = e_0^\perp$ we have $\gamma_\mu^{21} \wedge \lambda^2(v) = 0$ by (\[decom-L-W8\]) and hence,
$$\begin{aligned}
\gamma_\mu \wedge \lambda^2(v) =&\; \Phi \wedge \lambda^2(v_\mu) \wedge \lambda^2(v)\\
=&\; (e^0 \wedge \varphi + \ast_7 \varphi) \wedge (e^0 \wedge v_\mu^\flat + \imath_{v_\mu} \varphi) \wedge (e^0 \wedge v^\flat + \imath_v \varphi)\\
=&\; e^0 \wedge \varphi \wedge (\imath_{v_\mu} \varphi) \wedge (\imath_v \varphi) + \ast_7 \varphi \wedge e^0 \wedge v_\mu^\flat \wedge (\imath_v \varphi)\\
&\quad + \ast_7 \varphi \wedge (\imath_{v_\mu} \varphi) \wedge e^0 \wedge v^\flat\\
\stackrel{(\ref{eq:form-1}), (\ref{eq:form-0})}= &\; 6 {\langle}v_\mu, v {\rangle}{\mbox{\rm vol}}+ 3 {\langle}v_\mu, v {\rangle}{\mbox{\rm vol}}+ 3 {\langle}v_\mu, v {\rangle}{\mbox{\rm vol}}= 12 {\langle}v_\mu, v {\rangle}{\mbox{\rm vol}}.\end{aligned}$$
Thus, $$\label{eq:PPl}
\pi_7([P, P]_p^{FN}) = \frac1{12} \ast (\gamma_\mu \wedge \lambda^2(e_i))\; \Phi \wedge \lambda^2(e_i) \otimes e_\mu.$$
For arbitrary $v \in V_7 = e_0^\perp$ we compute
$$\begin{aligned}
\nonumber
\gamma_\mu \wedge \lambda^2(v) =&\; 2 ((\imath_{e_\nu} \Phi) \wedge \imath_{e_\mu} (\nabla_{e_\nu} \Phi) \wedge \lambda^2(v)\\
\nonumber
&\qquad + (\imath_{e_\mu} \imath_{e_\nu} \Phi) \wedge e^\rho \wedge (\imath_{e_\nu} \nabla_{e_\rho}\Phi) \wedge \lambda^2(v))\\
\nonumber
=&\; 2 ((\imath_{e_\nu} \Phi) \wedge \imath_{e_\mu} \lambda^4(T_p(e_\nu)) \wedge \lambda^2(v)\\
\nonumber
&\qquad + (\imath_{e_\mu} \imath_{e_\nu} \Phi) \wedge e^\rho \wedge (\imath_{e_{\nu}} \lambda^4(T_p(e_\rho))) \wedge \lambda^2(v))\\
\nonumber
\nonumber
\stackrel{(\ref{spinformula1}), (\ref{spinformula2})}=& 2 ((-4 \delta_{\mu \nu} {\langle}T(e_\nu), v{\rangle}{\mbox{\rm vol}}+ e^{\nu \mu} \wedge \lambda^4(T_p(e_\nu)) \wedge \lambda^2(v))\\
\nonumber
& + (-12 \delta_{\mu \rho} {\langle}T_p(e_\rho), v {\rangle}{\mbox{\rm vol}}+ e^{\rho \mu} \wedge \lambda^4(T_p(e_\rho)) \wedge \lambda^2(v))\\
\nonumber
=&\; 2 (-16 {\langle}T_p(e_\mu), v{\rangle}{\mbox{\rm vol}}- 2 e^{\mu \nu} \wedge \lambda^4(T_p(e_\nu)) \wedge \lambda^2(v)) \\
\nonumber
=&\; 2 \Big(-16 {\langle}T_p(e_\mu), v{\rangle}- 4 \sigma(e_\mu, T_p(e_\nu), e_\nu, v)\Big) {\mbox{\rm vol}}\\
\stackrel{(\ref{eq:phi-sigma})}=&\; -8 {\langle}(4 T_p + \phi_\sigma(T_p))(e_\mu), v{\rangle}{\mbox{\rm vol}},\end{aligned}$$
and this together with (\[eq:PPl\]) implies (\[eq:pi7-Phi\]) and completes the proof.
With this, we are now ready to prove the following which immediately implies Theorem \[thm:Brackets-Spin-intro\] from the introduction.
\[thm:Brackets-Spin\] Let $(M^8, \Phi)$ be a manifold with a ${\mbox{\rm Spin}(7)}$-structure $(M^8, \Phi)$, let $\nabla$ be the Levi-Civita connection of $g = g_\Phi$, and let $T \in {{\Omega}}^1(M^8, W_7(M^8))$ be its torsion endomorphism defined in Definition \[def:torsionSpin7\]. Then for every $p \in M^8$ the following are equivalent.
1. $T_p = 0 \in T_p^*M^8 \otimes W^7 (M)_p.$
2. The ${\mbox{\rm Spin}(7)}$-structure is torsion-free at $p$, i.e., $(\nabla \Phi)_p = 0$.
3. $\pi_7([P, P]_p^{FN}) = 0 \in \Lambda^6_7 T_p^*M^8 \otimes T_pM^8$.
4. $[P, P]_p^{FN} = 0 \in \Lambda^6 T_p^*M^8 \otimes T_pM^8$.
The equivalence of the first two statements was shown in [@Fernandez1986]. Also, $T_p = 0$ implies $(\nabla \Phi)_p = 0$, whence by Proposition \[prop:FN\], $[P, P]_p^{FN} = 0$, and this trivially implies $\pi_7([P, P]_p^{FN}) = 0$.
By (\[eq:pi7-Phi\]), $\pi_7([P, P]_p^{FN}) = 0$ iff $4 T_p + \phi_\sigma(T_p) = 0$, and since $\phi_\sigma$ does not have $-4$ as an eigenvalue by Lemma \[lem:phisigma\], this implies that $T_p = 0$.
The 16 classes of $G_2$- and 4 classes of ${\mbox{\rm Spin}(7)}$-structures {#subs:16classesG2}
===========================================================================
In this section, we shall interpret the classification of $G_2$-structures and of ${\mbox{\rm Spin}(7)}$-structures ([@FG1982] and [@Fernandez1986]) in terms of the Frölicher-Nijenhuis bracket.
For the $G_2$-case, this classification is given by determining which components of the torsion endomorphism $T$ vanish, where $T$ is regarded as a section of the endomorphism bundle ${\mbox{\rm End}}(TM^7)$ which is $G_2$-equivariantly isomorphic to $$\label{eq:decomp-endo}
V_7(M^7) \otimes V_7(M^7) \cong V_1(M^7) \oplus V_7(M^7) \oplus V_{14}(M^7) \oplus V_{27}(M^7).$$ Since this decomposition has $4$ summands, the classification consists of $2^4 = 16$ cases.
Observe that both $\Lambda^5_7 T^\ast M^7 \otimes TM^7$ and $\Lambda^6 T^\ast M^7 \otimes TM^7$ are $G_2$-equivariantly isomorphic to $V_7(M^7) \otimes V_7(M^7)$, where explicit isomorphisms are given by\
$$\begin{array}{cccccccc}
K: & \Lambda^5_7 T^\ast M^7 \otimes TM^7 & \ni & (\ast \varphi \wedge v^\flat) \otimes w & \longmapsto & v^\flat \otimes w & \in & T^\ast M^7 \otimes TM^7\\[2mm]
L: & \Lambda^6 T_p^\ast M^7 \otimes TM^7 & \ni & (\ast v^\flat) \otimes w & \longmapsto & v^\flat \otimes w & \in & T^\ast M^7 \otimes TM^7.
\end{array}$$
If $(M^7, \varphi)$ is a manifold with a $G_2$-structure and the cross products $Cr$ and $\chi$, then we define the sections $$K_{\pi_7([Cr, \chi]^{FN})},\; L_{[\chi, \chi]^{FN}} \in \Gamma({\mbox{\rm End}}(TM)).$$ Therefore, by Propositions \[prop:Cr-chi\], \[prop:chi-chi\] there are $G_2$-equivariant vector bundle isomorphisms $$\begin{aligned}
\tau_1, \tau_2: {\mbox{\rm End}}(TM^7) & \longrightarrow {\mbox{\rm End}}(TM^7)\end{aligned}$$ such that for the torsion endoromphism $T \in \Gamma({\mbox{\rm End}}(TM))$ we have $$\label{eq:tor-FNG2}
{}\tau_1(T) = K_{\pi_7([Cr, \chi]^{FN})} \qquad \mbox{and} \qquad \tau_2(T) = L_{[\chi, \chi]^{FN}},$$ where by a slight abuse of notation we denote the map $\tau_i: \Gamma({\mbox{\rm End}}(TM)) \to \Gamma({\mbox{\rm End}}(TM))$ applying $\tau_i$ pointwise by the same symbol.
For an element $A = a_{ij}e^i \otimes e_j \in {\mbox{\rm End}}(V_7)$ let us denote its skew-symmetrization by $$\sigma_A := a_{ij} e^{ij} \in \Lambda^2 V_7^\ast.$$ With this notation, it follows from (\[eq:pi-Crchi\]) and (\[eq:pi-chichi\]) that $\tau_1$ and $\tau_2$ take the form $$\begin{aligned}
\tau_1(T) &= 2 \Big(T - 2 T^\top - tr(T) id \Big),\\
\tau_2(T) &= -4 (T + T^\top) + 6 \ast (e^i \wedge \sigma_T \wedge \varphi) \otimes e_i,\end{aligned}$$ summing over some basis $(e_i)$ with dual basis $(e^i)$.
The $G_2$-equivariance of $\tau_1$ and $\tau_2$ and (\[eq:tor-FNG2\]) now implies that the $V_k(M)$-component of $T$ vanishes if and only if the $V_k(M)$-component of $K_{\pi_7([Cr, \chi]^{FN})}$ vanishes if and only if the $V_k(M)$-component of $L_{[\chi, \chi]^{FN}}$ vanishes. Since the cases in the Fernandez-Gray classification are determined by the vanishing of the components of $T$, we obtain the interpretation of these cases given in Table \[table-G21\].
Classes $\begin{array}{c}\mbox{Relation on}\\ A \in \{ T, K_{\pi_7([Cr, \chi]^{FN})}, L_{[\chi, \chi]^{FN}}\} \end{array}$
---------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------- -- --
$V_1(M) \oplus V_7(M) \oplus V_{14}(M) \oplus V_{27}(M)$ no relation
$V_7(M) \oplus V_{14}(M) \oplus V_{27}(M)$ $tr(A) = 0$
$V_1(M) \oplus V_{14}(M) \oplus V_{27}(M)$ $\sigma_A \in {{\Omega}}^2_{14}(M)$
$V_1(M) \oplus V_7(M) \oplus V_{27}(M)$ $\sigma_A \in {{\Omega}}^2_7(M)$
$V_1(M) \oplus V_7(M) \oplus V_{14}(M)$ $A + A^\top - \frac{2}{7} tr (A) {\rm id}_{TM} = 0$
$V_{14}(M) \oplus V_{27}(M)$ $\sigma_A \in {{\Omega}}^2_{14}(M), \quad tr(A) = 0$
$V_7(M) \oplus V_{27}(M)$ $\sigma_A \in {{\Omega}}^2_7(M), \quad tr(A) = 0$
$V_7(M) \oplus V_{14}(M)$ $A + A^\top=0$
$V_1(M) \oplus V_{27}(M)$ $A - A^\top=0$
$V_1(M) \oplus V_{14}(M)$ $A + A^\top - \frac{2}{7} tr (A) {\rm id}_{TM} = 0, \quad \sigma_A \in {{\Omega}}^2_{14}(M)$
$V_1(M) \oplus V_7(M)$ $A + A^\top - \frac{2}{7} tr (A) {\rm id}_{TM} = 0, \quad \sigma_A \in {{\Omega}}^2_7(M)$
$V_{27}(M)$ $A - A^\top=0, \quad tr(A) = 0$
$V_{14}(M)$ $A+A^\top = 0, \quad \sigma_A \in {{\Omega}}^2_{14}(M)$
$V_7(M)$ $A+A^\top = 0, \quad \sigma_A \in {{\Omega}}^2_7(M)$
$V_1(M)$ $A = \frac{1}{7} tr (A) {\rm id}_{TM}$
$\{ 0 \}$ $A = 0$
: []{data-label="table-G21"}
The interpretation of manifolds $(M^8, \Phi)$ with a ${\mbox{\rm Spin}(7)}$-structure is analogous. Again, the torsion $T$ and the projection $\pi_7([P, P]^{FN})$ are sections of the ${\mbox{\rm Spin}(7)}$-equivariantly isomorphic bundles $T^\ast M^8 \otimes W_7(M^8)$ and $\Lambda^6_7 T^\ast M^8 \otimes TM^8$, respectively, with an explicit identification given by $$H: \Lambda^6_7 T^\ast M^8 \otimes W_7(M^8) \ni \Phi \wedge (\lambda^2(a)) \otimes v \longmapsto
v^\flat \otimes a \in T^\ast M^8 \otimes W_7(M^8),$$ and if $(M^8, \Phi)$ is a manifold with a ${\mbox{\rm Spin}(7)}$-structure and the $3$-fold product $P$, then by (\[eq:pi7-Phi\]) $$H_{\pi_7([P, P]^{FN})} = \tau_3(T), \qquad \mbox{where} \qquad \tau_3(T) = -\dfrac23 (4T + \phi_\sigma(T)).$$ Here, by abuse of notation we regard $\phi_\sigma$ as the pointwise application of the map from (\[eq:phi-sigma\]) to sections of $W_7(M^8) \otimes T^\ast M^8 \cong Lin(W_8(M^8), W_7(M^8))$.
By (\[eq:W8W7\]), $W_7(M^8) \otimes T^\ast M^8$ can be decomposed as $W_8(M^8) \oplus W_{48}(M^8)$ whence by the ${\mbox{\rm Spin}(7)}$-equivariance of $\tau_3$, the $W_k(M^8)$-component of $T$ vanishes if and only if the $W_k(M^8)$-component of $H_{\pi_7([P,P]^{FN})}$ does. Since the classification of Fernández [@Fernandez1986] into $2^2 = 4$ different cases is given by the vanishing of the components of the torsion $T$, it follows that these cases can be also interpreted by the vanishing of the components of $H_{\pi_7([P,P]^{FN})}$, which leads to the interpretation of the classes of ${\mbox{\rm Spin}(7)}$-manifolds given in Table \[table-Spin7\], where $pr_k: W_7(M^8) \otimes T^\ast M^8 \to W_k(M^8)$ is the canonical projection.
Classes $\begin{array}{c}\mbox{Relation on}\\ A \in \{ T, H_{\pi_7([P, P]^{FN})}\} \end{array}$
--------------------- ----------------------------------------------------------------------------------------- -- --
$W_8 \oplus W_{48}$ no relation
$W_{48}$ $pr_8 (A) = 0$
$W_8$ $pr_{48} (A) = 0$
$\{ 0 \}$ $A = 0$
: []{data-label="table-Spin7"}
Appendix
========
In this appendix, we shall collect some of the formulas which we needed in the calculations in this paper. Most of them are known and can be found in a similar form e.g. in [@SW2017 Lemma 4.37], but we shall collect them here for the reader’s convenience.
\[formulas\] For all $u,v,w,r \in V_7$ and any orthonormal basis $(e_i)$ of $V_7$ the following identities hold. $$\begin{aligned}
\label{eq:form-0}
u^\flat \wedge (\imath_v \varphi) \wedge \ast \varphi =\;& 3 {\langle}u, v{\rangle}{\mbox{\rm vol}}\\
\label{eq:form-11}
u^\flat \wedge (\imath_v \ast \varphi) \wedge \varphi =\;& 4 {\langle}u, v {\rangle}{\mbox{\rm vol}}\\
\label{eq:form-1}
(\imath_u \varphi) \wedge (\imath_{v} \varphi) \wedge \varphi =\;& 6 {\langle}u, v{\rangle}{\mbox{\rm vol}}\\
\label{eq:form-2}
u^\flat \wedge (\imath_{v} \varphi) \wedge (\imath_{w} \varphi) \wedge (\imath_r \varphi) =\;& 2 \Big({\langle}v,w {\rangle}{\langle}u,r {\rangle}\\
\nonumber
&+ {\langle}u,v{\rangle}{\langle}w, r{\rangle}+ {\langle}u,w{\rangle}{\langle}v, r {\rangle}\Big) {\mbox{\rm vol}}\\
\label{eq:form-7}
(\imath_{u}\imath_{v} \ast \varphi) \wedge w^\flat \wedge \ast \varphi =\;& -2 u^\flat \wedge v^\flat \wedge w^\flat \wedge \ast \varphi\\
\label{eq:form-3}
(\imath_{u}\imath_{v} \varphi) \wedge w^\flat \wedge (\imath_r \varphi) \wedge \varphi =\;& 2 ({\langle}v,w{\rangle}{\langle}u,r{\rangle}- {\langle}u,w{\rangle}{\langle}v,r{\rangle}) {\mbox{\rm vol}}\\ \nonumber & - 2 u^\flat \wedge v^\flat \wedge w^\flat \wedge r^\flat \wedge \varphi\\
\label{eq:form-4}
(\imath_{u} \imath_{v} \ast \varphi) \wedge (\imath_{w} \ast \varphi) \wedge (\imath_r \varphi) =\;& 2({\langle}v,w{\rangle}{\langle}u,r{\rangle}- {\langle}u,w{\rangle}{\langle}v,r{\rangle}) {\mbox{\rm vol}}\\ \nonumber & + u^\flat \wedge v^\flat \wedge w^\flat \wedge r^\flat \wedge \varphi\\
\label{eq:form-5}
u^\flat \wedge v^\flat \wedge (\imath_w \ast \varphi) \wedge (\imath_r \varphi) =\;&2({\langle}v,w{\rangle}{\langle}u,r{\rangle}- {\langle}u,w{\rangle}{\langle}v,r{\rangle}) {\mbox{\rm vol}}\\ \nonumber & + u^\flat \wedge v^\flat \wedge w^\flat \wedge r^\flat \wedge \varphi\\
\label{eq:form-6}
u^\flat \wedge v^\flat \wedge (\imath_w \imath_r \ast \varphi) \wedge \varphi =\;&2({\langle}v,w{\rangle}{\langle}u,r{\rangle}- {\langle}u,w{\rangle}{\langle}v,r{\rangle}) {\mbox{\rm vol}}\\ \nonumber & - u^\flat \wedge v^\flat \wedge w^\flat \wedge r^\flat \wedge \varphi\\
\label{eq:form13}
(\imath_u \ast \varphi) \wedge (\imath_v \varphi) =\;&
-2 \ast ( u^\flat \wedge v^\flat )+ u^\flat \wedge v^\flat \wedge \varphi\\
\label{eq:form14}
\varphi \wedge (\imath_u \varphi) =\;& 2 u^\flat \wedge \ast \varphi\\
\label{formulas2} (\imath_{e_i} \varphi)^2 =\;& 6 \ast \varphi\\
\label{formulas-3} (\imath_u \imath_{e_i} \ast \varphi) \wedge (\imath_v \imath_{e_i} \ast \varphi) =\;& 2 (\imath_u \varphi) \wedge (\imath_v \varphi)\\
\label{formulas-4}
(\imath_u \imath_{e_i} \ast \varphi) \wedge (\imath_{e_i} \varphi) =\;& 3 u^\flat \wedge \varphi$$
For the proof of these identities, observe that the left hand side of each equation is a $G_2$-invariant element of some tensor power of $V_7$, and therefore it has to be a linear combination of the summands on the right hand side; the coefficients of this linear combination then can be determined by using the explicit formulas for $\varphi$ and $\ast \varphi$ in (\[varphi\]) and (\[varphi\*\]).
To pick one explicit example which is not among the identities shown in [@SW2017], observe that the left hand side of (\[eq:form-2\]) is an element of $(V_7 \otimes \odot^3 V_7)^{G_2}$. Since $\odot^3 V_7 \cong V_7 \oplus V_{77}$, we have $\dim (V_7 \otimes \odot^3 V_7)^{G_2} = 1$, and there is one $G_2$-invariant element of $V_7 \otimes \odot^3 V_7$ given by deriving the square of the scalar product which lies in $\odot^4 V_7$. Thus, $$\begin{aligned}
u^\flat \wedge (\imath_{v} \varphi) \wedge (\imath_{w} \varphi) \wedge (\imath_r \varphi) =\;& c \Big({\langle}v,w {\rangle}{\langle}u,r {\rangle}\\
\nonumber &+ {\langle}u,v{\rangle}{\langle}w, r{\rangle}+ {\langle}u,w{\rangle}{\langle}v, r {\rangle}\Big) {\mbox{\rm vol}}.\end{aligned}$$ for some constant $c \in {{\mathbb R}}$. Now setting $u = v = w = r =: e_1$ and using (\[varphi\]) implies that $c = 2$, showing (\[eq:form-2\]).
The remaining identities are shown in a similar fashion.
The following two decompositions of $G_2$- and ${\mbox{\rm Spin}(7)}$-representations is also well known, cf. [@SW2017 Theorem 8.5, 9.8], [@Kar2005 (4.7), (4.8)].
\[Lambda2-V7\] Decompose $\Lambda^2 V_7^\ast = \Lambda^2_7 V_7^\ast \oplus \Lambda^2_{14} V_7^\ast$ according to (\[decom-L-V7\]). Then $$\begin{aligned}
\Lambda^2_7 V_7^\ast &= \{ \alpha^2 \in \Lambda^2 V_7^\ast \mid \ast(\alpha^2 \wedge \varphi) = 2 \alpha^2\}, \quad \mbox{and}\\
\Lambda^2_{14} V_7^\ast &= \{ \alpha^2 \in \Lambda^2 V_7^\ast \mid \ast(\alpha^2 \wedge \varphi) = - \alpha^2\}.\end{aligned}$$ In particular, $$\label{eq:describe-Lamb2V7}
\begin{array}{lll}
\Lambda^2_7 V_7^\ast &= &\{ \alpha^2 + \ast(\alpha^2 \wedge \varphi) \mid \alpha^2 \in \Lambda^2 V_7^\ast\}, \quad \mbox{and}\\[2mm]
\Lambda^2_{14} V_7^\ast &= &\{ 2 \alpha^2 - \ast(\alpha^2 \wedge \varphi) \mid \alpha^2 \in \Lambda^2 V_7^\ast\}.
\end{array}$$
\[Kar47-48\] Decompose $\Lambda^2 W_8^\ast = \Lambda^2_7 W_8^\ast \oplus \Lambda^2_{21} W_8^\ast$ according to (\[decom1-L-W8\]). Then $$\begin{aligned}
\Lambda^2_7 W_8^\ast &= \{ \alpha^2 \in \Lambda^2 W_8^\ast \mid \Phi \wedge \alpha^2 = 3 \ast \alpha^2\}, \quad \mbox{and}\\
\Lambda^2_{21} W_8^\ast &= \{ \alpha^2 \in \Lambda^2 W_8^\ast \mid \Phi \wedge \alpha^2 = - \ast \alpha^2\}.\end{aligned}$$
We shall also need the following result.
For all $u, v \in W_7$ and $a, b \in W_8$ the following formulas hold: $$\begin{aligned}
\nonumber
&(\imath_a \Phi) \wedge (\imath_{b} \lambda^4(u)) \wedge \lambda^2(v) =\\
\label{spinformula1}&\qquad -4 {\langle}a, b {\rangle}_{W_8} {\langle}u,v {\rangle}_{W_7} {\mbox{\rm vol}}+ a^\flat \wedge b^\flat \wedge \lambda^4(u) \wedge \lambda^2(v),\\
\nonumber
&a^\flat \wedge (\imath_b \imath_{e_\mu} \Phi) \wedge (\imath_{e_\mu} \lambda^4(u)) \wedge \lambda^2(v) =\\
\label{spinformula2} &\qquad -12 {\langle}a, b {\rangle}_{W_8} {\langle}u,v {\rangle}_{W_7} {\mbox{\rm vol}}+ a^\flat \wedge b^\flat \wedge \lambda^4(u) \wedge \lambda^2(v)\end{aligned}$$ where in (\[spinformula2\]) the sum is taken over an orthonormal basis $(e_\mu)$ of $W_8$.
By (\[decom1-L-W8\]) and (\[decom-Sk-W7\]), the decomposition of $W_8 \otimes W_8$ and $W_7 \otimes W_7$ into ${\mbox{\rm Spin}(7)}$-irreducible summands yields $$\begin{aligned}
W_8 \otimes W_8 &= \odot^2 W_8 \oplus \Lambda^2 W_8 \cong (W_1 \oplus W_{35}) \oplus (W_7 \oplus W_{21}),\\
W_7 \otimes W_7 &= \odot^2 W_7 \oplus \Lambda^2 W_7 \cong (W_1 \oplus W_{27}) \oplus W_{21},\end{aligned}$$ so that there are two inequivalent summands in common and hence, the space of ${\mbox{\rm Spin}(7)}$-invariant tensors in $W_8 \otimes W_8 \otimes W_7 \otimes W_7$ is $2$-dimensional. Since the left hand sides of (\[spinformula1\]) and (\[spinformula2\]) describe such tensors, it follows that there must be constants $c_1, \ldots, c_4 \in {{\mathbb R}}$ such that $$\begin{aligned}
\nonumber
&(\imath_a \Phi) \wedge (\imath_{b} \lambda^4(u)) \wedge \lambda^2(v) =\\
\label{spinformula1c}&\qquad c_1 {\langle}a, b {\rangle}_{W_8} {\langle}u,v {\rangle}_{W_7} {\mbox{\rm vol}}+ c_2 a^\flat \wedge b^\flat \wedge \lambda^4(u) \wedge \lambda^2(v),\\
\nonumber
&a^\flat \wedge (\imath_b \imath_{e_\mu} \Phi) \wedge (\imath_{e_\mu} \lambda^4(u)) \wedge \lambda^2(v) =\\
\label{spinformula2c} &\qquad c_3 {\langle}a, b {\rangle}_{W_8} {\langle}u,v {\rangle}_{W_7} {\mbox{\rm vol}}+ c_4 a^\flat \wedge b^\flat \wedge \lambda^4(u) \wedge \lambda^2(v).\end{aligned}$$
In order to determine these constants, we first let $a = b := e_0$, so that $$\begin{aligned}
(\imath_{e_0} \Phi) \wedge (\imath_{e_0} \lambda^4(u)) \wedge \lambda^2(v) =& \varphi \wedge (\imath_u \ast_7 \varphi) \wedge (e^0 \wedge v^\flat + (\imath_v \varphi))\\
=& -e^0 \wedge v^\flat \wedge (\imath_u \ast_7 \varphi) \wedge \varphi\\
\stackrel{(\ref{eq:form-11})}=& -4 {\langle}u,v {\rangle}{\mbox{\rm vol}},\end{aligned}$$ and from this, $c_1 = -4$ follows. Whence if we let $a := e_0$ and $b, u, v \in W_7 = e_0^\perp$ then $$\begin{aligned}
(\imath_{e_0} \Phi) \wedge (\imath_b \lambda^4(u)) \wedge \lambda^2(v) \stackrel{(\ref{def:lambda})}=& \varphi \wedge (-e^0 \wedge (\imath_b \imath_u \ast_7 \varphi) - \imath_b (u^\flat \wedge \varphi))\\ & \qquad \wedge (e^0 \wedge v^\flat + (\imath_v \varphi))\\
=& - \varphi \wedge e^0 \wedge (\imath_b \imath_u \ast_7 \varphi) \wedge (\imath_v \varphi)\\
& \qquad + \varphi \wedge u^\flat \wedge (\imath_b \varphi) \wedge e^0 \wedge v^\flat\\
\stackrel{(\ref{eq:form14})}=& 2 e^0 \wedge (\imath_b \imath_u \ast_7 \varphi) \wedge v^\flat \wedge \ast_7 \varphi\\
& \qquad + 2 b^\flat \wedge \ast_7 \varphi \wedge u^\flat \wedge e^0 \wedge v^\flat\\
\stackrel{(\ref{eq:form-7})}=& -2 e^0 \wedge b^\flat \wedge u^\flat \wedge v^\flat \wedge \ast_7 \varphi\\
\stackrel{(\ref{eq:l4l2})}=& e^0 \wedge b^\flat \wedge \lambda^4(u) \wedge \lambda^2(v),\end{aligned}$$ so that $c_2 = 1$ follows. Now substituting $a = b := e_0$ and $u = v := e_1$ into (\[spinformula2c\]) and using the index $i$ to run from $1$ to $7$ yields $$\begin{aligned}
e^0 \wedge (-\imath_{e_i} \varphi) \wedge (\imath_{e_i} (-e^1 \wedge \varphi)) \wedge (\imath_{e_1} \varphi) =& e^0 \wedge (\imath_{e_i} \varphi) \wedge (\delta_{1i} \varphi - e^1 \wedge (\imath_{e_i} \varphi)) \wedge (\imath_{e_1} \varphi)\\
=& e^0 \wedge (\imath_{e_1} \varphi) \wedge (\imath_{e_1} \varphi) \wedge \varphi\\
&\qquad - e^0 \wedge e^1 \wedge (\imath_{e_i} \varphi) \wedge (\imath_{e_i} \varphi) \wedge \imath_{e_1} \varphi\\
\stackrel{(\ref{eq:form-1}),(\ref{formulas2})}=& 6 {\mbox{\rm vol}}- 6 e^0 \wedge e^1 \wedge \imath_{e_1} \varphi\wedge \ast \varphi\\
&\stackrel{(\ref{eq:form-0})}= -12 {\mbox{\rm vol}},\end{aligned}$$ so that $c_3 = -12$ follows. Finally, for $a := e_0$, $u := e_1$ and $b, v \in W_7 = e_0^\perp$ (\[spinformula2c\]) reads $$\begin{aligned}
e^0 \wedge (\imath_b \imath_{e_\mu} \Phi) \wedge &(\imath_{e_\mu} \lambda^4(e_1)) \wedge \lambda^2(v)\\
=&\; e^0 \wedge (\imath_b \varphi) \wedge (\imath_{e_1} \ast_7 \varphi) \wedge (\imath_v \varphi)\\
& + e^0 \wedge (\imath_b \imath_{e_i} \ast_7 \varphi) \wedge (-\imath_{e_i} (e^1 \wedge \varphi)) \wedge (\imath_v \varphi)\\
\stackrel{(*)}=&\; 0 + e^0 \wedge (\imath_b \imath_{e_i} \ast_7 \varphi) \wedge (-\delta_{1i} \varphi + e^1 \wedge (\imath_{e_i} \varphi)) \wedge (\imath_v \varphi)\\
=&\; -e^0 \wedge (\imath_b \imath_{e_1} \ast_7 \varphi) \wedge \varphi \wedge (\imath_v \varphi)\\
& + e^{01} \wedge (\imath_b \imath_{e_i} \ast_7 \varphi) \wedge (\imath_{e_i} \varphi) \wedge (\imath_v \varphi)\\
\stackrel{(\ref{eq:form14}), (\ref{formulas-4})}=&\; -2 e^0 \wedge (\imath_b \imath_{e_1} \ast_7 \varphi) \wedge v^\flat \wedge \ast_7 \varphi + 3 e^{01} \wedge b^\flat \wedge \varphi \wedge (\imath_v \varphi)\\
\stackrel{(\ref{eq:form14}), (\ref{eq:form-7})}=&\; 4 e^0 \wedge b^\flat \wedge e^1 \wedge v^\flat \wedge \ast_7 \varphi + 6 e^{01} \wedge b^\flat \wedge v^\flat \wedge \ast_7 \varphi\\
=&\; -2 e^0 \wedge b^\flat \wedge e^1 \wedge v^\flat \wedge \ast_7 \varphi\\
\stackrel{(\ref{eq:l4l2})}=& e^0 \wedge b^\flat \wedge \lambda^4(e_1) \wedge \lambda^2(v),\end{aligned}$$ so that $c_4 = 1$ follows. At $(*)$ we have used that the map $$(u,v,w) \longmapsto \ast \Big((\imath_u \varphi) \wedge (\imath_v \ast_7 \varphi) \wedge (\imath_w \varphi)\Big)$$ is a $G_2$-invariant element of $W_7 \otimes \odot^2 W_7$, and since by (\[decom-Sk-W7\]) $\odot^2 W_7 \cong W_1 \oplus W_{27}$ has no irreducible component isomorphic to $W_7$, there is no such element other than $0$.
Acknowledgement {#acknowledgement .unnumbered}
---------------
A part of this project has been discussed during HVL’s visit to the Osaka City University in December 2015. She thanks Professor Ohnita for his invitation to Osaka and his hospitality. HVL and LS also thank the Max Planck Institute for Mathematics in the Sciences in Leipzig for its hospitality during extended visits. We also thank the referee for many helpful comments which enabled us to significantly improve the manuscript.
[99999]{} , Vector cross product, Comment. Math. Helv. 42(1967), 222-236.
, Metrics with exceptional holonomy, Ann. of Math. 126(1987), 525-576.
, Some remarks on $G_2$-structures, Proceedings of Gökova Geometry-Topology Conference 2005, 75-109, Gökova Geometry/Topology Conference (GGT), Gökova, 2006.
, A Classification of Riemannian Manifolds with Structure Group ${\mbox{\rm Spin}(7)}$. Annali di Mat. Pura Appl, 143(1986), pp. 101-122.
, Riemannian manifolds with structure group $G_{2}$, Annali di Mat. Pura Appl. 32 (1982), 19-45.
, Theory of vector-valued differential forms. I. Derivations of the graded ring of differential forms., Indag. Math. 18 (1956), 338-359.
, Some new cohomology invariants for complex manifolds. I, II, Indag. Math. 18 (1956), 540-552, 553-564.
, Calibrated geometry, Acta Math. 148(1982), 47-157.
, Introduction to Lie algebras and representation theory, second printing, Graduate Texts in Mathematics, 9. Springer-Verlag, New York-Berlin, 1978.
, Deformations of $G_2$ and ${\mbox{\rm Spin}(7)}$-structures, Can. J. Math. 57(2005), 1012-1055.
, Deformations of homogeneous associative submanifolds in nearly parallel $G_2$-manifolds, to appear in Asian J. Math., arXiv:1407.8046.
, Natural operators in differential geometry, Springer 1993.
, McLean’s second variation formula revisited, J. Geom. Phys, 113(2017), 188-196, arXiv:1605.01267.
, Deformations of Calibrated Submanifolds, Comm. in Analysis and Geom. 6 (1998), 705-747.
, Notes on the octonions, arXiv: 1005.2820.
[^1]: The first named author is supported Grant-in-Aid for JSPS fellows (26-7067), the second named author is partially supported by RVO: 67985840
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Frances Y. Kuo'
- Dirk Nuyens
- 'Frances Y. Kuo[^1]'
- 'Dirk Nuyens[^2]'
bibliography:
- 'mybibfile.bib'
title:
- 'Hot New Directions for Quasi-Monte Carlo Research in Step with Applications'
- |
Hot New Directions for Quasi-Monte Carlo Research\
in Step with Applications
---
Introduction
============
High dimensional computation is a new frontier in scientific computing, with applications ranging from financial mathematics such as option pricing or risk management, to groundwater flow, heat transport, and wave propagation. A tremendous amount of progress has been made in the past two decades on the theory and application of *quasi-Monte Carlo* (*QMC*) *methods* for approximating high dimensional integrals. See e.g., the classical references [@Nie92; @SJ94] and the recent books [@DP10; @Lem09; @LP14]. One key element is the *fast component-by-component construction* [@CKN06; @Nuy14; @NC06a; @NC06b] which provides parameters for first order or *higher order QMC methods* [@DKS13; @DP10] for sufficiently smooth functions. Another key element is the careful selection of parameters called *weights* [@SWW04; @SW98] to ensure that the worst case errors in an appropriately weighted function space are bounded independently of the dimension. The dependence on dimension is very much the focus of the study on *tractability* [@NW10] of multivariate problems.
We are particularly keen on the idea that new theory and methods for high dimensional computation are developed not in isolation, but in close connection with applications. The theoretical QMC convergence rates depend on the appropriate pairing between the function space and the class of QMC methods. Practitioners are free to choose the theoretical setting or pairing that is most beneficial for their applications, i.e., to achieve the best possible convergence rates under the weakest assumptions on the problems. As QMC researchers we take application problems to be our guide to develop new theory and methods as the needs arise. This article provides an overview of some interfaces between such theory and applications.
We begin in Section \[sec:settings\] by summarizing three theoretical settings. The first setting is what we consider to be the standard QMC setting for integrals formulated over the unit cube. Here the integrand is assumed to have square-integrable mixed first derivatives, and it is paired with *randomly shifted lattice rules* [@SKJ02b] to achieve first order convergence. The second setting is for integration over $\bbR^s$ against a product of univariate densities. Again the integrands have square-integrable mixed first derivatives and we use randomly shifted lattice rules to achieve first order convergence. The third setting returns to the unit cube, but considers integrands with higher order mixed derivatives and pairs them with *interlaced polynomial lattice rules* [@God15] which achieve higher order convergence. These three settings are discussed in more detail in [@KN16].
Next in Section \[sec:apps\] we outline three applications of QMC methods: option pricing, GLMM (generalized linear mixed models) maximum likelihood, PDEs with random coefficients – all with quite different characteristics and requiring different strategies to tackle them. We explain how to match each example application with an appropriate setting from Section \[sec:settings\]. In the option pricing application, see e.g., [@ABG98; @GKSW08; @GilWat09], none of the settings is applicable due to the presence of a *kink*. We discuss the strategy of *smoothing by preintegration* [@GKLS], which is similar to the method known as conditional sampling [@ACN13]. In the maximum likelihood application [@KDSWW08], the change of variables plays a crucial role in a similar way to importance sampling for Monte Carlo methods. In the PDE application, see e.g., [@CD15; @GWZ14; @KN16; @SG11], the *uniform* and the *lognormal* cases correspond to integration over the unit cube and $\bbR^s$, respectively, and the two cases tap into different QMC settings. For the lognormal case we briefly contrast three ways to generate the random field: *Karhunen–Loève expansion*, *circulant embedding* [@DN97; @GKNSS11; @GKNSS-paper1; @GKNSS-paper2], and *$H$-matrix technique* [@FKS; @Hac15].
Then in Section \[sec:savings\] we discuss three different cost saving strategies that can be applied to all of the above applications. First, *multi-level methods* [@Gil15] restructure the required computation as a telescoping sum and tackle different levels separately to improve the overall cost versus error balance, while more general *multi-index methods* [@HNT16] allow different criteria to be considered simultaneously in a multi-index telescoping sum. Second, the *multivariate decomposition methods* [@KNPSW; @KSWW10a; @Was13] work in a similar way by making an explicit decomposition of the underlying function into functions of only subsets of the variables [@KSWW10b]. The third strategy is *fast QMC matrix-vector multiplication* which carries out the required computation for multiple QMC samples at the same time using an FFT [@DKLS15].
We provide pointers to some software resources in Section \[sec:software\] and conclude the article in Section \[sec:summary\] with a summary and an outlook to future work. An overview of the various components of this article is given in Figure \[fig:overview\].
\[fig:overview\]
{width="100.00000%"} {width="80.00000%"}
Three Settings {#sec:settings}
==============
Here we describe three theoretical function space settings paired with appropriate QMC methods. These three setting are also covered in [@KN16]. Of course these three pairs are not the only possible combinations. We selected them due to our preference for constructive QMC methods that achieve the best possible convergence rates, with the implied constant independent of dimension, under the weakest possible assumptions on the integrands.
Setting 1: Standard QMC for the Unit Cube {#sec:setting1}
-----------------------------------------
For $f$ a real-valued function defined over the $s$-dimensional unit cube $[0,1]^s$, with $s$ finite and fixed, we consider the integral $$\begin{aligned}
\label{eq:int1}
I(f)
\,=\, \int_{[0,1]^s} f(\bsy)\, \rd\bsy\,.\end{aligned}$$
### Weighted Sobolev spaces {#weighted-sobolev-spaces .unnumbered}
We assume in this standard setting that the integrand $f$ belongs to a *weighted Sobolev space of smoothness one* in the unit cube $[0,1]^s$. Here we focus on the *unanchored* variant in which the norm is defined by, see also [@SWW04], $$\begin{aligned}
\label{eq:norm1}
\|f\|_\bsgamma
\,=\,
\Bigg[
\sum_{\setu\subseteq\{1:s\}}
\frac{1}{\gamma_\setu}
\int_{[0,1]^{|\setu|}}
\bigg(\int_{[0,1]^{s-|\setu|}}
\frac{\partial^{|\setu|}f}{\partial \bsy_\setu}(\bsy)
\,\rd\bsy_{\{1:s\}\setminus\setu}
\bigg)^2
\rd\bsy_\setu
\Bigg]^{1/2},\end{aligned}$$ where $\{1:s\}$ is a shorthand notation for the set of indices $\{1,2,\ldots,s\}$, $(\partial^{|\setu|}f)/(\partial \bsy_\setu)$ denotes the mixed first derivative of $f$ with respect to the “active” variables $\bsy_\setu = (y_j)_{j\in\setu}$, while $\bsy_{\{1:s\}\setminus\setu} =
(y_j)_{j\in\{1:s\}\setminus\setu}$ denotes the “inactive” variables.
There is a weight parameter $\gamma_\setu\ge 0$ associated with each subset of variables $\bsy_\setu$ to model their relative importance. We denote the weights collectively by $\bsgamma$. Special forms of weights have been considered in the literature. *POD weights* (*product and order dependent weights*), arisen for the first time in [@KSS12], take the form $$\gamma_\setu \,=\, \Gamma_{|\setu|}\,\prod_{j\in\setu} \Upsilon_j\,,$$ which is specified by two sequences $\Gamma_0= \Gamma_1=1,
\Gamma_2,\Gamma_3,\ldots\ge 0$ and $\Upsilon_1\ge\Upsilon_2\ge \cdots
> 0$. Here the factor $\Gamma_{|\setu|}$ is said to be order dependent because it is determined solely by the cardinality of $\setu$ and not the precise indices in $\setu$. The dependence of the weight $\gamma_\setu$ on the indices $j\in\setu$ is controlled by the product of terms $\Upsilon_j$. Each term $\Upsilon_j$ in the sequence corresponds to one coordinate direction; the sequence being non-increasing indicates that successive coordinate directions become less important. Taking all $\Gamma_{|\setu|}=1$ or all $\Upsilon_j = 1$ corresponds to the weights known as *product weights* or *order dependent weights* in the literature [@SWW04; @SW98].
### Randomly shifted lattice rules {#randomly-shifted-lattice-rules .unnumbered}
We pair the weighted Sobolev space with *randomly shifted lattice rules*; the complete theory can be found in [@DKS13]. Randomly shifted lattice rules approximate the integral by $$\begin{aligned}
\label{eq:lat}
Q(f)
\,=\, \frac{1}{n} \sum_{i=1}^n f(\bst_i),\qquad
\bst_i \,=\, \left\{\frac{i\,\bsz}{n} + \bsDelta\right\}\,,\end{aligned}$$ where $\bsz\in \bbZ^s$ is known as the *generating vector*, $\bsDelta$ is a *random shift* drawn from the uniform distribution over $[0,1]^s$, and the braces indicate that we take the fractional parts of each component in a vector.
A randomly shifted lattice rule provides an unbiased estimate of the integral, i.e., $\bbE[Q(f)] = I(f)$, where the expectation is taken with respect to the random shift $\bsDelta$. Its quality is determined by the choice of the generating vector $\bsz$. By analyzing the quantity known as *shift-averaged worst case error*, it is known that good generating vectors can be obtained using a *CBC construction* (*component-by-component construction*), determining the components of $\bsz$ one at a time sequentially, to achieve nearly $\calO(n^{-1})$ convergence rate which is optimal in the weighted Sobolev space of smoothness one, and the implied constant in the big $\calO$ bound can be independent of $s$ under appropriate conditions on the weights $\bsgamma$.
More precisely, if $n$ is a power of $2$ then we know that the CBC construction yields the root-mean-square error bound, for all $\lambda\in
(1/2,1]$, $$\begin{aligned}
\label{eq:error1}
\sqrt{\bbE\left[ |I(f) - Q(f)|^2 \right]}
&\,\le\,
\Bigg(
\frac{2}{n}
\sum_{\emptyset\ne\setu\subseteq\{1:s\}} \gamma_\setu^\lambda\,
[\vartheta(\lambda)]^{|\setu|}
\Bigg)^{1/(2\lambda)}
\,\|f\|_\bsgamma\,,\end{aligned}$$ where $\vartheta(\lambda) := 2\zeta(2\lambda)/(2\pi^2)^\lambda$, with $\zeta(a) := \sum_{k=1}^\infty k^{-a}$ denoting the Riemann zeta function. A similar result holds for general $n$. The best rate of convergence clearly comes from choosing $\lambda$ close to $1/2$, but the advantage is offset by the fact that $\zeta(2\lambda)\to\infty$ as $\lambda\to
(1/2)_+$.
### Choosing the weights {#choosing-the-weights .unnumbered}
To apply this abstract theory to a given practical integrand $f$, we need to first obtain an estimate of the norm $\|f\|_\bsgamma$. Remember that at this stage we do not yet know how to choose the weights $\gamma_\setu$. Assuming that bounds on the mixed first derivatives in can be obtained so that $$\label{eq:norm-bound}
\|f\|_\bsgamma \,\le\,
\left(\sum_{\setu\subseteq\{1:s\}} \frac{B_\setu}{\gamma_\setu}\right)^{1/2},$$ we can substitute into and then, with $\lambda$ fixed but unspecified at this point and $A_\setu =
[\vartheta(\lambda)]^{|\setu|}$, we choose the weights $\gamma_\setu$ to minimizing the product $$C_\bsgamma \,:=\, \left(\sum_{\setu\subseteq\{1:s\}} \gamma_\setu^\lambda\,A_\setu\right)^{1/(2\lambda)}
\left(\sum_{\setu\subseteq\{1:s\}} \frac{B_\setu}{\gamma_\setu}\right)^{1/2}.$$ Elementary calculus leads us to conclude that we should take $$\label{eq:weights}
\gamma_\setu \,:=\, \left(\frac{B_\setu}{A_\setu}\right)^{1/(1+\lambda)},$$ which yields $$C_\bsgamma \,=\,
\Bigg(
\sum_{\setu\subseteq\{1:s\}} A_\setu^{1/(1+\lambda)} B_\setu^{\lambda/(1+\lambda)}
\Bigg)^{(1+\lambda)/(2\lambda)}.$$ We then specify a value of $\lambda$, as close to $1/2$ as possible, to ensure that $C_\bsgamma$ can be bounded independently of $s$. This in turn determines the theoretical convergence rate which is $\calO(n^{-1/(2\lambda)})$.
The chosen weights $\gamma_\setu$ are then fed into the CBC construction to produce generating vectors for randomly shifted lattice rules that achieve the desired theoretical error bound for this integrand. This strategy for determining weights was first considered in [@KSS12].
*Fast* CBC constructions (using FFT) can produce generating vectors for an $n$-point rule in $s$ dimensions in $\calO(s\,n\log n)$ operations in the case of product weights [@Nuy14], and in $\calO(s\,n\log n +
s^2\,n)$ operations in the case of POD weights [@KSS11]. Note that these are considered to be pre-computation costs. The actual cost for generating the points on the fly is $\calO(s\,n)$ operations, no worse than Monte Carlo simulations. Strategies to improve on the computational cost of approximating the integral are discussed in Section \[sec:savings\].
The CBC construction yields a lattice rule which is extensible in dimension $s$. We can also construct *lattice sequences* which are extensible or embedded in the number of points $n$, at the expense of increasing the implied constant in the error bound [@CKN06; @DPW08; @HHLL00; @HN03].
Setting 2: QMC Integration over $\bbR^s$ {#sec:setting2}
----------------------------------------
QMC approximation to an integral which is formulated over the Euclidean space $\bbR^s$ can be obtained by first mapping the integral to the unit cube as follows: $$\begin{aligned}
\label{eq:int-Rs}
I(f) \,=\, \int_{\bbR^s} f(\bsy)\,\prod_{j=1}^s \phi(y_j)\,\rd\bsy
&\,=\, \int_{[0,1]^s} f(\Phi^{\mbox{-}1}(\bsw))\,\rd\bsw \\
&\,\approx\, \frac{1}{n} \sum_{i=1}^n f(\Phi^{\mbox{-}1}(\bst_i)) \,=\, Q(f) \,. \nonumber\end{aligned}$$ (With a slight abuse of notation we have reused $I(f)$ and $Q(f)$ from the previous subsection for integration over $\bbR^s$ in this subsection.) Here $\phi$ can be *any general univariate probability density function*, and $\Phi^{\mbox{-}1}$ denotes the component-wise application of the inverse of the cumulative distribution function corresponding to $\phi$. Note that in many practical applications we need to first apply some clever transformation to convert the integral into the above form; some examples are discussed in Section \[sec:apps\]. The transformed integrand $f\circ \Phi^{\mbox{-}1}$ arising from practical applications typically does not belong to the Sobolev space defined over the unit cube due to the integrand being unbounded near the boundary of the cube, or because the mixed derivatives of the transformed integrand do not exist or are unbounded. Thus the theory in the preceding subsection generally does not apply in practice. Some theory for QMC on singular integrands is given in [@Owe06].
We summarize here a special weighted space setting in $\bbR^s$ for which randomly shifted lattice rules have been shown to achieve nearly the optimal convergence rate of order one [@KSWWat10; @NK14]. The norm in this setting is given by $$\begin{aligned}
\label{eq:norm2}
\|f\|_\bsgamma
&\,=\,
\Bigg[
\sum_{\setu\subseteq\{1:s\}} \frac{1}{\gamma_\setu}
\int_{\bbR^{|\setu|}}
\bigg(
\int_{\bbR^{s-|\setu|}}
\frac{\partial^{|\setu|} f}{\partial \bsy_\setu}(\bsy)
\bigg(\prod_{j\in\{1:s\}\setminus\setu} \phi(y_j)\bigg)
\,\rd\bsy_{\{1:s\}\setminus\setu}
\bigg)^2 \nonumber\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\;\times
\bigg(\prod_{j\in\setu} \varpi_j^2(y_j)\bigg)
\,\rd\bsy_\setu \Bigg]^{1/2} \,.\end{aligned}$$ Comparing with , apart from the difference that the integrals are now over the unbounded domain, there is a probability density function $\phi$ as well as additional *weight functions* $\varpi_j$ which can be chosen to moderate the tail behavior of the mixed derivatives of $f$.
The convergence results for the CBC construction of randomly shifted lattice rules in this general setting depend on the choices of $\phi$ and $\varpi_j$. For $n$ a power of $2$, the root-mean-square error bound takes the form, for all $\lambda\in (1/(2r),1]$, $$\sqrt{\bbE\left[ |I(f) - Q(f)|^2 \right]}
\,\le\,
\Bigg(
\frac2n
\sum_{\emptyset\ne \setu\subseteq\{1:s\}}
\gamma_\setu^\lambda\, \prod_{j\in\setu} \vartheta_j(\lambda)\Bigg)^{1/(2\lambda)}\,
\,
\|f\|_{\bsgamma}\,,$$ with $r$ (appearing in the applicable lower bound on $\lambda$) and $\vartheta_j(\lambda)$ depending on $\phi$ and $\varpi_j$, see [@NK14 Theorem 8]. Some special cases have been analyzed:
- See [@GKNSSS15 Theorem 15] or [@KN16 Theorem 5.2] for $\phi(y) = \phi_{\rm nor}(y) = \exp(-y^2/2)/\sqrt{2\pi}$ being the standard normal density and $\varpi_j^2(y_j) = \exp (-2\,\alpha_j\,
|y_j|)$ with $\alpha_j > 0$.
- See [@KN16 Theorem 5.3] for $\phi = \phi_{\rm nor}$ and $\varpi_j^2(y_j) = \exp (-\alpha\, y_j^2)$ with $\alpha < 1/2$.
- See [@SKS13 Theorem 2] for $\phi$ being a logistic, normal, or Student density and $\varpi_j = 1$.
To apply this abstract theory to a practical integral over $\bbR^s$, it is important to realize that the choice of $\phi$ can be tuned as part of the process of transformation to express the integral in the form . (This point will become clearer when we describe the maximum likelihood application in Subsection \[sec:app2\].) Then the choice of weight functions $\varpi_j$ arises as part of the process to obtain bounds on the norm of $f$, as in . (This point will become clearer when we describe the PDE application in Subsection \[sec:app3\].) Finally we can choose the weights $\gamma_\setu$ as in but now with $A_\setu =
\prod_{j\in\setu} \vartheta_j(\lambda)$ for the appropriate $\vartheta_j(\lambda)$ corresponding to the choice of $\phi$ and $\varpi_j$. The choice of density $\phi$, weight functions $\varpi_j$, and weight parameters $\gamma_\setu$ then enter the CBC construction to obtain the generating vector of good randomly shifted lattice rules that can achieve the theoretical error bound for this integrand.
In practice, it may well be that the weights $\gamma_\setu$ obtained in this way are not sensible because we were working with theoretical upper bounds on the error that may be too pessimistic. It may already be so in the standard setting of the previous subsection, but is more pronounced in the setting for $\bbR^s$ due to the additional complication associated with the presence of $\phi$ and $\varpi_j$.
Setting 3: Smooth Integrands in the Unit Cube {#sec:setting3}
---------------------------------------------
Now we return to the integration problem over the unit cube and outline a weighted function space setting from [@DKLNS14] for smooth integrands of order $\alpha$. The norm is given by $$\begin{aligned}
\label{eq:norm3}
\|f\|_{\bsgamma}
\,=\,
\sup_{\setu\subseteq\{1:s\}}
\sup_{\bsy_\setv \in [0,1]^{|\setv|}}
\frac{1}{\gamma_\setu}
\sum_{\setv\subseteq\setu} \,
\sum_{\bstau_{\setu\setminus\setv} \in \{1:\alpha\}^{|\setu\setminus\setv|}}
\bigg|\int_{[0,1]^{s-|\setv|}}
(\partial^{(\bsalpha_\setv,\bstau_{\setu\setminus\setv},\bszero)} f)(\bsy) \,
\rd \bsy_{\{1:s\} \setminus\setv}
\bigg|\,
.\end{aligned}$$ Here $(\bsalpha_\setv,\bstau_{\setu\setminus\setv},\bszero)$ denotes a multi-index $\bsnu$ with $\nu_j = \alpha$ for $j\in\setv$, $\nu_j =
\tau_j$ for $j\in\setu\setminus\setv$, and $\nu_j = 0$ for $j\notin\setu$. We denote the $\bsnu$-th partial derivative of $f$ by $\partial^\bsnu f
\,=\,
(\partial^{|\bsnu|}f)/(\partial^{\nu_1}_{y_1}\partial^{\nu_2}_{y_2}\cdots\partial^{\nu_s}_{y_s})$.
This function space setting can be paired with *interlaced polynomial lattice rules* [@God15; @GD15] to achieve higher order convergence rates in the unit cube. A *polynomial lattice rule* [@Nie92] is similar to a lattice rule (see without the random shift $\bsDelta$), but instead of a generating vector of integers we have a generating vector of polynomials, and thus the regular multiplication and division are replaced by their polynomial equivalents. We omit the technical details here. An interlaced polynomial lattice rule with $n =
2^m$ points in $s$ dimensions with interlacing factor $\alpha$ is obtained by taking a polynomial lattice rule in $\alpha\,s$ dimensions and then interlacing the bits from every successive $\alpha$ dimensions to yield one dimension. More explicitly, for $\alpha = 3$, given three coordinates $x = (0.x_1 x_2 \ldots x_m)_2$, $y = (0.y_1 y_2 \ldots y_m)_2$ and $z =
(0.z_1 z_2 \ldots z_m)_2$ we interlace their bits to obtain $w = (0.x_1
y_1 z_1 x_2 y_2 z_2 \ldots x_m y_m z_m)_2$.
An interlaced polynomial lattice rule with interlacing factor $\alpha\ge
2$, with irreducible modulus polynomial of degree $m$, and with $n=2^m$ points in $s$ dimensions, can be constructed by a CBC algorithm such that, for all $\lambda \in (1/\alpha,1]$, $$\begin{aligned}
|I(f) - Q(f)|
\,\le\,
\left(\frac{2}{n} \sum_{\emptyset\ne\setu\subseteq{\{1:s\}}}
\gamma_\setu^\lambda\, [\vartheta_{\alpha}(\lambda)]^{|\setu|}\right)^{1/\lambda}\,
\|f\|_{\bsgamma}\,,\end{aligned}$$ where $\vartheta_{\alpha}(\lambda) :=
2^{\alpha\lambda(\alpha-1)/2}
([1+1/(2^{\alpha\lambda}-2)]^\alpha-1)$. This result can be found in [@KN16 Theorem 5.4], which was obtained from minor adjustments of [@DKLNS14 Theorem 3.10].
Given a practical integrand $f$, if we can estimate the corresponding integrals involving the mixed derivatives in , then we can choose the weights $\gamma_\setu$ so that every term in the supremum is bounded by a constant, say, $c$. This strategy in [@DKLNS14] led to a new form of weights called *SPOD weights* (*smoothness-driven product and order dependent weights*); they take the form $$\gamma_\setu
\,=\,
\sum_{\bsnu_\setu \in \{1:\alpha\}^{|\setu|}}
\Gamma_{|\bsnu_\setu|} \prod_{j\in\setu} \Upsilon_j(\nu_j)\,.$$ If the weights $\bsgamma$ are SPOD weights, then the fast CBC construction of the generating vector has cost $\calO(\alpha\,s\, n\log n +
\alpha^2\,s^2 n)$ operations. If the weights $\bsgamma$ are product weights, then the CBC algorithm has cost $\calO(\alpha\,s\, n\log n)$ operations.
Three Applications {#sec:apps}
==================
Integrals over $\bbR^s$ often arise from practical applications in the form of multivariate expected values $$\label{eq:QoI}
\bbE_\rho [q] \,=\, \int_{\bbR^s} q(\bsy) \, \rho(\bsy) \, \rd \bsy\,,$$ where $q$ is some quantity of interest which depends on a vector $\bsy =
(y_1,\ldots,y_s)$ of parameters or variables in $s$ dimensions, and $\rho$ is some multivariate probability density function describing the distribution of $\bsy$, *not necessarily a product of univariate functions* as we assumed in , and so we need to make an appropriate transformation to apply our theory. Below we discuss three motivating applications with quite different characteristics, and we will explain how to make use of the different settings in Section \[sec:settings\].
Application 1: Option Pricing {#sec:app1}
-----------------------------
Following the Black–Scholes model, integrals arising from option pricing problems take the general form , with $$q(\bsy) = \max(\mu(\bsy),0)
\quad\mbox{and}\quad
\rho(\bsy) = \frac{\exp(-\frac{1}{2}\bsy^{\tt T}\Sigma^{-1}\bsy)}{\sqrt{(2\pi)^s \det(\Sigma)}},$$ where the variables $\bsy = (y_1,\ldots,y_s)^{\tt T}$ correspond to a discretization of the underlying Brownian motion over a time interval $[0,T]$, and the covariance matrix has entries $\Sigma_{ij} =
(T/s)\min(i,j)$. For example, in the case of an *arithmetic average Asian call option* [@ABG98; @GKSW08; @GilWat09], the payoff function $q$ depends on the smooth function $\mu(\bsy) = (1/s)\sum_{j=1}^s
S_{t_j}(\bsy) - K$ which is the difference between the average of the asset prices $S_{t_j}$ at the discrete times and the strike price $K$.
The widely accepted strategy to rewrite these option pricing integrals from the form to the form with product densities is to take a factorization $\Sigma = AA^{\tt T}$ and apply a change of variables $\bsy = A\bsy'$. This yields an integral of the form with $$f(\bsy') = q(A\bsy')
\quad\mbox{and}\quad
\phi = \phi_{\rm nor}.$$ The choice of factorization therefore determines the function $f$. For example, $A$ can be obtained through Cholesky factorization (commonly known as the *standard construction*; in this case it is equivalent to generating the Brownian motions sequentially in time), through *Brownian bridge construction* [@CMO97], or eigenvalue decomposition sometimes called the *principal component construction* [@ABG98]. Note that in practice these factorizations are not carried out explicitly due to the special form of the covariance matrix. In fact, they can be computed in $\calO(s)$, $\calO(s)$ and $\calO(s\,\log s)$ operations, respectively [@GKSW08].
The success of QMC for option pricing cannot be explained by most existing theory due to the *kink* in the integrand induced by the maximum function. However, for some factorizations it is shown in [@GKS13] that all *ANOVA terms* of $f$ are smooth, with the exception of the highest order term. This hints at a *smoothing by preintegration* strategy, where a coordinate with some required property is chosen, say $y_k$, and we integrate out this one variable (either exactly or numerically with high precision) to obtain a function in $s-1$ variables $$P_k(f) \,:=\, \int_{-\infty}^\infty f(\bsy)\,\phi_{\rm nor}(y_k)\,\rd y_k\,.$$ Under the right conditions (e.g., integrating with respect to $y_1$ in the case of the principal components construction), this new function is smooth and belongs to the function space setting of Subsection \[sec:setting2\] (with one less variable) [@GKLS]. This strategy is related to the method known as *conditional sampling* [@ACN13].
Application 2: Maximum Likelihood {#sec:app2}
---------------------------------
Another source of integrands which motivated recent developments in the function space setting of Subsection \[sec:setting2\] is a class of generalized linear mixed models (GLMM) in statistics, as examined in [@KDSWW08; @KSWWat10; @SKS13]. A specific example of the Poisson likelihood time series model considered in these papers involves an integral of the form , with $$q(\bsy) = \prod_{j=1}^s \frac{\exp(\tau_j (\beta+y_j) - e^{\beta + y_j})}{\tau_j !}
\quad\mbox{and}\quad
\rho(\bsy) = \frac{\exp(-\frac{1}{2}\bsy^{\tt T}\Sigma^{-1}\bsy)}{\sqrt{(2\pi)^s\det(\Sigma)}}.$$ Here $\beta\in\bbR$ is a model parameter, $\tau_1,\ldots,\tau_s\in
\{0,1,\ldots\}$ are the count data, and $\Sigma$ is a Toeplitz covariance matrix with $\Sigma_{ij}=\sigma^2\varkappa^{|i-j|}/(1-\varkappa^2)$, where $\sigma^2$ is the variance and $\varkappa\in(-1,1)$ is the autoregression coefficient.
An obvious way to rewrite this integral in the form with product densities is to factorize $\Sigma$ as discussed in the previous subsection for the option pricing applications, but this would yield a very spiky function $f$. Instead, the strategy developed in [@KDSWW08] recenters and rescales the exponent $T(\bsy)$ of the product $q(\bsy)\rho(\bsy) =: \exp(T(\bsy))$ as follows:
1. Find the unique stationary point $\bsy^*$ satisfying $\nabla
T(\bsy^*)=0$.
2. Determine the matrix $\Sigma^*=(-\nabla^2T(\bsy^*))^{-1}$ which describes the convexity of $T$ around the stationary point.
3. Factorise $\Sigma^*=A^*{A^*}^{{\tt T}}$.
4. Apply a change of variables $\bsy=A^*\bsy'+\bsy^*$.
5. Multiply and divide the resulting integrand by the product $\prod_{j=1}^s \phi(y_j')$ where $\phi$ is any univariate density (not necessarily the normal density).
These steps then yield an integral of the form with $$f(\bsy') \,=\, \frac{c\,\exp (T(A^*\bsy'+\bsy^*))}{\prod_{j=1}^s \phi(y_j')}$$ for some scaling constant $c>0$. Note that the choice of $A^*$ and $\phi$ determines $f$.
The paper [@SKS13] provides careful estimates of the norm of the resulting integrand $f$ in the setting of Subsection \[sec:setting2\] corresponding to three different choices of density $\phi$, with the weight functions taken as $\varpi_j = 1$, and gives the formula for the weight parameters $\gamma_\setu$ that minimize the overall error bound.
These GLMM problems are extremely challenging not only for QMC but also in general the tools are still lacking. There is still lots of room to develop new QMC methods and theory for these problems.
Application 3: PDEs with Random Coefficients {#sec:app3}
--------------------------------------------
Our third application is motivated by fluid flow through a porous medium, typically modelled using Darcy’s Law, with random coefficients. A popular toy problem is the elliptic PDE with a random coefficient [@CD15; @GWZ14; @SG11] $$-\nabla \cdot (a(\bsx, \omega)\,\nabla u(\bsx, \omega))
= \kappa(\bsx)\quad \text{for $\bsx\in D\subset\bbR^d$ and almost all $\omega\in\Omega$, }$$ with $d\in\{1,2,3\}$, subject to homogeneous Dirichlet boundary conditions. The coefficient $a(\bsx, \omega)$ is assumed to be a random field over the spatial domain $D$ (e.g., representing the permeability of a porous material over $D$), and $\Omega$ is the probability space. The goal is to compute the expected values $\bbE [ G(u) ]$ of some bounded linear functional $G$ of the solution $u$ over $\Omega$.
For practical reasons it is often assumed that $a(\bsx,\omega)$ is a *lognormal random field*, that is, $a(\bsx, \omega) = \exp( Z(\bsx,
\omega) )$, where $Z(\bsx, \omega)$ is a Gaussian random field with a prescribed mean and covariance function. This is known as the *lognormal case*. However, researchers often analyze a simpler model known as the *uniform case*.
### The uniform case {#the-uniform-case .unnumbered}
In the uniform case, we consider the parametric PDE $$\label{eq:PDE-par}
-\nabla \cdot (a(\bsx, \bsy)\,\nabla u(\bsx, \bsy))
= \kappa(\bsx)\quad \text{for $\bsx\in D\subset\bbR^d$},$$ together with $$\label{eq:uniform}
a(\bsx, \bsy) \,=\, a_0(\bsx) + \sum_{j\ge1} y_j \psi_j(\bsx)\,,$$ where the parameters $y_j$ are independently and uniformly distributed on the interval $[-\frac{1}{2},\frac{1}{2}]$, and we assume that $0<
a_{\min}\le a(\bsx,\bsy)\le a_{\max} < \infty$ for all $\bsx$ and $\bsy$.
A (single-level) strategy for approximating $\bbE[G(u)]$ is as follows:
1. Truncate the infinite sum in to $s$ terms.
2. Solve the PDE using finite element methods with meshwidth $h$.
3. Approximate the resulting $s$-dimensional integral using QMC with $n$ points.
So the error is a sum of truncation error, discretization error, and quadrature error.
For the QMC quadrature error in Step 3, we have the integral with $$f(\bsy) \,=\, G(u^s_h(\cdot,\bsy - \tfrac{\boldsymbol{1}}{\boldsymbol{2}})),$$ where $u^s_h$ denotes the finite element solution of the truncated problem, and the subtraction by $\tfrac{\boldsymbol{1}}{\boldsymbol{2}}$ takes care of the translation from the usual unit cube $[0,1]^s$ to $[-\frac{1}{2},\frac{1}{2}]^s$. By differentiating the PDE , we can obtain bounds on the mixed derivatives of the PDE solution with respect to $\bsy$. This leads to bounds on the norm of the integrand $f$ and so we can apply the theoretical setting of Subsection \[sec:setting1\] to obtain up to first order convergence for QMC. Under appropriate assumptions and with first order finite elements, we can prove that the total error for the above 3-step strategy is of order [@KSS12] $$\calO(s^{-2(1/p_0-1)} + h^{2} + n^{-\min(1/p_0-1/2,1-\delta)})\,,\quad \delta\in (0,\tfrac{1}{2})\,,$$ where $p_0\in (0,1)$ should be as small as possible while satisfying $\sum_{j\ge 1} \|\psi_j\|_{L_\infty}^{p_0} < \infty$. This part is presented as a step-by-step tutorial in the article [@KN-tutorial] from this volume.
The bounds on the derivatives of the PDE with respect to $\bsy$ also allow us to obtain bounds on the norm and so we can also apply the theoretical setting of Subsection \[sec:setting3\] to obtain higher order convergence [@DKLNS14]. Specifically, the $\calO(n^{-\min(1/p_0-1/2,1-\delta)})$ term can be improved to $\calO(n^{-1/p_0})$. Also the $\calO(h^2)$ term can be improved by using higher order finite elements. See [@KN16; @KN-tutorial] for more details.
### The lognormal case with Karhunen–Loève expansion {#the-lognormal-case-with-karhunenloève-expansion .unnumbered}
In the lognormal case, we have the same parametric PDE , but now we use the *Karhunen–Loève expansion* (*KL expansion*) of the Gaussian random field (in the exponent) to write $$a(\bsx,\bsy) \,=\, a_0(\bsx)\,\exp\left(\sum_{j\ge1} y_j\,\sqrt{\mu_j} \, \xi_j(\bsx)\right)\,,$$ where $a_0(\bsx)>0$, the $\mu_j$ are real, positive and non-increasing in $j$, the $\xi_j$ are orthonormal in $L_2(D)$, and the parameters $y_j\in\bbR$ are standard $\calN(0,1)$ random variables. Truncating the infinite series in $a(\bsx,\bsy)$ to $s$ terms and solving the PDE with a finite element method as in the uniform case, we have now an integral of the form with $$f(\bsy) \,=\, G(u^s_h(\cdot,\bsy))
\quad\mbox{and}\quad \phi = \phi_{\rm nor}\,.$$ One crucial step in the analysis of [@GKNSSS15] is to choose suitable weight functions $\varpi_j$ so that the function $f$ has a finite and indeed small norm , so that the theoretical setting of Subsection \[sec:setting2\] can be applied. Again see [@KN16; @KN-tutorial] for more details.
In this lognormal case with KL expansion (and also the uniform case), the cost per sample of the random field is $\calO(s\,M)$ operations, where $M$ is the number of finite element nodes. This dominates the cost in evaluating the integrand function under the assumption that assembling the stiffness matrix to solve the PDE (which depends on the random field) is higher than the cost of the PDE solve which is $\calO(M\,\log M)$. When $s$ is large the cost of sampling the random field can be prohibitive, and this is why the following alternative strategies emerged.
### The lognormal case with circulant embedding {#the-lognormal-case-with-circulant-embedding .unnumbered}
Since we have a Gaussian random field we can actually sample the random field exactly on any set of $M$ spatial points. This leads to an integral of the form with (assuming the field has zero mean) $$q(\bsy) \,=\, G(u^s_h(\cdot,\bsy))
\quad\mbox{and}\quad
\rho(\bsy) = \frac{\exp(-\frac{1}{2}\bsy^{\tt T} R^{-1}\bsy)}{\sqrt{(2\pi)^s \det(R)}},$$ where $R$ is an $M\times M$ covariance matrix, and initially we have $s=
M$. (Note the subtle abuse of notation that the second argument in $u^s_h(\bsx,\cdot)$ has a different meaning to the KL case, in the sense that there the covariance is already built in.) This integral can be transformed into the form with a factorization $R =
AA^{\tt T}$ and a change of variables $\bsy = A\bsy'$, as in the option pricing example, to obtain $$f(\bsy') \,=\, G(u^s_h(\cdot,A\bsy'))
\quad\mbox{and}\quad
\phi = \phi_{\rm nor}\,.$$ The advantage of this discrete formulation is that there is no error arising from the truncation of the KL expansion. However, the direct factorization and matrix-vector multiplication require $\calO(M^3)$ operations which can be too costly when $M$ is large.
The idea of *circulant embedding* [@DN97; @GKNSS11; @GKNSS-paper1; @GKNSS-paper2] is to sample the random field on a regular grid and to embed the covariance matrix of these points into a larger $s\times s$ matrix which is *nested block circulant with circulant blocks*, so that FFTs can be used to reduce the per sample cost to $\calO(s\,\log s)$ operations. Values of the random field at the finite element quadrature nodes can be obtained by interpolation. Note that this turns the problem into an even higher dimensional integral, and we can have $s\gg M$. For this strategy to work we need to use regular spatial grid points to sample the field and a stationary covariance function (i.e., the covariance depends only on the relative distance between points). An additional difficulty is to ensure positive definiteness of the extended matrix; this is studied in [@GKNSS-paper1].
### The lognormal case with $H$-matrix technique {#the-lognormal-case-with-h-matrix-technique .unnumbered}
Another approach for the discrete matrix formulation of the lognormal case is to first approximate $R$ by an $H$-matrix [@Hac15] and make use of $H$-matrix techniques to compute the matrix-vector multiplication with the square-root of this $H$-matrix at essentially linear cost $\calO(M)$. Two iterative methods have been proposed in [@FKS] to achieve this (one is based on a variant of the *Lanczos iteration* and the other on the *Schultz iteration*), with full theoretical justification for the error incurred in the $H$-matrix approximation. An advantage of this approach over circulant embedding is that it does not require the spatial grid to be regular nor that the covariance be stationary.
### Other developments {#other-developments .unnumbered}
A different QMC analysis for the lognormal case has been considered in [@HPS17]. QMC for holomorphic equations was considered in [@DLS16], and for Baysesian inversion in [@DGLS; @SST17]. Recently there is also QMC analysis developed for the situation where the functions in the expansion of $a(\bsx,\bsy)$ have local support, see [@GHS; @HerSch; @Kaz].
Three Cost Saving Strategies {#sec:savings}
============================
In this section we discuss the basic ideas of three different kinds of cost saving strategies that can be applied to QMC methods, without going into details. Actually, the circulant embedding and $H$-matrix technique discussed in the previous section can also be considered as cost saving strategies. These strategies are not mutually exclusive, and it may be possible to mix and match them to benefit from compound savings.
Saving 1: Multi-level and Multi-index {#sec:sav1}
-------------------------------------
The *multi-level* idea [@Gil15] is easy to explain in the context of numerical integration. Suppose that there is a sequence $(f_\ell)_{\ell\ge 0}$ of approximations to an integrand $f$, with increasing accuracy and cost as $\ell$ increases, such that we have the telescoping sum $$f \,=\, \sum_{\ell=0}^\infty (f_\ell - f_{\ell-1}), \quad
f_{-1}:=0.$$ For example, the different $f_\ell$ could correspond to different number of time steps in option pricing, different number of mesh points in a finite element solve for PDE, different number of terms in a KL expansion, or a combination of aspects. A multi-level method for approximating the integral of $f$ is $$A_{\rm ML}(f) \,=\, \sum_{\ell=0}^L Q_\ell(f_\ell - f_{\ell-1}),$$ where the parameter $L$ determines the number of *levels*, and for each level we apply a different quadrature rule $Q_\ell$ to the difference $f_\ell - f_{\ell-1}$.
The integration error (in this simple description with deterministic quadrature rules) satisfies $$|I(f) - A_{\rm ML}(f)| \;\le\;
\underbrace{|I(f) - I_L(f_L)|\vphantom{\sum_{\ell=0}^\infty }}_{\le\, \varepsilon/2}
\;+\; \underbrace{\sum_{\ell=0}^L |(I_\ell-Q_\ell)(f_\ell - f_{\ell-1})|}_{\le\, \varepsilon/2}.$$ For a given error threshold $\varepsilon>0$, the idea (as indicated by the underbraces) is that we choose $L$ to ensure that the first term (the truncation error) on the right-hand side is $\le\varepsilon/2$, and we specify parameters for the quadrature rules $Q_\ell$ so that the second term (the quadrature error) is also $\le\varepsilon/2$. The latter can be achieved with a Lagrange multiplier argument to minimize cost subject to the given error threshold. Our hope is that the successive differences $f_\ell-f_{\ell-1}$ will become smaller with increasing $\ell$ and therefore we would require less quadrature points for the more costly higher levels.
The *multi-index* idea [@HNT16] generalizes this from a scalar level index $\ell$ to a vector index $\bsell$ so that we can vary a number of different aspects (e.g., spatial/temporal discretization) simultaneously and independently of each other. It makes use of the sparse grid concept so that the overall cost does not blow up with respect to the dimensionality of $\bsell$, i.e., the number of different aspects being considered. A simple example is that we use different finite element meshwidths for different spatial coordinates. This is equivalent to applying sparse finite element methods within a multilevel algorithm, see the article [@GilKS] in this volume.
Multi-level and multi-index extensions of QMC methods for the applications from Section \[sec:apps\] include e.g., [@DKLS16; @GilWat09; @KSSSU; @KSS15; @RNV].
Saving 2: Multivariate Decomposition Method {#sec:sav2}
-------------------------------------------
In the context of numerical integration, the *multivariate decomposition method* (*MDM*) [@GilWas17; @KNPSW; @KSWW10a; @Was13] makes use of a decomposition of the integrand $f$ of the form $$f \,=\, \sum_{|\setu|<\infty} f_\setu\,,$$ where the sum is over all finite subsets $\setu\subset\{1,2,\ldots\}$ and each function $f_\setu$ depends only on the integration variables with indices in the set $\setu$. Then MDM takes the form $$A_{\rm MDM}(f) \,=\, \sum_{\setu\in\calA} Q_\setu(f_\setu)\,$$ where $\calA$ is known as the *active set* of subsets of indices, and for each $\setu$ in the active set we apply a different quadrature rule $Q_\setu$ to $f_\setu$.
Analogously to the multi-level idea, the error of MDM satisfies $$|I(f) - A_{\rm MDM}(f)| \;\le\;
\underbrace{\sum_{\setu\notin\calA}|I_\setu(f_\setu)|}_{\le\, \varepsilon/2}
\;+\; \underbrace{\sum_{\setu\in\calA} |(I_\setu-Q_\setu)(f_\setu)|}_{\le\, \varepsilon/2}\,,$$ where we choose the active set $\calA$ to ensure that the truncation error is $\le\varepsilon/2$, and we use a Lagrange multiplier argument to specify parameters for the quadrature rules so that the quadrature error is also $\le\varepsilon/2$. Our hope is that, although the cardinality of the active set $\calA$ might be huge (e.g., tens of thousands), the cardinality of the individual subsets $\setu\in\calA$ might be relatively small (e.g., at most $8$ or $10$), and therefore we transfer the problem into that of solving a large number of low dimensional integrals.
There are many important considerations in the implementation of MDM [@GKNW]. First, we need to decide on how to decompose the integrand $f$ so that values of the functions $f_\setu$ can be computed. One obvious choice is known as the *anchored decomposition* which can be computed via the explicit formula [@KSWW10b] $$\label{eq:anchored}
f_\setu(\bsy_\setu) \,=\,
\sum_{\setv\subseteq\setu} (-1)^{|\setu|-|\setv|}
f(\bsy_\setv;\bsa),$$ where $\bsa$ is an *anchor* and $(\bsy_\setv;\bsa)$ denotes a vector obtained from $\bsy$ by replacing the component $y_j$ with the corresponding component $a_j$ when the index $j$ does not belong to the subset $\setv$. (This is similar to the well-known ANOVA decomposition which, however, involves integrals that cannot be computed in practice.) Second, we need to specify and construct the active set $\calA$ and have an efficient data structure to store the sets for later traversing. Third, we need to explore nestedness or embedding in the quadrature rules, taking into account the sum in and develop efficient ways to reuse function evaluations.
Saving 3: Fast QMC Matrix-vector Multiplication {#sec:sav3}
-----------------------------------------------
There is a certain structure in some QMC methods that can allow for fast matrix-vector multiplication using FFT. This structure has been exploited in the fast CBC construction of lattice rules and polynomial lattice rules [@Nuy14]. We now explain how this same structure can also be used in more general circumstances [@DKLS15].
For notational convenience, we denote all QMC points $\bst_i$ as row vectors in this subsection. Given an arbitrary matrix $A$, suppose we want to $$\mbox{compute}\quad
\bsy_i\,A
\quad\mbox{for all}\quad i=1,\ldots,n\,,$$ with the row vectors $\bsy_i = \chi(\bst_i)$, where $\chi$ denotes an arbitrary univariate function that is applied to every component of the QMC point $\bst_i$. Typically we have $\bst_n \equiv \bst_0 = \bszero$ so we can leave it out. Consider for simplicity the case $n$ is prime and suppose we can write $$Y\,:=\,
\begin{bmatrix} \bsy_1 \\\vdots \\ \bsy_{n-1}\end{bmatrix} \,=\, C\, P\,$$ where $C$ is an $(n-1)\times(n-1)$ circulant matrix, while $P$ is a matrix containing a single $1$ in each column and $0$ everywhere else. Then we can compute $Y\bsa$ in $\calO(n\log n)$ operations for any column $\bsa$ of $A$.
The desired factorization $Y = CP$ is possible if we have deterministic lattice points or deterministic polynomial lattice points, and if we apply the inverse cumulative distribution function mapping or tent transform [@CKNS16; @DNP14; @Hic02]. However, it does *not* work with random shifting, scrambling [@Owe97], or interlacing. This strategy can be used to generate normally distributed points with a general covariance matrix (no need for stationarity as in circulant embedding), solving PDEs with uniform random coefficients, or solving PDEs with lognormal random coefficients involving finite element quadratures.
Software Resources {#sec:software}
==================
We provide some software resources for the practical application of QMC methods:
- *The Magic Point Shop*: a collection of QMC point generators and generating vectors.\
<https://people.cs.kuleuven.be/~dirk.nuyens/qmc-generators/>
- *Fast component-by-component constructions*: a collection of software routines for fast CBC constructions of generating vectors.\
<https://people.cs.kuleuven.be/~dirk.nuyens/fast-cbc/>
- *QMC4PDE*: accompanying software package for the survey [@KN16] on using QMC methods for parametrized PDE problems.\
<https://people.cs.kuleuven.be/~dirk.nuyens/qmc4pde/>
- *A practical guide to QMC methods*: a non-technical introduction of QMC methods with software demos.\
<https://people.cs.kuleuven.be/~dirk.nuyens/taiwan/>
Summary and Outlook {#sec:summary}
===================
In this article we summarized three QMC theoretical settings: randomly shifted lattice rules achieving first order convergence in the unit cube and in $\bbR^s$, and interlaced polynomial lattice rules achieving higher order convergence in the unit cube. One important feature is that the error bound can be independent of the dimension under appropriate conditions on the weights. Another important feature is that these QMC methods can be constructed by fast CBC algorithms.
We outlined three different applications and explained how they can be pre-processed to make use of the different theory. We also discussed three cost saving strategies that can be combined with QMC in these applications.
This paper is not meant to be a comprehensive survey on QMC methods. There are of course many other significant developments on QMC methods and their applications. For example, we did not discuss *tent transformation* (also known as the baker’s transform), which can yield second order convergence for randomly shifted rules or first order convergence for deterministic lattice rules [@CKNS16; @DNP14; @Hic02]. We also did not discuss *scrambling* [@Owe97], which is a well-known randomization method that can potentially improve the convergence rates by an extra half order.
For the future we would like to see QMC in new territories, to tackle a significantly wider range of more realistic problems. Some emerging new application areas of QMC include e.g., Bayesian inversion [@DGLS; @SST17], stochastic wave propagation [@GH15; @GanKS], quantum field theory [@AGHJLV16; @JLGM14], and neutron transport [@GGKSS; @GPS].
Looking ahead into future QMC developments, what would be on the top of our wish list? We would very much like to have a *“Setting 4” where we have QMC methods that achieve higher order convergence in $\bbR^s$, with error bounds that are independent of $s$, and for which fast constructions are possible.* This open problem has seen some partial solutions [@DILP; @NN] but there is more to be done!
The authors acknowledge the financial supports from the Australian Research Council (FT130100655 and DP150101770) and the KU Leuven research fund (OT:3E130287 and C3:3E150478).
[999]{}
[Achtsis, N., Cools, R., Nuyens, D.:]{} [Conditional sampling for barrier option pricing under the Heston model]{}, in: Monte Carlo and Quasi-Monte Carlo Methods 2012 (J. Dick, F. Y. Kuo, G. Peters, and I. H. Sloan, eds), Springer, Berlin, pp. 253–269 (2013)
[Acworth, P., Broadie, M., Glasserman, P.:]{} [A comparison of some Monte Carlo and quasi-Monte Carlo techniques for option pricing]{}, in: Monte Carlo and quasi-Monte Carlo methods 1996 (P. Hellekalek, G. Larcher, H. Niederreiter, and P. Zinterhof, eds.), Springer, Berlin, pp. 1–18 (1998)
[Ammon, A., Hartung, T., Jansen, K., Leövey, H., Vollmer, J.:]{} [On the efficient numerical solution of lattice systems with low-order couplings]{}, [Comput. Phys. Commun.]{} [**198**]{}, 71–81 (2016)
[Caflisch, R.E., Morokoff, W., Owen, A.B.:]{} [Valuation of mortgage backed securities using Brownian bridges to reduce effective dimension]{}, [J. Comput. Finance]{} [**1**]{}, 27–46 (1997)
[Cohen, A., DeVore, R.:]{} [Approximation of high-dimensional parametric PDEs]{}, [Acta Numer.]{} [**24**]{}, 1–159 (2015)
[Cools, R., Kuo, F.Y., Nuyens, D.:]{} [Constructing embedded lattice rules for multivariate integration]{}, [SIAM J. Sci. Comput.]{} [**28**]{}, 2162–2188 (2006)
[Cools, R., Kuo, F.Y., Nuyens, D., Suryanarayana, G.:]{} [Tent-transformed lattice rules for integration and approximation of multivariate non-periodic functions]{}, [J. Complexity]{} [**36**]{}, 166–181 (2016)
[Dick, J., Gantner, R.N., Le Gia, Q.T., Schwab, Ch.:]{} [Higher order Quasi-Monte Carlo integration for Bayesian estimation]{} (in review)
[Dick, J., Irrgeher, Ch., Leobacher, G., Pillichshammer, F.:]{} [On the optimal order of integration in Hermite spaces with finite smoothness]{} (in review)
[Dick, J., Kuo, F.Y., Le Gia, Q.T., Nuyens, D., Schwab, Ch.:]{} [Higher order QMC Galerkin discretization for parametric operator equations]{}, [SIAM J. Numer. Anal.]{} [**52**]{}, 2676–2702 (2014)
[Dick, J., Kuo, F.Y., Le Gia, Q.T., Schwab, Ch.:]{} [Fast QMC matrix-vector multiplication]{}, [SIAM J. Sci. Comput.]{} [**37**]{}, A1436–A1450 (2015)
[Dick, J., Kuo, F.Y., Le Gia, Q.T., Schwab, Ch.:]{} [Multi-level higher order QMC Galerkin discretization for affine parametric operator equations]{}, [SIAM J. Numer. Anal.]{}, [**54**]{}, 2541–2568 (2016)
[Dick, J., Kuo, F.Y., Sloan, I.H.:]{} [High-dimensional integration: the Quasi-Monte Carlo way]{}, [Acta Numer.]{} **22**, 133–288 (2013)
[Dick, J., Le Gia, Q.T., Schwab, Ch.:]{} Higher order Quasi-Monte Carlo integration for holomorphic, parametric operator equations, [SIAM/ASA Journal on Uncertainty Quantification]{}, [**4**]{}, 48–79 (2016)
[Dick, J., Nuyens, D., Pillichshammer, F.:]{} Lattice rules for nonperiodic smooth integrands, [Numer. Math.]{} **126**, 259–291 (2014)
[Dick, J., Pillichshammer, F.:]{} [Digital Nets and Sequences]{}, Cambridge University Press (2010)
[Dick, J., Pillichshammer, F., Waterhouse, B.J.:]{} [The construction of good extensible rank-$1$ lattices]{}, [Math. Comp.]{} [**77**]{}, 2345–2374 (2008)
[Dietrich, C.R., Newsam, G.H.:]{} [Fast and exact simulation of stationary Gaussian processes through circulant embedding of the covariance matrix]{}, [SIAM J. Sci. Comput.]{} **18**, 1088–1107 (1997)
[Feischl, M., Kuo, F.Y., Sloan, I.H.:]{} [Fast random field generation with $H$-matrices]{} (in review)
[Ganesh, M., Hawkins, S.C.:]{} [A high performance computing and sensitivity analysis algorithm for stochastic many-particle wave scattering]{}, [SIAM J. Sci. Comput.]{} [**37**]{}, A1475–A1503 (2015)
[Ganesh, M., Kuo, F.Y., Sloan, I.H.:]{} [Quasi-Monte Carlo finite element methods for stochastic heterogeneous wave progagation models]{} (in preparation)
[Gantner, R.N., Herrmann, L., Schwab, Ch.:]{} [Quasi-Monte Carlo integration for affine-parametric, elliptic PDEs: local supports imply product weights]{} (in review)
[Gilbert, A.D., Graham, I.G., Kuo, F.Y., Scheichl, R., Sloan, I.H.:]{} [Quasi-Monte Carlo theory for an eigenproblem with a random coefficient]{} (in preparation)
[Gilbert, A.D., Kuo, F.Y., Nuyens, D., Wasilkowski, G.W.:]{} [Efficient implementation of the multivariate decomposition method]{} (in preparation)
[Gilbert, A.D., Wasilkowski, G.W.:]{} [Small superposition dimension and active set construction for multivariate integration under modest error demand]{}, [J. Complexity]{} [**42**]{}, 94–109 (2017)
[Giles, M.B.:]{} [Multilevel Monte Carlo methods]{}, [Acta Numer.]{} [**24**]{}, 259–328 (2015)
[Giles, M.B., Kuo, F.Y., Sloan, I.H.:]{} [Combining sparse grids, multilevel MC and QMC for elliptic PDEs with random coefficients]{} (submitted to this volume)
[Giles, M.B., Kuo, F.Y., Sloan, I.H., Waterhouse, B.J.:]{} [Quasi-Monte Carlo for finance applications]{}, [ANZIAM J.]{} **50**, C308–C323 (CTAC2008) (2008)
[Giles, M.B., Waterhouse, B.J.:]{} [Multilevel quasi-Monte Carlo path simulation]{}. [Radon Series Comp. Appl. Math.]{} [**8**]{}, 1–18 (2009)
[Goda, T.:]{} [Good interlaced polynomial lattice rules for numerical integration in weighted Walsh spaces]{}, [J. Comput. Appl. Math.]{} [**285**]{}, 279–294 (2015)
[Goda, T., Dick, J.:]{} [Construction of interlaced scrambled polynomial lattice rules of arbitrary high order]{}, [Found. Comput. Math.]{} [**15**]{}, 1245–1278 (2015)
[Graham, I.G., Kuo, F.Y., Nichols, J.A., Scheichl, R., Schwab, Ch., Sloan, I.H.:]{} [QMC FE methods for PDEs with log-normal random coefficients]{}, [Numer. Math.]{} [**131**]{}, 329–368 (2015)
[Graham, I.G., Kuo, F.Y., Nuyens, D., Scheichl, R., Sloan, I.H.:]{} [Quasi-Monte Carlo methods for elliptic PDEs with random coefficients and applications]{}, [J. Comput. Phys.]{} [**230**]{}, 3668–3694 (2011)
[Graham, I.G., Kuo, F.Y., Nuyens, D., Scheichl, R., Sloan, I.H.:]{} [Analysis of circulant embedding methods for sampling stationary random fields]{} (in preparation)
[Graham, I.G., Kuo, F.Y., Nuyens, D., Scheichl, R., Sloan, I.H.:]{} [Circulant embedding with QMC – analysis for elliptic PDE with lognormal coefficients]{} (in preparation)
[Graham, I.G., Parkinson, M.J., Scheichl, R.:]{} [Modern Monte Carlo variants for uncertainty quantification in neutron transport]{} (in review)
[Griebel, M., Kuo, F.Y., Sloan, I.H.:]{} [The smoothing effect of integration in $\mathbb{R}^d$ and the ANOVA decomposition]{}, [Math. Comp.]{} [**82**]{} (2013), 383–400; and the note in [Math. Comp.]{} [**86**]{}, 1847–1854 (2017)
[Griewank, A., Kuo, F.Y., Leövey, H., Sloan, I.H.:]{} [High dimensional integration of kinks and jumps – smoothing by preintegration]{} (in review)
[Gunzburger, M., Webster, C., Zhang, G.:]{} Stochastic finite element methods for partial differential equations with random input data, [Acta Numer.]{} [**23**]{}, 521–650 (2014)
[Hackbusch, W:]{} [Hierarchical matrices: algorithms and analysis]{}, Springer, Heidelberg (2015)
[Haji-Ali, A.L., Nobile, F., Tempone, R.:]{} [Multi-index Monte Carlo: when sparsity meets sampling]{}, [Numer. Math.]{} **132**, 767–806 (2016)
[Harbrecht, H., Peters, M., Siebenmorgen, M.:]{} [On the quasi-Monte Carlo method with Halton points for elliptic PDEs with log-normal diffusion]{}, [Math. Comp.]{} [**86**]{}, 771–797 (2017)
[Herrmann, L., Schwab, Ch.:]{} [QMC integration for lognormal-parametric, elliptic PDEs: local supports imply product weights]{} (in review)
[Hickernell, F.J.:]{} [Obtaining $O(N^{-2+\epsilon})$ convergence for lattice quadrature rules]{}, in: Monte Carlo and Quasi-Monte Carlo Methods 2000 (K. T. Fang, F. J. Hickernell, and H. Niederreiter, eds.), Springer, Berlin, pp. 274–289 (2002)
[Hickernell, F.J., Hong, H.S., LÉcuyer, P., Lemieux, C.:]{} [Extensible lattice sequences for quasi-Monte Carlo quadrature]{}, [SIAM J. Sci. Comput.]{} **22**, 1117–1138 (2000)
[Hickernell, F.J., Niederreiter, H.:]{} [The existence of good extensible rank-$1$ lattices]{}, [J. Complexity]{} **19**, 286–300 (2003)
[Jansen, K., Leövey, H., Griewank, A., Müller-Preussker, M.:]{} [Quasi-Monte Carlo methods for lattice systems: a first look]{}, [Comput. Phys. Commun.]{} [**185**]{}, 948–959 (2014)
[Kazashi, Y.:]{} [Quasi-Monte Carlo integration with product weights for elliptic PDEs with log-normal coefficients]{} (in review)
[Kuo, F.Y., Dunsmuir, W.D.M., Sloan, I.H., Wand, M.P., Womersley, R.S.:]{} [Quasi-Monte Carlo for highly sturctured generalised response models]{} [Methodol. Comput. App.]{} [**10**]{}, 239–275 (2008)
[Kuo, F.Y., Nuyens, D.:]{} [Application of quasi-Monte Carlo methods to elliptic PDEs with random diffusion coefficients – a survey of analysis and implementation]{}, [Found. Comput. Math.]{} [**16**]{}, 1631–1696 (2016)
[Kuo, F.Y., Nuyens, D.:]{} [Application of quasi-Monte Carlo methods to PDEs with random coefficients – an overview and tutorial]{} (to appear in this volume)
[Kuo, F.Y., Nuyens, D., Plaskota, L., Sloan, I.H., Wasilkowski, G.W.:]{} [Infinite-dimensional integration and the multivariate decomposition method]{}, J. Comput. Appl. Math. [**326**]{}, 217–234 (2017)
[Kuo, F.Y., Scheichl, R., Schwab, Ch., Sloan, I.H., Ullmann, E.:]{} [Multilevel Quasi-Monte Carlo methods for lognormal diffusion problems]{}, [Math. Comp.]{} [**86**]{}, 2827–2860 (2017)
[Kuo, F.Y., Schwab, Ch., Sloan, I.H.:]{} [Quasi-Monte Carlo methods for high dimensional integration: the standard weighted-space setting and beyond]{}, [ANZIAM J.]{} [**53**]{}, 1–37 (2011)
[Kuo, F.Y., Schwab, Ch., Sloan, I.H.:]{} [Quasi-Monte Carlo finite element methods for a class of elliptic partial differential equations with random coefficient]{}, [SIAM J. Numer. Anal.]{} [**50**]{}, 3351–3374 (2012)
[Kuo, F.Y., Schwab, Ch., Sloan, I.H.:]{} [Multi-level quasi-Monte Carlo finite element methods for a class of elliptic partial differential equations with random coefficient]{}, [Found. Comput. Math.]{} [**15**]{}, 411–449 (2015)
[Kuo, F.Y., Sloan, I.H., Wasilkowski, G.W., Waterhouse, B.J.:]{} [Randomly shifted lattice rules with the optimal rate of convergence for unbounded integrands]{}, [J. Complexity]{} [**26**]{}, 135–160 (2010)
[Kuo, F.Y., Sloan, I.H., Wasilkowski, G.W., Woźniakowski, H.:]{} [Liberating the dimension]{}, [J. Complexity]{} [**26**]{}, 422–454 (2010)
[Kuo, F.Y., Sloan, I.H., Wasilkowski, G.W., Woźniakowski, H.:]{} [On decompositions of multivariate functions]{}, [Math. Comp.]{} [**79**]{}, 953–966 (2010)
[Lemieux, C.:]{} [Monte Carlo and Quasi-Monte Carlo Sampling]{}, Springer, New York (2009)
[Leobacher, G., Pillichshammer, F.:]{} [Introduction to Quasi-Monte Carlo Integration and Applications]{}, Springer (2014)
[Nguyen, D.T.P., Nuyens, D.:]{} [Multivariate integration over $\mathbb{R}^s$ with exponential rate of convergence]{}, [J. Comput. Appl. Math.]{} [**315**]{}, 327–342 (2017)
[Nichols, J.A., Kuo F.Y.:]{} [Fast CBC construction of randomly shifted lattice rules achieving $\mathcal{O}(N^{-1+\delta})$ convergence for unbounded integrands in $\mathbb{R}^s$ in weighted spaces with POD weights]{}, [J. Complexity]{} [**30**]{}, 444–468 (2014)
[Niederreiter, H.:]{} [Random Number Generation and Quasi-Monte Carlo Methods]{}, SIAM, Philadelphia (1992)
[Novak, E., Woźniakowski, H.:]{} [Tractability of Multivariate Problems, II: Standard Information for Functionals]{}, European Mathematical Society, Zürich (2010)
[Nuyens, D.:]{} The construction of good lattice rules and polynomial lattice rules, in: Uniform Distribution and Quasi-[M]{}onte [C]{}arlo Methods (P. Kritzer, H. Niederreiter, F. Pillichshammer, A. Winterhof, eds.), Radon Series on Computational and Applied Mathematics Vol. 15, De Gruyter, pp. 223–256 (2014)
[Nuyens, D., Cools, R.:]{} [Fast algorithms for component-by-component construction of rank-$1$ lattice rules in shift-invariant reproducing kernel Hilbert spaces]{}, [Math. Comp.]{} [**75**]{}, 903–920 (2006)
[Nuyens, D., Cools, R.:]{} [Fast component-by-component construction of rank-$1$ lattice rules with a non-prime number of points]{}, [J. Complexity]{} [**22**]{}, 4–28 (2006)
[Owen, A.B.:]{} [Scrambled net variance for integrals of smooth functions]{}, [Ann. Statist.]{} **25**, 1541–1562 (1997)
[Owen, A.B.:]{} [Halton sequences avoid the origin]{}, [SIAM Rev.]{} **48**, 487–503 (2006)
[Robbe, P., Nuyens, D., Vandewalle, S.:]{} [A multi-index quasi-Monte Carlo algorithm for lognormal diffusion problems]{}, SIAM J. Sci. Comput. (to appear)
[Scheichl, R., Stuart, A., Teckentrup, A.L.:]{} [Quasi-Monte Carlo and multilevel Monte Carlo methods for computing posterior expectations in elliptic inverse problems]{}, [SIAM/ASA Journal on Uncertainty Quantification]{} [**5**]{}, 493–518 (2017)
[Schwab, Ch., Gittelson, C.J.:]{} [Sparse tensor discretizations of high-dimensional parametric and stoch astic PDEs]{}, [Acta Numer.]{} [**20**]{}, 291–467 (2011)
[Sinescu, V., Kuo, F.Y., Sloan, I.H.:]{} [On the choice of weights in a function space for quasi-Monte Carlo methods for a class of generalised response models in statistics]{}, in: Monte Carlo and Quasi-Monte Carlo Methods 2012 (J. Dick, F. Y. Kuo, G. Peters, and I. H. Sloan, eds), Springer Verlag, Heidelberg, pp. 631–647 (2013)
[Sloan, I.H., Joe, S.:]{} [Lattice Methods for Multiple Integration]{}, Oxford University Press, Oxford (1994)
[Sloan, I.H., Kuo, F.Y., Joe, S.:]{} [Constructing randomly shifted lattice rules in weighted Sobolev spaces]{}, [SIAM J. Numer. Anal.]{} [**40**]{}, 1650–1665 (2002)
[Sloan, I.H., Wang, X., Woźniakowski, H.:]{} [Finite-order weights imply tractability of multivariate integration]{}, [J. Complexity]{} [**20**]{}, 46–74 (2004)
[Sloan, I.H., Woźniakowski, H.:]{} [When are quasi-Monte Carlo algorithms efficient for high-dimensional integrals?]{}, [J. Complexity]{} [**14**]{}, 1–33 (1998)
[Wasilkowski, G.W.:]{} [On tractability of linear tensor product problems for $\infty$-variate classes of functions]{}, [J. Complexity]{}, [**29**]{}, 351–369 (2013)
[^1]: Frances Y. Kuo (): School of Mathematics and Statistics, University of New South Wales, Sydney NSW 2052, Australia, [`[email protected]`]{}
[^2]: Dirk Nuyens: Department of Computer Science, KU Leuven, Celestijnenlaan 200A, 3001 Leuven, Belgium, [`[email protected]`]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Consider a face-to-face parallelohedral tiling of $\mathbb R^d$ and a $(d-k)$-dimensional face $F$ of the tiling. We prove that the valence of $F$ (i.e. the number of tiles containing $F$ as a face) is not greater than $2^k$. If the tiling is affinely equivalent to a Voronoi tiling for some lattice (the so called Voronoi case), this gives a well-known upper bound for the number of vertices of a Delaunay $k$-cell. Yet we emphasize that such an affine equivalence is not assumed in the proof.'
address:
- 'Steklov Mathematical Institute of the Russian Academy of Sciences, 8 Gubkina street, Moscow 119991, Russia'
- 'Yaroslavl State University, 14 Sovetskaya street, Yaroslavl 150000, Russia'
author:
- Alexander Magazinov
title: An upper bound for a valence of a face in a parallelohedral tiling
---
tiling,parallelohedron,Voronoi conjecture
Introduction
============
The central point of the parallelohedra theory is the famous Voronoi conjecture.
\[voronoi\]
Every $d$-dimensional parallelohedron $P$ is affinely equivalent to a Dirichlet-Voronoi domain for some $d$-dimensional lattice.
Although the conjecture was posed in 1909 in the paper [@vor], it has not been proved or disproved so far in the general case. However, several significant partial solutions were obtained [@del; @erd; @ord; @vor; @zhi].
Let $\mathcal T(P)$ be a face-to-face tiling of $\mathbb R^d$ by parallel copies of a parallelohedron $P$. Choose an arbitrary $(d-k)$-dimensional face $F$ of the tiling.
Denote by $\pi$ the orthogonal projection along $\operatorname{\it lin}F$ onto the complementary affine space $(\operatorname{\it lin}F)^\bot$. Then there exists a complete $k$-dimensional cone fan $\operatorname{\it Fan}(F)$ ([*the fan of $F$*]{}) that splits $(\operatorname{\it lin}F)^\bot$ into convex polyhedral cones with vertex $\pi(F)$, and a neighborhood $U = U(\pi(F))$ such that every face $F'\supset F$ corresponds to a cone $C\in \operatorname{\it Fan}(F)$ satisfying $$\pi(F')\cap U = C\cap U.$$ Speaking informally, $\operatorname{\it Fan}(F)$ has the same combinatorics as the transversal section of $\mathcal T(P)$ in a small neighborhood of $F$.
The definition above is equivalent to the definition of a [*star*]{} of a face $F$, introduced by Ryshkov and Rybnikov [@rry], yet seems to be more formal.
Studying the combinatorial structure of such cone fans proved to be an effective tool to verify the Voronoi conjecture in special cases [@del; @ord; @zhi].
For fixed $k$ and varying $d$, it is a complicated problem to classify all possible combinatorial types of $\operatorname{\it Fan}(F)$ if a positive answer to conjecture \[voronoi\] is not preassumed. The classification for $k = 2$ has been obtained in [@min] and consists of 2 combinatorial types of cone fans shown in Figure \[f1\].
\[f1\]
for $k = 3$ Delaunay [@del] proved that every fan of a $(d-3)$-face of a parallelohedral tiling belongs to one of the 5 types shown in Figure \[f2\].
\[f2\]
The case $k = 4$ has been partially considered in [@ord], but the complete classification was not obtained, and for $k>4$ almost nothing is known. The main goal of this paper is to show that a classification is possible [*in principle*]{} for every $k$, i.e. the list of all combinatoral types of $\operatorname{\it Fan}F$, where $F$ is a $(d-k)$-face, is finite. The idea is to obtain an upper bound for the number of tiling parallelohedra incident to a face $F$.
Let $$\nu(F) = card\,\{ P'\in \mathcal T(P) : F\subset P' \}.$$ $\nu(F)$ will be called the [*valence*]{} of the face $F$.
\[2k\]
Let $F$ be a $(d-k)$-dimensional face of a parallelohedral tiling of $\mathbb R^d$. Then $$\nu(F)\leq 2^k.$$
The upper bound $2^k$ is sharp for all integer $d, k$ satisfying $0<k\leq d$; for example, for every $(d-k)$-face $F$ of a cubic tiling of $\mathbb R^d$ holds $$\nu(F) = 2^k.$$
Theorem \[2k\] immediately implies
\[finiteness\]
Given $k\in \mathbb N$, there exists a set of complete $k$-dimensional cone fans $$\{\mathcal C^k_1, \mathcal C^k_2, \ldots, \mathcal C^k_{N(k)} \}$$ such that for every $d$, every $d$-parallelohedron $P$ and every $(d-k)$-face $F$ of $\mathcal T(P)$ the fan of $F$ is isomorphic to some $\mathcal C^k_i$.
Obviously, there is only a finite number of combinatorial types of complete $k$-dimensional cone fans splitting $\mathbb R^k$ into no more than $2^k$ full-dimensional convex polyhedral cones. According to theorem \[2k\], the fan of $F$ necessarily belongs to one of those combinatorial types.
For a centrally symmetric polytope $Q$ denote by $c(Q)$ its center of symmetry.
To proceed with the proof of theorem \[2k\] recall several basic properties of parallelohedral tilings.
1. A parallelohedron $P$ is a centrally symmetric polytope (see [@min]).
2. The set $$\Lambda = \{c(P'): P'\in \mathcal T(P)\}$$ is a lattice (see also [@min]). Under assumption $\mathbf 0 \in \Lambda$, one can also treat $\Lambda$ as a translation group.
3. \[c3\] If $P_1, P_2 \in \mathcal T(P)$ and $P_1\cap P_2 \neq \varnothing$, then $P_1\cap P_2$ is a centrally symmetric face of $\mathcal T(P)$. Moreover, $$c(P_1\cap P_2) = \frac {c(P_1)+c(P_2)}{2}.$$ A face $F$ of $\mathcal T(P)$ representable in the form $F = P_1\cap P_2$, where $P_1, P_2 \in \mathcal T(P)$, is called [*standard*]{} (see [@dol]).
The Voronoi case
================
For a better understanding of the aim, first restrict oneself to the Voronoi case. Start considering this case with a folklore lemma.
\[parity\] Let $P_1, P_2 \in \mathcal T(P)$ be the 2 distinct parallelohedra such that $c(P_1)$ and $c(P_2)$ belong to the same class modulo $2\Lambda$. (In other words, $P_1$ and $P_2$ belong to the same parity class.) Then $P_1\cap P_2 = \varnothing$.
According to property 3 of a parallelohedral tiling (see page ), $$\label{50}
\frac{c(P_1) + c(P_2)}{2} \in (P_1\cap P_2).$$
On the other hand, since $c(P_1)$ and $c(P_2)$ belong to the same class $\mod{2\Lambda}$, $$\frac{c(P_1) + c(P_2)}{2} \in \Lambda.$$ Therefore $\frac{c(P_1) + c(P_2)}{2}$ is a center of some parallelohedron $P_3$ different from $P_1$ and $P_2$. Thus $$\frac{c(P_1) + c(P_2)}{2} \in int\, P_3,$$ which is a contradiction to (\[50\]). Lemma is proved.
The technique used in the proof is rather standard. For example, similar methods were used in [@dsa]. (See also [@vor part 1, p. 277], the description of parallelohedra of a given parity class adjoint to a given parallelohedron by a hyperface.)
Let $\mathcal T_V(\Lambda)$ be the Voronoi tiling for some $d$-dimensional lattice $\Lambda \subset \mathbb R^d$. Then $\mathcal T_V(\Lambda)$ has a dual [*Delaunay*]{} tiling $\mathcal D(\Lambda)$. Every $(d-k)$-face $F$ in $\mathcal T_V(\Lambda)$ has a $k$-dimensional dual face $D(F)$ in $\mathcal D(\Lambda)$ such that $$\label{51}
D(F) = \operatorname{\it conv}\{ c(P): P\in \mathcal T_V(\Lambda)\; \text{and} \; F\in P \}.$$ Moreover, the combinatorial structure of $D(F)$ completely describes the combinatorial structure of $\operatorname{\it Fan}(F)$. In particular, $\nu(F)$ is equal to the number of vertices of $D(F)$.
\[vcase\] The statement of theorem \[2k\] holds for Voronoi tilings.
Following the argument above, it is sufficient to prove that $D(F)$ has at most $2^k$ vertices. According to the properties of a lattice Delaunay tiling, $$dim\, \operatorname{\it aff}D(F) = k,$$ so the set of vertices of $D(F)$ is a subset of some $k$-dimensional lattice $\Lambda(F) \subset \Lambda$.
By lemma \[parity\], every 2 vertices of $D(F)$ belong to different classes $\mod{2\Lambda}$, otherwise 2 parallelohedra of the same parity class had to intersect at $F$. Therefore every 2 vertices of $D(F)$ belong to different classes $\mod{2\Lambda(F)}$. Since $\Lambda(F)$ has exactly $2^k$ classes $\mod{2\Lambda(F)}$, there are at most $2^k$ vertices of $D(F)$. Proposition is proved.
Another approach to estimate the number of vertices of $D(F)$ is described in [@dla Proposition 13.2.8].
Thus, if the Voronoi conjecture (conjecture \[voronoi\]) has a positive answer, theorem \[2k\] is proved.
In the way similar to corollary \[finiteness\], proposition \[vcase\] implies that there are finitely many combinatorial types of $D(F)$ for fixed $k$ and regardless of $d$. Moreover, an algorithm for classification all possible lattice Delaunay $k$-cells is given in [@dsv] An idea of such algorithms certainly belongs to Voronoi [@vor] who constructed an algorithm to classify all possible combinatorial types of Voronoi parallelohedra.
For an arbitrary parallelohedral tiling $\mathcal T(P)$ and its $(d-k)$-face $F$ it is also possible to introduce in the similar way as (\[51\]) the set $$D(F^{d-k}) = \operatorname{\it conv}\{c(P') : P'\in \mathcal T(P) \; \text{and} \; F^{d-k}\subset P' \}.$$ As far as author knows, there are no satisfactory results on $\dim\operatorname{\it aff}D(F)$ in the general case.
In the case $$\label{10}
dim\,\operatorname{\it aff}D(F) \leq k$$ theorem \[2k\] easily follows from lemma \[parity\]. Yet the inequality (\[10\]) remains an open problem.
Outline of the proof
====================
The methods described above essentially involve the assumption that conjecture \[voronoi\] has a positive answer. However, theorem \[2k\] can be proved in non-Voronoi case as well, by exploiting several other ideas.
The proof of theorem \[2k\] consists of the 3 main steps.
1. Construct a set $$\{ F_1 = F, F_2, \ldots, F_m \}$$ of all faces of a parallelohedron $P_0 \in \mathcal T(P)$ such that every 2 faces of the set are equivalent by a $\Lambda$-translation. Prove that $\nu(F_1) = m$.
2. Refine the notion of [*an antipodal set*]{} given in [@dgr]. Prove that the set $$W = \{\pi(F_i) : i = 1, 2, \ldots, m\},$$ is antipodal (here $\pi$ is the above defined projection along $\operatorname{\it lin}F$ onto the complementary space $(\operatorname{\it lin}F)^\bot$).
3. Estimate the cardinality of an arbitrary antipodal set in $\mathbb R^k$.
The third step uses the technique introduced by Danzer and Grünbaum in [@dgr]. However, [@dgr] deals with antipodal full-dimensional point sets in $\mathbb R^k$, i.e. the sets $W$ satisfying $dim \operatorname{\it aff}W = k$. Although Danzer and Grünbaum’s theorem cannot be used directly here, it is not hard to extend the technique to the class of antipodal sets satisfying the refined definition.
Faces equivalent by translation
===============================
To introduce a uniform notation, put $F_1 = F$. Choose a parallelohedron $P_0 \in \mathcal T(P)$ such that $F_1 \subset P_0$. Let $$\{F_1, F_2, \ldots, F_m\}$$ be the set of all faces of $P_0$ equivalent to $F_1$ up to a $\Lambda$-translation.
Denote by $\mathbf t_{ij}$ the vector of $\Lambda$-translation such that $$F_i + \mathbf t_{ij} = F_j.$$ For every $i,j \in \{1, 2, \ldots, m\}$ define $$P_{ij} = P_0 + \mathbf t_{ij}.$$ Clearly, $P_{ij}\in \mathcal T$.
\[valence\]
$\nu(F_1) = m$.
Since $F_i \subset P_0$, one has $$F_1 = F_i + \mathbf t_{i1} \subset P_0 + \mathbf t_{i1} = P_{i1}.$$
Therefore there are at least $m$ parallelohedra of $\mathcal T(P)$ meeting at $F_1$, because the parallelohedra $$P_0 = P_{11}, P_{21}, P_{31}, \ldots, P_{m1}$$ are pairwise different.
Suppose there exists one more parallelohedron $P'\in \mathcal T(P)$ such that $F_1\subset P'$. Thus there is a non-zero $\Lambda$-translation $\mathbf t$ such that $P' = P_0 + \mathbf t$. Obviously, $\mathbf t \neq \mathbf t_{i1}$, so $-\mathbf t\neq \mathbf t_{1i}$ for every $i=1,2,\ldots, m$.
Let $F' = F_1 - \mathbf t$. Since $F_1 \subset P'$, $$F' = F_1 - \mathbf t \subset P' - \mathbf t = P_0.$$ Consequently, $F'$ is a face of $P_0$ equivalent to $F_1$ up to a $\Lambda$-translation, hence $F' = F_i$ for some $i\in \{2,3,\ldots, m\}$. By definition of $\mathbf t_{1i}$, one has $$F_1 - \mathbf t = F' = F_i = F_1 + \mathbf t_{1i}.$$ Therefore $-\mathbf t = \mathbf t_{1i}$, which is a contradiction. So, there are exactly $m$ parallelohedra of $\mathcal T(P)$ meeting at $F_1$, namely $$P_0, P_{21}, P_{31}, \ldots, P_{m1}.$$ Thus $\nu(F_0) = m$.
Constructing an antipodal set
=============================
Call a finite set $W\subset \mathbb R^k$ [*antipodal*]{} if for every pair of distinct points $x, y \in W$ there exists a pair of [*distinct*]{} parallel hyperplanes $\beta, \gamma$ such that $$x\in \beta, \quad y\in \gamma,$$ and $W$ lies between $\beta$ and $\gamma$.
Recall that $\pi$ is a projection along $\operatorname{\it lin}F_1$ onto the complementary space $\mathbb R^k$.
\[antipod\]
Let $w_i = \pi(F_i)$. Then $$W = \{ w_i: i = 1, 2, \ldots, m \}$$ is an antipodal set in $\mathbb R^k$.
It is sufficient to show that for every integer $i,j, 1\leq i<j\leq m$ there exist two distinct parallel hyperplanes $\gamma_{ij}$ and $\gamma_{ji}$ such that $$w_i \in \gamma_{ij}, \quad w_j\in \gamma_{ji},$$ and $W$ lies between these hyperplanes.
In $\mathbb R^d$ take the hyperplane $\Gamma_{ij}$ that separates the parallelohedra $P_0$ and $P_{ji}$ from each other. ($\Gamma_{ij}$ has to be a supporting hyperplane to each of these 2 parallelohedra.) Since $F_j\subset P_0$, $$F_i = F_j + \mathbf t_{ji} \subset P_0 + \mathbf t_{ji} = P_{ji},$$ and therefore $F_i \subset P_0\cap P_{ji} \subset \Gamma_{ij}$.
By definition, put $\Gamma_{ji} = \Gamma_{ij} - \mathbf t_{ji}$. The hyperplane $\Gamma_{ij}$ is supporting to $P_{ji}$, so the hyperplane $\Gamma_{ij}$ is supporting to $P_{ji} - \mathbf t_{ji} = P_0$. Also, since $F_i \subset \Gamma_{ij}$, $$F_j = F_i - \mathbf t_{ji} \subset \Gamma_{ij} - \mathbf t_{ji} = \Gamma_{ji}.$$
Thus $\Gamma_{ij}$ and $\Gamma_{ji}$ are two parallel supporting hyperplanes satisfying $$F_i \subset \Gamma_{ij}, \quad F_j\subset \Gamma_{ji}.$$
Now define $$\gamma_{ij} = \pi (\Gamma_{ij}) \quad \text{and} \quad \gamma_{ji} = \pi (\Gamma_{ji}).$$ Since $P_0$ lies between $\Gamma_{ij}$ and $\Gamma_{ji}$, then $W$ lies between $\gamma_{ij}$ and $\gamma_{ji}$. The hyperplanes $\gamma_{ij}$ and $\gamma_{ji}$ are distinct because the point $\pi(c(P_0))$ lies strictly between them. Thus the required hyperplanes are constructed for every pair of points $(w_i, w_j)$. Lemma \[antipod\] is proved.
Cardinality of an antipodal set
===============================
\[card\]
Let $W = \{w_1, w_2, \ldots, w_m\} \subset \mathbb R^k$ be an antipodal set. Then $m\leq 2^k$.
Denote by $H_x^a$ the homothety with center $x$ and coefficient $a$.
Take an arbitrary $a\in (0, \frac 12)$ and prove that $$H_{w_i}^a(\operatorname{\it conv}W) \cap H_{w_j}^a(\operatorname{\it conv}W) = \varnothing.$$
Indeed, let $\beta_{ij}$ be the hyperplane parallel and equidistant to the hyperplanes $\gamma_{ij}$ and $\gamma_{ji}$ . Then the sets $H_{w_i}^a(\operatorname{\it conv}W)$ and $H_{w_j}^a(\operatorname{\it conv}W)$ lie in different open half-spaces in respect to $\beta_{ij}$ and therefore do not intersect.
Let $k'\leq k$ be the affine dimension of $W$. Since $$H_{w_i}^a(\operatorname{\it conv}W)\subset \operatorname{\it conv}W$$ and because the sets $H_{w_i}^a(W)$ are pairwise non-intersecting, one has $$\label{60}
vol_{k'}(\operatorname{\it conv}W)\geq \sum\limits_{i=1}^m vol_{k'}(H_{w_i}^a(\operatorname{\it conv}W)) = m\cdot a^{k'}\,vol_{k'}(\operatorname{\it conv}W),$$ where $vol_{k'}$ stands for the $k'$-dimensional volume.
Further, because $vol_{k'}(W)>0$, (\[60\]) implies $$\label{61}
m \leq a^{-k'}.$$ The inequality (\[61\]) holds for every $0<a<\frac 12$, so it holds also for $a = \frac 12$: $$m \leq 2^{k'}\leq 2^k,$$
Lemma \[card\] is proved.
Now the statement of theorem \[2k\] is easily obtained by combining lemma \[antipod\] and lemma \[card\].
Acknowledgements {#acknowledgements .unnumbered}
================
There are a number of people who participated in the discussion of ideas and results, namely, R. Erdahl, F. Vallentin, A. Schürmann, A. Garber, A. Gavrilyuk and M. Kozachok.
The research was partially done at Fields Institute, Toronto, Canada during the Thematic Semester of Discrete Geometry and Applications and at Queen’s University, Kingston, Canada with an invitation of prof R. Erdahl. Author also appreciates the effort by the Laboratory of Geometrical Methods in Mathematical Physics at MSU, Moscow to make possible the visit to Canada.
Finally, author acknowledges prof. N. Dolbilin for scientific guidance, an introduction to the theory of parallelohedra and numerous useful comments on the text.
[99]{}
L. Danzer, B. Grünbaum, Über zwei Probleme bezüglich konvexer Körpern von P. Erdös und von V. L. Klee, Math. Z., 79 (1962), 95 – 99.
B. N. Delaunay, Sur la partition régulière de l’espace à 4 dimensions, Izv. Acad. sci. of the USSR. Ser. VII. Sect. of phys. and math. sci., 1 – 2 (1929), 79 – 100, 147 – 164.
B. N. Delaunay, N. N. Sandakova, Theory of stereohedra, Proc. of Steklov Math. Inst., 64 (1961), 28 – 51. (In Russian)
M. M. Deza, M. Laurent, Geometry of Cuts and Metrics, Springer, Berlin – Heidelberg – New York, 1997.
N. P. Dolbilin, Properties of faces of parallelohedra, Proc. Steklov Inst. of Math., 266 (2009), 112 – 126.
M. Dutour Sikirić, A. Schürmann, F. Vallentin, Complexity and algorithms for computing Voronoi cells of lattices, Mathematics of computation, 78 (2009), 1713 – 1731.
R. Erdahl, Zonotopes, dicings, and Voronoi’s conjecture on parallelohedra, European Journal of Combinatorics, 20(6), 1999, 527 – 549.
H. Minkowski, Allgemeine Lehrsätze über die konvexe Polyeder, Nach. Ges. Wiss., Göttingen, 1897, 198 – 219.
A. Ordine, Proof of the Voronoi conjecture on parallelotopes in a new special case, Ph.D. Thesis, Queen’s University, Ontario, 2005.
S. S. Ryshkov, K. A. Rybnikov Jr., The theory of quality translations with applications to tilings, Europ. J. Combinatorics, 18 (1997), 431 – 444.
G. Voronoi, Nouvelles applications des paramètres continus à la théorie des formes quadratiques. Deuxième mémoire. Recherches sur les paralléloèdres primitifs, J. Reine Angew. Math., 134 (1908), 198 – 287 and 136 (1909), 67 – 178.
O. K. Zhitomirskii, Verschärfung eines Satzes von Voronoi, J. of Leningrad Math. Soc., 2 (1929), 131 – 151.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We establish the existence of two rigid triples of conjugacy classes in the algebraic group $G_{2}$ in characteristic $5$, complementing results of the second author with Liebeck and Marion. As a corollary, the finite groups $G_{2}(5^n)$ are not $(2,4,5)$-generated, confirming a conjecture of Marion in this case.'
address:
- 'M.J. Conder, Department of Pure Mathematics and Mathematical Statistics, Centre for Mathematical Sciences, University of Cambridge, Wilberforce Road, Cambridge, CB3 0WB, United Kingdom'
- 'A.J. Litterick, Fakultät für Mathematik, Ruhr-Universität Bochum, Universitätsstra[ß]{}e 150, D-44780 Bochum, Germany Fakultät für Mathematik, Universität Bielefeld, Postfach 100131, D-33501 Bielefeld, Germany'
author:
- Matthew Conder
- Alastair Litterick
title: 'Further Rigid Triples of Classes in [$G_{2}$]{}'
---
[^1]
[^2]
Introduction
============
Let $G$ be a connected simple algebraic group over an algebraically closed field $K$, and let $C_{1}$, $\ldots$, $C_{s}$ be conjugacy classes of $G$. Following [@Strambach1999], we say the $s$-tuple $\mathbf{C} = (C_{1},\ldots,C_{s})$ is *rigid* in $G$ if the set $$\mathbf{C}_{0} \stackrel{\textup{def}}{=} \{ (x_{1},\ldots,x_{s}) \in C_{1} \times \ldots \times C_{s} \, : \, x_{1}x_{2}\ldots x_{s} = 1 \}$$ is non-empty and forms a single orbit under the action of $G$ by simultaneous conjugation.
Some well-known examples of rigid tuples of classes in simple algebraic groups are the Belyi triples and Thompson tuples, defined in [@Voelklein1998]. Other rigid triples are known, see for instance [@Dettweiler1999; @Feit1985; @Guralnick2014; @Liebeck2011; @Thompson1985]. Rigid tuples of classes are interesting in the context of the inverse Galois problem [@Malle1999], and also arise naturally in the theory of ordinary differential equations [@Katz1996].
Recall that a group is $(a,b,c)$-generated if it is generated by elements $x$, $y$ and $z$, of respective orders $a$, $b$ and $c$, such that $xyz = 1$. The group is then called an $(a,b,c)$-group, and the triple $(x,y,z)$ is called an $(a,b,c)$-triple of the group. The theory of $(a,b,c)$-generation of finite groups has close connections to rigidity, for instance it is a basic observation that given a rigid tuple $\mathbf{C}$ of classes of $G$, all subgroups $\left<x_1,\ldots,x_s\right>$ for $(x_1,\ldots,x_s) \in \mathbf{C}_{0}$ are conjugate in $G$, so that there is at most one $r > 0$ such that the finite subgroup $G(p^{r})$ is generated by elements in such an $s$-tuple.
Let $K = \bar{\mathbb{F}}_{5}$ be the algebraic closure of the field of five elements. In [@Liebeck2011] it is shown that the simple algebraic group $G = G_{2}(K)$ has a rigid triple of conjugacy classes of elements of orders $2$, $5$ and $5$, and any triple of elements $(x_1,x_2,x_3)$ in the corresponding set $\mathbf{C}_{0}$ generates a copy of $\textup{Alt}_{5}$. This is then used to show that none of the groups $G_{2}(5^n)$, $SL_{3}(5^n)$ or $SU_{3}(5^n)$ is a $(2,5,5)$-group.
Here we produce two further rigid triples of classes in $G = G_2(K)$, closely related to the triple above. Recall from [@Chang1974] that $G$ has a unique class of involutions, with representative $t$, say, and $C_{G}(t) = A_{1}\tilde{A}_{1}$ is a central product of two subgroups $SL_{2}(K)$, where $A_{1}$ (resp. $\tilde{A}_{1}$) is generated by a long (resp. short) root subgroup of $G$. There also exist two classes of elements of order $4$, with representatives $s_{1}$ and $s_{2}$, such that $C_{G}(s_{1}) = A_{1}T'$ and $C_{G}(s_{2}) = \tilde{A}_{1}T''$, where $T'$ and $T''$ are $1$-dimensional tori. Finally, recall from [@Lawther1995] that $G$ has three classes of unipotent elements of order 5: the long and short root elements, and the class labelled $G_{2}(a_1)$, with representative $u = x_{\beta}(1)x_{3\alpha + \beta}(1)$, where $\alpha$ (resp. $\beta$) is the short (resp. long) simple root of $G$. From [@Liebeck2012 Table 22.1.5], the centraliser $C_{G}(u) = U_{4}.\textup{Sym}_{3}$, where $U_{4}$ is a $4$-dimensional connected unipotent group.
\[thm:main\]
- The triples of classes $\mathbf{C} = (t^{G},s_{1}^{G},u^{G})$ and $\mathbf{D} = (t^{G},s_{2}^{G},u^{G})$ are rigid in $G = G_{2}(K)$.
- Every triple of elements in $\mathbf{C}_{0}$ or $\mathbf{D}_{0}$ generates a subgroup isomorphic to the symmetric group $\textup{Sym}_{5}$.
- None of the groups $G_{2}(5^n)$ are a $(2,4,5)$-group for any $n$. Neither are the groups $SL_{3}(5^n)$ or $SU_{3}(5^n)$.
1. Each subgroup $\textup{Sym}_{5}$ in part (ii) here contains a subgroup $\textup{Alt}_{5}$ arising from [@Liebeck2011 Theorem 1(ii)].
2. Keeping track of details in the proof in [@Liebeck2011] shows that $G_{2}(K)$ has a unique class of subgroups $\textup{Alt}_{5}$. These subgroups have centraliser $\textup{Sym}_{3}$, and by Lang’s theorem these split into three classes in $G_{2}(5^r)$, with centraliser orders $6$, $3$ and $2$. Similarly, if $S$ and $S'$ are representatives of the two subgroup classes in part (ii) here, then $C_{G}(S) \cong \textup{Sym}_{3}$, while $C_{G}(S')$ is cyclic of order $2$. It follows that the class of $S$ (resp. $S'$) splits into $3$ (resp. $2$) classes of subgroups in $G_{2}(5^r)$, with centralisers of order $6$, $3$, $2$ (resp. $2$ and $2$).
3. A conjecture of Marion [@Marion2010] states that, for a simple algebraic group $G$ in characteristic $p$, if $\delta_{i}$ denotes the dimension of the variety of elements of $G$ of order $i$ and if $\delta_{a} + \delta_{b} + \delta_{c} = 2\, \textup{dim}(G)$, then at most finitely many of the finite groups $G(p^{r})$ are $(a,b,c)$-groups. For $G$ of type $G_{2}$, this criterion holds precisely when $(a,b,c) = (2,4,5)$ or $(2,5,5)$. Hence part (iii), together with [@Liebeck2011 Theorem 1(iii)], verifies the conjecture for $G = G_{2}$ in characteristic $5$. A non-constructive proof of this fact is given in [@Jambor2017 Proposition 3.7(i)], where it is shown that every $(2,4,5)$-subgroup and $(2,5,5)$-subgroup of $G_{2}(K)$ is reducible on the natural $7$-dimensional module, by considering the dimensions of $SL_{7}(K)$-conjugacy classes of elements in the relevant $(a,b,c)$-triples.
Proof of the Theorem
====================
We proceed in the manner of [@Liebeck2011]. Let $G = G_{2}(K)$ and $t$, $u$, $s_{1}$, $s_{2} \in G$ as above. If $\sigma$ is a Frobenius morphism of $G$ induced from the field map $x \mapsto x^{5}$ of $K$, then $$G = \bigcup_{n=1}^{\infty} G_{\sigma^{n}} = \bigcup_{n = 1}^{\infty} G_{2}(5^n).$$ The element $u = x_{\beta}(1)x_{3\alpha + \beta}(1)$ is a regular unipotent element in a subgroup $A_{2} = SL_{3}(K)$ of $G$ generated by long root groups, and therefore lies in a subgroup $\Omega_{3}(5) \cong \textup{Alt}_{5}$ of $G$, which we denote by $A$. Now, let $S = N_{A_2}(A) = SO_{3}(5) \cong \textup{Sym}_{5}$. Following the proof given in [@Liebeck2011] we find that $N_{G}(A) = S \times C_{G}(A)$ and $C_{G}(A) = \left<z,\tau\right> \cong \textup{Sym}_{3}$, where $\left<z\right>$ is the centre of $A_{2}$ and $\tau$ is an outer involution in $N_{G}(A_{2}) = A_{2}.2$. Note that $C_{A_2}(\tau) = SO_{3}(K)$, so $\tau \in C_{G}(S)$.
Let $v$ be an involution in $S \setminus A$, so that $S = \left<A,v\right>$, and define $S' = \left<A,v\tau\right>$, so that $S' \cong \textup{Sym}_{5}$ also. Then $C_{G}(S)$, $C_{G}(S') \le C_{G}(A) = \left<z,\tau\right>$ and therefore $$\begin{aligned}
C_{G}(S) &= \left<z,\tau\right>, \label{cgs} \\
C_{G}(S') &= \left<\tau\right>. \label{cgs'}\end{aligned}$$ In particular $S$ and $S'$ are not conjugate in $G$.
Next consider the set of $(2,4,5)$-triples of $\textup{Sym}_{5}$. It is straightforward to show that there are exactly $120$ such triples, and that $\textup{Sym}_{5}$ acts transitively on these by simultaneous conjugation.
Now let $\mathbf{C} = (t^{G},s_1^{G},u^{G})$ and $\mathbf{D} = (t^{G},s_{2}^{G},u^{G})$, and for $q$ a fixed power of $5$ let $\mathbf{C}_{0}(q) = \mathbf{C}_{0} \cap G_{2}(q)^{3}$ and $\mathbf{D}_{0}(q) = \mathbf{D}_{0} \cap G_{2}(q)^{3}$. We now show that $|\mathbf{C}_{0}(q)| = |\mathbf{D}_{0}(q)| = |G_{2}(q)|$. For this we require the character table of $G_{2}(q)$, given in [@Chang1974] and available in the CHEVIE [@Geck1996] computational package. Since $C_{G}(u)/C_{G}(u)^{\circ} = S_{3}$, an application of Lang’s theorem [@Malle2011 Theorem 21.11] shows that $u^{G} \cap G_{2}(q)$ splits into three classes, with representatives denoted in [@Chang1974] by $u_{3}$, $u_{4}$ and $u_{5}$, and respective centraliser orders $6q^{4}$, $3q^{4}$ and $2q^{4}$. For $x,y,z \in G_{2}(q)$ let $a_{xyz}$ be the corresponding class algebra constant. Calculations with the character table show that $$\begin{aligned}
a_{ts_{i}u_{j}} &= \left\{
\begin{array}{cl}
q^{4} & \textup{ if } i = 1,\ j \in \{3,4,5\} \textup{ or } i = 2,\ j = 4, \\
3q^{4} & \textup{ if } i = 2,\ j = 3, \\
0 & \textup{ if } i = 2, j = 5.
\end{array}
\right.\end{aligned}$$ and it follows that $$\begin{aligned}
|\mathbf{C}_{0}(q)| &= \sum_{j = 3}^{5} |u_{j}^{G_2(q)}|a_{ts_{1}u_{j}} = |G_{2}(q)| \left(\frac{q^{4}}{6q^{4}} + \frac{q^{4}}{3q^{4}} + \frac{q^{4}}{2q^{4}}\right) = |G_{2}(q)|, \\
|\mathbf{D}_{0}(q)| &= \sum_{j = 3}^{5} |u_{j}^{G_2(q)}|a_{ts_{2}u_{j}} = |G_{2}(q)| \left(\frac{3q^{4}}{6q^{4}} + \frac{q^{4}}{2q^{4}}\right) = |G_{2}(q)|.\end{aligned}$$
Now let $\mathbf{E}$ denote (resp. $\mathbf{E}'$) denote the set of triples $(x_1,x_2,x_3) \in \mathbf{C}_{0} \cup \mathbf{D}_{0}$ which generate a conjugate of $S$ (resp. a conjugate of $S'$). Then $G$ is transitive on both $\mathbf{E}$ and $\mathbf{E}'$, since if $\left<x_1,x_2,x_3\right> = \left<y_1,y_2,y_3\right>^{g}$ are each isomorphic to $\textup{Sym}_{5}$, then $(x_1^{g},x_2^{g},x_3^{g})$ and $(y_1,y_2,y_3)$ are $(2,4,5)$ triples in a fixed copy of $\textup{Sym}_{5}$, hence conjugate in $\textup{Sym}_{5}$ by the observation above. Moreover both $\mathbf{E}$ and $\mathbf{E}'$ are non-empty, since $S$ and $S'$ each contain $(2,4,5)$-triples and a unique conjugacy class of unipotent elements, whose elements are conjugate to an element of $A$ and therefore are conjugate to $u$. By (\[cgs\]) and (\[cgs’\]) the stabiliser of a point in $\mathbf{E}$ is isomorphic to $\textup{Sym}_{3}$, and the stabiliser of a point in $\mathbf{E}'$ is cyclic of order $2$. Hence applying Lang’s theorem shows that $\mathbf{E}(q) = \mathbf{E} \cap G_{2}(q)^{3}$ splits into three $G_{2}(q)$-orbits, of orders $|G_{2}(q)|/r$ for $r = 2,3,6$, and similarly $\mathbf{E}'(q) = \mathbf{E}' \cap G_{2}(q)^{3}$ splits into two orbits, each of order $|G_{2}(q)|/2$. Therefore, $$\begin{aligned}
|\mathbf{E}(q)| + |\mathbf{E}'(q)| = |G_{2}(q)|\left(\frac{1}{6} + \frac{1}{3} + \frac{1}{2} + \frac{1}{2} + \frac{1}{2}\right) = |\mathbf{C}_{0}(q)| + |\mathbf{D}_{0}(q)|\end{aligned}$$ and it follows that $\mathbf{C}_{0}(q) \cup \mathbf{D}_{0}(q) = \mathbf{E}(q) \cup \mathbf{E}'(q)$ for each $q$. Therefore $$\mathbf{C}_{0} \cup \mathbf{D}_{0} = \bigcup_{n = 1}^{\infty}\mathbf{C}_{0}(5^n) \cup \mathbf{D}_{0}(5^n) = \bigcup_{n = 1}^{\infty}\mathbf{E}(5^n) \cup \mathbf{E}'(5^n) = \mathbf{E} \cup \mathbf{E}'$$ Hence $G$ has exactly two orbits on $\mathbf{C}_{0} \cup \mathbf{D}_{0}$. A triple in $\mathbf{C}_{0}$ cannot lie in the same orbit as a triple in $\mathbf{D}_{0}$ since the corresponding elements of order $4$ are not $G$-conjugate, and it follows that the two $G$-orbits are $\mathbf{C}_{0}$ and $\mathbf{D}_{0}$.
This proves parts (i) and (ii) of the Theorem. For part (iii), suppose that $G_{2}(5^n)$, $SL_{3}(5^n)$ or $SU_{3}(5^n)$ is a $(2,4,5)$-group, with corresponding set of generators $x_1$, $x_2$, $x_3$. Since $L(G_2) \downarrow A_{2}$ is a direct sum of $L(A_2)$ and two 3-dimensional irreducible $A_{2}$-modules (cf. [@Liebeck1996a Table 8.5]), it follows that $C_{L(G_2)}(x_1,x_2,x_3) = 0$. An application of a result of Scott [@Scott1977] to the module $L(G)$, as in the proof of [@Strambach1999 Corollary 3.2], then yields $$\textup{dim}(x_{1}^{G}) + \textup{dim}(x_{2}^{G}) + \textup{dim}(x_{3}^{G}) \ge 2\, \textup{dim}(G) = 28,$$ implying $(x_{1}^{G},x_{2}^{G},x_{3}^{G}) = \mathbf{C}$ or $\mathbf{D}$, which contradicts part (ii) of the Theorem.
[99]{} Bomshik Chang and Rimhak Ree, *The characters of [$G_{2}(q)$]{}*, Symposia [M]{}athematica, [V]{}ol. [XIII]{} ([C]{}onvegno di [G]{}ruppi e loro [R]{}appresentazioni, [INDAM]{}, [R]{}ome,1972), Academic Press, London, 1974, pp. 395–413. [MR ]{}[0364419 (51 \#673)]{}
Michael Dettweiler and Stefan Reiter, *On rigid tuples in linear groups of odd dimension*, J. Algebra **222** (1999), no. 2, 550–560. [MR ]{}[1734230 (2000m:12007)]{}
W. Feit and P. Fong, *Rational rigidity of [$G_2(p)$]{} for any prime [$p >
5$]{}*, Proceedings of the [R]{}utgers group theory year, 1983–1984 ([N]{}ew [B]{}runswick, [N]{}.[J]{}., 1983–1984), Cambridge Univ. Press, Cambridge, 1985, pp. 323–326. [MR ]{}[817266 (87e:12006b)]{}
Meinolf Geck, Gerhard Hiss, Frank L[ü]{}beck, Gunter Malle, and G[ö]{}tz Pfeiffer, *C[HEVIE]{}—a system for computing and processing generic character tables*, Appl. Algebra Engrg. Comm. Comput. **7** (1996), no. 3, 175–210, Computational methods in Lie theory (Essen, 1994). [MR ]{}[1486215 (99m:20017)]{}
Robert Guralnick and Gunter Malle, *Rational rigidity for [$E_8(p)$]{}*, Compos. Math. **150** (2014), no. 10, 1679–1702. [MR ]{}[3269463]{}
Sebastian Jambor, Alastair Litterick, and Claude Marion, *On finite simple images of triangle groups*, To appear in Israel Journal of Mathematics.
Nicholas M. Katz, *Rigid local systems*, Annals of Mathematics Studies, vol. 139, Princeton University Press, Princeton, NJ, 1996. [MR ]{}[1366651 (97e:14027)]{}
R. Lawther, *Jordan block sizes of unipotent elements in exceptional algebraic groups*, Comm. Algebra **23** (1995), no. 11, 4125–4156. [MR ]{}[1351124]{}
Martin W. Liebeck, Alastair J. Litterick, and Claude Marion, *A rigid triple of conjugacy classes in [$G_2$]{}*, J. Group Theory **14** (2011), no. 1, 31–35. [MR ]{}[2764920 (2011m:20115)]{}
Martin W. Liebeck and Gary M. Seitz, *Reductive subgroups of exceptional algebraic groups*, Mem. Amer. Math. Soc. **121** (1996), no. 580, vi+111. [MR ]{}[1329942 (96i:20059)]{}
[to3em]{}, *Unipotent and nilpotent classes in simple algebraic groups and [L]{}ie algebras*, Mathematical Surveys and Monographs, vol. 180, American Mathematical Society, Providence, RI, 2012. [MR ]{}[2883501]{}
Gunter Malle and B. Heinrich Matzat, *Inverse [G]{}alois theory*, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 1999. [MR ]{}[1711577 (2000k:12004)]{}
Gunter Malle and Donna Testerman, *Linear algebraic groups and finite groups of [L]{}ie type*, Cambridge Studies in Advanced Mathematics, vol. 133, Cambridge University Press, Cambridge, 2011. [MR ]{}[2850737 (2012i:20058)]{}
Claude Marion, *On triangle generation of finite groups of [L]{}ie type*, J. Group Theory **13** (2010), no. 5, 619–648. [MR ]{}[2720195 (2011j:20031)]{}
Leonard L. Scott, *Matrices and cohomology*, Ann. of Math. (2) **105** (1977), no. 3, 473–492. [MR ]{}[0447434 (56 \#5746)]{}
Karl Strambach and Helmut V[ö]{}lklein, *On linearly rigid tuples*, J. Reine Angew. Math. **510** (1999), 57–62. [MR ]{}[1696090 (2000e:20075)]{}
J. G. Thompson, *Rational rigidity of [$G_2(5)$]{}*, Proceedings of the [R]{}utgers group theory year, 1983–1984 ([N]{}ew [B]{}runswick, [N]{}.[J]{}., 1983–1984), Cambridge Univ. Press, Cambridge, 1985, pp. 321–322. [MR ]{}[817265 (87e:12006a)]{}
Helmut V[ö]{}lklein, *Rigid generators of classical groups*, Math. Ann. **311** (1998), no. 3, 421–438. [MR ]{}[1637911 (99g:12005)]{}
[^1]: The first author is jointly supported by the Cambridge and Woolf Fisher Trusts.
[^2]: The second author is supported by the Alexander von Humboldt Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The recently reported magnetic ordering in insulating two-dimensional (2D) materials, such as chromium triiodide (CrI$_3$) and chromium tribromide (CrBr$_3$), opens new possibilities for the fabrication of magneto-electronic devices based on 2D systems. Inevitably, the magnetization and spin dynamics in 2D magnets are strongly linked to Joule heating. Therefore, understanding the coupling between spin, charge and heat, i.e. spin caloritronic effects, is crucial. However, spin caloritronics in 2D ferromagnets remains mostly unexplored, due to their instability in air. Here we develop a fabrication method that integrates spin-active contacts with 2D magnets through hBN encapsulation, allowing us to explore the spin caloritronic effects in these materials. The angular dependence of the thermal spin signal of the CrBr$_3$/Pt system is studied, for different conditions of magnetic field and heating current. We highlight the presence of a significant magnetic proximity effect from CrBr$_3$ on Pt revealed by an anomalous Nernst effect in Pt, and suggest the contribution of the spin Seebeck effect from CrBr$_3$. These results pave the way for future magnonic devices using air-sensitive 2D magnetic insulators.'
author:
- Tian Liu
- Julian Peiro
- 'Dennis K. de Wal'
- 'Johannes C. Leutenantsmeyer'
- 'Marcos H.D. Guimarães'
- 'Bart J. van Wees'
title: 'Spin caloritronics in a CrBr$_3$ based magnetic van der Waals heterostructure'
---
[^1]
[^2]
Introduction
============
The search for magnetism in 2D systems has been a non-trivial topic for decades. Recently, 2D magnetism was demonstrated in an insulating material CrI$_3$ [@huang2017layer], which shows antiferromagnetic exchange between the layers, resulting in zero (non-zero) net magnetization for even (odd) number of layers. It has been shown that CrBr$_3$ exhibits ferromagnetism when exfoliated down to a few layers [@Kim2019evolution] and monolayers [@Kim2019a] while preserving its magnetic order.
This discovery offers us a platform to explore magnonics in 2D systems. Magnonics refers to spintronics based on magnons, which are quantized spin waves, i.e. collective excitations of ordered electron spins in magnetic materials [@bloch1930theorie; @Kittel2004; @shan2018coupled]. Magnonic spin transport has been extensively studied in various ways in 3D magnetic insulators, e.g. spin pumping [@tserkovnyak2002enhanced], Spin Seebeck Effect (SSE) [@uchida2010spin] and electrical injection and detection of magnons [@cornelissen2015long]. The outstanding magnon transport properties of the ferrimagnetic insulator yttrium iron garnet (YIG) and the robustness and fast dynamic of AFM materials like iron oxide [@lebrun2018tunable] and nickel oxide [@hahn2014conduction] triggered the development of the first magnon transport device prototypes for application using these materials [@cornelissen2015long; @Wu2018; @Chumak2014]. The predicted novel physical phenomena [@Pershoguba2018; @Cheng2016; @Nakata2017; @Xu2019; @Ghazaryan2018] hosted by low-dimensional magnon systems represent a strong potential for 2D magnonics. Thermally excited magnon transport was reported recently in an AFM vdW 2D material MnPS$_3$ [@xing2019magnon]. However, magnonics in 2D van der Waals magnetic systems still remains mostly unexplored, especially for 2D ferromagnetic (FM) systems.
One of the difficulties to study such phenomena is the easy degradation in air of the magnetic 2D materials, bringing extra technical challenges for integrating magnonic circuits with these materials. Here, we introduce a technique of bottom metallic contacts on an air-sensitive material CrBr$_3$, aiming at preliminary study of magnonics in 2D ferromagnetic materials. We select CrBr$_3$ as a medium for 2D magnonics study [@tsubokawa1960magnetic] as its FM order is independent on the number of layers and thus it simplifies the device fabrication. The Curie temperature of CrBr$_3$ is about T$_c$=37 $K$ [@tsubokawa1960magnetic] in bulk, reducing to 27 K for monolayers [@Kim2019a]. CrBr$_3$ presents Perpendicular Magnetic Anisotropy (PMA) [@Kim2019evolution] with an out-of-plane coercive field of 4 mT and an in-plane saturation field of 400 mT for a few layers [@Kim2019a]. The saturation magnetization of about 271 kA/m is reported nearly equal for in-plane and out-of-plane orientation in bulk and differs by less than 20% for 3-layer CrBr$_3$ [@Kim2019evolution; @Richter2018].
Device Geometry and Measurements
================================
In this work we employ non-local angular-dependent magnetoresistance (nlADMR) measurements on a hBN-encapsulated CrBr$_3$ flake contacted by Pt strips. ADMR measurements have been widely used to characterize the spin Hall Magnetoresistance (SMR) in local geometries [@nakayama2013spin] or the spin Seebeck effect (SSE) in non-local geometry [@cornelissen2015long] and identify them out of other caloritronics effects [@KikkawaPRB2013; @Meier2015; @Meyer2017; @Avci2020]. We fabricated a device where Pt strips (5.5 nm thick) are deposited into a pre-etched 16.6 nm-thick hBN flake on top of a silicon oxide substrate. A 6.5 nm-thick top hBN flake is used to pick up and fully cover a 7 nm-thick CrBr$_3$ flake (about 10 layers)[@zomer2011transfer]. A schematic of the device and non-local measurement geometry is shown in Fig.1a.
{width="440pt"}
In this system, a gradient of temperature is created by the Joule heating from a remote Pt heater which generates a magnon-mediated spin flow due to the magnon density dependence on the temperature [@LudoThesis], i.e. the SSE. At the interface between a magnet and a non-magnetic material, a transfer of magnon spin ($+\hbar$) from the CrBr$_3$ to the Pt is possible by spin flip of a $-\hbar/2$ spin to a $+\hbar/2$ spin in the Pt. The spin current generated this way in the Pt contact converts into a charge current by inverse spin Hall effect (ISHE) and can be measured as a voltage difference. In the geometry defined in Fig.\[fig:1\], the ADMR is then sensitive to the $x$ component of the magnetization of CrBr$_3$, $\mathbf{M}_x$. In the in-plane ADMR configuration (Fig.\[fig:1\]a and d), the orientation of the magnetization with regard to the detection contact drives the angular dependence, therefore a $cos(\varphi)$ dependence is expected.
All data shown in the main text was measured on a pair formed of a 310 nm-wide injector and a 520 nm-wide detector, spaced by 500 nm edge to edge, and at a base temperature of 5K under a reference magnetic field of 4T, unless specifically mentioned. We separate different harmonics by using a standard low frequency (6Hz to 13Hz) lock-in techniques. The voltage response is composed of different orders and are expanded as: $ V\left(t\right) = R_1I\left(t\right) + R_2I\left(t\right)^2 + \cdots $ [@cornelissen2015long], where $R_i$ is the $i$th-order response [@bakker2010interplay] to the applied AC current $I\left(t\right)$. As the electrical magnon injection scales linearly with current, its response is expected in the first harmonic signal. The thermal injection depends quadratically on the applied current and the response appears in the second harmonic signal.
![The dependence of second harmonic signals on applied current through the injector. (a) Top panel: low bias signals with cos($\varphi$) fitting measured at 20 $\upmu$A, with a fitted amplitude (-29 $\pm$ 1 V/A$^2$); bottom panel: high bias signals with cos($\varphi$) fitting measured at 140 $\upmu$A, with a fitted amplitude (0.64 $\pm$ 0.03 V/A$^2$). (b) Bias dependence of $R_{\textrm{nl}}^{2\omega}$. Bias dependence shown in these graphs were measured at 5 K under a magnetic field of 4 T. The inner figure presents the zoom-in data of $R_{\textrm{nl}}^{2\omega}$, for the applied current from 100 $\upmu$A to 300 $\upmu$A.[]{data-label="fig:2"}](fig2.pdf){width="250pt"}
{width="440pt"}
First and second harmonic responses of the non-local signal have been measured simultaneously all along this study. The first order angular dependence is expected to obey the relation $R^{1\omega} = V/I = R_{0}^{1\omega} + R_{\textrm{nl}}^{1\omega}cos^2{(\varphi)}$ [@cornelissen2015long], where $R_{0}^{1\omega}$ is an offset resistance, and $R_{\textrm{nl}}^{1\omega}$ is the magnitude of the first harmonic signal. However, we do not observe the expected $cos^2{(\varphi)}$ modulation in the first harmonic signal, as the fitted first order resistance $R_{\textrm{nl}}^{1\omega}$ is only detected in the order of 0.01 m$\Omega$ which is comparable to the standard deviation. An example of measured first harmonic signal can be found in Fig.1(c). Yet, this value is at least 3 orders smaller than the $R_{\textrm{nl}}^{1\omega}$ reported for the Pt/YIG system [@cornelissen2015long]). The measurements are carried out over a wide range of applied currents and magnetic fields, and with the maximum lock-in detection sensitivity. A typical measurement of first harmonic non-local signal is shown in Fig.\[fig:1\]c, for a current of 20 $\upmu$A at 5 K. In contrast, the non-local second harmonic signals exhibit a clear sinusoidal behavior (Fig.\[fig:1\]e) under an in-plane rotating magnetic field. The magnitudes of non-local signals were fitted with: $$\label{eqn:2}
R^{2\omega} = \frac{V}{I^2} = R_{0}^{2\omega} + R_{\textrm{nl}}^{2\omega}cos{(\varphi)},$$ where $R_{0}^{2\omega}$ is the offset resistance for the second harmonic signal. A non-zero offset $R_{0}^{2\omega}$ is always present, possibly from unintended Seebeck contribution in the detector[@sierra2018thermoelectric]. $R_{\textrm{nl}}^{2\omega}$ is the magnitude of the second harmonic signal. For the corresponding second harmonic measured in Fig.\[fig:1\]e, we extract an amplitude $R_{\textrm{nl}}^{2\omega}=-36\ \pm \ 1 \ V/A^2$, which is comparable to the magnitude of room-temperature non-local SSE measured on bulk Pt/YIG samples [@cornelissen2015long] with equal angular dependence. If we compare to the typical top contact geometry used to detect SSE from YIG [@cornelissen2015long], the same SSE detected here in bottom contact geometry should produce a spin current in the opposite direction. Therefore, the ISHE induced in Pt is reversed compared to the top Pt on YIG, hence, we expect an opposite sign of the signal. The negative sign observed here would correspond to the positive sign measured in [@cornelissen2015long] and, if attributed to SSE, reveals a transfer of magnon-spin from CrBr$_3$ to the Pt top surface. However, at this point, we cannot rule out other effects like proximity induced Anomalous Nernst Effect (pANE) in Pt [@KikkawaPRL2013]. We discuss about relevant effects later (see Fig\[fig:4\]c, rotation of out-of-plane magnetic field).
The current dependence of $R_{\textrm{nl}}^{2\omega}$ is plotted in Fig.\[fig:2\], for a contact pair with distance of 950 nm center to center (edge to edge distance of 500 nm). $R_{\textrm{nl}}^{2\omega}$ depends on the applied current non-linearly, and a sign reversal of $R_{\textrm{nl}}^{2\omega}$ occurs between 40 and 100 $\upmu$A. For data measured at 60 $\upmu$A and 80 $\upmu$A, an angular modulation of the second harmonic signal is still observed but it is not described by a simple cosine function [@SuppMat]. An example of the negative $R_{\textrm{nl}}^{2\omega}$ at low current is shown in Fig.\[fig:2\]a (top panel), and an example of the positive $R_{\textrm{nl}}^{2\omega}$ at high current is plotted in Fig.\[fig:2\]a (bottom panel). The absolute amplitude $\left|R_{\textrm{nl}}^{2\omega}\right|$ in general decreases with increasing current at the heater, as plotted in Fig.\[fig:2\]b. Its value for positive amplitude at high current is one to two orders of magnitude lower than its value for negative amplitude at low current, depending on the applied current.
To get better insight of the role of the complex temperature distribution in our device for this non-linear behavior, we employ a 2-dimensional finite element model (FEM) simulating a geometry of the $x$-$z$ plane. Indeed the full hBN-encapsulation of the CrBr$_3$ flake in this device brings inevitable additional heat conduction paths resulting in strong current-dependent thermal gradients in both x and z directions ($\partial_x T$ and $\partial_z T$ respectively). As $\kappa_{\textrm{CrBr}_3}$, the thermal conductivity of CrBr$_3$, is unknown, we ran the computation for different thermal conductance ratios $\eta_{\textrm{K}}$ so that $\kappa_{\textrm{CrBr}_3}\left(T\right)=\eta_{\textrm{K}} \,\kappa_{\textrm{hBN}}\left(T\right)$, with $\kappa_{\textrm{hBN}}$ the thermal conduction of hBN, and taking into account the highly temperature dependent thermal conduction of the materials (see Supplemental Material VII [@SuppMat]). This modeling reveals a strong dependence of the temperature profile as a function of the heating current. It qualitatively supports that the main contribution of the thermal gradient in the Pt detector is in $x$ direction ($\partial_x T$). Yet there also is a non-negligible thermal gradient in $z$ direction ($\partial_z T$), in the CrBr$_3$ as well as in the Pt detector, allowing for SSE and possible unintended effects occurring in the Pt detector that will be discussed below.
{width="\textwidth"}
The in-plane magnetic field dependence on the second order nlADMR amplitude $R_{\textrm{nl}}^{2\omega}$ is plotted in Fig.\[fig:3\]a. We apply a range of fields from 0 T to 7 T for the in-plane rotation measurements at 5 K. At low current (20 $\upmu$A), we observe a linear increase of $|R_{\textrm{nl}}^{2\omega}|$ from 0 T to 3 T. After 4 T, the magnitude tends to saturate showing only a slight decay (Fig.\[fig:3\]a and Fig.\[fig:3\]c). At high current (160 $\upmu$A), we also observe a linear increase of $R_{\textrm{nl}}^{2\omega}$ from 0 T to 4 T, but with magnitudes about 50 times smaller than $|R_{\textrm{nl}}^{2\omega}|$ for low current. After 4 T, the magnitude still increases but at a lower rate (Fig.\[fig:3\]a and Fig.\[fig:3\]d). The lower magnitude at high current is consistent with the reduction of the magnetization expected for a temperature increase due to Joule heating. The origin of the magnetic field dependence remains unclear. As the saturation of the magnetization of tri-layer CrBr$_3$ in its hard-plane is reported to occur at 400 mT [@Kim2019evolution], the linear increases cannot be simply explained by the saturation of the magnetization as from an isolated CrBr$_3$ layer and reveals the contribution of additional field dependent effects.
The second order nlADMR is also measured at three different temperatures, 5 K, 10 K and 60 K, and the fitted amplitudes of $R_{\textrm{nl}}^{2\omega}$ are shown in Fig.\[fig:3\]b for low (20 $\upmu$A) and high current (160 $\upmu$A) measured under 4 T. Compared with the signal at 5K, the fitted cosine amplitude at 10K decreases for both low and high bias. Far above T$_c$ at 60 K, a very small but non-zero value of $R_{\textrm{nl}}^{2\omega}$ is observed in our measurements (0.08 $\pm$ 0.03 $V/A^2$ at 160 $\upmu$A and -3 $\pm$ 2 $V/A^2$ at 20 $\upmu$A). We attribute this small non-zero value to an artifact from the measurement setup [@SuppMat].
We present hereafter a series of out-of-plane nlADMR (OOP-nlADMR) measurements i.e. fixing $\varphi=0^{\circ}$ and varying $\theta$ by rotating the magnetic field in the $x$-$z$ plane, as defined on Fig.\[fig:1\]. Some examples and the current dependence of this OOP-nlADMR are summarized in Fig.\[fig:4\]. The first observation, with Fig.\[fig:4\]b and \[fig:4\]d as examples, is that all OOP-nlADMR signals exhibit a non-zero angular phase shift varying with the heating current. We investigated the origin of this phase considering the various effects that could add to the SSE signal. Nernst, Seebeck, Spin Nernst Magnetoresistance (SNMR)[@Meyer2017; @Kim2017] effects are discarded as major contributions, either due to the probing geometry, or their angular dependence, a detailed description is given in [@SuppMat]. However the anomalous Nernst effect (ANE), which has already been reported as a possible effect, arising from a proximity induced ferromagnetism into the Pt [@Leutenantsmeyer2016; @Zollner2016; @Meier2015; @Guo2014; @Lu2013; @KikkawaPRL2013], cannot be ruled out.
Considering a proximity ANE (pANE) in Pt, a transverse pANE voltage $\Delta V_{\textrm{pANE}}$ reads : $$\frac{\Delta V_{\textrm{pANE}}}{L_{\textrm{Pt}}}=\left |\mathbf{\nabla} V \right |_y = \left |-S_{\textrm{pANE}}\left(\mathbf{m}\times \left[-\mathbf{\nabla} T\right]\right)\right |_y
\label{eqn:ANE}$$ Where $S_{\textrm{pANE}}$ is the pANE coefficient, $\mathbf{m}$ is the unit vector of direction of the magnetization and $L_{\textrm{Pt}}$ is the $y$-axis length of the contact area of Pt with CrBr$_3$. As the magnetization of CrBr$_3$ is expected to saturate for fields beyond 1 T in the hard plane [@Richter2018; @Kim2019evolution], we also assume the proximity induced magnetization parallel to the magnetic field at 4 T. Then, two contributions of the pANE are distinguished (Fig.\[fig:4\]a) : the pANE signal caused by the IP gradient $\partial_x T$, pANE$_x$ which varies as $sin\left( \theta \right)$, and the pANE signal caused by the OOP gradient $\partial_z T$, pANE$_z$ which varies as $cos\left(\theta\right)$.
The pANE induced by the temperature gradient along $x$ (pANE$_x$) can be isolated from the other signals by changing the heat flow direction. By interchanging the heater and detector contacts, the heat flow direction along the $x$ axis ($\propto\partial_x T$) is reversed, but the heat flow direction along the $z$ axis ($\propto\partial_z T$) remains the same. Hence, the pANE$_z$ contribution will stay unchanged while the pANE$_x$ will reverse its sign. In Fig.\[fig:4\]b, we provide a normalized second order nlADMR $R^{2\omega}_N=R^{2\omega} A_{\textrm{Pt}}/L_{\textrm{Pt}}$, with $A_{\textrm{Pt}}$ the Pt electrode cross-section, at 20 $\upmu$A and 4 T, for the configuration forward defined in Fig.\[fig:1\], and the nlADMR from a reversed geometry where heater and detector are interchanged. As the width and length of the two electrodes are different, as well as their interface with CrBr$_3$ possibly, the heating power injected will differ by a small factor. Therefore our comparison remains only qualitative. Nevertheless, the amplitudes and offsets are alike and the two traces differ mainly by the apparent opposite phase shift.
If both pANE$_x$ and pANE$_z$ contributions are significant in our system, the difference between the forward geometry (Fig.\[fig:4\]b) signal and the reverse geometry (Fig.\[fig:4\]b) signal will reveal the $sin\left(\theta\right)$ behavior, and the sum of these two signals will reveal the $cos\left(\theta\right)$ behavior. As a result, we obtain the respectives traces shown in Fig.\[fig:4\]c. The good agreement of the fittings on both curves is a confirmation that the pANE is present in the Pt detector.
Based on this observation, we extracted the two contributions for every ADMR at different current and at a constant magnetic field of 4 T, by fitting the expression $R^{2\omega}=R_{0}^{2\omega}+R_{\textrm{SSE}+\textrm{pANE}z}\, cos \theta +R_{\textrm{pANE}x}\, sin \theta$. The measurements at 40, 100 and 280 $\upmu$A are shown in Fig.\[fig:4\]c, and the fitted sinusoidal curve presents the phase shift in each case. The current dependence of the extracted amplitudes is provided in Fig.\[fig:4\]d. The $R_{\textrm{SSE}+\textrm{pANE}z}$ and $R_{\textrm{pANE}x}$ contributions both follow a similar decreasing trend with applied current. While $R_{\textrm{SSE}+\textrm{pANE}z}$ dominates at 20 and 40 $\upmu$A, $R_{\textrm{pANE}x}$ becomes close to twice $R_{\textrm{SSE}+\textrm{pANE}z}$ at higher current. The variation of the amplitude of $R_{\textrm{SSE}+\textrm{pANE}z}$ at low currents follows the variation of the signal for IP field rotation in Fig \[fig:2\]b, however the sign reversal for the derived $R_{\textrm{SSE}+\textrm{pANE}z}$ does not occur in the OOP configuration. To elucidate the contribution of the spin Seebeck, we introduce the ratio $\xi=-R_{\textrm{SSE}+\textrm{pANE}z}/R_{\textrm{pANE}x}=-\left(R_{\textrm{SSE}}+R_{\textrm{pANE}z}\right)/R_{\textrm{pANE}x}$ of the two contributions (inset of Fig.\[fig:4\]e), the ratio $\delta = S_{\textrm{pANE}}^z$/$S_{\textrm{pANE}}^x$ to account for any difference between the IP ($S_{\textrm{pANE}}^x$) and OOP ($S_{\textrm{pANE}}^z$) proximity anomalous Nernst coefficients, as well as the ratio $\gamma=\partial_z T/\partial_x T$ of the temperature gradients in Pt. As a result, the SSE contribution to the measured signal simply reads (demonstration in [@SuppMat]): $$R_{\textrm{SSE}}=R_{\textrm{pANE}x} \left(\delta \gamma - \xi \right)
\label{equ:evalSSE}$$ Based upon the fact that the saturated magnetization of CrBr$_3$ has been reported to be of same magnitude when oriented IP or OOP, we assume $\delta\approx 1$, i.e. $S_{\textrm{pANE}}^z \approx S_{\textrm{pANE}}^x$. Following this assumption, the estimated ratio of the two contributions $\gamma$ lays between $-0.20$ and $0.15$, according to our FEM simulation based on thermal conduction properties of CrBr$_3$ and hBN layers (i.e. the ratio $\eta_{\textrm{K}}=\kappa_{\textrm{CrBr}_3}/\kappa_{\textrm{hBN}}$ ) (see details in [@SuppMat]). Even using $\delta \gamma = \pm0.5$ accounting for the possible underestimation of $\partial_z T$ due to the omission of a small heat leakage via the Pt/Au contacts leads on SiO$_2$, we extract a significant SSE contribution to the nlADMR signal at low heating current, as plotted in Fig.\[fig:4\]f. We provide the magnetic field dependence of the OOP-nlADMR in Fig.\[fig:5\]. Fig.\[fig:5\]a shows examples of the evolution of the OOP-nlADMR for 1, 4, 7 T, for a current fixed to 20 $\upmu$A. The same operation to separate pANE$_z$+SSE from pANE$_x$ is applied to this measurement set and the amplitude variation of each component is shown in Fig.\[fig:5\]b for magnetic fields from 0 to 7 T. The pANE$_z$+SSE variation is comparable to the one measured in in-plane rotation configuration (Fig.\[fig:3\]a, blue curve), except that we do not observe the high field saturation decrease. The dependence of the pANE$_x$ trace follows a similar increase until 2 T, but shows a slight decrease for 3 and 4 T and increases again to reach the same value as pANE$_z$+SSE at 7T. This behaviour is captured into the $\xi$ ratio that shows a peak above 1.5 for 3 and 4 T and a value remaining around 1 for other fields strengths. As the temperature profile is fixed, the difference between SSE+pANE$_z$ and pANE$_x$ must be strongly linked to the magnetic properties of the CrBr$_3$/Pt structure.
![(a) SSE angular dependence shown for 1, 4 and 7 T, with current fixed to 20 $\upmu$A at 5K. (b) Magnetic field dependence of pANE$_x$ and SSE + pANE$_z$ signal amplitude for the forward configuration. In the inset of (b) the ratio $\xi = -(R_{\textrm{SSE}}+R_{\textrm{pANE}z})/R_{\textrm{pANE}x}$ is given. See [@SuppMat] for the data extraction in detail.[]{data-label="fig:5"}](Fig5_OoP_BfieldDep_3c.pdf){width="240pt"}
Discussion
==========
By analyzing the OOP-nlADMR, we show that pANE$_x$ presents a different angular dependence than SSE and pANE$_z$ allowing to separate the two contributions. Despite the lack of insight on the mechanism inducing the magnetization in Pt, this assumption is based on that the magnetic moments emerging on the Pt atoms are imprinted by the moments of CrBr$_3$. Yet the saturated magnetization of CrBr$_3$ has been measured to differ by less than 20% between the orientation along the easy axis and the orientation in the hard plane. Therefore, the induced magnetization in Pt is expected to behave accordingly, leading to a comparable anomalous Nernst coefficient depending on the magnetization value but weakly on its orientation.
A pANE contribution to the ADMR has been identified in Pt/YIG systems as well, but the pANE$_z$ represents at most 5% of the voltage signal, the left 95% being attributed to SSE induced ISHE [@KikkawaPRL2013]. Because of the significant magnetic exchange field already noticed in CrBr$_3$ [@Tang2019; @Kim2019a] as well as the strong temperature gradients involved (beyond 2 orders of magnitude higher than in [@KikkawaPRL2013]), in our CrBr$_3$/Pt system, the pANE cannot be neglected and the SSE signal is at best comparable with the pANE$_z$.
In Pt/YIG system, the magnon SSE signal decreases with the magnetic field [@cornelissen2016field]. In Fig\[fig:3\].a, we notice that, after 3T, the fitted amplitude of the low current curve does not change with magnetic field, but $R_{\textrm{nl}}^{2\omega}$ of the high current curve increases linearly with magnetic field. In other words, $R_{\textrm{nl}}^{2\omega}$ at low current tends to decrease where SSE contributes most, compared to the amplitude at high current where the SSE contributes less. Hence, our measurements, with support of a temperature distribution simulation, suggest that the high amplitude signal observed at low current is dominated by SSE from CrBr$_3$.
According to the expected angular dependence of the SSE and ANE, the SSE+pANE$_z$ signal should appear in both IP-nlADMR and OOP-nlADMR while pANE$_x$ should be only detected in OOP-nlADMR. Therefore, the same current dependence of SSE+pANE$_z$ in both configurations is expected. According to the FEM simulation, a reversal of $\partial_z T$ occurs at sufficiently high current, simultaneously in CrBr$_3$ and Pt at the detection interface (see [@SuppMat]). This leads to a reversal of the SSE+pANE$_z$, most likely dominated by ANE in the high current range. However, the sign reversal is only observed in the IP measurements (in Fig\[fig:2\]b), not showing in the OOP measurements after the separation (in Fig\[fig:4\]e). As the IP and OOP measurements were performed with different cool-down processes, the thermal conductivity is possibly changed at the interface. This implies that the sign reversal current is possibly shifted to a much higher value, therefore not observed in the OOP measurements.
Furthermore, we also suggest that a quantitative discrimination between pANE and SSE is possible. We provide an indicative estimation of the magnitude of the SSE based on the assumption that the pANE coefficient is equal for ANE$_x$ and ANE$_z$. By far, we are limited by the current knowledge on the material properties of the 2D magnet. However, if the thermal conduction profiles and the magnetization dynamics are characterized concretely, a more accurate separation of the two spin-caloritronic effects can be realized.
Nevertheless, the magnetic field dependence of pANE$_x$ and SSE+pANE$_z$ and the difference between them bring new questions. The ANE scales with the magnetization via the coefficient $S_{\textrm{pANE}}$. The non-monotonic field dependence of pANE$_x$ suggests a complex evolution of the induced magnetization in Pt, due to either the presence of magnetic domains or any additional interaction at the interface.
Conclusions
===========
To conclude, we demonstrate the relevance of the full hBN encapsulation and the bottom contacting design to enable the integration of air-reactive materials such as CrBr$_3$, for studying spin-caloritronic effects in 2D magnets. By using second order nlADMR measurement on such an encapsulated CrBr$_3$/Pt device, we reveal, by detecting the presence of a proximity ANE voltage, a significant proximity induced magnetism from CrBr$_3$ into the adjacent Pt contacts. With reasonable assumptions, we conjecture about the presence of a weak SSE, dominating the signal in the low current regime, while the pANE prevails for currents above 60$\upmu $A. The non-trivial magnetic field dependence of the separated effects leaves open questions as for the current understanding of magnetic effects at the interface of heavy metal and 2D magnets. The encapsulation shows itself an elegant technique to address these questions in deeper investigations of air-degradable 2D materials, and opens the way to future magnon transport studies.
Methods
=======
CrBr$_3$ and hBN crystals are provided by a commercial company HQgraphene. CrBr$_3$ is an air sensitive material. To study magnonics with CrBr$_3$ in a non-local geometry, we encapsulate a 7 nm-thin chromium tribromide flake and platinum (Pt) strips into two hexagonal Boron Nitride (hBN) layers (top layer and bottom layer). The stacking of van der Waals materials was performed in a glove box filled with inert gas argon by using standard PC/PDMS dry transfer method. Pt strips were first grown on bottom hBN. After that CrBr$_3$ with a top hBN thin layer was transferred on top of the Pt strips. See [@SuppMat] for more details in fabrication process.
The authors thank Prof. Justin Ye and Puhua Wan for granting us access to their transfer system in an Ar glove box. We thank Johan G. Holstein, Hans de Vries, Herman Adema, and Tom J. Schouten for technical assistance. We thank fruitful discussions with Geert R. Hoogeboom, Alexey A. Kaverzin and Jing Liu. This project has received funding from the Dutch Foundation for Fundamental Research on Matter (FOM, now known as NWO-I) as a part of the Netherlands Organization for Scientific Research (NWO), the European Union’s Horizon 2020 research and innovation programme under grant agreement No 696656 and 785219 (Graphene Flagship Core 1 and Core 2) and Zernike Institute for Advanced Materials. MHDG acknowledges support from NWO VENI 15093.
[46]{} 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 [****, ()](https://doi.org/10.1038/nature22391) [****, ()](https://doi.org/10.1073/pnas.1902100116) [****, ()](https://doi.org/10.1038/s41928-019-0302-6) @noop [****, ()]{} [**](http://www.amazon.com/Introduction-Solid-Physics-Charles-Kittel/dp/047141526X/ref=dp_ob_title_bk), ed. (, ) @noop [**]{} (, ) @noop [****, ()]{} [****, ()](https://doi.org/10.1038/nmat2856) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevLett.120.097205) [****, ()](https://doi.org/10.1038/ncomms5700) [****, ()](https://doi.org/10.1103/PhysRevX.8.011010) [****, ()](https://doi.org/10.1103/PhysRevLett.117.217202) [****, ()](https://doi.org/10.1103/PhysRevB.96.224414) [****, ()](https://doi.org/10.1021/acs.nanolett.9b03739), [****, ()](https://doi.org/10.1038/s41928-018-0087-z) @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevMaterials.2.024004) @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevB.88.214403) [****, ()](https://doi.org/10.1038/ncomms9211) [****, ()](https://doi.org/10.1038/nmat4964) [****, ()](https://doi.org/10.1103/PhysRevLett.124.027701) @noop [****, ()]{} [****, ()](https://doi.org/10.1088/0022-3727/48/2/025001) , @noop [Ph.D. thesis]{}, () @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevLett.110.067207) [](---link---) [****, ()](https://doi.org/10.1038/s41467-017-01493-5) [****, ()](https://doi.org/10.1088/2053-1583/4/1/014001) [****, ()](https://doi.org/10.1103/PhysRevB.94.155441) [****, ()](https://doi.org/10.1103/PhysRevB.89.214406) [****, ()](https://doi.org/10.1103/PhysRevLett.110.147207) [ ()](http://arxiv.org/abs/1910.10411), @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevB.91.024402) [****, ()](https://doi.org/10.1103/PhysRevB.13.4607) [****, ()](https://doi.org/10.1016/j.ssc.2012.04.022) [****, ()](https://doi.org/10.1088/0370-1328/72/5/429) [****, ()](https://doi.org/10.1103/PhysRevB.87.184421) [****, ()](https://doi.org/10.1002/adfm.201501429)
[^1]: Contributed equally to this work
[^2]: Contributed equally to this work
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper is devoted to the problem of sample path large deviations for multidimensional queueing models with feedback. We derive a new version of the contraction principle where the continuous map is not well-defined on the whole space: we give conditions under which it allows to identify the rate function. We illustrate our technique by deriving a large deviation principle for a class of networks that contains the classical Jackson networks.'
author:
- Marc Lelarge
bibliography:
- 'ex.bib'
title: Sample path large deviations for queueing networks with Bernoulli routing
---
Introduction {#introduction .unnumbered}
============
This paper is concerned with the theory of large deviations of stochastic processes related to discrete event systems. As opposed to classical stochastic dynamical systems, for which the evolution is continuous and described by a stochastic differential equation, discrete event systems are characterized by synchronization mechanisms that prevent most of the classical tools to apply. We present here a new approach for the analysis of the sample path large deviations of such processes. Unlike standard methods that require establishing upper and lower bounds, our method relies on the analogy between the theory of weak convergence and the theory of large deviations. This analogy is well-known and has been studied by many authors, we refer to the recent book of Feng and Kurtz [@fenkur] that surveys this field. We should in particular quote the work of Puhalskii [@Puh:pro], [@Puh:book] quite similar to our approach. However, we will not use the framework of idempotent measures developed by Puhalskii. We discuss in more details our general methodology and its relation with the existing literature after the description of the queueing networks we consider.
To apply our method we choose a class of queueing networks with Bernoulli routing, where feedback is allowed. The discontinuous dynamic of queueing networks makes it hard to study and large deviations results in the literature are treated on a case by case basis as in the work of Ganesh and Anantharam [@ganan], Bertsimas, Paschalidis and Tsitsiklis [@bpt] or Ramanan and Dupuis [@ramdup]. As we will see, adding the possibility of feedback makes the problem much harder. For queueing networks with feedback, existing large deviations results are restricted to networks described by finite-dimensional Markov processes, see the works of Dupuis, Ellis and Weiss [@dupelwei], Dupuis and Ellis [@dupel] and Ignatiouk-Robert [@ignaop]. In this paper, we consider networks where the output process of a queue is modeled by a reflection mapping. This class contains the classical Jackson networks and our large deviations results extend existing results for this class obtained by Atar and Dupuis [@atardupuis] and Ignatiouk-Robert [@ign]. Our technique allows to obtain large deviations results under non-exponential assumptions. This case corresponds to networks with autonomous service and gives an approximation for queueing networks where each station acts as a standard single server queue. While preparing this paper, the author became aware of the work of Puhalskii [@Puh:jack] who considers generalized Jackson networks. The form of the rate function for the queue length process obtained in [@Puh:jack] coincides with our result, which confirms the intuition that in the large deviation regime, networks with autonomous service approximate well generalized Jackson networks. We will discuss more carefully this result in Section \[sec:result\].
In the next section, we give an overview of the general methodology and then introduce the general notation. Section \[sec:contr\] gives an extension of the contraction principle that will allow us to identify the rate function. Our result is stated without any reference to any specific discrete event system and could be applied to other systems. In Sections \[sec:extpsi\] and \[sec:ldp\], we apply our method to the case of a queueing network.
General methodology {#general-methodology .unnumbered}
-------------------
For simplicity, we adopt here the notation corresponding to our example of queueing network. As in [@lel] or [@maj-fluid], we define the arrival and departure processes $\A$ and $\D$ of each station of the network as the solution of the fixed point equation $$\begin{aligned}
\label{fixpointeq}\left\{\begin{array}{lcl}\A &=& \Gamma(\D,\JN),\\
\D &=& \Phi(\A,\JN),\end{array}\right.&\Leftrightarrow &
(\A,\D)=:\Psi(\JN).\end{aligned}$$ Here $\JN$ is a process that describes all the primitives of the network as service times at the different stations, routing decisions, arrival times in the network. The maps $\Gamma$ and $\Phi$ describe the dynamic of the network.
We consider a sequence of queueing networks $\{\JN_n\}_n$, where the primitives are counting processes (i.e. $\JN_n$ belongs to a space denoted by $\Ecal$) and satisfy a large deviation principle (LDP). We denote by $\Isf^\JN$ the corresponding rate function. It is known that the map $\Psi$ is well defined if the primitives of the network are counting processes, see [@cm:disc] or [@lel] and we denote $(\A_n,\D_n)=\Psi(\JN_n)$. It is natural to ask whether $\Psi$ is well defined for processes in $\Dbb$, the space of cadlag non-decresing functions or at least for absolutely continuous functions. If this was true, and if $\Psi$ was shown to be continuous, then we would get thanks to the contraction principle that the sequence of processes $\{(\A_n,\D_n)\}_n$ satisfies a LDP with good rate function $$\begin{aligned}
\label{eq:quote}"\Isf^{\A,\D}(\A,\D)=\inf\left\{ \Isf^{\JN}(\JN),\:
\Psi(\JN)=(\A,\D)\right\}."\end{aligned}$$ However, the map $\Psi$ turns out not to be well defined for all possible limits of a sequence of networks $\{\JN_n\}_n\in
\Ecal^\nat$ as defined previously. In particular, the fixed point equation (\[fixpointeq\]) can very well be stated for processes in $\Dbb$ but then may have several different solutions as noted by Majewski [@maj-fluid]. We give in the appendix a simple example.
To circumvent this difficulty, we adopt the following strategy. We find a domain $\Dcal_\JN\subset\Ecal$ satisfying the following constraints:
- the map $\Psi$ is well defined on $\Dcal_\JN$;
- any solution $(\A,\D)$ of the fixed point equation (\[fixpointeq\]) associated with a “continuous” Jackson network $\JN$ can be approximated by a sequence $\{\JN_n\}\in
\Dcal_\JN^\nat$ such that $$\begin{aligned}
\label{a}\JN_n&\rightarrow& \JN,\\
\label{b}\Psi(\JN_n)&\rightarrow& (\A,\D),\\
\label{c}\Isf^\JN(\JN_n)&\rightarrow& \Isf^\JN(\JN).\end{aligned}$$
Hence in order to remove the quote from (\[eq:quote\]), we follow a quite standard method of proofs for large deviations of stochastic processes analogue with the theory of weak convergence [@fenkur]: it consists of first verifying a compactness condition and then showing that there is only one possible limit. In our context, we proceed as follows:
1. we show that our sequence of processes is exponentially tight;
2. we use $\Dcal_\JN$ to determine the rate function.
In Section \[sec:contr\], we give the theoretical framework that shows how any domain verifying assumptions (\[a\]), (\[b\]) and (\[c\]) determines the rate function. This result is stated in great generality (without any reference to our specific problem) and could be of independent interest since this method of proof could be applied to other dynamical systems (with discontinuous statistics).
Notation {#notation .unnumbered}
--------
For $(E,d,\leq)$ a complete, separable metric space with partial order $\leq$, we denote by $\Dbb(E)$ the space of cadlag non-decreasing $E$-valued functions defined on $\real_+$ with Skorohod ($J_1$) topology and by $\Cbb(E)$ the space of continuous non-decreasing $E$-valued functions defined on $\real_+$. Restricted to $\Cbb(E)$ the Skorohod topology is just the compact uniform topology.
For $x,y\in \real^K$, we write $x\leq y$ if $x^{(i)}\leq y^{(i)}$ for all $i$. We denote by $\wedge$ the minimum and by $\vee$ the maximum in $\real^K$. For $\Xbf, \Ybf \in \Dbb(\real_+^K)$, we write $\Xbf\leq \Ybf$ if $\Xbf(t)\leq \Ybf(t)$ for all $t\geq 0$ and for maps $F,G \in \Dbb(\real_+^K)^2$, we denote $F\leq G$ if $F(\Xbf)\leq G(\Xbf)$ for all $\Xbf\in \Dbb(\real_+^K)$. For $x\in
\real^K$, we denote $\|x\|=\vee_{i=1}^K x^{(i)}$ and for $\Xbf\in\Dbb(\real_+^K)$, we denote $\|\Xbf\| = \sup_t \|\Xbf(t)\|$. We denote $\Dbb_0(E)=\{f\in \Dbb(E),\:f(0)=0\}$ and $\Cbb_0(E)=\{f\in \Cbb(E),\:f(0)=0\}$.
A piecewise linear function is a continuous function such that there exists a partition $\tau=(t_0=0<t_1<\dots)$ with $t_k\rightarrow
\infty$ and such that the function is linear on each interval $(t_k,t_{k+1})$. For any function $f\in \Dbb(\real_+^K)$, we define the polygonal approximation of $f$ with step $1/n$ as the (piecewise linear) function $$\begin{aligned}
f_n(t) = f\left(\frac{\lfloor nt\rfloor}{n}\right)+\left(nt- \lfloor
nt\rfloor\right)\left(f\left(\frac{\lfloor
nt\rfloor+1}{n}\right)-f\left(\frac{\lfloor
nt\rfloor}{n}\right)\right)\end{aligned}$$
$\Mbb^K$ is the set of substochastic matrices of size $K\times K$. For $M\in \Mbb^K$, we denote by $\rho(M)$ its spectral radius, by $M^t$ its transpose and $M^{(i)}$ denotes the line $M^{(i)}=
(M^{(i,1)},\dots M^{(i,K)})$. In particular, we will identify a function $\Pbf\in \Dbb(\Mbb^K)$ with its $K$ components $\Pbf^{(i)}\in \Dbb(\real_+^K)$, where $\Pbf^{(i)}(t)=
(\Pbf^{(i,1)}(t),\dots \Pbf^{(i,K)}(t))$ with $\sum_{j}\Pbf^{(i,j)}(t)\leq 1$ for all $t\geq 0$ and all $i$. Note that for $M,N\in \Mbb^K$, we have $M\leq N$ if $M^{(i,j)}\leq
N^{(i,j)}$ for all $i$ and $j$.
We will use the Kullback-Leibler information divergence, which is a nonsymmetric measure of distance between distributions in the sense that for any two distributions $P$ and $R$ on $\Xcal^k$ where $\Xcal$ is a finite set, $$\begin{aligned}
\Dsf(P\|R) = \sum_{x\in \Xcal^k}P(x)\log\left(
\frac{P(x)}{R(x)}\right),\end{aligned}$$ is nonnegative and equals $0$ if and only if $P=R$. We use the standard notational conventions $\log 0=-\infty$, $\log\frac{1}{0}=\infty$ and $0 \log0=0\log\frac{0}{0}=0$. For any fixed $R$, the divergence $\Dsf(P\|R)$ is a continuous function of $P$ restricted to $\{P,\:S(P)\subset S(Q)\}$ where $S(P)$ denotes the support of $P$ (see [@ccc]).
For $P\in \Mbb^K$, we denote by $\tilde{P}$ the $K\times (K+1)$ stochastic matrix obtained as follows: for all $i,j\leq K$, $\tilde{P}^{(i,j)}=P^{(i,j)}$ and $\tilde{P}^{(i,K+1)}=1-\sum_{k=1}^K P^{(i,k)}$. For $P,R\in \Mbb^K$, we will denote $$\begin{aligned}
\tilde{\Dsf}(P\|R)&:=& \Dsf(\tilde{P}\|\tilde{R})\\
&=& \sum_{i,j\leq K} P^{(i,j)}\log\left(
\frac{P^{(i,j)}}{R^{i,j)}}\right)+\sum_{i\leq K}\left(1-\sum_k
P^{(i,k)}\right)\log\left(\frac{1-\sum_k P^{(i,k)}}{1-\sum_k
R^{(i,k)}} \right)\\
&=:& \sum _{i=1}^K \tilde{\Dsf}(P^{(i)}\|R^{(i)}).\end{aligned}$$
An extension of the contraction principle {#sec:contr}
=========================================
Let $\Ecal,\Fcal$ be complete separable metric spaces. Let $G:\Ecal\times \Fcal\rightarrow \real$ be a continuous function. We assume that there exists $\Dcal\subset \Ecal$, such that for all $x\in \Dcal$, there exists an unique $y\in \Fcal$ such that $G(x,y)=0$. We denote it by, $y=H(x)$ where $H:\Dcal\rightarrow
\Fcal$, $$\begin{aligned}
\forall x\in \Dcal,\quad G(x,y)=0 \:\Leftrightarrow \: y=H(x).\end{aligned}$$
\[prop:contraction\] Let $\{X_n\}_n$ be a sequence of $\Ecal$-valued random variables and $\{Y_n\}_n$ be a sequence of $\Fcal$-valued random variables. We assume that each sequence is exponentially tight. Assume that the sequence $\{X_n\}_n$ satisfies a LDP with good rate function $I^X$ and that $G(X_n,Y_n)=0$ a.s. for all $n$.
We assume that for all $(x,y)$ such that $G(x,y)=0$ and $\Isf^X(x)<\infty$, there exists a sequence $\{\Scal_n(x,y)\}_n\in \Dcal^\nat$ such that
- $\Scal_n(x,y) \rightarrow x$;
- $H(\Scal_n(x,y)) \rightarrow y$;
- $\Isf^X(\Scal_n(x,y))\rightarrow \Isf^X(x)$.
$x_n\rightarrow x$, such that $x_n\in \Dcal$ for all $n$, $H(x_n)\rightarrow y$ and $\Isf^X(x_n)\rightarrow \Isf^X(x)$. We denote by $\Scal(x,y)=\{x_n\}_n$ this sequence. If $G(x,y)\neq 0$ or $\Isf^X(x)=\infty$, we take $\Scal(x,y)=\emptyset$ and we denote $\Scal(y) = \cup_x\{\Scal(x,y)\}$ (which might be empty).
Then the sequence $\{X_n,Y_n\}_n$ satisfies a LDP with good rate function: $$\begin{aligned}
\label{ratefunction}\Isf^{X,Y}(x,y):= \left\{\begin{array}{ll} \Isf^X(x), & G(x,y)=0,\\
\infty,& \mbox{otherwise.}
\end{array}\right.\end{aligned}$$
In particular, if $X_n\in \Dcal$ for all $n$ and if the sequence $\{H(X_n)\}_n$ is exponentially tight, then it satisfies a LDP in $\Fcal$ with good rate function: $$\begin{aligned}
\label{eq:rateH}\Isf^{H(X)}(y):= \inf \{ \lim_{n\rightarrow
\infty}\Isf^X(x_n),\: \{x_n\}_n\in \Scal(y)\}.\end{aligned}$$
\[rem1\]
- There are alternative ways of expressing the rate function, $$\begin{aligned}
\Isf^{H(X)}(y)= \inf\{ \Isf^X(x), \: y\in H^x\},\end{aligned}$$ where $H^x:=\{y\in \Fcal,\: \exists x_n\rightarrow x,\:
H(x_n)\rightarrow y\}$. $\Isf^{H(X)}$ is the lower semicontinuous regularization of the following function defined for $y\in
H(\Dcal)\subset \Fcal$, $$\begin{aligned}
\tilde{\Isf}^{H(X)}(y):=\inf\{ \Isf^X(x), \: y=H(x)\}.\end{aligned}$$ The main interest of the definition (\[eq:rateH\]) is that the rate function is computed only thanks to the sequences $\Scal(x,y)\in \Dcal^\nat$.
- Note that if $H(\Dcal)$ is closed (in particular if $\Dcal=\Ecal$) then this proposition follows from the contraction principle (for an extensive discussion of this principle, see the work of Garcia [@garcia]). Roughly speaking, Proposition \[prop:contraction\] tells us that if $\Dcal$ is dense in a certain sense in $\Ecal$, then the contraction principle still holds for the map $H$.
Thanks to Lemma 3.6 of [@fenkur], the sequence $\{X_n,Y_n\}_n$ is exponentially tight. Then by Theorem 3.7 of [@fenkur], there exists a subsequence $\{n_k\}$ along which the sequence $\{X_{n_k},Y_{n_k}\}_{n_k}$ satisfies a LDP with a good rate function. If we can prove that there is a unique possible rate function (that does not depend on the subsequence $\{n_k\}$) then the proposition will follow.
Hence, for simplicity of notations, we still denote the extracted subsequence by $\{X_n,Y_n\}_n$ and we assume that $\{X_n,Y_n\}_n$ satisfies a LDP with good rate function $\tilde{\Isf}^{X,Y}$. We will show that $\tilde{\Isf}^{X,Y}=\Isf^{X,Y}$ given by (\[ratefunction\]).
Consider the continuous mappings $H_1$ and $H_2$ from $\Ecal\times
\Fcal$ to $\Ecal\times \Fcal \times \real$, $$\begin{aligned}
H_1(x,y):=(x,y,G(x,y)), \quad H_2(x,y):=(x,y,0).\end{aligned}$$
We have clearly $H_1(X_n,Y_n)=H_2(X_n,Y_n)$ a.s. Moreover thanks to the contraction principle, $\{H_1(X_n,Y_n)\}_n$ and $\{H_2(X_n,Y_n)\}_n$ satisfy LDPs with the good rate functions $$\begin{aligned}
\Isf^{H_1}(x,y,z) =\inf\{\tilde{\Isf}^{X,Y}(x,y), z=G(x,y)\}\quad
\Isf^{H_2}(x,y,z) =\inf\{\tilde{\Isf}^{X,Y}(x,y), z=0\},\end{aligned}$$ where $\inf\emptyset=\infty$. Since $H_1(X_n,Y_n)=H_2(X_n,Y_n)$, we have $\Isf^{H_1}=\Isf^{H_2}$. Now we have, $$\begin{aligned}
\tilde{\Isf}^{X,Y}(x,y)=\inf_z\{\Isf^{H_1}(x,y,z)\}=\inf\{\tilde{\Isf}^{X,Y}(x,y),
G(x,y)=0\},\end{aligned}$$ hence $\tilde{\Isf}^{X,Y}(x,y)=\infty$ as soon as $G(x,y)\neq 0$. It remains to show that $G(x,y)=0$ implies $\tilde{\Isf}^{X,Y}(x,y) =
\Isf^X(x)$. We have clearly $\Isf^X(x)\leq \tilde{\Isf}^{X,Y}(x,y)$ for all $(x,y)$ since $\{X_n\}$ satisfies a LDP with good rate function $$\begin{aligned}
\Isf^{X}(x)=\inf\{\tilde{\Isf}^{X,Y}(x,y),\:y\in \Fcal,
\:G(x,y)=0\}.\end{aligned}$$
In particular, the definition of $\Dcal$ implies $\Isf^X(x)=\tilde{\Isf}^{X,Y}(x,H(x))$ for $x\in \Dcal$.
For all $y\in F$, we assume that if $x$ is such that $G(x,y)=0$ and $I^X(x)<\infty$, then there exists $x_n\rightarrow
x$, such that $x_n\in D$ for all $n$, $H(x_n)\rightarrow y$ and $I^X(x_n)\rightarrow I^X(x)$. We denote $$\begin{aligned}
\tilde{I}^{X,Y}(x,y):= \left\{\begin{array}{ll} I^X(x), & G(x,y)=0,\\
\infty,& \mbox{otherwise.}
\end{array}\right.\end{aligned}$$
Since $I^{X}(x)\leq \tilde{I}^{X,Y}(x,y)$, we have $I^{X,Y}(x,y)\leq\tilde{I}^{X,Y}(x,y)$.
Take $(x,y)$ such that $G(x,y)=0$ and $\Isf^{X}(x)<\infty$. There exists $x^*_n\rightarrow x$ with $x^*_n\in \Dcal$, $H(x^*_n)\rightarrow y$ and $\Isf^X(x^*_n)\rightarrow \Isf^X(x)$. Thanks to the lower semicontinuity property of $\tilde{\Isf}^{X,Y}$, we can find for any $\delta>0$, an $\epsilon>0$ such that $$\begin{aligned}
\frac{1}{\delta}\wedge
\left(\tilde{\Isf}^{X,Y}(x,y)-\delta\right)\leq \inf_{z\in
B(y,\epsilon)}\tilde{\Isf}^{X,Y}(x,z),\end{aligned}$$ where $B(y,\epsilon)$ is the closed ball in $\Fcal$ of center $y$ and radius $\epsilon$.
Thanks to the lower semicontinuity of the function $x\mapsto\inf_{z\in B(y,\epsilon)}\tilde{\Isf}^{X,Y}(x,z)$, we have $$\begin{aligned}
\inf_{z\in B(y,\epsilon)}\tilde{\Isf}^{X,Y}(x,z)&\leq&
\liminf_{x_n\rightarrow
x}\inf_{z\in B(y,\epsilon)}\tilde{\Isf}^{X,Y}(x_n,z)\\
&\leq& \liminf_{n\rightarrow \infty}\inf_{z\in
B(y,\epsilon)}\tilde{\Isf}^{X,Y}(x^*_n,z)\\
&\leq&\lim_{n\rightarrow \infty}\Isf^X(x^*_n)=\Isf^X(x),\end{aligned}$$ because $H(x^*_n)\in B(y,\epsilon)$ for sufficiently large $n$. Hence we proved that for any $\delta>0$, $\frac{1}{\delta}\wedge
\left(\tilde{\Isf}^{X,Y}(x,y)-\delta\right)\leq \Isf^X(x)$ for $(x,y)$ such that $G(x,y)=0$ and $\Isf^{X}(x)<\infty$, this concludes the proof of (\[ratefunction\]).
The various expressions of $\Isf^{H(X)}$ are now quite easy to obtain from $$\begin{aligned}
\label{eq:rateH1}\Isf^{H(X)}(y) = \inf\{ \Isf^X(x), \: G(x,y)=0\}.\end{aligned}$$ For (\[eq:rateH\]), note that since the set $\{x,\: G(x,y) = 0\}$ is closed the minimum in (\[eq:rateH1\]) (if it is finite) is attained for a certain $x^*$ with $G(x^*,y)=0$ and $\Isf^X(x^*)<\infty$.
We prove now that $$\begin{aligned}
\inf\{\Isf^X(x),\:y\in H^x\}= \inf\{\Isf^X(x),\: G(x,y)=0\}.\end{aligned}$$ If $y\in H^x$, then there exists $x_n\rightarrow x$ such that $H(x_n)\rightarrow y$. Hence by continuity of $G$, we have $G(x,y)=0$. Now if $G(x,y)=0$ and $\Isf^X(x)<\infty$, it follows from the assumptions that $y\in H^x$.
To see that the last expression in Remark \[rem1\] is true, we show that for any open set $O\subset \Fcal$, we have, $$\begin{aligned}
\label{eq:ineqI}\inf_{y\in O}\Isf^{H(X)}(y) = \inf_{y\in O}
\{\Isf^X(x), \:y=H(x)\}.\end{aligned}$$ For $y\in O$ and any $x$ such that $G(x,y)=0$, there exists $x_n\rightarrow x$, such that $H(x_n)\rightarrow y$ and $\Isf^X(x_n)\rightarrow \Isf^X(x)$. Hence for $n$ sufficiently large, we have $H(x_n)\in O$ and then $$\begin{aligned}
\inf_{y\in O} \{\Isf^X(x), \:y=H(x)\}\leq \inf_n \Isf^X(x_n)\leq
\Isf^X(x).\end{aligned}$$ Taking the minimum over all $x$ such that $G(x,y)=0$ gives the $\geq$ inequality in (\[eq:ineqI\]), the converse inequality is obvious.
Queueing networks with Bernoulli routing: description and large deviations results {#sec:network}
==================================================================================
General setting and notation
----------------------------
We start with the basic model for an isolated queue and refer to [@ABR] for more details on the relationship with other models of the literature.
The model for an isolated queue is in term of two primitive quantities belonging to $\Dbb(\real_+)$: the arrival process $\A$ and the service process $\Sbf$. The departure process $\D$ is a derived quantity that is obtained as a functional of the arrival and service processes as follows: $$\begin{aligned}
\label{ssq}\D(t) := \inf_{0\leq s\leq t}\left\{
\Sbf(t)-\Sbf(s)+\A(s)\right\}\wedge \Sbf(t).\end{aligned}$$
From a mathematical point of view, if $\Rcal:\Dbb\rightarrow \Dbb$ (where $\Dbb$ is the space of cadlag $\real$-valued functions defined on $\real_+$) is the one-dimensional Skorohod’s reflection map defined by $\Rcal(\Xbf)(t):=\sup_{0\leq s\leq
t}\left\{\Xbf(t)-\Xbf(s)\right\}\vee \Xbf(t)$. We have $\D=\A-\Rcal(\A-\Sbf)$. It is easy to see that $\D\in \Dbb(\real_+)$ and $\D\leq \A$.
The queue length process is defined as the difference of the arrival process and the departure process, $$\begin{aligned}
\Q(t):= \A(t)-\D(t)=\sup_{0\leq s\leq t}\left\{
\A(t)-\A(s)-(\Sbf(t)-\Sbf(s))\right\}\vee \left(
\A(t)-\Sbf(t)\right).\end{aligned}$$
If the arrival process $\A$ and the service process $\Sbf$ are counting processes, this model is called a single queue with autonomous service: the queue length is increased by one whenever there is an arrival from the arrival process and the queue length is decreased by one whenever there is an arrival from the service process and the queue is not empty (see [@Borovkov]). Note in particular that in the case where the process $\Sbf$ is a Poisson point process, then this model is a standard $./M/1$ queue.
We now consider networks obtained by interconnecting queues modeled by (\[ssq\]) when the departure process of one queue is randomly routed to the other queues as for Jackson networks. The networks we consider are characterized by the fact that service times and routing decisions are associated with stations and not with customers. This means that we associate to each of the $K$ stations three predefined counting processes: an arrival process, a service process and a routing process. The arrival process and the service process of station $k$ are described by the sequences of exogenous arrival times $\{T^{(k)}_j\}_{j\geq 1}$ and service times $\{\sigma^{(k)}_j\}_{j\geq 1}$. If there is no exogenous arrival at station $k$, we use the convention $T^{(k)}_j=\infty$ for all $j$. When the $j$-th customer has completed his service at station $k$, he is sent to station $\nu^{(k)}_j$ (or leaves the network if $\nu^{(k)}_j=K+1$) and is put at the end of the queue on this station, where $\{\nu^{(k)}_j\}_{j\geq 1}$ is also a predefined sequence, called the routing sequence. The sequences $\{T^{(k)}_j\}_{j\geq 1}$, $\{\sigma^{(k)}_j\}_{j\geq 1}$ and $\{\nu^{(k)}_j\}_{j\geq 1}$, where $k$ ranges over the set of stations, are called the driving sequences of the network. A network will be defined by $\left\{\{\sigma^{(k)}_j\}_{j\geq 1},\:
\{\nu^{(k)}_j\}_{j\geq 1},\:\{T^{(k)}_j\}_{j\geq 1},\: n^{(k)},\:
1\leq k\leq K\right\}$, where $(n^{(1)},\ldots,n^{(K)})$ describes the initial condition. The interpretation is as follows: at time $t=0$, in node $k$, there are $n^{(k)}$ customers with service times $\sigma^{(k)}_1,\dots, \sigma^{(k)}_{n^{(k)}}$ (if appropriate, $\sigma^{(k)}_1$ may be interpreted as a residual service time). In particular at time $0$, the total number of customers in the network is $n^{(1,K)}=n^{(1)}+\dots n^{(K)}$.
In what follows, we will describe the driving sequences thanks to their associated counting functions. We will use the following notation: $\sigma^{(k)}(1,n) = \sum_{j=1}^n\sigma^{(k)}_j$, for $0\leq k\leq K$.
We define the sequence of networks ${\JN}_n=\{\Sbf_n(t),
\Pbf_n(t),\Nbf_n(t)\}$ with $$\begin{aligned}
\Nbf^{(i)}_n(t) &=&\frac{1}{n}\left(n^{(i)}_n+\sum_k \ind_{\{T^{(i)}_k\leq nt\}}\right),\\
\Sbf^{(i)}_n(t) &=& \frac{1}{n} \sum_k
\ind_{\{\sigma^{(i)}(1,k)\leq nt\}},\\
\Pbf_n^{(i,j)}(t) &=& \frac{1}{n} \sum_{k\leq
nt}\ind_{\{\nu^{(i)}_k=j\}}.\end{aligned}$$
Note that we allow the initial queue length to depend on $n$, $\Nbf^{(i)}_n(0)=n^{(i)}_n$ but the other driving sequences describing the arrival times, the service times and the routing decisions do not depend on $n$. Note also that if there is no exogenous arrival at station $i$, we have $\Nbf^{(i)}_n(t)=\Nbf^{(i)}_n(0)$ for all $t\geq 0$.
For the network $\JN_n$, we denote the corresponding input and output processes of each queue $k$ of the network by $\A_n^{(k)}$ and $\D_n^{(k)}$ respectively. We will use the following notation $\A_n =(\A_n^{(1)},\dots, \A_n^{(K)})$ and $\D_n=(\D_n^{(1)},\dots,
\D_n^{(K)})$. We now describe how the processes $\A_n$ and $\D_n$ are obtained form $\JN_n$.
We define the map $\Gamma:\Dbb_0(\real^K_+)\times \Dbb_0(
\Mbb^K)\times \Dbb(\real_+^K)\rightarrow\Dbb(\real^K_+)$ as follows: $$\begin{aligned}
\Gamma(\Xbf,\Pbf,\Nbf)^{(i)}(t) := \Nbf^{(i)}(t) +
\sum_{j=1}^K\Pbf^{(j,i)}(\Xbf^{(j)}(t)).\end{aligned}$$
The following lemma is straightforward.
\[lem:contGamma\]The map $\Gamma$ is continuous for the compact uniform topology and non-decreasing in its first argument.
We define the map $\Phi:\Dbb(\real^K_+)\times\Dbb_0(\real^K_+)\rightarrow\Dbb_0(\real^K_+)$ as follows: $$\begin{aligned}
\Phi(\Xbf,\Ybf)^{(i)}(t):= \inf_{0\leq s\leq t}\left\{
\Ybf^{(i)}(t)-\Ybf^{(i)}(s)+\Xbf^{(i)}(s)\right\}\wedge
\Ybf^{(i)}(t).\end{aligned}$$
\[lem:contPhi\] The map $\Phi$ is continuous for the compact uniform topology and non-decreasing in its first argument.
We can clearly consider the map $\Phi$ with $K=1$ only. Let $\Rcal$ be the one-dimensional reflection map, we have $\Phi(\Xbf,\Ybf)=\Xbf-\Rcal(\Xbf-\Ybf)$. It is easy to see that for any $T>0$, $$\begin{aligned}
\sup_{0\leq t\leq T}|\Rcal(\Xbf)(t) -\Rcal(\Xbf')(t)|\leq 2
\sup_{0\leq t\leq T}|\Xbf(t) -\Xbf'(t)|,\end{aligned}$$ from which the continuity of $\Phi$ follows. Its monotonicity is obvious.
\[rem:conc\]Consider the mapping $\Phi$ with $K=1$ and $\Ybf(t)=\mu t$, with $\mu\geq 0$. If $\mu=0$, since $\Phi(\Xbf,\Ybf)\leq \Ybf$, we have $\Phi(\Xbf,\Ybf)(t)=0$ for all $t$. If $\mu\not= 0$, we have $\Phi(\Xbf,\Ybf)(t) = \inf_{0\leq
s\leq t} \{\Xbf(s)+\mu(t-s)\}$. Moreover if $\Xbf$ is a concave function, then this equation reduces to $\Phi(\Xbf,\Ybf)(t)=
\Xbf(t)\wedge \mu t$. Hence we can write $$\begin{aligned}
\mbox{$\Ybf(t)=\mu t$, with $\mu\geq 0$} &\Rightarrow &\Phi(\Xbf,\Ybf)(t) = \mu t \wedge \inf_{0\leq s\leq t} \{\Xbf(s)+\mu(t-s)\},\\
\mbox{if moreover $\Xbf$ is a concave function } & \Rightarrow &
\Phi(\Xbf,\Ybf)(t) = \mu t \wedge \Xbf(t).\end{aligned}$$
It is easy to adapt the proof of Theorem 2.1 of [@cm:disc] or Proposition 2.1 of [@lel] to show that the following fixed-point equation: $$\begin{aligned}
\label{fixpointeq1}
\left\{\begin{array}{lcl}\A_n &=& \Gamma(\D_n,\Pbf_n,\Nbf_n )=\Gamma(\D_n,\JN_n),\\
\D_n &=& \Phi(\A_n,\Sbf_n)=\Phi(\A_n,\JN_n),\end{array}\right.\end{aligned}$$ has an unique solution when each component of $n\Sbf_n$, $n\Pbf_n$ and $n\Nbf_n$ is a counting function (i.e. non-decreasing function of $\Dbb(\real^K_+)$ or $\Dbb(\Mbb^K)$ that is piece-wise constant with jumps of size one). In this case the corresponding functions $n\A_n$ and $n\D_n$ are also counting functions and we denote the solution of (\[fixpointeq1\]) by $\Psi(\Sbf_n,\Pbf_n,\Nbf_n)=\Psi(\JN_n)$.
Note that the only difference between our model and generalized Jackson networks as described in [@lel] resides in the queueing mechanism (\[ssq\]) which is sometimes called autonomous. Consider a network $\JN=\{\Sbf,\Pbf, \Nbf\}$ where the processes are counting processes. Then due to some monotonicity arguments, it is possible to relate (see [@CTK]):\
- the process $(\tilde{\A},\tilde{\D})$ associated to $\JN$ with the dynamic described in [@lel];\
- the processes $\Psi(\JN)=(\A,\D)$ solution of the fixed point equation.\
Note that in the case where the process $\Sbf$ is a Poisson point process, our model is exactly a Jackson network (see [@Apq]).
We will see in the next section that our stochastic assumptions allow us to cover this case. Hence our large deviation result is a generalization of [@atardupuis] and [@ign].
Stochastic assumptions {#sec:sto}
----------------------
In what follows, it will be important to distinguish the nodes of the network that do not receive any exogenous customer, i.e. the nodes $i\in \Scal^c$ with $\Scal=\{i,\:T^{(i)}_1\leq \infty \}$. A network $\JN=\{\Sbf,\Pbf,\Nbf\}$ is an object in $\Ecal\subset
\Dbb_0(\real_+^K) \times \Dbb_0(\Mbb^K)\times \Dbb(\real_+^K)$, with the additional constraints:
1. $\Nbf^{(i)}(t)=\Nbf^{(i)}(0)$ for all $t$, for $i\notin
\Scal$;
2. for all $0\leq v\leq u$, we have $\sum_{j=1}^K \Pbf^{(i,j)}(u)-\Pbf^{(i,j)}(v)\leq (u-v)$.
Note that $\Ecal$ is closed in $\Dbb_0(\real_+^K) \times
\Dbb_0(\Mbb^K)\times \Dbb(\real_+^K)$.
We define for $(s^{(1)},\dots, s^{(K)})\in \real_+^K$ and $(n^{(1)},\dots, n^{(K)})\in \real_+^K$, the functions $$\begin{aligned}
\Isf^\Sbf(s^{(1)},\dots, s^{(K)}) &=& \sum_{i=1}^K
\Isf^{\Sbf^{(i)}}(s^{(i)}),\\
\Isf^\Nbf(n^{(1)},\dots, n^{(K)}) &=& \sum_{i\in \Scal}
\Isf^{\Nbf^{(i)}}(n^{(i)})+\infty\ind_{\{n^{(i)}>0, \:i\notin
\Scal\}},\end{aligned}$$ where each $\Isf^{\Sbf^{(i)}}$(resp. $\Isf^{\Nbf^{(i)}}$ for $i\in \Scal$) is a $[0,\infty]$-valued convex good rate function, attaining zero on $\real_+$ admitting a unique minimum at the point $\mu^{(i)}$ (resp. $\lambda^{(i)}$ for $i\in \Scal$) and with a domain open on the right.
We assume that the sequence ${\JN}_n=\{\Sbf_n(t),
\Pbf_n(t),\Nbf_n(t)\}$ satisfies a LDP in the space $\Ecal$ with a good rate function $\Isf^{\JN}$ given by $$\begin{aligned}
\label{ass:rate}\Isf^{\JN}(\Sbf, \Pbf,\Nbf) :=
\Isf^0(\Nbf(0))+\int_0^\infty
\Isf^\Sbf(\dot{\Sbf}(t))+\tilde{\Dsf}(\dot{\Pbf}(t)\|R)+\Isf^\Nbf(\dot{\Nbf}(t))
dt,\end{aligned}$$ if the argument functions are absolutely continuous and equal to infinity otherwise. We make the following assumptions on the matrix $R$:
1. We assume that $\rho(R)<1$.
2. We assume that for all $1\leq i\leq K$, we have $$\begin{aligned}
\label{eq:RO}(\Ncal+\Ncal R+\dots+\Ncal R^K)^{(i)}>0,\end{aligned}$$ where $\Ncal$ is the line vector of $\real_+^K$ defined by $\Ncal^{(i)}=\ind_{\{i\in \Scal\}}$.
We show now that our stochastic assumptions cover the case where $\Sbf^{(k)}$ and $\Nbf^{(i)}$ (with $i\in \Scal$) are independent and correspond to renewal processes and where the routing is a Bernouilli routing associated with the matrix $R$ that satisfies previous assumption.
We recall here some results of Puhalskii [@puh:ssq] concerning large deviations of renewal processes and show that our assumptions on the rate function (\[ass:rate\]) are satisfied in the i.i.d case. Denote by $\{\zeta_i,\:i\geq 1\}$ a sequence of non-negative i.i.d. random variables with positive mean. Let $$\begin{aligned}
\nonumber \alpha(\theta) &=& \log \esp\left[ e^{\theta \zeta_1}\right],\\
\nonumber \theta^* &=& \sup\{\theta>0, \:\alpha(\theta)<\infty \},\\
\label{eq:equal}\alpha^*(x) &=& \sup_{\theta}\{\theta x-\alpha(\theta)\}=\sup_{\theta< \theta^*}\{\theta x-\alpha(\theta)\},\\
\nonumber g(x)&=& x\alpha^*(1/x)=\sup_{\theta<\theta^*}\{\theta
-x\alpha(\theta)\}.\end{aligned}$$ Note that the function $\alpha$ is a convex function and differentiable on $(-\infty,\theta^*)$ with $\alpha'(0)=\esp[\zeta_1]>0$. In particular, we have $\lim_{\theta\uparrow\theta^*}\alpha(\theta)=\infty$, from which we get the equality in (\[eq:equal\]). The functions $\alpha^*$ and $g$ are convex rate functions. Introduce the sequence of processes $\{\Cbf_n\}_n$: $$\begin{aligned}
\Cbf_n(t) = \frac{1}{n} \sum_{i}\ind_{\{\sum_{j=1}^i\zeta_j\leq
nt\}}.\end{aligned}$$ Then Theorem 3.1 of [@puh:ssq] gives: If $\prob(\zeta_1>0)=1$, then the sequence $\{\Cbf_n\}_n$ satisfies a LDP in $\Dbb(\real_+)$ with the good rate function $$\begin{aligned}
\Isf^\Cbf(\x) =\left\{\begin{array}{ll} \int_0^\infty
g(\dot{\x}(t))dt,
&\mbox{if $\x\in \Cbb(\real_+)$ is absolutely continuous,}\\
\infty,&\mbox{otherwise.}
\end{array}\right.\end{aligned}$$ It then follows that $g$ is a good rate function. Moreover, we have $\essinf \zeta_1 =0$ if and only if $g(x)$ is finite for all $x\geq
\esp[\zeta_1]^{-1}$ (note in particular, that in this case, the domain of $g$ is open on the right). The proof of this fact can be found in [@Puh:jack] and follows the argument: from $\alpha(\theta) \geq \esp[\zeta_1]\theta$, we have $g\left(\esp[\zeta_1]^{-1}\right)=0$ and for all $x\geq
\esp[\zeta_1]^{-1}$, we have $g(x) = \sup_{\theta\leq
0}\left\{\theta -x\alpha(\theta) \right\}$. If $\essinf \zeta_1=0$, we have for arbitrary $\epsilon>0$ and for $\theta\leq 0$, $$\begin{aligned}
\alpha(\theta) = \log\esp\left[ e^{\theta \zeta_1}\right]\geq \theta
\epsilon+\log \prob(\zeta_1<\epsilon),\end{aligned}$$ hence for $x>\epsilon^{-1}$, we have $g(x)\leq -\log
\prob(\zeta_1<\epsilon)$. It is clear that if $\essinf \zeta_1>0$, then for any $x>\essinf \zeta_1^{-1}$, we have $g(x)=\infty$.
Concerning the large deviations of the routing processes given in term of the Kullback-Leibler information divergence, it follows directly from Corollary 6.1 of [@funcldp] in the case of Bernouilli routing, i.e. when the sequences $\{\nu^{(k)}_j\}_{j\geq
1}$ are sequences of i.i.d. random variables in $[1,K]$ and independent in $k$ such that $$\begin{aligned}
\prob(\nu^{(k)}_1=i)=R^{(k,i)}.\end{aligned}$$
Sample path large deviations for the queue length process {#sec:result}
---------------------------------------------------------
We now return to the sequence of queueing networks defined in Section \[sec:network\]. Recall that $(\A_n,\D_n)$ correspond to the arrival and departure processes from each station. We now give our theorem for the queue length process defined as $\Q_n=\A_n-\D_n$.
\[th:qu\] The sequence of processes $\{\Q_n\}_n$ satisfies a LDP in $\Dbb(\real^K_+)$ with good rate function that is finite for $\Q$ absolutely continuous given by: $$\begin{aligned}
\Isf^{0}(\Q(0))+\Isf^\Q_{\Q(0)}(\Q),\end{aligned}$$ where for $q\geq 0$, $\Isf^\Q_q(.)$ is a good rate function that is finite for absolutely continuous $\Q$ such that $\Q(0)=q$ and given by: $$\begin{aligned}
\Isf^\Q_q(\Q) := \int_0^\infty \Hsf^\Q(\Q(s),\dot{\Q}(s))ds,\end{aligned}$$ where $\Hsf^\Q$ is given by, $$\begin{aligned}
\Hsf^\Q(Q,\dot{Q}):=\inf\left\{\sum_{i\in
E(Q)}\Isf^{\Sbf^{(i)}}(D^{(i)})\ind_{\{D^{(i)}>\mu^{(i)}\}}+\sum_{i\notin
E(Q)}\Isf^{\Sbf^{(i)}}(D^{(i)}) +\sum_i D^{(i)}
\tilde{\Dsf}(P^{(i)}\|R^{(i)})+\Isf^\Nbf(N)\right\}\end{aligned}$$ where $E(Q)=\{i,\: Q^{(i)}=0\}$ and the infimum is taken over the set of $(D,P,N)\in \real_+^K\times \Mbb^K\times \real_+^K$ such that $$\begin{aligned}
\dot{Q} = N+(P^t-Id)D.$$
In [@Puh:jack], Puhalskii obtains a LDP for the queue length process of a generalized Jackson network with a rate function that coincides with Theorem \[th:qu\]. Note that our model is slightly different here since we model the dynamic of a queue by a reflection mapping. Still in the case of Poisson processes for the inputs, both models correspond to the (exponential distribution) Jackson network. Recall that the rate function for a Poisson process of rate $\lambda$ is given by (we keep the same notation as in \[sec:sto\]), $$\begin{aligned}
\label{eq:pois}
\Isf^{\Cbf}(\xbf)=\int_0^\infty \lambda
\dot{\xbf}(t)\log\frac{\dot{\xbf}(t)}{\lambda}-\dot{\xbf}(t)+\lambda\:
dt,\end{aligned}$$ for absolutely continuous functions $\xbf\in \Cbb(\real_+)$. Hence if we replace (\[eq:pois\]) in the expression of $I^\Q_q$, we obtain the rate function for the large deviations of a Jackson network. In this specific case, the rate function has been obtained in different forms by Atar and Dupuis [@atardupuis] and Igniatiouk-Robert [@ign] and some bounds have been computed by Majewski [@maj-ldbound]. Compare to these results, our representation has the advantage of being quite intuitive, in the sense that each term is easy to interpret. If we interpret $D,P,N$ as instantaneous departure, routing and exogenous arrival rates, then $N+(P^t-Id)D$ is just the vector of rates at which the queue lengths vary. Hence given a rate of change of $\Q$, the system behaves in such a way to minimize the instantaneous “costs” of departure, routing and exogenous arrival rates over all the rates that yield the desired $\dot{\Q}$.
From a methodological point of view, the argument of [@Puh:jack] is quite different from ours since the density condition (that we could compare to our Proposition \[prop:contraction\]) is verified on the rate function $\Isf^{\Q}_q$ (see condition (D) in [@Puh:jack]) whereas we are checking the density argument on the rate function of the inputs.
Extension of $\Psi$ to piece-wise linear networks {#sec:extpsi}
=================================================
In this section we consider processes that are continuous, i.e. in $\Cbb(E)$, hence topological concepts refer to the compact uniform topology.
We first recall Proposition 3.2 of [@lel],
\[prop:fixpoint:lel\] Given a $K\times K$ substochastic matrix $P$ with $\rho(P)<1$ and vectors $(\alpha,y)\in \real_+^{2K}$, the fixed point equation $$\begin{aligned}
x^{(i)}&=& \alpha^{(i)}+\sum_{j=1}^K P^{(j,i)}\left(x^{(j)}\wedge
y^{(j)}\right),\end{aligned}$$ has a unique solution $x(y,P,\alpha)$. Moreover, $(y,\alpha)\mapsto
x(y,P,\alpha)$ is a continuous non-decreasing function.
We first consider a linear network $\JN$ and show that the mapping $\Psi$ (defined as the solution of the fixed-point Equation (\[fixpointeq1\])) is well defined for such a network. By linear, we mean the following $\Nbf^{(i)}(t)=N^{(i)}+\lambda^{(i)}t$, with $\lambda^{(i)}\geq 0$ and $N^{(i)}\in \real_+$, $\Sbf^{(i)}(t)
=\mu^{(i)}t$, with $\mu^{(i)}\geq 0$, and $\Pbf^{(i,j)}(t) =
P^{(i,j)}t$. We assume that $\rho(P)<1$.
\[lem:linear\]Under previous assumptions, the fixed point equation (\[fixpointeq\]) has an unique solution $\Xbf_f[\mu,P,N,\lambda](t)=x(\mu t,P,N+\lambda t)$, where $\mu
=(\mu^{(i)})_i$, $N=(N^{(i)})_i$ and $\lambda=(\lambda^{(i)})_i$.
Since $\mu,P,N,\lambda$ are fixed here, we omit to explicitly write the dependence in these variables. In this case, the fixed point equation (\[fixpointeq\]) reduces to (see Remark \[rem:conc\]) $$\begin{aligned}
\label{fixpointeq:lin}\left\{\begin{array}{lcl}\A^{(i)}(t) &=& N^{(i)}+\lambda^{(i)}t+\sum_{j=1}^K P^{(j,i)}\D^{(j)}(t),\\
\D^{(i)}(t) &=& \mu^{(i)}t \wedge \inf_{0\leq s\leq t}\{
\A^{(i)}(s)+\mu^{(i)}(t-s)\}.\end{array}\right.\end{aligned}$$
Thanks to Proposition \[prop:fixpoint:lel\], $\Xbf_f(t)=x(\mu
t,P,N+\lambda t)$ is the unique solution of the fixed point equation $$\begin{aligned}
\label{fixpointeq:lin2}\left\{\begin{array}{lcl}\A^{(i)}(t) &=& N^{(i)}+\lambda^{(i)}t+\sum_{j=1}^K P^{(j,i)}\D^{(j)}(t),\\
\D^{(i)}(t) &=& \A^{(i)}(t) \wedge \mu^{(i)}t.\end{array}\right.\end{aligned}$$ We prove now that $\Xbf_f$ is the unique solution of the fixed point equation (\[fixpointeq:lin\]).
For simplicity, we denote the fixed point equation (\[fixpointeq:lin\]), resp. (\[fixpointeq:lin2\]), by $\A=F(\A)$, resp. by $\A=\tilde{F}(\A)$. Note that these functions are non-decreasing, continuous and such that $F\leq \tilde{F}$.
From $\0\leq \Xbf_f$, we get $\0\leq F(\0)\leq \tilde{F}(\0)\leq
\tilde{F}(\Xbf_f)$. Hence $F^n(\0)\nearrow \Lbf\leq \Xbf_f$ and $F(\Lbf)=\Lbf$. Moreover for any solution $\Ybf$ of the fixed point equation (\[fixpointeq:lin\]), we have $\Lbf\leq \Ybf \leq \Xbf_f$ because $\Ybf=F(\Ybf)\leq \tilde{F}(\Ybf)$ and $\tilde{F}^n(\Ybf)\nearrow \Xbf_f$.
Since $\0$ is a concave function, we have $F(\0)=\tilde{F}(\0)$ and hence it is still a concave function. Hence we have $\tilde{F}^n(\0)=F^n(\0)$ since the image by $\tilde{F}$ of a concave function is a concave function and $F=\tilde{F}$ on the subspace of concave functions. Hence we have $\Lbf = \Xbf_f$ which concludes the proof.
In order, to extend $\Psi$ to piece-wise linear networks, we proceed step by step on each interval where the driving functions $\Sbf,\Pbf,\Nbf$ are linear. The following lemma allows to glue the constructed solution on each adjacent interval. In a queueing context, this lemma says that the output of a single server queue fed by the arrival process $\A$ and service time process $\Sbf$ viewed from time $u$ is just the same as the output process of a single server queue that we start at time $u$ with arrival process $\tilde{\A}(t)=\A(t+u)-\A(u)+\A(u)-\D(u)$ (i.e. with the same increment as the original process on this period of time plus an additional bulk corresponding to the queue length at time $u$) and with service time process $\tilde{\Sbf}(t)$.
\[lem:trans\] Let $\A,\Sbf\in \Dbb(\real_+)\times \Dbb_0(\real_+)$ and $\D=\Phi(\A,\Sbf)$. Define $\tilde{\A},\tilde{\Sbf}\in
\Dbb(\real_+)\times \Dbb_0(\real_+)$ as follows $$\begin{aligned}
\tilde{\A}(t) &:=& \A(t+u)-\D(u),\\
\tilde{\Sbf}(t) &:=& \Sbf(t+u)-\Sbf(u).\end{aligned}$$ Let $\tilde{\D}=\Phi(\tilde{\A},\tilde{\Sbf})$, then we have $$\begin{aligned}
\tilde{\D}(t) = \D(t+u)-\D(u).\end{aligned}$$
We show that for $\D=\Phi(\A,\Sbf)$, we have $$\begin{aligned}
\D(t+u)-\D(u) = \inf_{u\leq s\leq t+u}\left\{ \Sbf(t+u)
-\Sbf(s)+\A(s)-\D(u)\right\}\wedge \left\{\Sbf(t+u)-\Sbf(u)
\right\},\end{aligned}$$ from which the lemma follows.
We write $$\begin{aligned}
\D(t+u)-\D(u) &=& \inf_{0\leq s\leq u}\left\{ \Sbf(t+u)
-\Sbf(s)+\A(s)-\D(u)\right\}\\
&&\wedge \inf_{u\leq s\leq t+u}\left\{ \Sbf(t+u)
-\Sbf(s)+\A(s)-\D(u)\right\}\wedge \left\{\Sbf(t+u)-\D(u) \right\},\end{aligned}$$ Since $\D(u)\leq \Sbf(u)$, we have to prove that $$\begin{aligned}
\Sbf(t+u)-\Sbf(u) \geq \inf_{0\leq s\leq u}\left\{ \Sbf(t+u)
-\Sbf(s)+\A(s)-\D(u)\right\}\wedge \left\{\Sbf(t+u)-\D(u) \right\}.\end{aligned}$$ This will follow from, $$\begin{aligned}
\inf_{0\leq s\leq u}\left\{ \Sbf(t+u) -\Sbf(s)+\A(s)-\D(u)\right\}
&=&\Sbf(t+u)-\Sbf(u) +\inf_{0\leq s\leq u}\left\{ \Sbf(u)
-\Sbf(s)+\A(s)\right\}-\D(u)\\
&\leq&\Sbf(t+u)-\Sbf(u).\end{aligned}$$
We consider now piece-wise linear networks: the functions $u\mapsto\Nbf^{(i)}(u), u\mapsto\Sbf^{(i)}(u)$ and $u\mapsto
\Pbf^{(i,j)}(u)$ are continuous piece-wise linear functions such that $\Nbf^{(i)}(0)\in\real_+$ and $\Sbf^{(i)}(0)=\Pbf^{(i,j)}(0)=0$ and $\rho(\dot{\Pbf}(t))<1$ for all $t\geq 0$.
\[prop:JNlinear\] For a piece-wise linear network, there exists an unique solution of the fixed point equation (\[fixpointeq\]). We still denote by $\Psi$ the mapping that to any piece-wise linear network $\JN$ associates the corresponding couple $(\A,\D)$.
The existence is a direct consequence of monotonicity properties and continuity of the maps $\Gamma$ and $\Phi$. We define the sequence of processes $\{\A[k],\D[k]\}_{k\geq 0}$ with the recurrence equation: $$\begin{aligned}
\left\{ \begin{array}{l}
\A[k+1] = \Gamma(\D[k],\JN),\\
\D[k+1] = \Phi(\A[k+1],\JN),
\end{array}\right.\end{aligned}$$ and with initial condition $\D[0] = \0$. By the monotonicity properties of $\Phi$ and $\Gamma$, we have $$\begin{aligned}
\0\leq\A[1] &\Rightarrow& \Phi(\0,\JN)=\0=\D[0] \leq
\Phi(\A[1],\JN)=\D[1]\\
&\Rightarrow& \Gamma(\D[0],\JN)=\A[1]\leq \Gamma(\D[1],\JN)=\A[2],\end{aligned}$$ and the sequence $\{\A[k],\D[k]\}_{k\geq 0}$ is increasing. Note that $\D[k]\leq \Sbf$ and hence the following limits are well defined $$\begin{aligned}
\lim_{k\rightarrow \infty} \A[k] = \A\quad \mbox{and,}\quad
\lim_{k\rightarrow \infty} \D[k] = \D.\end{aligned}$$ Since $\Gamma$ and $\Phi$ are continuous, $(\A,\D)$ is a solution of the fixed point equation (\[fixpointeq\]).
We now prove uniqueness. First recall that we call $\alpha$, a partition of $\real_+$, any increasing sequence of points $\alpha=\{a_n\}_n$ with $a_0=0$ and $a_n\rightarrow \infty$. For two partitions $\alpha=\{a_n\}_n$ and $\beta=\{b_n\}_n$, we say that $\gamma=\{g_n\}_n$ is the union of $\alpha$ and $\beta$ if $\gamma$ is a partition such that for all $n$ there exists $m$ such that either $g_n=a_m$ or $g_n=b_m$.
Let $\tau=\{t_n\}_n$ be the union of the partitions associated with each function $\Sbf, \Pbf, \Nbf$. We define for $x\in \real_+$, $d(x,\tau)=\min_n\{t_n-x,\: t_n> x\}>
0$. Assume that we are given two solutions of the fixed point equation (\[fixpointeq\]): $(\A_1,\D_1)$ and $(\A_2,\D_2)$. First note that thanks to Lemmas \[lem:wPhi\] and \[lem:wGamma\], any solution of (\[fixpointeq\]) is absolutely continuous. Let $z=\inf\{t, \A_1(t)\neq \A_2(t)\}$, in particular, we have $\A_1(t)=\A_2(t)$ and $\D_1(t)=\D_2(t)$ for all $t\leq z$.
Define $u=\min_i d(\D_\bullet^{(i)}(z),\tau)\wedge d(z,\tau)>0$, where the notation $_\bullet$ can be replaced either by $_1$ or by $_2$. We have that for $t\in [0,u]$, $$\begin{aligned}
\tilde{\Sbf}^{(i)}(t)&:=&\Sbf^{(i)}(z+t)-\Sbf^{(i)}(z)=t\mu^{(i)}, \\
\tilde{\Pbf}^{(i,j)}(t)&:=&\Pbf^{(i,j)}(\D_\bullet^{(i)}(z)+t)-\Pbf^{(i,j)}(\D_\bullet^{(i)}(z))=t P^{(i,j)}, \\
\tilde{\Nbf}^{(i)}&:=&\Nbf^{(i)}(z+t)-\Nbf^{(i)}(z)+\A_\bullet^{(i)}(z)-\D_\bullet^{(i)}(z)=t\lambda^{(i)}+\A_\bullet^{(i)}(z)-\D_\bullet^{(i)}(z),\end{aligned}$$ Let $\tilde{\A}(t)=\Xbf_f[\mu,P,\A_\bullet(z)-\D_\bullet(z),\lambda](t)$ be the unique solution associated to the infinite horizon linear network defined above. The associated departure process is $\tilde{\D}(t) = \tilde{\A}(t)\wedge\mu t$. Let $v=\inf\{t,\:
\inf_i\tilde{\D}^{(i)}(t)=u\}$, in particular since $\tilde{\D}^{(i)}(t)\leq \mu^{(i)}t$, we have $v>0$. In view of Lemma \[lem:trans\], we have for $t\in (0,v)$, $$\begin{aligned}
\A_\bullet(t+z) = \tilde{\A}(t)+\D(z),\quad \D_\bullet(t+z)=
\tilde{\D}(t)+\D(z)\end{aligned}$$ this contradicts the fact that $z<\infty$ and concludes the proof.
Let $\Ecal \subset \Dbb_0(\real_+^K) \times \Dbb_0(\Mbb^K)\times
\Dbb(\real_+^K)$ as defined at the beginning of Section \[sec:sto\] and $\Fcal=\Dbb(\real_+^K)\times \Dbb_0(\real_+^K)$.
For $\JN\in \Ecal$ and $(\A,\D)\in \Fcal$, we define the function $$\begin{aligned}
G(\JN,\A,\D)= \|(\A-\Gamma(\D,\JN),\D-\Phi(\A,\JN))\|.\end{aligned}$$ The function $G$ is continuous and such that $$\begin{aligned}
G(\JN,\A,\D)= 0 \Leftrightarrow\left\{\begin{array}{lcl}\A &=& \Gamma(\D,\JN),\\
\D &=& \Phi(\A,\JN).\end{array}\right.\end{aligned}$$
Let $\Dcal_\JN$ be the subspace of $\Ecal$ of piecewise linear networks: namely $\JN=(\Sbf,\Pbf,\Nbf)\in \Dcal_\JN$ if the functions $u\mapsto\Nbf^{(i)}(u), u\mapsto\Sbf^{(i)}(u)$ and $u\mapsto \Pbf^{(i,j)}(u)$ are piecewise linear non-decreasing functions such that $\rho(\dot{\Pbf}(t))<1$ for all $t\geq 0$ and $\Nbf^{(i)}=\0$ for $i\notin \Scal$. We denote $\dot{\JN}=(\dot{\Sbf},\dot{\Pbf},\dot{\Nbf})$.
We proved that $$\begin{aligned}
\forall \JN\in \Dcal_\JN,\quad G(\JN,\A,\D)=0 \Leftrightarrow
(\A,\D)=\Psi(\JN),\end{aligned}$$ where $\Psi$ has been explicitly defined above. We are exactly in the framework of Section \[sec:contr\]. In the next section we construct the mapping $\Scal: \Ecal\times \Fcal\rightarrow
\Dcal_\JN^\nat$.
Sample path large deviations {#sec:ldp}
============================
In order to simplify the notations, we assume that $\Nbf_n(0)=0$ for all $n$. This condition can be weakened to the standard condition: $$\begin{aligned}
\lim_{n\rightarrow
\infty}\frac{1}{n}\log\prob(\Nbf_n(0)>\epsilon)=0,\end{aligned}$$ for all $\epsilon>0$. In this case, we have $\Isf^0(x)=\infty$ for all $x\neq0$ and $\Isf^0(0)=0$.
It is possible to deal with the case where the initial condition satisfies a LDP as assumed in Theorem \[th:qu\] by using a standard conditioning argument (as done in [@puh:ssq] for example).
Construction of the approximating sequence
------------------------------------------
This section is devoted to the proof of the following proposition:
\[prop:JNapprox\] We consider $\JN=(\Sbf,\Pbf,\Nbf)\in \Ecal$ such that $I^\JN(\JN)<\infty$ and such that there exists $(\A,\D)\in \Fcal$ that satisfies the fixed point equation (\[fixpointeq\]) given by, $$\begin{aligned}
\left\{\begin{array}{lcl}\A &=&\Gamma(\D,\JN),\\
\D &=& \Phi(\A,\JN).\end{array}\right.\end{aligned}$$ There exists a sequence $\{\JN_n\}_n=\Scal(\JN,\A,\D)$ such that $$\begin{aligned}
\label{eq:S1}\JN_n&\in& \Dcal_\JN\quad \mbox{for all $n$;}\\
\label{eq:S2}\JN_n &\rightarrow & \JN;\\\label{eq:S3}\Psi(\JN_n)&\rightarrow& (\A,\D);\\
\label{eq:S4}\Isf^\JN(\JN_n)&\rightarrow& \Isf^\JN(\JN).\end{aligned}$$
First note that since $\Isf^\JN(\JN)<\infty$, each process $\Sbf,\Pbf,\Nbf$ is absolutely continuous and $\dot{\JN}$ is well-defined. Moreover thanks to Lemma \[lem:propPhi\], the processes $\A$ and $\D$ are absolutely continuous too.
The idea to construct the sequence $\{\JN_n\}_n$ is to consider the piecewise approximation of the fixed point equation (\[fixpointeq\]). First consider the routing equation $\A=\Gamma(\D,\JN)$ for times $t$ such that $nt\in \nat$, $$\begin{aligned}
\underbrace{\A^{(i)}(t+1/n)-\A^{(i)}(t)}_{\Delta^{(i)}_n(\A)(t)}
&=&\underbrace{\Nbf^{(i)}(t+1/n)-\Nbf^{(i)}(t)}_
{\Delta^{(i)}_n(\Nbf)(t)}+\sum_{j=1}^K
\dot{\tilde{\Pbf}}_n^{(j,i)}(\D^{(j)}(t+))\underbrace{(\D^{(j)}(t+1/n)-\D^{(j)}(t))}_{\Delta^{(j)}_n(\D)(t)},\end{aligned}$$ where we define the piece-wise linear process $\tilde{\Pbf}_n^{(j,i)}(t)$ as follows, for $s\in
(\D^{(j)}(t),\D^{(j)}(t+1/n))$, $$\begin{aligned}
\dot{\tilde{\Pbf}}_n^{(j,i)}(s)
:=\frac{\Pbf^{(j,i)}(\D^{(j)}(t+1/n))-\Pbf^{(j,i)}(\D^{(j)}(t))}{\D^{(j)}(t+1/n)-\D^{(j)}(t)},\end{aligned}$$ if $\D^{(j)}(t+1/n)\neq \D^{(j)}(t)$, and we take $\dot{\tilde{\Pbf}}_n^{(j,i)}(\D^{(j)}(t))=0$ otherwise. In other words, we have $$\begin{aligned}
\tilde{\Pbf}_n^{(j,i)}(\D^{(j)}(t+1/n))-\tilde{\Pbf}_n^{(j,i)}(\D^{(j)}(t))&=&\dot{\tilde{\Pbf}}_n^{(j,i)}(\D^{(j)}(t+))(\D^{(j)}(t+1/n)-\D^{(j)}(t))\\
&=&\Pbf^{(j,i)}(\D^{(j)}(t+1/n))-\Pbf^{(j,i)}(\D^{(j)}(t))\end{aligned}$$
Note that $\{\dot{\tilde{\Pbf}}_n^{(j,i)}(t)\}_{i,j}\in \Mbb^K$ since we have by the definition of $\Ecal$, $$\begin{aligned}
\sum_i\Pbf^{(j,i)}(\D^{(j)}(t+1/n))-\Pbf^{(j,i)}(\D^{(j)}(t))\leq\D^{(j)}(t+1/n)-\D^{(j)}(t),\end{aligned}$$ but the matrix $(\dot{\tilde{\Pbf}}_n^{(j,i)}(\D^{(j)}(t+)))_{i,j}$ may not be of spectral radius less than $1$.
To circumvent this difficulty, we modify slightly the processes as follows, (the variables $\eta,\epsilon_n,\delta$ will be made precise latter) $$\begin{aligned}
\label{eq:approx}\Delta_n^{(i)}(\A)+\frac{\eta^{(i)}}{n}&=&\Delta_n^{(i)}(\Nbf)+\frac{\delta^{(i)}}{n}\\
\nonumber&&+\sum_{j=1}^K
\left((1-\epsilon_n^{(j)})\dot{\tilde{\Pbf}}_n^{(j,i)}+\epsilon_n^{(j)}
R^{(j,i)}
\right)\left(\Delta_n^{(j)}(\D)+\frac{\eta^{(j)}}{n}\right),\end{aligned}$$ where we omit to write the time $t$ and use the simplified notation $\dot{\tilde{\Pbf}}_n^{(j,i)}=\dot{\tilde{\Pbf}}_n^{(j,i)}(\D^{(j)}(t+))$.
We have to find $\eta,\epsilon_n,\delta$ such that (\[eq:approx\]) holds with $\eta^{(i)},\epsilon^{(i)}_n,\delta^{(i)}$ non-negative and $\delta^{(i)}=0$ for $i\not\in \Scal$. These constraints are satisfied by the following choice: first take $\delta$ such that $\delta^{(i)}>0$ for all $i\in \Scal$ and $\delta^{(i)}=0$ for $i\not\in \Scal$. Let $\eta(\delta)=\eta$ be the unique solution in $\real_+^K$ of the following equation (recall that $\rho(R)<1$), $$\begin{aligned}
\eta^{(i)} = \delta^{(i)}+\sum_{j=1}^K \eta^{(j)}R^{(j,i)}.\end{aligned}$$ Note that $\eta^{(i)}>0$ for all $i$ thanks to (\[eq:RO\]). Finally let define $\epsilon_n(\delta)=\epsilon_n$ as follows $\epsilon_n^{(i)} =
\frac{\eta^{(i)}}{n\Delta_n^{(i)}(\D)+\eta^{(i)}}\in (0,1]$ (note that $\epsilon^{(i)}_n=1$ if and only if $\Delta_n^{(i)}(\D)=0$).
It is easy to see that (\[eq:approx\]) holds since we have $$\begin{aligned}
(1-\epsilon_n^{(j)})\left(\Delta_n^{(j)}(\D)+\frac{\eta^{(j)}}{n}\right)
= \Delta_n^{(j)}(\D),&\mbox{or,}&
\epsilon_n^{(j)}\left(\Delta_n^{(j)}(\D)+\frac{\eta^{(j)}}{n}\right)
= \frac{\eta^{(j)}}{n},\end{aligned}$$ which imply respectively that $$\begin{aligned}
\Delta_n^{(i)}(\A)&=&\Delta_n^{(i)}(\Nbf)+\sum_{j=1}^K
(1-\epsilon_n^{(j)})\dot{\tilde{\Pbf}}_n^{(j,i)}\left(\Delta_n^{(j)}(\D)+\frac{\eta^{(j)}}{n}\right)\quad\mbox{and,}\\
\frac{\eta^{(i)}}{n}&=&\frac{\delta^{(i)}}{n}+\sum_{j=1}^K
\epsilon_n^{(j)} R^{(j,i)}
\left(\Delta_n^{(j)}(\D)+\frac{\eta^{(j)}}{n}\right),\end{aligned}$$ and summing these two equalities gives (\[eq:approx\]).
For $\delta$ fixed, we define for $s\in
(\D^{(j)}(t)+t\eta(\delta),\D^{(j)}(t+1/n)+(t+1/n)\eta(\delta))$, $$\begin{aligned}
\dot{\Pbf}_{n,\delta}^{(j,i)}(s)
=(1-\epsilon_n^{(j)})\dot{\tilde{\Pbf}}_n^{(j,i)}(\D^{(j)}(t+))+\epsilon_n^{(j)}
R^{(j,i)},\end{aligned}$$ where $\epsilon_n(\delta)$ is defined as above. In view of Lemma \[lem:specrad\], the matrix $\dot{\Pbf}_{n,\delta}^{(j,i)}(s)$ is of spectral radius less than one since $\epsilon_n^{(j)}>0$ for all $j$. Then as a direct consequence of (\[eq:approx\]), we have for $nt\in \nat$, $$\begin{aligned}
\label{eq:approx2}\A^{(i)}(t) + t\eta(\delta) = \Nbf^{(i)}(t)
+t\delta + \sum_{j=1}^K
\Pbf^{(j,i)}_{n,\delta}(\D^{(j)}(t)+t\eta(\delta)).\end{aligned}$$
If $\Nbf_{n,\delta}$ is the polygonal approximation of $t\rightarrow\Nbf(t)+t\delta$ with step $1/n$, we have clearly $\dot{\Nbf}_{n,\delta}\rightarrow \dot{\Nbf}+\delta$ as $n$ tends to infinity. Similarly, we have as $n$ tends to infinity, $$\begin{aligned}
\dot{\Pbf}_{n,\delta}^{(j,i)}(\D^{(j)}(t)+t\eta(\delta))\rightarrow
\left\{\begin{array}{ll}
(1-\epsilon^{(j)}(t))\dot{\Pbf}^{(j,i)}(\D^{(j)}(t))\frac{(\dot{\D}^{(j)}(t)+\eta(\delta))}{\dot{\D}^{(j)}(t)}&\\
\quad \quad
+\epsilon^{(j)}(t)R^{(j,i)}(\dot{\D}^{(j)}(t)+\eta(\delta))&\mbox{if
$\dot{\D}^{(j)}(t)>0$},\\
R^{(j,i)}\eta(\delta)&\mbox{otherwise,}
\end{array} \right.\end{aligned}$$ where $\epsilon^{(j)}(t) =
\eta^{(j)}(\delta)/(\eta^{(j)}(\delta)+\dot{\D}^{(j)}(t))<1$. Hence when $n$ tends to infinity and $\delta$ tends to zero, we have $\dot{\Nbf_{n,\delta}}\rightarrow \dot{\Nbf}$ and $\dot{\Pbf}_{n,\delta}^{(j,i)}\rightarrow\dot{\Pbf}^{(j,i)}$.
We consider now the queueing equation $\D=\Phi(\A,\Sbf)$ and construct the approximating sequence for $\Sbf$.
We begin with a first general lemma: given three processes $\A \leq
\D$ and $\Sbf$, we construct a piecewise linear function $\Sbf_n$ (with step $1/n$) as follows (with $nt\in \nat$):
- if $\A(t)=\D(t)$ and $\A(t+1/n)=\D(t+1/n)$, then $\Sbf_n(t+1/n)-\Sbf_n(t)=\Sbf(t+1/n)-\Sbf(t)$;
- otherwise, $\Sbf_n(t+1/n)-\Sbf_n(t)=\D(t+1/n)-\D(t)$.
We will denote this construction by $\Sbf_n = \Upsilon_n(\A,
\D,\Sbf)$.
\[lem:approx\] Let $(\A,\D,\Sbf)$ be absolutely continuous functions of $\Cbb(\real_+^K)\times \Cbb_0(\real_+^K) \times
\Cbb_0(\real_+^K)$such that $\Phi(\A,\Sbf)=\D$. We denote $\Sbf_n= \Upsilon_n(\A, \D,\Sbf)$. We have $\D_n=\Phi(\A_n,\Sbf_n)$ where $(\A_n,\D_n)$ is the polygonal approximation of $(\A,\D)$ with step $1/n$ and we have the following convergence as $n$ tends to infinity: $\Sbf_n\to \Sbf$, $\dot{\Sbf}_n\rightarrow \dot{\Sbf}$ and $\int_0^\infty
\Isf^{\Sbf}(\dot{\Sbf}_{n}(t))dt\to\int_0^\infty
\Isf^{\Sbf}(\dot{\Sbf}(t))dt$ .
We denote $\tilde{\D}_n=\Phi(\A_n,\Sbf_n)$. From the proof of Lemma \[lem:trans\], we have $$\begin{aligned}
\tilde{\D}_n(t+1/n)-\tilde{\D}_n(t) = \inf_{t\leq s\leq
t+1/n}\left\{ \Sbf_n(t+1/n)
-\Sbf_n(s)+\A_n(s)-\tilde{\D}_n(t)\right\}\wedge
\left\{\Sbf_n(t+1/n)-\Sbf_n(t) \right\},\end{aligned}$$ since all the functions are linear on the interval $(t,t+1/n)$, we have (with $nt\in \nat$), $$\begin{aligned}
\tilde{\D}_n(t+1/n)-\tilde{\D}_n(t) &=& \left\{
\A_n(t+1/n)-\tilde{\D}_n(t)\right\}\wedge
\left\{\Sbf_n(t+1/n)-\Sbf_n(t) \right\}.\end{aligned}$$ If $\tilde{\D}_n(t)=\D_n(t)$, then we have clearly $\tilde{\D}_n(t+1/n)=\D_n(t+1/n)$ since
- if $\A_n(t)=\D_n(t)$ and $\A_n(t+1/n)=\D_n(t+1/n)$, then we have $\Sbf(t+1/n)-\Sbf(t)\geq
\D_n(t+1/n)-\D_n(t)=\A_n(t+1/n)-\tilde{\D}_n(t)$ see (\[ineq:lem\]) for the inequality;
- otherwise, $\Sbf_n(t+1/n)-\Sbf_n(t)=\D_n(t+1/n)-\D_n(t)$ by definition and $\A_n(t+1/n)\geq
\D_n(t+1/n)$.
This proves the first part of the lemma. Moreover it follows directly form the definition of $\Upsilon$ that $\Sbf_n\left(t+\frac{1}{n}\right)-\Sbf_n(t) \leq
\Sbf\left(t+\frac{1}{n}\right)-\Sbf(t)$, hence we have for all $t$, $\limsup_{n\to \infty}\Sbf_n(t)\leq \Sbf(t)$ by a continuity argument. The fact that $\Sbf_n\to\Sbf$ follows directly from Fatou’s Lemma and the fact that $\dot{\Sbf}_n\to\dot{\Sbf}$. We now prove this last fact, let $C=\{t,\: \A(t)=\D(t)\}$. $C$ is a closed set and according to Lemma \[lem:propPhi\], we have for all $t\in
C^c$ (the complementary set of $C$), $\dot{\Sbf}(t)=\dot{\D}(t)$. For such $t\in C^c$, we have for $\epsilon>0$ sufficiently small and for sufficiently large $n$, $\A_n(u)\neq \D_n(u)$ for all $|u-t|\leq
\epsilon$. Hence we have $\dot{\Sbf}_n(t)=\dot{\D}_n(t)\rightarrow
\dot{\D}(t)$. Now for $t\in C^o$ in the interior of $C$, we have clearly $\dot{\Sbf}_n(t)\rightarrow \dot{\Sbf}(t)$. Hence we have $\dot{\Sbf}_n(t)\rightarrow \dot{\Sbf}(t)$ for $t\in C^o\cup C^c$.
We prove the last statement of the lemma. Since any open set of $\real$ is a countable union of disjoint intervals, $$\begin{aligned}
\int_{C^c} \Isf^{\Sbf}(\dot{\Sbf}_n(t))dt = \int_{C^c}
\Isf^{\Sbf}(\dot{\D}_n(t))dt &\leq& \int_{C^c}
\Isf^{\Sbf}(\dot{\D}(t))dt,\quad \mbox{by Jensen's
inequality}\\
&=&\int_{C^c} \Isf^{\Sbf}(\dot{\Sbf}(t))dt,\end{aligned}$$ and also directly still by Jensen’s inequality $\int_{C^o}
\Isf^{\Sbf}(\dot{\Sbf}_n(t))dt\leq \int_{C^o}
\Isf^{\Sbf}(\dot{\Sbf}(t))dt$. The convergence then follows from $$\begin{aligned}
\liminf_{n\rightarrow \infty}\int_0^\infty
\Isf^{\Sbf}(\dot{\Sbf}_{n}(t))dt&\geq&
\int_0^\infty\liminf_{n\rightarrow
\infty} \Isf^{\Sbf}(\dot{\Sbf}_{n}(t))dt\\
&\geq&\int_0^\infty \Isf^{\Sbf}(\dot{\Sbf}(t))dt,\end{aligned}$$ where the first inequality is due to Fatou’s Lemma and the second one to the lower semicontinuity of $\Isf^{\Sbf}$.
We define the sequence $\JN_{n,\delta}=(\Sbf_{n,\delta},\Pbf_{n,\delta},\Nbf_{n,\delta})$ where $\Sbf_{n,\delta}(t)=\Upsilon_n(\A(t)+\eta t, \D(t)+\eta
t,\Sbf(t)+\eta t)$. Note that we have $\D(t) +\eta t=\Phi(\A(t)+\eta
t,\Sbf(t) +\eta t)$, hence Lemma \[lem:approx\] applies, in particular, we have $\dot{\Sbf}_{n,\delta}(t)\rightarrow
\dot{\Sbf}(t)+\eta(\delta)$ as $n$ tends to infinity.
We have $\JN_{n,\delta}\in \Dcal_\JN$ by construction and the sequence $\{\JN_{n,\delta_n}\}_n$ satisfies (\[eq:S2\]) for some $\delta_n\rightarrow 0$. Moreover, we have thanks to (\[eq:approx2\]) and Lemma \[lem:approx\], $$\begin{aligned}
\left\{\begin{array}{lcl}\A_{n,\delta}&=&\Gamma(\D_{n,\delta},\JN_{n,\delta}),\\
\D_{n,\delta} &=&
\Phi(\A_{n,\delta},\JN_{n,\delta}),\end{array}\right.\Leftrightarrow
(\A_{n,\delta},\D_{n,\delta})=\Psi(\JN_{n,\delta}),\end{aligned}$$ where $\A_{n,\delta}$ and $\D_{n,\delta}$ are the polygonal approximation of $\A(t)+\eta t$ and $\D(t)+\eta t$ with step $1/n$ and $\Psi$ has been defined in Section \[sec:extpsi\].
For $n\rightarrow \infty$ and $\delta\rightarrow 0$, we have $(\A_{n,\delta},\D_{n,\delta})\rightarrow (\A,\D)$, hence we have $\Psi(\JN_{n,\delta})\rightarrow (\A,\D)$, i.e. the sequence $\{\JN_{n,\delta_n}\}_n$ satisfies (\[eq:S3\]).
We now show that (\[eq:S4\]) is also satisfied. We fix $T>0$ and prove first that we have, for $\delta$ sufficiently small, $$\begin{aligned}
\label{eq:approxI}\lefteqn{\int_0^T
\Isf^{\Sbf}(\dot{\Sbf}(t))+\tilde{\Dsf}(\dot{\Pbf}(t)\|R)+\Isf^{\Nbf}(\dot{\Nbf}(t))dt
-er(\delta)T}\\
\nonumber&\leq& \liminf_{n\to \infty} \int_0^T
\Isf^{\Sbf}(\dot{\Sbf}_{n,\delta}(t))+\tilde{\Dsf}(\dot{\Pbf}_{n,\delta}(t)\|R)+\Isf^{\Nbf}(\dot{\Nbf}_{n,\delta}(t))dt\\
\nonumber&\leq& \limsup_{n\to \infty} \int_0^T
\Isf^{\Sbf}(\dot{\Sbf}_{n,\delta}(t))+\tilde{\Dsf}(\dot{\Pbf}_{n,\delta}(t)\|R)+\Isf^{\Nbf}(\dot{\Nbf}_{n,\delta}(t))dt\\
\label{eq:approxII}&\leq& \int_0^T
\Isf^{\Sbf}(\dot{\Sbf}(t))+\tilde{\Dsf}(\dot{\Pbf}(t)\|R)+\Isf^{\Nbf}(\dot{\Nbf}(t))dt
+er(\delta)T,\end{aligned}$$ where $er(\delta)$ tends to zero as $\delta$ tends to zero, from which (\[eq:S4\]) follows by monotonicity.
We first deal with the case of the sequence of processes $\{\Sbf_{n,\delta}\}_n$ (we can restrict ourselves to the one dimensional case). We denote $\Sbf_\delta(t)=\Sbf(t)+\eta(\delta)t$.
We define $\varsigma=\esssup\{\dot{\Sbf}(t),\: t\leq
T\}=\inf\{u,\:Leb[t\leq T,\:\dot{\Sbf}(t)>u ]=0\}$, where $Leb$ is for the Lebesgue measure. Since $\int_0^T
\Isf^\Sbf(\dot{\Sbf}(t))dt<\infty$, $\varsigma$ belongs to the domain of $\Isf^\Sbf$ which is open on the right. Hence we can find $\epsilon>0$ such that $\varsigma+\epsilon$ still belongs to this domain and take $\delta$ such that $\eta(\delta)<\epsilon$. Moreover, since $\Isf^\Sbf$ is convex, it is uniformly continuous on $[0,\varsigma+\epsilon]$. Hence, we can assume that we have $\beta(\alpha)\to 0$ as $\alpha\to 0$ such that, $$\begin{aligned}
\forall x,y\in [0,\varsigma+\epsilon], \:|x-y|<\alpha \Rightarrow
|\Isf^\Sbf(x)-\Isf^\Sbf(y)|\leq \beta(\alpha).\end{aligned}$$
From Lemma \[lem:approx\], we have $$\begin{aligned}
\nonumber\int_0^T \Isf^{\Sbf}(\dot{\Sbf}(t))dt -\beta(\eta(\delta)) T\leq&&\\
\nonumber\lim_{n\rightarrow \infty}\int_0^T
\Isf^{\Sbf}(\dot{\Sbf}_{n,\delta}(t))dt&=& \int_0^T
\Isf^{\Sbf}(\dot{\Sbf}_\delta(t))dt\\
\label{eq:ineqsup}&&\leq \int_0^T \Isf^{\Sbf}(\dot{\Sbf}(t))dt
+\beta(\eta(\delta)) T.\end{aligned}$$ and the converse inequality, $$\begin{aligned}
\liminf_{n\rightarrow \infty}\int_0^\infty
\Isf^{\Sbf}(\dot{\Sbf}_{n,\delta}(t))dt &\geq&\int_0^T
\Isf^{\Sbf}(\dot{\Sbf}_\delta(t))dt\geq \int_0^T
\Isf^{\Sbf}(\dot{\Sbf}(t))dt -\beta(\eta(\delta)) T.\end{aligned}$$ Hence we proved (\[eq:approxI\]) and (\[eq:approxII\]) for $\Isf^\Sbf$.
The same kind of arguments can be repeated for $\Nbf_{n,\delta}$ which is just the polygonal approximation of $t\mapsto\Nbf(t)+\delta
t$. Note that $\{\JN_{n,\delta}\}_n\in\Dcal_\JN^\nat$ implies that $\Nbf^{(i)}_{n,\delta}(t)=0$ for all $i\notin \Scal$. For $i\in
\Scal$, we can use the fact that the domain of $\Isf^{\Nbf^{(i)}}$ is open as previously. In the case of $\Pbf_{n,\delta}$, we can not use the argument on the openness of the domain, but we have $\tilde{\Dsf}(R^{(i)}\|R^{(i)})=0$ and then the convexity of $\tilde{\Dsf}$ directly implies that $\tilde{\Dsf}(\dot{\Pbf}_{n,\delta}^{(i)}\|R^{(i)})\leq
\tilde{\Dsf}(\dot{\Pbf}^{(i)}\|R^{(i)})$, from which we derive an equivalent of (\[eq:ineqsup\]).
Exponential tightness
---------------------
We first recall some definitions. A sequence of random variables $\{X_n\}_n\in (\real^K)^\nat$ is exponentially tight if $$\begin{aligned}
\lim_{M\rightarrow \infty}\limsup_{n\rightarrow \infty} \frac{1}{n}
\log \prob(\|X_n\|>M)= -\infty.\end{aligned}$$ For $\delta>0$ and $T>0$, define the modulus of continuity in $\Dbb(E)$ by $$\begin{aligned}
w'(\Xbf,\delta,T):= \inf_{\{t_i\}}\max_i
\sup_{s,t\in[t_{i-1},t_i)}d(\Xbf(s),\Xbf(t)),\end{aligned}$$ where the infimum is over $\{t_i\}$ satisfying $$\begin{aligned}
0=t_0< t_1<\dots <t_{m-1}<T\leq t_m\end{aligned}$$ and $\min_{1\leq i\leq n}(t_i-t_{i-1})>\delta$.
Theorem 4.1 of [@fenkur] tells us: let $\Tcal_0$ be a dense subset of $\real_+$. Suppose that for each $t\in \Tcal_0$, $\{\Xbf_n(t)\}_n$ is exponentially tight. Then $\{\Xbf_n\}_n$ is exponentially tight in $\Dbb(E)$ if and only if for each $\epsilon>0$ and $T>0$, $$\begin{aligned}
\label{exptight}\lim_{\delta\rightarrow 0}\limsup_{n\rightarrow
\infty}\frac{1}{n}\log \prob(w'(\Xbf_n,\delta,T)>\epsilon)=-\infty.\end{aligned}$$
A sequence of stochastic processes $\{\Xbf_n\}_n$ that is exponentially tight in $\Dbb(E)$ is $C$-exponentially tight if for each $\eta>0$ and $T>0$, $$\begin{aligned}
\label{Cexptight}\limsup_{n\rightarrow\infty} \frac{1}{n}
\log\prob(\sup_{s\leq T} d(\Xbf_n(s),\Xbf_n(s-))\geq \eta)=-\infty.\end{aligned}$$
Then Theorem 4.13 of [@fenkur] gives: an exponentially tight sequence $\{\Xbf_n\}_n$ in $\Dbb(E)$ is $C$-exponentially tight if and only if each rate function $\Isf$ that gives the LDP for a subsequence $\{\Xbf_{n(k)}\}_{n(k)}$, satisfies $\Isf(\x)=\infty$ for each $\x\in \Dbb(E)$ such that $\x\notin \Cbb(E)$.
The stochastic assumptions of Section \[sec:sto\] ensure that the sequence of processes $\{\JN_n\}_n$ satisfies a LDP with good rate function (this implies that the sequence is exponentially tight) giving an infinite mass to discontinuous path. Hence the sequence of processes $\{\JN_n\}_n$ is $C$-exponentially tight.
We have to show that the sequence of processes $\{(\A_n,\D_n)\}_n$ is exponentially tight. The fact of dealing with non-decreasing processes simplifies the definitions. For $\Xbf\in \Dbb(\real_+^K)$ (or $\Dbb(\Mbb^K)$) non-decreasing, $\delta>0$ and $T>0$, we define $w_\delta(\Xbf,T)=\sup_{t\in[0,T]}\|\Xbf(t+\delta)-\Xbf(t)\|$. We have clearly $w'(\Xbf,\delta,T) = w_\delta(\Xbf,T)$ and if $\{\Xbf_n(0)\}_n$ is exponentially tight then (\[exptight\]) implies that $\{\Xbf_n(t)\}_n$ is exponentially tight for each $t>0$. Lemmas \[lem:wPhi\] and \[lem:wGamma\] show that conditions (\[exptight\]) and (\[Cexptight\]) are satisfied for the sequence of processes $\{(\A_n,\D_n)\}_n$. The exponential tightness of $\{(\A_n(0),\D_n(0))\}_n$ is clear since $\A_n(0)=\D_n(0)=0$.
Large deviations results
------------------------
\[prop:sldp\] The sequence of processes $\{(\A_n, \D_n)\}_n$ satisfies a LDP in $\Dbb(\real^K_+)\times \Dbb(\real_+^K)$ with good rate function $\Isf^{\A,\D}$. For $\A,\D$ absolutely continuous and such that $\A(0)=\D(0)=0$ and $\A\geq \D$, $\Isf^{\A,\D}$ is given by $$\begin{aligned}
\label{eq:LDP}\Isf^{\A,\D}(\A, \D)=\int_0^\infty
H(\A(s),\D(s),\dot{\A}(s),\dot{\D}(s))ds,\end{aligned}$$ where $H(A,D,\dot{A},\dot{D}):=\inf_{P,N}
h(A,D,\dot{A},\dot{D},P,N)$, with $h$ given by, $$\begin{aligned}
\lefteqn{h(A,D,\dot{A},\dot{D},P,N):=}\\
&&\sum_{i\in
E(A,D)}\Isf^{\Sbf^{(i)}}(\dot{D}^{(i)})\ind_{\{\dot{D}^{(i)}>\mu^{(i)}\}}+\sum_{i\notin
E(A,D)}\Isf^{\Sbf^{(i)}}(\dot{D}^{(i)}) +\sum_i \dot{D}^{(i)}
\tilde{\Dsf}(P^{(i)}\|R^{(i)})+\Isf^\Nbf(N)\end{aligned}$$ where $E(A,D)=\{i,\: A^{(i)}=D^{(i)}\}$ and with the infimum taken over the set of $(P,N)\in \Mbb^K\times \real_+^K$ such that $$\begin{aligned}
\dot{A} &=& N+P^t\dot{D}.$$ For all other $\A,\D$, we have $\Isf^{\A,\D}(\A,\D)=\infty$.
Thanks to the results of previous sections, conditions of Proposition \[prop:contraction\] are satisfied and we define $$\begin{aligned}
\label{eq:ratetilde}\tilde{\Isf}^{\A,\D}(\A, \D)=\inf
\left\{\lim_{n\rightarrow \infty}\Isf^\JN(\JN_n),\: \{\JN_n\}_n\in
\Scal(\A,\D)\right\},\end{aligned}$$ where we recall that $\Scal(\A,\D)=\cup_\JN \Scal(\JN,\A,\D)$, and $\Scal(\JN,\A,\D)$ is defined in Proposition \[prop:JNapprox\]. We have to show that $\tilde{\Isf}^{\A,\D}=\Isf^{\A,\D}$ given by (\[eq:LDP\]).
We define $$\begin{aligned}
\label{eq:ratetilde}\tilde{I}^{\A,\D}(\A, \D)=\inf
\left\{I^\JN(\JN),\: (\A,\D) =
(\Gamma(\D,\JN),\Phi(\A,\JN))\right\}.\end{aligned}$$ We have to show that $\tilde{I}^{\A,\D}=I^{\A,\D}$ given by (\[eq:LDP\]). Note that $\tilde{I}^{\A,\D}$ is the lower semicontinuous regularization of $\bar{I}^{\A,\D}$ defined as in (\[eq:ratetilde\]) but where the infimum is restricted to piece-wise linear Jackson networks with routing matrix having a derivative of spectral radius less than one. Hence it is sufficient to prove that $I^{\A,\D}=\bar{I}^{\A,\D}$ and to show that $I^{\A,\D}$ is a good rate function.
Consider $\JN\in \Dcal_\JN$ and let $(\A,\D)=\Psi(\JN)$. Let $\tau=\{0=t_0<t_1<\dots \}$ be such that the processes $\A,\D, \Sbf,
\Nbf$ and $\D\circ\Pbf$ have a constant derivative on each $(t_k,t_{k+1})$. Then from $\A=\Gamma(\D,\JN)$, we derive $$\begin{aligned}
\dot{\A}^{(i)}(t) = \dot{\Nbf}^{(i)}(t) +\sum_j \dot{\D}^{(j)}(t)
\dot{\Pbf}^{(j,i)}(\D^{(j)}(t)).\end{aligned}$$ From $\D = \Phi(\A,\Sbf)$, we get the following constraints:
- if $\A^{(i)}(t_k)>\D^{(i)}(t_k)$ or $\A^{(i)}(t_{k+1})>\D^{(i)}(t_{k+1})$, then we have $\dot{\D}^{(i)}(t) = \dot{\Sbf}^{(i)}(t)$ for $t\in (t_k,t_{k+1})$;
- otherwise $\A^{(i)}(t)=\D^{(i)}(t)$ for $t\in (t_k,t_{k+1})$ and we have $\dot{\Sbf}^{(i)}(t)\geq
\dot{\A}^{(i)}(t)=\dot{\D}^{(i)}(t)$ for $t\in (t_k,t_{k+1})$.
Now we can compute $\Isf^\JN(\JN)$ as follows $$\begin{aligned}
\Isf^\JN(\JN) &=&\int_0^\infty\sum_{i\in
E(A,D)}\Isf^{\Sbf^{(i)}}(\dot{\Sbf}^{(i)}(s))+\sum_{i\notin
E(A,D)}\Isf^{\Sbf^{(i)}}(\dot{\D}^{(i)}(s))+\Isf^\Nbf(\dot{\Nbf}(s)) ds\\
&&+\int_0^\infty\sum_j \dot{\D}^{(j)}(s)
\tilde{\Dsf}(\dot{\Pbf}^{(j)}(s)\|R^{(j)})ds\\
&\geq& \int_0^\infty
h(\A(s),\D(s),\dot{\A}(s),\dot{\D}(s),\dot{\Pbf}(s),\dot{\Nbf}(s))ds
\geq \Isf^{\A,\D}(\Psi(\JN)),\end{aligned}$$ since for $i\in E(\A(s),\D(s))$, we have $\Isf^{\Sbf^{(i)}}(\dot{\Sbf}^{(i)}(s))\geq
\Isf^{\Sbf^{(i)}}(\dot{\D}^{(i)}(s))\ind_{\{\dot{\D}^{(i)}(s)>\mu^{(i)}\}}$ because $\dot{\Sbf}^{(i)}(s)\geq \dot{\D}^{(i)}(s)$ and $\Isf^{\Sbf^{(i)}}$ is non-negative, convex with $\mu^{(i)}$ as unique zero. Hence, we have $\tilde{\Isf}^{\A,\D}\geq \Isf^{\A,\D}$.
Consider now $(\A,\D)$ such that $\Isf^{\A,\D}(\A,\D)<\infty$, then we denote by $(\pbf(s),\nbf(s))$ the argument that achieves the minimum in $H(\A(s),\D(s),\dot{\A}(s),\dot{\D}(s))$ for any fixed $s$ (note that $h$ is a good rate function). Let $\Pbf(\D(t))
=\int_0^t \pbf(s) ds$ and $\Nbf(t) = \int_0^t \nbf(s) ds$, note that $\pbf$ and $\nbf$ are measurable since $H$ is a good rate function. We have $\A = \Gamma(\D,\Pbf,\Nbf)$. Now define $\sbf(s)$ as follows:
- if $\A^{(i)}(s) =\D^{(i)}(s)$ then $\sbf^{(i)}(s)=\dot{\D}^{(i)}(s)
\vee \mu^{(i)}$;
- if $\A^{(i)}(s) > \D^{(i)}(s)$ then $\sbf^{(i)}(s)=\dot{\D}^{(i)}(s)$.
We have $\D=\Phi(\A,\Sbf)$ with $\Sbf(t)=\int_0^t \sbf(s) ds$. Hence we have $(\A,\D) = (\Gamma(\D,\JN),\Phi(\A,\JN))$ for $\JN=(\Sbf,\Pbf,\Nbf)$ and $\Isf^\JN(\JN)=\Isf^{\A,\D}(\A,\D)<\infty$ by construction. Hence the sequence $\Scal(\JN,\A,\D)=\{\JN_n\}_n$ is well-defined and we have $\tilde{\Isf}^{\A,\D}(\A,\D)\leq
\lim_{n\rightarrow\infty}\Isf^\JN(\JN_n)=\Isf^\JN(\JN)=\Isf^{\A,\D}(\A,\D)$.
From this proposition, it is quite easy to derive a LDP for the process $\Q_n(t) :=\A_n(t)-\D_n(t)$ counting the number of customers in each queue. Thanks to the contraction principle, we have $$\begin{aligned}
\Isf^\Q(\Q) = \inf\{\Isf^{\A,\D}(\A,\D), \Q=\A-\D \},\end{aligned}$$ which gives directly Theorem \[th:qu\].
Acknowledgements {#acknowledgements .unnumbered}
----------------
I am thankful to Anatolii Puhalskii for insightful comments and for providing me with a copy of [@Puh:jack]
Appendix
========
Properties of the map $\Gamma$ and $\Phi$
-----------------------------------------
For $\Xbf\in \Dbb(\real_+^K)$, $\delta>0$ and $T>0$, we define $w_\delta(\Xbf,T)=\sup_{t\in[0,T]}\|\Xbf(t+\delta)-\Xbf(t)\|$.
\[lem:wPhi\] We have $$\begin{aligned}
w_\delta(\Phi(\Xbf,\Ybf),T)\leq w_\delta(\Ybf,T).\end{aligned}$$
It is clearly sufficient to consider the case $K=1$. We will prove that $$\begin{aligned}
\label{ineq:lem}\Phi(\Xbf,\Ybf)(t+\delta)-\Phi(\Xbf,\Ybf)(t) \leq
\Ybf(t+\delta)-\Ybf(t),\end{aligned}$$ from which the lemma follows. If $\Phi(\Xbf, \Ybf)(t)=\Ybf(t)$, then we have $\Phi(\Xbf, \Ybf)(t+\delta) \leq \Ybf(t+\delta)$ and (\[ineq:lem\]) is clear.
Assume now that $\Phi(\Xbf, \Ybf)(t) =\inf_{0\leq
s<t}\left\{\Ybf(t)-\Ybf(s)+\Xbf(s)\right\}<\Ybf(t)$. We have $$\begin{aligned}
\Ybf(t+\delta)-\Ybf(s)+\Xbf(s)=
\Ybf(t)-\Ybf(s)+\Xbf(s)+\Ybf(t+\delta)-\Ybf(t),\end{aligned}$$ and (\[ineq:lem\]) follows by taking the minimum in $s\in[0,t]$ and observing that $\Phi(\Xbf, \Ybf)(t+\delta)\leq \inf_{0\leq
s<t}\left\{\Ybf(t+\delta)-\Ybf(s)+\Xbf(s)\right\}$.
The following lemma is clear:
\[lem:wGamma\] We have $$\begin{aligned}
w_\delta(\Gamma(\Xbf,\Pbf,\Nbf),T)\leq
w_\delta(\Nbf,T)+w_\delta(\Pbf,\|\Xbf(T)\|).\end{aligned}$$
\[lem:propPhi\] Assume $\Sbf\in \Dbb_0(\real_+)$ is absolutely continuous, then for any $\A\in \Dbb(\real_+)$, we have $\D:=\Phi(\A, \Sbf)$ is absolutely continuous and,
- for all $t$ such that $\A(t)>\D(t)$, we have $\dot{\D}(t)=\dot{\Sbf}(t)$;
- if $\A(t)=\D(t)$ for $t\in (u,v)$ with $u<v$, then we have $\dot{\Sbf}(t)\geq
\dot{\A}(t)=\dot{\D}(t)$ for $t\in (u,v)$.
It follows directly form (\[ineq:lem\]) that if $\Sbf$ is absolutely continuous, then $\Phi(\Xbf,\Sbf)$ is absolutely continuous for any $\Xbf$. The rest of the lemma is obvious.
Auxiliary results
-----------------
\[lem:specrad\] Given a substochastic matrix $R$ such that $\rho(R)<1$ and a substochastic matrix $P$ such that the support of $P$ is included in the support of $R$, i.e. $R^{(i,j)}=0\Rightarrow P^{(i,j)}=0$. Then for any $\epsilon$ such that $0< \epsilon^{(i)}\leq 1$ for all $i$, the matrix with coefficients $M^{(i,j)}=(1-\epsilon^{(i)})
P^{(i,j)}+\epsilon^{(i)} R^{(i,j)}$ is of spectral radius less than $1$.
By a suitable permutation of rows and columns, we can assume that $R$ is given in its canonical form $$\begin{aligned}
\label{eq:canonic}R =\left(\begin{array}{ccccc}
S_1(R)&*&*&*\\
0&S_2(R)&*&*\\
0&0&\ddots&*\\
0&0&0&S_n(R)\end{array} \right),\end{aligned}$$ where each $S_i(R)$ is an irreducible matrix. We have $\rho(R)<1$ if and only if each $S_i(R)$ is not a stochastic matrix.
In view of the assumption on the support of $P$, the matrix $P$ has the same structure as (\[eq:canonic\]) and we have with the same notation as above, $S_i(M)$ which is an irreducible and not stochastic matrix.
Consider a continuous Jackson network $\JN=(\Sbf,\Pbf,\Nbf)$, i.e. such that $$\begin{aligned}
\Nbf\in \Cbb(\real_+^K),\quad \Sbf\in \Cbb_0(\real_+^K), \quad
\Pbf\in \Cbb_0(\Mbb^K).\end{aligned}$$ We assume that the functions $\Sbf$, $\Pbf$ and $\Nbf$ are absolutely continuous with a right-derivative at each $t\geq 0$ such that $\dot{\Sbf}(t)\geq 0$, $\dot{\Nbf}(t)\geq 0$ and $\rho(\dot{\Pbf}(t))<1$ for all $t$.
Then the fixed point equation (\[fixpointeq\]) restricted to functions with a right-derivative at each $t\geq 0$ has an unique solution, that we still denote $\Psi(\JN)$. Moreover for any sequence of Jackson networks $\JN_n$ that converges to $\JN$, where $\JN$ satisfies previous assumptions, we have $\Psi(\JN_n)\rightarrow \Psi(\JN)$.
This proposition extends Proposition 3.3 of [@lel].
First note that thanks to Lemmas \[lem:wPhi\] and \[lem:wGamma\], any solution of (\[fixpointeq\]) is absolutely continuous.
For any continuous Jackson network $\JN$, we can find a sequence of piece-wise linear Jackson networks $\JN_n$ that tends u.o.c. to $\JN$.
For any fixed $n\geq 1$, we define the sequences of processes $\{\A^t_n[k],\D^t_n[k]\}_{k\geq 0}$ and $\{\A^b_n[k],\D^b_n[k]\}_{k\geq 0}$ with the same recurrence equation: $$\begin{aligned}
\left\{ \begin{array}{l}
\A_n[k+1] = \Gamma(\D_n[k],\JN_n),\\
\D_n[k+1] = \Phi(\A_n[k+1],\JN_n),
\end{array}\right.\end{aligned}$$ but with different initial conditions $\D^t_n[0] = \Sbf_n$ and $\D^b_n[0] = \0$.\
We will show that the following diagram is commutative: $$\begin{aligned}
\begin{array}{ccc}
(a)\left\{\begin{array}{l}
\A_n^\bullet[k+1] = \Gamma(\D^\bullet_n[k],\JN_n),\\
\D_n^\bullet[k+1] = \Phi(\A^\bullet_n[k+1],\JN_n).
\end{array}\right.
&\xrightarrow[(1)]{k\rightarrow \infty}& (b)\left\{\begin{array}{l}
\A_n = \Gamma(\D_n,\JN_n),\\
\D_n = \Phi(\A_n,\JN_n).
\end{array}\right.\\
\\
(2)\Big{\downarrow} n\rightarrow\infty&&(3)\Big{\downarrow} n\rightarrow\infty\\
\\
(c)\left\{\begin{array}{l}
\A^\bullet[k+1] = \Gamma(\D^\bullet[k],\JN),\\
\D^\bullet[k+1] = \Phi(\A^\bullet[k+1],\JN).
\end{array}\right.&\xrightarrow[(4)]{k\rightarrow \infty}&
(d)\left\{\begin{array}{l}
\A = \Gamma(\D,\JN),\\
\D = \Phi(\A,\JN).
\end{array}\right.
\end{array}\end{aligned}$$ We denote by $.^\bullet$ (for $.^t$ or $.^b$) the dependence in the initial conditions ($t$ for top and $b$ for bottom). Limit (1) and the fact that the limit does not depend on the initial conditions are direct consequences of Proposition \[prop:JNlinear\]. Limit (2) is due to continuity properties for $\Gamma$ and $\Phi$.\
We now prove that limit (4) holds thanks to monotonicity properties.
We have $$\begin{aligned}
\D_n^b[0] \leq \D_n \leq \D_n^t[0],\end{aligned}$$ hence by monotonicity of $\Gamma$, we have $$\begin{aligned}
\A_n^b[1] \leq \A_n \leq \A_n^t[1],\end{aligned}$$ and then by monotonicity of $\Phi$, $$\begin{aligned}
\D_n^b[0] \leq\D_n^b[1] \leq \D_n \leq \D_n^t[1]\leq \D_n^t[0].\end{aligned}$$ Hence, from the monotonicity of the mappings $\Gamma$ and $\Phi$ in their first argument, we derive: $$\begin{aligned}
\A_n^b[k]\leq \A_n^b[k+1] \leq &\A_n& \leq \A_n^t[k+1] \leq \A_n^t[k],\\
\D_n^b[k]\leq \D_n^b[k+1] \leq &\D_n& \leq \D_n^t[k+1] \leq
\D_n^t[k].\end{aligned}$$ Hence making $n\rightarrow \infty$, we have $$\begin{aligned}
\A^b[k]\leq \A^b[k+1] \leq &\liminf_n\A_n\leq \limsup_n \A_n& \leq \A^t[k+1] \leq \A^t[k],\\
\D^b[k]\leq \D^b[k+1] \leq &\liminf_n\D_n\leq \limsup_n\D_n& \leq
\D^t[k+1] \leq \D^t[k].\end{aligned}$$ Now making $k\rightarrow \infty$, we have (by monotone convergence) $$\begin{aligned}
\lim_{k\rightarrow \infty}\A^b[k]=\A^b \leq &\underline{\A}\leq \overline{\A}& \leq \A^t=\lim_{k\rightarrow\infty}\A^t[k],\\
\lim_{k\rightarrow\infty}\D^b[k]= \D^b \leq &\underline{\D}\leq
\overline{\D}& \leq \D^t=\lim_{k\rightarrow\infty} \D^t[k].\end{aligned}$$ Moreover, $(\A^b,\D^b)$ and $(\A^t,\D^t)$ are respectively the minimal and maximal solution of the fixed point equation (\[fixpointeq\]). Hence to finish the proof, we have to prove that $\A^t=\A^b$ and $\D^t=\D^b$. Let $u=\inf\{s,\: \A^t(s)>\A^b(s)\}$. If $u>0$, we have $\A^t(s)=\A^b(s):=\A^{(i)}(s)$ (and hence $\D^t(s)=\D^b(s):=\D^{(i)}(s)$) for all $0\leq s\leq u$. Let $E:=\{i,\: \D^{(i),t}(u)=\A^{(i),t}(u)\}=\{i,\:
\D^{(i),b}(u)=\A^{(i),b}(u)\}$ (the set of empty queues). Note that for $i\notin E$, we have (for both top and bottom processes) $\D^{(i),\bullet}(u) < \A^{(i),\bullet}(u)$. Since we are dealing with continuous functions, there exists $\epsilon>0$ such that for $u\leq t\leq u+\epsilon$, we have $\D^{(i),\bullet}(t)
<\A^{(i),\bullet}(u)\leq \A^{(i),\bullet}(t)$. For all $u\leq s\leq
t\leq u+\epsilon$, we have $$\begin{aligned}
\Sbf^{(i)}(t)-\Sbf^{(i)}(s)+\A^{(i),\bullet}(s)\geq
\A^{(i),\bullet}(u)>\D^{(i),\bullet}(t),\end{aligned}$$ hence for $u\leq t\leq u+\epsilon$, we have $$\begin{aligned}
\D^{(i),\bullet}(t) = \inf_{0\leq s\leq
u}\left\{\Sbf^{(i)}(t)-\Sbf^{(i)}(s)+\A^{(i)}(s) \right\}\wedge
\Sbf^{(i)}(t).\end{aligned}$$
Hence for all $i\notin E$, we have $\D^{(i),t}(t)=\D^{(i),b}(t):=\D^{(i)}(t)$ for $t\leq u+\epsilon$. Moreover note that we have $\dot{\D}^{(i)}(t) = \dot{\Sbf}^{(i)}(t)$ for $t\leq u+\epsilon$. Then we have $$\begin{aligned}
\A^{(i),\bullet}(t) = \underbrace{\Nbf^{(i)}(t) +\sum_{j\notin E}
\Pbf^{(j,i)}(\D^{(j)}(t))}_{\Lambda^{(i)}(t)} +\sum_{j\in E}
\Pbf^{(j,i)}(\D^{(j),\bullet}(t)).\end{aligned}$$
$$\begin{aligned}
\left\{\begin{array}{lcl} a^{(i)}(u)&=&
\dot{\Lambda}^{(i)}(u)+\sum_{j\in
E}d^{(j)}(u)\dot{\Pbf}^{(j,i)}(\A^{(i)}(u)),\\
d^{(i)}(u) &=& a^{(i)}(u)\wedge
\dot{\Sbf}^{(i)}(u).\end{array}\right.\end{aligned}$$
Denote $\overline{a}^{(i),\bullet}(u) =
\limsup_{x\rightarrow
0}\frac{\A^{(i),\bullet}(u+x)-\A^{(i),\bullet}(u)}{x}$ and $\overline{d}^{(i),\bullet}(u)=\limsup_{x\rightarrow
0}\frac{\D^{(i),\bullet}(u+x)-\D^{(i),\bullet}(u)}{x}$. We have $$\begin{aligned}
\overline{a}^{(i),\bullet}(u)&=& \dot{\Lambda}^{(i)}(u)+\sum_{j\in
E}\overline{d}^{(j),\bullet}(u)\dot{\Pbf}^{(j,i)}(\A^{(i)}(u)).\end{aligned}$$
Hence we have $E\neq \emptyset$ and there exists $i\in E$ such that $\D^{(i),t}(u+\epsilon)>\D^{(i),b}(u+\epsilon)$ for all $\epsilon>0$ sufficiently small. Let $F^\bullet$ be the set of indices such that $\D^{(i),\bullet}(t)=\A^{(i),\bullet}(t)$ for all $i\in E$ and for all $t<u+\epsilon$. Then we have $$\begin{aligned}
\A^{(i),\bullet}(t) = \Nbf^{(i)}(t) +\sum_{j\notin E}
\Pbf^{(j,i)}(\D^{(j),\bullet}(t)) +\sum_{j\in F^\bullet}
\Pbf^{(j,i)}(\A^{(j),\bullet}(t))+\sum_{j\in E\backslash F^\bullet}
\Pbf^{(j,i)}(\D^{(j),\bullet}(t)),\end{aligned}$$ taking the derivative determines uniquely $\A^{\bullet}$.
For $i\in E\backslash F^\bullet$, we have for $t< u+\epsilon$, $$\begin{aligned}
\D^{(i),\bullet}(t)=\Sbf^{(i)}(t-u+\Sbf^{(i)\leftarrow}(\A^{(i),\bullet}(u))).\end{aligned}$$ Since $\A^{(i),t}(u)=\A^{(i),b}(u)$, we have $\D^{(i),t}(t)=\D^{(i),b}(t)$. This contradicts the fact that $u>0$ but if $u=0$ previous proof applies with $E=[1,K]$. This establishes the existence of $\Psi$.
In order to prove that $\Psi$ is continuous, we see that previous proof applies exactly in the same way but now the fact limit (1) of the diagram exists and does not depend on the initial conditions is due to the first part of the property.
An example
----------
In this section, we construct 2 different sequences of Jackson networks $\JN_n^1$ and $\JN_n^2$ such that their fluid limits are the same $$\begin{aligned}
\JN_n^1\rightarrow \JN \quad \mbox{and}\quad \JN_n^2\rightarrow \JN,\end{aligned}$$ but such that $$\begin{aligned}
(\A_n^1,\D_n^1)=\Psi(\JN_n^1)&\rightarrow& (\A^1,\D^1),\\
(\A_n^2,\D_n^2)=\Psi(\JN_n^2)&\rightarrow& (\A^2,\D^2),\end{aligned}$$ with $(\A^1,\D^1)\not= (\A^2,\D^2)$.
We consider a toy example with only one station (hence we omit the superscript $.^{(1)}$ that refers to that only station). Once a customer is served, he can either go out of the network or go back to this same node. We define the following driving sequences: $$\begin{aligned}
T^{n} &=&
(\underbrace{1,\dots,1}_{n},n,\underbrace{1,\dots,1}_{n},n,\dots),\\
\sigma^{n}&=& \alpha(1,1,\dots),\end{aligned}$$ with $\alpha<1$. We define now two different routing sequences $$\begin{aligned}
\nu^{n} &=&
(\underbrace{2,\dots,2}_{n+1},\underbrace{1,\dots,1}_{n+1},\dots),\\
\nu^{n}(x) &=& (\underbrace{2,\dots,2}_{\lfloor
xn\rfloor},1,\underbrace{2,\dots,2}_{n-\lfloor xn
\rfloor},\underbrace{1,\dots,1}_{\lfloor
xn\rfloor},2,\underbrace{1,\dots,1}_{n-\lfloor xn\rfloor}, \dots),\end{aligned}$$ where $x<1$. We denote by $\JN_n^1=\left\{ \sigma^n,\nu^n,T^n
\right\}$ and $\JN_n^2=\left\{ \sigma^n,\nu^n(x),T^n \right\}$. $\nu^n(x)$ is obtained from $\nu^n$ by only interchanging a 1 and a 2. Hence we have $$\begin{aligned}
\JN_n^1\rightarrow \JN \quad \mbox{and}\quad \JN_n^2\rightarrow \JN.\end{aligned}$$ Indeed the fluid network $\JN$ is given on Figure \[fig:JN\].
(0,120) (-200,0)
\
In the fluid limit, in case 1, the queue is always empty and the departure process is the same as the arrival process $\Nbf$. In case 2, the fluid limit of the departure process and the queue length process is given on Figure \[fig:dep\].
(0,140) (-100,0)
\
To explain $\D^2$, we write for each arrival (number on the left) the couple corresponding to:\
the inter-arrival time $|$ the routing decision ($1$ means that the customer goes back in the queue and $2$ means that the customer leaves the network): $$\begin{aligned}
\begin{array}{ccccc}
1&\rightarrow&1&|&2\\
2&\rightarrow&1&|&2\\
3&\rightarrow&1&|&2\\
&\vdots&&&\\
\lfloor xn\rfloor&\rightarrow&1&|&2\\
\lfloor xn\rfloor+1&\rightarrow&1&|&1,2\\
\lfloor xn\rfloor+2&\rightarrow&1&|&2\\
&\vdots&&&\\
n&\rightarrow&1&|&2\\
n+1&\rightarrow&n&|&\underbrace{1,\dots,1}_{\lfloor
xn\rfloor},2\\
n+2&\rightarrow&1&|&\underbrace{1,\dots,1}_{n-\lfloor
xn\rfloor},2\\
n+3&\rightarrow&1&|&2\\
&\vdots&&&\iffalse\\
n+\lfloor xn\rfloor+1&\rightarrow&1&|&2\\
n+\lfloor xn\rfloor+2&\rightarrow&1&|&1,2\\
n+\lfloor xn\rfloor+3&\rightarrow&1&|&2\\
&\vdots&&&\\
2n+1&\rightarrow&1&|&2\\
2(n+1)&\rightarrow&n&|&\underbrace{1,\dots,1}_{\lfloor
xn\rfloor},2\\
&\vdots&&& \fi
\end{array}\end{aligned}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove existence and uniqueness results for (mild) solutions to some non-linear parabolic evolution equations with a rough forcing term. Our method of proof relies on a careful exploitation of the interplay between the spatial and time regularity of the solution by capitialising some of Kato’s ideas in semigroup theory. Classical Young integration theory is then shown to provide a means of interpreting the equation. As an application we consider the three dimensional Navier-Stokes system with a stochastic forcing term arising from a fractional Brownian motion with $h>1/2$.'
author:
- 'Thomas Cass, Zhongmin Qian, Jan Tudor'
title: 'Non-linear evolution equations driven by rough paths'
---
Introduction
============
In this paper we study the initial-value problem of the following non-linear evolution equation of parabolic type $$\frac{d}{dt}u+Au+\tilde{Q}(u)=F(u)\dot{w} \label{ev-01}$$ in a separable Hilbert space $X$ with an initial value $u_{0}\in X$, where $%
-A$ is the infinitesimal generator of an analytic semigroup $\left\{
P_{t}\right\} _{t\geq0}$, $\tilde{Q}:D(\tilde{Q})\subset X\rightarrow X$ is a non-linear operator, and $F:X\rightarrow L_{2}(Z,X)$ satisfies a Lipschitz type condition which we specify later. The driving path for the perturbation $w=(w_{t})_{t\geq0}$ is an $\alpha$-Hölder continuous path in a separable Hilbert space $Z$, where $\alpha\in(\frac{1}{2},1]$. The class of paths includes sample paths of fractional Brownian Motion with Hurst parameter $h>\frac{1}{2}$.
Our motivation for studying equations of this form is the three dimensional Navier-Stokes system
$$\begin{aligned}
\frac{\partial }{\partial t}u+u.\nabla u &=&\Delta u-\nabla p+F(u)\dot{w} \\
\nabla \cdot u &=&0\end{aligned}$$
in a bounded domain $\Omega $ with compact, smooth boundary $\Gamma $. The force term $F(u)\dot{w}$ is modeled by a fractional Brownian Motion in the Hilbert space $Z$, and $u_{t}$ describes the velocity of the fluid flow under the influence of a stochastic force $F(u)\dot{w}$ and subject to the no-slip condition. By projecting to the $L^{2}$-space of solenoidal vector fields on $\Omega $ via the Leray-Hopf projection $P_{\infty }$, the above equation can be written as the following evolution equation $$\frac{d}{dt}u+Au+P_{\infty }(u.\nabla u)=P_{\infty }F(u)\dot{w}
\label{eq:NS-EV}$$where $A=-P_{\infty }\circ \Delta $ is the Stokes operator, which is a self adjoint operator.
If we let $\left\{ P_{t}\right\} _{t\geq 0}$ be the semigroup generated by $%
-A$, $h(s)=P_{t-s}u(s)$ and we consider the formal calculation $$\begin{aligned}
F^{\prime }(s) &=&(AP_{t-s})u(s)+P_{t-s}u^{\prime }(s) \\
&=&(AP_{t-s})u(s)-(P_{t-s}A)u(s) \\
&&-P_{t-s}\tilde{Q}(u(s))+P_{t-s}\left( F(u(s))\dot{w}_{s}\right) \\
&=&-P_{t-s}\tilde{Q}(u(s))+P_{t-s}F(u(s))\dot{w}_{s}\text{,}\end{aligned}$$then integrating from $0$ to $t$ one obtains $$u(t)=P_{t}u_{0}-\int_{0}^{t}P_{t-s}\tilde{Q}(u(s))ds+%
\int_{0}^{t}P_{t-s}F(u(s))\dot{w}_{s}ds\text{.} \label{mi-1}$$The first integral on the right-hand side is considered as the Bochner integral, the second one needs to be interpreted as some sort of stochastic integration. Hence the previous has to be written as $$u(t)=P_{t}u_{0}-\int_{0}^{t}P_{t-s}\tilde{Q}(u(s))ds+%
\int_{0}^{t}P_{t-s}F(u(s))dw_{s}$$so that if $u$ satisfies the above, we call it a mild solution to ([ev-01]{}). For Navier-Stokes equations in the classical case that $w$ is differentiable Kato made the following observation (see [@fujita]) which proves significant. If $t-s>0$ and $x\in X$, then $P_{t-s}x$ belongs to the domain $D(A^{\tau })$ for any real $\tau $. On the other hand, if $\tau >0$, $A^{-\tau }$ is bounded, the integral involving the non-linear terms of ([mi-1]{}) can be rewritten to enhance the regularity. For example $$\int_{0}^{t}P_{t-s}\tilde{Q}(u(s))ds=\int_{0}^{t}A^{\tau }P_{t-s}A^{-\tau }%
\tilde{Q}(u(s))ds\text{.}$$As long as $A^{-\tau }\tilde{Q}(u(s))$ is bounded on $[0,t]$, then, since $%
||A^{\tau }P_{t-s}||\leq C(t-s)^{-\tau }$ which is still integrable on $%
[0,t] $ for any $\tau <1$. Therefore, we may consider the non-linear operator $Q=A^{-\tau }\tilde{Q}$ instead of $\tilde{Q}$, where $%
Q(x)=A^{-\tau }\tilde{Q}(x)$ for $x\in D(Q)$, but in general the domain of definition $D(Q)$ can be extended to be a little bit larger than $D(\tilde{Q}%
)$ due to the fact that $A^{-\tau }$ is a bounded linear operator. Essentially the same idea applies to the stochastic case, although the difficulty is in the stochastic term.
Now, according to Sobolevski [@sobolev], we have the estimate $$||A^{-\frac{1}{4}}P_{\infty }(w.\nabla u)||\leq C||\sqrt{A}u||||\sqrt{A}w||
\label{eq:sobiq}$$where $||\cdot ||$ is the $L^{2}$-norm and $A$ is the Stokes operator (with Dirichlet condition) in a bounded domain with smooth boundary $\Gamma $. Hence as we are motivated by the stochastic version of this situation, we find fixed points in a certain space of the map $$\mathbb{L}u(t)=P_{t}u_{0}-\int_{0}^{t}A^{\tau
}P_{t-s}Q(u(s))ds+\int_{0}^{t}P_{t-s}F(u(s))dw_{s}$$where $w$ is a fractional Brownian Motion.
Several existing papers study evolution equations driven by a fractional Brownian Motion (fBM): we summarize some of the main results here. In [maslowski]{} the authors prove path-wise existence and uniqueness of mild solutions to equations of the form (\[ev-01\]) where the perturbation operator $\tilde{Q}:X\rightarrow X$ is taken to be continuous with linear growth. Additionally they assume the stochastic force term $F$ satisfies $%
F:X\rightarrow L\left( X\right) $ and the composition $P_{t}F\left( \cdot
\right) $ has linear growth and is Lipschitz with a constant proportional to $t^{\gamma _{1}},$ $t^{-\gamma _{2}}$ respectively for $\gamma _{1},\gamma
_{2}\in \left[ 0,1\right) $. The solutions are obtained from fractional calculus methods for fBM with Hurst parameter $h>\frac{1}{2}$ and are elements of the space of $X$ valued paths such that the norm $$\left\Vert u\right\Vert :=\sup_{t\in \left[ 0,T\right] }\left\Vert u\left(
t\right) \right\Vert _{X}+\sup_{t\in \left[ 0,T\right] }\int_{0}^{t}\frac{%
\left\Vert u\left( t\right) -u\left( s\right) \right\Vert _{X}}{\left(
t-s\right) ^{1+\alpha }}ds,$$$\alpha \in \left( 1-h,\frac{1}{2}\right) ,$ is finite. Both of the cases $h>%
\frac{1}{2}$ and $h<\frac{1}{2}$ are treated in [@tindel] where necessary and sufficient conditions are obtained for existence of solutions to stochastically forced linear evolution equations with linear noise, i.e. of the form $$\frac{d}{dt}u+Au=F\dot{w} \label{lineqn}$$where $F$ does not depend on $u$. Here the authors do not solve path-wise but instead use Skorohod type integration to produce a solution which is a square integrable (in the stochastic sense) $X$ valued process and then show space regularity of the solution as a continuous map into the domain of $-A$. More recently, in [@gubinelli], the equation (\[lineqn\]) with nonlinear $F$ is studied (i.e. $F:X\rightarrow L_{2}\left( Z,X\right) $ with some Lipschitz type conditions) and path-wise mild solutions are obtained using Young integration. In this paper we treat a non-linear evolution equation forced by non-linear noise. The ideas used are straight forward while the difficulty lies in the delicate interplay between the space and time regularity of the paths involved.
The paper is organized as follows. In the second section, we set out the conditions on the operators involved and define the space in which we seek a mild solution. In section 3, we make some initial estimates which follow from our conditions and are used in the fixed point proof; in the fourth and fifth sections we discuss the Bochner and Young integral terms of our fixed point map. Subsequently we prove the existence of a mild solution and finally discuss the application to the Navier-Stokes equations.
Preliminaries
=============
We would like to set up the technical conditions on various terms appearing in the previous evolution equation. The model example is randomly enforced Navier-Stokes equation in two or three dimensions, in which $A=-P_{\infty
}\circ \Delta $ (or $A=-P_{\infty }\circ \Delta +I$ where $I$ is the identity operator) is the Stokes operator on a bounded domain $\Omega $ together with Dirichlet boundary conditions, and $\tilde{Q}(u)=P_{\infty
}\left( u.\nabla u\right) $. Therefore our assumptions will be motivated with the aim of applying the abstract setting to this case.
With this in mind, we assume $A$ is a positive-definite, self-adjoint operator on $X$, with positive spectral gap, so that the spectrum of $A$ lies in the half line $[\lambda _{0},\infty )$. Its domain is denoted by $%
D(A)$. For every real $\epsilon $, $A^{\epsilon }$ is again a self-adjoint operator, and $A^{\epsilon }$ is bounded if $\epsilon <0$. The domain $%
D(A^{\epsilon })$ is decreasing, so that $D(A)\subset D(A^{\epsilon })$ for $%
\epsilon \in \lbrack 0,1]$ and $D(A^{\epsilon })$ is a Hilbert space under the norm $||A^{\epsilon }x||$.
Let $P_{t}=e^{-tA}$ be the $C_{0}$-semigroup of contractions on $X$ generated by $-A$. We will frequently use the following facts: for any $t>0$ and $x\in X$, $P_{t}x\in D\left(A^{\epsilon}\right)$ for any $\epsilon\leq1$ and $||A^{\epsilon}P_{t}||\leq\frac{C}{t^{\epsilon}}$ for some constant $C$ depending on $\epsilon$.
For every $t>0$ and $x\in D(A^{\epsilon })$ for some $\epsilon \in (0,1]$, we have $$\begin{aligned}
x-P_{t}x &=&\int_{0}^{t}AP_{s}xds \\
&=&\int_{0}^{t}A^{1-\epsilon }P_{s}A^{\epsilon }xds\end{aligned}$$so that $$\begin{aligned}
||x-P_{t}x|| &\leq &||A^{\epsilon }x||\int_{0}^{t}||A^{1-\epsilon }P_{s}||ds
\notag \\
&\leq &C(\epsilon )||A^{\epsilon }x||\int_{0}^{t}s^{-(1-\epsilon )}ds \notag
\\
&=&\frac{C(\epsilon )}{\epsilon }t^{\epsilon }||A^{\epsilon }x||
\label{regularity}\end{aligned}$$an estimate which will be needed in order to derive useful a priori estimates. Now we are in a position to formulate the technical conditions on the non-linear term $Q$ as follows.
\[Qcond\]Let $\tau ,\delta \in \lbrack 0,1)$, $\delta \leq \alpha \in (%
\frac{1}{2},1]$ be three parameters such that $\delta +\tau <1$, $\alpha
+\tau <1$, $\alpha +\delta >1$, and let $K_{j}:[0,\infty )\rightarrow
\lbrack 0,\infty )$ ($j=0,1,2$) be three increasing functions.
1. The domain of definition of $Q$, $D(Q)=D(A^{\delta })$.
2. $Q$ is bounded on $D(A^{\delta })$, and $||Q(x)||\leq
K_{2}(||A^{\delta }x||)$.
3. $Q$ is weakly Gateaux differentiable on $D(A^{\delta })$: if $x\in
D(A^{\delta })$, then there is a linear operator $DQ(x):D(A^{\delta
})\rightarrow X$, such that $\varepsilon \rightarrow \langle Q(x+\varepsilon
\xi ),z\rangle $ is differentiable for any $\xi \in D(A^{\delta })$, $z\in
X^{\ast }$, $$\langle DQ(x)\xi ,z\rangle =\left. \frac{d}{d\varepsilon }\right\vert
_{\varepsilon =0}\langle H(x+\varepsilon \xi ),z\rangle$$and $||DQ(x)\xi ||\leq K_{1}(||A^{\delta }x||)||A^{\delta }\xi ||$. That is, $DQ(x)$ is a bounded linear operator in $D(A^{\delta })$.
If $A=-P_{\infty}\circ\Delta$ is the Stokes operator with Dirichlet boundary condition on a bounded domain with smooth boundary $\Gamma$, and $Q(u)=A^{-%
\frac{1}{4}}P_{\infty}(u.\nabla u)$, on the Hilbert space $K_{2}(\Omega)$, then we can choose $\delta=\frac{1}{2}$, $\tau=\frac{1}{4}$.
Now let us set out the conditions on the non-linear operator $%
F:X\rightarrow L_{2}(Z,X)$.
\[Fcond\]We assume the following for $\varepsilon =\max \left\{ \alpha
+\delta ,2\alpha \right\} $.
1. For every $x\in X,$ $\xi \in Z$ we have $F\left( x\right) \xi \in
D\left( A^{\varepsilon }\right) $ and $A^{\varepsilon }F$ is globally Lipschitz in the sense that $$||A^{\varepsilon }\left[ F(x)-F\left( y\right) \right] ||_{L_{2}\left(
Z,X\right) }\leq C||x-y||_{X}\mbox{ }\forall x,y\in X. \label{eq:cond1}$$
2. The composition $A^{\delta +\varepsilon }F$ is globally relatively Lipschitz in the sense that $$||A^{\delta +\varepsilon }\left[ F(x)-F\left( y\right) \right]
||_{L_{2}\left( Z,X\right) }\leq C||A^{\delta }\left[ x-y\right] ||_{X}\text{
}\forall x,y\in D\left( A^{\delta }\right) \text{.} \label{eq:cond2}$$
3. Additionally, we assume that $A^{\varepsilon }F:X\rightarrow
L_{2}\left( Z,X\right) $ is weakly Gateaux differentiable, i.e. $\tau
\rightarrow \left\langle \left\langle A^{\varepsilon }F\left( x+\tau
y\right) ,\xi \right\rangle \right\rangle $ is differentiable and there exists a linear operator $DA^{\varepsilon }F\left( x\right) :X\rightarrow
L_{2}\left( Z,X\right) $ such that $\left. \frac{d}{d\tau }\left\langle
\left\langle A^{\varepsilon }F\left( x+\tau y\right) ,\xi \right\rangle
\right\rangle \right\vert _{\tau =0}=\left\langle \left\langle
DA^{\varepsilon }F\left( x\right) y,\xi \right\rangle \right\rangle $ for all $\xi \in L_{2}\left( Z,X\right) ^{\prime }$, and satisfies the bounds $$\sup_{x}\left\Vert DA^{\varepsilon }F\left( x\right) \right\Vert _{L\left(
X,L_{2}\left( Z,X\right) \right) }<\infty$$and $$\left\Vert DA^{\varepsilon }F\left( x\right) -DA^{\varepsilon }F\left(
y\right) \right\Vert _{L\left( X,L_{2}\left( Z,X\right) \right) }\leq
C\left\Vert x-y\right\Vert .$$
Throughout this paper, the non-negative constants $C_{j}$ may be different from place to place, and $C_{j}$ may depend on only the parameters $\tau $, $%
\delta $, $\alpha $ and the operator $A$, but are independent of $u$, $w$, and $T$.
Initial Estimates
=================
The main goal of this section is to derive technical estimates which are used to show the existence and uniqueness of solutions to the evolution equation (\[ev-01\]). Let us define $\mathbb{H}_{T}$ to be the collection of all $\alpha $-Hölder continuous paths $u=(u_{t})$ in $X$ such that for each $t\in \lbrack 0,T]$, $u_{t}\in D(A^{\delta })$ and $t\rightarrow
A^{\delta }u_{t}$, $(s,t)\rightarrow \frac{u_{t}-u_{s}}{|t-s|^{\alpha }}$ are bounded. If $u\in \mathbb{H}_{T}$ we define the norm $$||u||_{\mathbb{H}_{T}}=\sup_{t\in \lbrack 0,T]}||A^{\delta }u_{t}||+\sup
_{\substack{ s,t\in \lbrack 0,T] \\ s\neq t}}\frac{||u_{t}-u_{s}||}{%
|t-s|^{\alpha }}\text{.} \label{c1-a}$$By definition, if $||u||_{\mathbb{H}_{T}}\leq \beta $ then $$||A^{\delta }u_{t}||\leq \beta \text{, \ \ }||u_{t}-u_{s}||\leq \beta
\left\vert t-s\right\vert ^{\alpha }$$for any $s,t\in \lbrack 0,T]$. And since $||x||\leq C_{\delta }||A^{\delta
}x||$ we also have $||u_{t}||\leq C_{\delta }\beta $. It is also obvious that $(\mathbb{H}_{T},||\cdot ||_{\mathbb{H}_{T}})$ is a Banach space.
\[lip-0\]$Q$ is locally Lipschitz continuous with respect to $A^{\delta
} $: $$||Q(x)-Q(y)||\leq K_{1}(||A^{\delta }x||+||A^{\delta }y||)||A^{\delta
}(x-y)||\text{.} \label{mi-3}$$
Let $\theta (s)=sx+(1-s)y$. Then $\theta \in C^{1}([0,1],X)$, $\theta
^{\prime }(s)\in D(A^{\delta })$ and $||A^{\delta }\theta (s)||\leq
||A^{\delta }x||+||A^{\delta }y||$. For every $\xi \in X^{\ast }$ $$\begin{aligned}
\langle Q(x)-Q(y),\xi \rangle &=&\int_{0}^{1}\frac{d}{ds}\langle Q(\theta
(s)),\xi \rangle \\
&=&\int_{0}^{1}\langle DQ(\theta (s))\theta ^{\prime }(s),\xi \rangle\end{aligned}$$which yields that $$\begin{aligned}
\left\vert \langle Q(x)-Q(y),\xi \rangle \right\vert &\leq
&\int_{0}^{1}||DQ(\theta (s))\theta ^{\prime }(s)||||\xi || \\
&\leq &\int_{0}^{1}K_{1}(||A^{\delta }x||+||A^{\delta }y||)||A^{\delta
}\theta ^{\prime }(s)||||\xi ||ds\end{aligned}$$so (\[mi-3\]) follows.
We turn to the operator $F.$
\[Flemma\]If $F$ satisfies Condition \[Fcond\] then we have for every $%
x,y\in X$ and $\xi \in Z$ $$\left\vert \left\vert A^{\epsilon }\left[ F\left( x\right) -F\left( y\right) %
\right] \xi \right\vert \right\vert _{X}\leq C\left\vert \left\vert
x-y\right\vert \right\vert _{X}\left\vert \left\vert \xi \right\vert
\right\vert _{Z} \label{consequence1}$$
and for all $x\in D\left( A^{\delta }\right) $, $\xi \in Z$$$\left\vert \left\vert A^{\delta +\epsilon }F\left( x\right) \xi \right\vert
\right\vert _{X}\leq C_{1}\left\vert \left\vert A^{\delta }x\right\vert
\right\vert \left\vert \left\vert \xi \right\vert \right\vert
_{Z}+C_{1}\left\vert \left\vert \xi \right\vert \right\vert _{Z}.
\label{consequence2}$$
Since the operator norm is bounded above by the Hilbert-Schmidt norm it follows immediately from (\[eq:cond1\]) that$$\left\vert \left\vert A^{\epsilon }\left[ F\left( x\right) -F\left( y\right) %
\right] \xi \right\vert \right\vert _{X}\leq C\left\vert \left\vert
A^{\epsilon }\left[ F\left( x\right) -F\left( y\right) \right] \right\vert
\right\vert _{L_{2}\left( Z,X\right) }\left\vert \left\vert \xi \right\vert
\right\vert _{Z}.$$
Similarly, using (\[eq:cond2\]) we have for any $y\in X$$$\begin{aligned}
\left\vert \left\vert A^{\delta +\epsilon }F\left( x\right) \xi \right\vert
\right\vert _{X} &\leq &\left\vert \left\vert A^{\delta +\epsilon }\left[
F\left( x\right) -F\left( y\right) \right] \xi \right\vert \right\vert
_{X}+\left\vert \left\vert A^{\delta +\epsilon }F\left( y\right) \xi
\right\vert \right\vert _{X} \\
&\leq &C\left\vert \left\vert A^{\delta }x\right\vert \right\vert
_{X}\left\vert \left\vert \xi \right\vert \right\vert +C\left\vert
\left\vert A^{\delta }y\right\vert \right\vert _{X}\left\vert \left\vert \xi
\right\vert \right\vert +\left\vert \left\vert A^{\delta +\epsilon }F\left(
y\right) \right\vert \right\vert _{L_{2}\left( Z,X\right) }\left\vert
\left\vert \xi \right\vert \right\vert \\
&\leq &C_{1}\left\vert \left\vert A^{\delta }x\right\vert \right\vert
_{X}\left\vert \left\vert \xi \right\vert \right\vert +C_{2}.\end{aligned}$$
Suppose that $F:X\rightarrow L_{2}\left( Z,X\right) $ satisfies Condition \[Fcond\]. If $u,v\in \mathbb{H}_{T}$ we have for $0\leq s<t\leq T$ $$\begin{aligned}
&&\left\vert \left\vert \left[ A^{\epsilon }F\left( u_{t}\right)
-A^{\epsilon }F\left( u_{s}\right) \right] -\left[ A^{\epsilon }F\left(
v_{t}\right) -A^{\epsilon }F\left( v_{s}\right) \right] \right\vert
\right\vert _{L_{2}\left( Z,X\right) } \\
&\leq &C\left( \sup_{x\in X}\left\vert \left\vert DA^{\epsilon }F\left(
x\right) \right\vert \right\vert +\frac{3}{2}\left\vert \left\vert
v\right\vert \right\vert _{\mathbb{H}_{T}}\right) \left\vert \left\vert
u-v\right\vert \right\vert _{\mathbb{H}_{T}}\left( t-s\right) ^{\alpha },\end{aligned}$$
which implies, in particular, that $A^{\epsilon }P_{t-\cdot }\left[
F\left( u_{\cdot }\right) -A^{\epsilon }F\left( v_{\cdot }\right) \right] :%
\left[ 0,t\right] \rightarrow L_{2}\left( Z,X\right) $ is $\alpha -$Hölder continuous with $$\left\vert \left\vert A^{\epsilon }P_{t-\cdot }\left[ F\left( u_{\cdot
}\right) -A^{\epsilon }F\left( v_{\cdot }\right) \right] \right\vert
\right\vert _{\alpha -\text{H\"{o}l;}\left[ 0,T\right] }\leq \left(
\sup_{x\in X}\left\vert \left\vert DA^{\epsilon }F\left( x\right)
\right\vert \right\vert +\frac{3}{2}\left\vert \left\vert v\right\vert
\right\vert _{\mathbb{H}_{T}}\right) \left\vert \left\vert u-v\right\vert
\right\vert _{\mathbb{H}_{T}}.$$
Here as elsewhere we denote the $\alpha -$Hölder norm of a path $u:\left[
s,t\right] \rightarrow W$ in a Banach space $\left( W,\left\vert \left\vert
\cdot \right\vert \right\vert _{W}\right) $ by $$\left\vert \left\vert u\right\vert \right\vert _{\alpha -\text{H\"{o}l;}%
\left[ s,t\right] }=\sup_{\substack{ t_{1}<t_{2}, \\ t_{1},t_{2}\in \left[
s,t\right] }}\frac{\left\vert \left\vert u\left( t_{2}\right) -u\left(
t_{1}\right) \right\vert \right\vert _{W}}{\left( t_{2}-t_{1}\right)
^{\alpha }}.$$
Let $\theta _{\tau }^{u}=\tau u_{t}+\left( 1-\tau \right) u_{s},$ $\theta
_{\tau }^{v}=\tau v_{t}+\left( 1-\tau \right) v_{s}$ and $\xi \in
L_{2}\left( Z,X\right) ^{\prime }$ then we have $$\begin{aligned}
&&\left\langle \left\langle \left[ A^{\epsilon }F\left( u_{t}\right)
-A^{\epsilon }F\left( u_{s}\right) \right] -\left[ A^{\epsilon }F\left(
v_{t}\right) -A^{\epsilon }F\left( v_{s}\right) \right] ,\xi \right\rangle
\right\rangle \\
&=&\int_{0}^{1}\frac{d}{d\tau }\left\langle \left\langle A^{\epsilon
}F\left( \theta _{\tau }^{u}\right) ,\xi \right\rangle \right\rangle d\tau
-\int_{0}^{1}\frac{d}{d\tau }\left\langle \left\langle A^{\epsilon }F\left(
\theta _{\tau }^{v}\right) ,\xi \right\rangle \right\rangle d\tau \\
&=&\int_{0}^{1}\left\langle \left\langle DA^{\epsilon }F\left( \theta _{\tau
}^{u}\right) \left( \frac{d}{d\tau }\left[ \theta _{\tau }^{u}-\theta _{\tau
}^{v}\right] \right) ,\xi \right\rangle \right\rangle d\tau \\
&&+\int_{0}^{1}\left\langle \left\langle \left[ DA^{\epsilon }F\left( \theta
_{\tau }^{u}\right) -DA^{\epsilon }F\left( \theta _{\tau }^{v}\right) \right]
\frac{d}{d\tau }\theta _{\tau }^{v},\xi \right\rangle \right\rangle d\tau .\end{aligned}$$
This implies $$\begin{aligned}
&&\left\vert \left\langle \left\langle \left[ A^{\epsilon }F\left(
u_{t}\right) -A^{\epsilon }F\left( u_{s}\right) \right] -\left[ A^{\epsilon
}F\left( v_{t}\right) -A^{\epsilon }F\left( v_{s}\right) \right] ,\xi
\right\rangle \right\rangle \right\vert \\
&\leq &\int_{0}^{1}\left\vert \left\vert DA^{\epsilon }F\left( \theta _{\tau
}^{u}\right) \left( \frac{d}{d\tau }\left[ \theta _{\tau }^{u}-\theta _{\tau
}^{v}\right] \right) \right\vert \right\vert \left\vert \left\vert \xi
\right\vert \right\vert d\tau \\
&&+\int_{0}^{1}\left\vert \left\vert \left[ DA^{\epsilon }F\left( \theta
_{\tau }^{u}\right) -DA^{\epsilon }F\left( \theta _{\tau }^{v}\right) \right]
\frac{d}{d\tau }\theta _{\tau }^{v}\right\vert \right\vert \left\vert
\left\vert \xi \right\vert \right\vert d\tau \\
&\leq &\int_{0}^{1}\sup_{x\in X}\left\vert \left\vert DA^{\epsilon }F\left(
x\right) \right\vert \right\vert \left\vert \left\vert \left[ u_{t}-u_{s}%
\right] -\left[ v_{t}-v_{s}\right] \right\vert \right\vert \left\vert
\left\vert \xi \right\vert \right\vert d\tau \\
&&+\int_{0}^{1}\left\vert \left\vert \theta _{\tau }^{u}-\theta _{\tau
}^{v}\right\vert \right\vert \left\vert \left\vert v_{t}-v_{s}\right\vert
\right\vert \left\vert \left\vert \xi \right\vert \right\vert d\tau \\
&=&\int_{0}^{1}\sup_{x\in X}\left\vert \left\vert DA^{\epsilon }F\left(
x\right) \right\vert \right\vert \left\vert \left\vert \left[ u_{t}-u_{s}%
\right] -\left[ v_{t}-v_{s}\right] \right\vert \right\vert \left\vert
\left\vert \xi \right\vert \right\vert d\tau \\
&&+\int_{0}^{1}\left[ \tau \left\vert \left\vert \left( u_{t}-u_{s}\right)
-\left( v_{t}-v_{s}\right) \right\vert \right\vert +\left\vert \left\vert
u_{s}-v_{s}\right\vert \right\vert \right] \left\vert \left\vert
v_{t}-v_{s}\right\vert \right\vert \left\vert \left\vert \xi \right\vert
\right\vert d\tau \\
&\leq &\sup_{x\in X}\left\vert |DA^{\epsilon }F\left( x\right) \right\vert
\left\vert \left\vert u-v\right\vert \right\vert _{\mathbb{H}_{T}}\left(
t-s\right) ^{\alpha }\left\vert \left\vert \xi \right\vert \right\vert \\
&&+\int_{0}^{1}\left[ \tau \left\vert \left\vert \left( u_{t}-u_{s}\right)
-\left( v_{t}-v_{s}\right) \right\vert \right\vert +\left\vert \left\vert
u_{s}-v_{s}\right\vert \right\vert \right] \left\vert \left\vert
v\right\vert \right\vert _{\mathbb{H}_{T}}\left( t-s\right) ^{\alpha
}\left\vert \left\vert \xi \right\vert \right\vert d\tau \\
&\leq &\left( \sup_{x\in X}\left\vert |DA^{\epsilon }F\left( x\right)
\right\vert +\frac{3}{2}\left\vert \left\vert v\right\vert \right\vert _{%
\mathbb{H}_{T}}\right) \left\vert \left\vert u-v\right\vert \right\vert _{%
\mathbb{H}_{T}}\left( t-s\right) ^{\alpha }\left\vert \left\vert \xi
\right\vert \right\vert\end{aligned}$$
and the result follows.
The Bochner Integral
====================
Let $u\in \mathbb{H}_{T}$ such that $\sup_{t\in (0,T]}||A^{\delta
}u_{t}||<\infty $. Let $\varepsilon \in \lbrack 0,1)$. Then
\(1) For any $t\in (0,T]$, $s\rightarrow A^{\varepsilon }P_{t-s}Q(u(s))$ is continuous on $(0,t)$.
\(2) For every $s<t\in (0,T]$, $\int_{s}^{t}A^{\varepsilon }P_{t-r}Q(u(r))dr$ exists and $$\left\Vert \int_{s}^{t}A^{\varepsilon }P_{t-r}Q(u(r))dr\right\Vert \leq
CK_{2}\left( \sup_{t\in (0,T]}||A^{\delta }u_{t}|\right) \left( t-s\right)
^{1-\epsilon }\text{.} \label{fe1}$$
Let $t\in (0,T]$ and consider $f(s)=A^{\varepsilon }P_{t-s}Q(u(s))$. Then for $s_{i}\in (0,t)$, $s_{1}>s_{2}$, one has $$\begin{aligned}
||f(s_{1})-f(s_{2})|| &\leq &||A^{\varepsilon }P_{t-s_{1}}\left(
Qu(s_{1})-Qu(s_{2})\right) || \\
&&+||A^{\varepsilon }P_{t-s_{1}}\left( I-P_{s_{1}-s_{2}}\right) Qu(s_{2})||
\\
&\leq &C(t-s_{1})^{-\varepsilon }||Qu(s_{1})-Qu(s_{2})|| \\
&&+C(t-s_{1})^{-\varepsilon }||\left( I-P_{s_{1}-s_{2}}\right) Qu(s_{2})|| \\
&\leq &C(t-s_{1})^{-\varepsilon }K_{1}\left( 2\sup_{t\in (0,T]}||A^{\delta
}u_{t}|\right) ||A^{\delta }\left( u(s_{1})-u(s_{2})\right) || \\
&&+C(t-s_{1})^{-\varepsilon }||\left( I-P_{s_{1}-s_{2}}\right) Qu(s_{2})||\end{aligned}$$Since $A^{\delta }u\in C((0,T],X)$ so is $u$ and letting $s_{1}\rightarrow
s_{2}\in (0,t)$ the continuity of the semigroup and of $u$ give $%
\lim_{s_{1}\rightarrow s_{2}}||f(s_{1})-f(s_{2})||=0$, i.e. $f$ is continuous on $(0,t)$. Moreover, $$\begin{aligned}
||f(s)|| &\leq &C(t-s)^{-\varepsilon }||Qu(s)|| \\
&\leq &C(t-s)^{-\varepsilon }K_{2}(||A^{\delta }u(s)||) \\
&\leq &CK_{2}\left( \sup_{t\in (0,T]}||A^{\delta }u_{t}|\right)
(t-s)^{-\epsilon }\text{,}\end{aligned}$$
from which (\[fe1\]) follows immediately.
For $u\in \mathbb{H}_{T}$ we define $$\mathbb{L}u(t)=P_{t}u_{0}-\int_{0}^{t}A^{\tau
}P_{t-s}Q(u(s))ds+\int_{0}^{t}P_{t-s}F(u(s))dw_{s} \label{mi-2}$$where $w=(w_{t})$ is an $\alpha $-Hölder continuous path in $Z$, where $%
\alpha \in (\frac{1}{2},1]$ such that $\alpha +\tau <1$, so that $2\alpha >1$. With these constraints, and recalling that $\delta \leq \alpha $ we see that from the previous proposition that $$\sup_{t\in \lbrack 0,T]}||A^{\delta }\int_{0}^{t}A^{\tau
}P_{t-s}Q(u(s))ds||\leq CK_{2}(\beta )T^{1-\delta -\tau }$$and$$\begin{aligned}
&&||\int_{0}^{t_{2}}A^{\tau }P_{t_{2}-s}Q(u(s))ds-\int_{0}^{t_{1}}A^{\tau
}P_{t_{1}-s}Q(u(s))ds|| \\
&=&\left\Vert \int_{0}^{t_{1}}\left( P_{t_{2}-t_{1}}-I\right) A^{\tau
}P_{t_{1}-s}Q\left( u\left( s\right) \right) ds+\int_{t_{1}}^{t_{2}}A^{\tau
}P_{t_{2}-s}Q\left( u\left( s\right) \right) ds\right\Vert \\
&\leq &C_{1}\left( t_{2}-t_{1}\right) ^{\alpha }K_{2}\left( \sup_{t\in \left[
0,T\right] }A^{\delta }u_{t}\right) t_{1}^{1-\left( \alpha +\tau \right)
}+C_{2}K_{2}\left( \sup_{t\in \left[ 0,T\right] }A^{\delta }u_{t}\right)
\left( t_{2}-t_{1}\right) ^{1-\tau } \\
&\leq &CT^{1-\left( \alpha +\tau \right) }K_{2}(\beta )\left(
t_{2}-t_{1}\right) ^{\alpha },\end{aligned}$$so that the Bochner integral appearing in (\[mi-2\]) has the right properties to be in $\mathbb{H}_{T}.$ We now turn our attention to the the properties of the Young integral featuring in (\[mi-2\]).
The Young Integral
==================
We will use the Young integral to make a rigorous path-wise definition of the stochastic integral above provided the process has sample paths of the correct regularity. The example to keep in mind is fractional Brownian motion. The following is due to L.C. Young from the 1930s when he made the definition following investigations into the convergence of Fourier series (see [@young]). It is an extension of Stieltjes integration to the case of paths of finite $p$ variation for $p>1$.
A continuous path $f:\left[ 0,T\right] \rightarrow X$ where $X$ is a Banach space is said to have finite $p$ variation on $\left[ 0,T\right] $ if $$\left\Vert f\right\Vert _{p,\left[ 0,T\right] }:=\left[ \sup_{D}\sum_{k}%
\left\Vert f_{t_{k}}-f_{t_{k-1}}\right\Vert _{X}^{p}\right] ^{\frac{1}{p}%
}<\infty$$where by $\sup_{D}$ we understand supremum over all partitions of $\left[ 0,T%
\right] $.
In analogy with the Riemann sums for Stieltjes integrals of paths of finite variation, we have the following general definition of the integral in the case of real valued functions $f$ and $w$ defined on the interval $\left[ 0,T%
\right] $.
It is obvious that if a path $f$ is Hölder continuous with exponent $%
h\in \left( 0,1\right) $ on the interval $\left[ s,t\right] $ then $f$ has finite $\frac{1}{h}$ variation and$$\left\Vert f\right\Vert _{\frac{1}{h},\left[ s,t\right] }\equiv \left(
\sup_{D\left[ s,t\right] }\sum_{k}\left\vert \left\vert
f_{t_{k}}-f_{t_{k-1}}\right\vert \right\vert ^{\frac{1}{h}}\right) ^{h}\leq
\left\Vert f\right\Vert _{h-H\ddot{o}l,\left[ s,t\right] }\left( t-s\right)
^{h}$$ We will use this fact several times in later proofs.
The Stieltjes integral $$\int_{0}^{T}f\left( t\right) dw_{t}$$is said to exist in the Riemann sense with value $I$ provided that for all $%
\delta >0$ there exists a $\varepsilon _{\delta }>0$ such that if the partition $0=t_{0}<t_{1}<\cdots <t_{N-1}<t_{N}=T$ satisfies $\left\vert
t_{i}-t_{i-1}\right\vert <\delta $ for all $i$, then$$\left\vert \sum_{k=0}^{N-1}f_{t_{k}}\left( w_{t_{k+1}}-w_{t_{k}}\right)
-I\right\vert <\varepsilon _{\delta }$$with $\varepsilon _{\delta }\rightarrow 0$ as $\delta \rightarrow 0$.
If the real valued functions $f$ and $w$ have finite $p$ and $q$ variation respectively such that $p,q>0$ and $\frac{1}{p}+\frac{1}{q}>1$, then the Stieltjes integral $$\int_{0}^{T}f\left( t\right) dw_{t}$$exists in the Riemann sense.
In fact the definition of the integral and Young’s theorem can be extended to the infinite dimensional case when $w:\left[ 0,T\right] \rightarrow Z$ and $f:\left[ 0,T\right] \rightarrow L\left( Z,X\right) $ are bounded paths with finite $p$ and $q$ variation respectively. In this case, the integral $%
\int_{0}^{t}f\left( s\right) dw_{s}:\left[ 0,T\right] \rightarrow X$ is a bounded path of finite $q$ variation and we have the estimate $$\left\Vert \int_{0}^{\cdot }\left( f\left( s\right) -f\left( 0\right)
\right) dw_{s}\right\Vert _{p,\left[ 0,T\right] }\leq C\left\Vert
f\right\Vert _{q,\left[ 0,T\right] }\left\Vert w\right\Vert _{p,\left[ 0,T%
\right] }$$known as Young’s inequality. The details can be found in [@caruana]. Using this, we can in a path-wise way define the integral $%
\int_{0}^{t}A^{\varepsilon }P_{t-s}F\left( u\left( s\right) \right) dw_{s}$ for processes $w$ with sufficiently regular sample paths.
Let $u\in \mathbb{H}_{T},$ $w:\left[ 0,T\right] \rightarrow Z$ be $%
\alpha -$Hölder continuous, $\epsilon \in \left[ 0,\alpha \right] $ and $%
\delta +\alpha >1$. If $F:X\rightarrow L_{2}\left( Z,X\right) $ satisfies conditions (\[eq:cond1\]) and (\[eq:cond2\]) then $$\left\vert \left\vert A^{\epsilon }P_{t-s}F\left( u_{s}\right) -A^{\epsilon
}P_{t-r}F\left( u_{r}\right) \right\vert \right\vert _{L_{2}\left(
Z,X\right) }\leq C\left( \sup_{r\in (0,T]}\left\vert \left\vert A^{\delta
}u_{r}\right\vert \right\vert +T^{\alpha -\delta }+1\right) \left\vert
s-r\right\vert ^{\delta } \label{holder bound}$$
for all $r,s\in \left[ 0,t\right] ,t\in \left[ 0,T\right]
$. In particular, $A^{\epsilon }P_{t-\cdot }F\left( u_{\cdot }\right) :\left[
0,t\right] \rightarrow L_{2}\left( Z,X\right) $ is $\delta $-Hölder continuous and hence the Young integral $$\int_{0}^{t}A^{\epsilon }P_{t-s}F\left( u_{s}\right) dw_{s}$$
exists and satisfies $$\left\vert \left\vert \int_{0}^{t}A^{\epsilon }P_{t-s}F\left( u_{s}\right)
dw_{s}-A^{\epsilon }P_{t}F\left( u_{0}\right) \left( w_{t}-w_{0}\right)
\right\vert \right\vert _{X}\leq C\left( \sup_{r\in (0,T]}\left\vert
\left\vert A^{\delta }u_{r}\right\vert \right\vert +T^{\alpha -\delta
}+1\right) . \label{young's inequality}$$
For $0\leq r<s\leq t$ let us observe by using (\[eq:cond1\]) and ([consequence2]{}) $$\begin{aligned}
&&\left\vert \left\vert \left( A^{\epsilon }P_{t-s}F\left( u_{s}\right)
-A^{\epsilon }P_{t-r}F\left( u_{r}\right) \right) \right\vert \right\vert
_{L_{2}\left( Z,X\right) } \\
&\leq &C\left( \left\vert \left\vert \left( I-P_{s-r}\right) A^{\epsilon
}F\left( u_{r}\right) \right\vert \right\vert _{L_{2}\left( Z,X\right)
}+\left\vert \left\vert A^{\epsilon }\left[ F\left( u_{s})-F(u_{r}\right) %
\right] \right\vert \right\vert _{L_{2}\left( Z,X\right) }\right) \\
&\leq &\frac{C\left( \delta \right) }{\delta }\left( s-r\right) ^{\delta
}\left\Vert A^{\epsilon +\delta }F\left( u_{s}\right) \right\Vert
_{L_{2}\left( Z,X\right) }+C\left\vert \left\vert u_{s}-u_{r}\right\vert
\right\vert \\
&\leq &\left( s-r\right) ^{\delta }\left( C_{1}\left\vert \left\vert
A^{\delta }u_{s}\right\vert \right\vert +C_{2}\right) +C_{3}T^{\alpha
-\delta }\left( s-r\right) ^{\delta }.\end{aligned}$$
From which it follows that $$\left\vert \left\vert \left( A^{\epsilon }P_{t-s}F\left( u_{s}\right)
-A^{\epsilon }P_{t-r}F\left( u_{r}\right) \right) \right\vert \right\vert
\leq C\left( \sup_{r\in (0,T]}\left\vert \left\vert A^{\delta
}u_{r}\right\vert \right\vert +T^{\alpha -\delta }+1\right) \left(
s-r\right) ^{\delta }.$$
Estimate$\ $(\[holder bound\]) follows at once. From [caruana]{} this is sufficient to guarantee the existence of the Young integral and (\[young’s inequality\]) follows from Young’s inequality.
The non-Linear mapping $\mathbb{L}$
===================================
We now prove that the non-linear mapping $\mathbb{L}$ is a (non-linear) bounded operator. More precisely:
\[invariance\]If $u\in \mathbb{H}_{T}$ and $u_{0}\in D(A^{\alpha })$, then $\mathbb{L}u\in \mathbb{H}_{T}$ and the following estimates hold: if $%
||u||_{\mathbb{H}_{T}}\leq \beta $ $$\begin{aligned}
||\mathbb{L}u||_{\mathbb{H}_{T}} &\leq &C_{1}\left( 1+T^{\alpha }\right)
||A^{\alpha }u_{0}||+C_{2}\left( T^{1-\delta -\tau }+T^{1-\tau -\alpha
}\right) K_{2}(\beta ) \\
&&+C_{3}\left[ \beta +1\right] \left( T^{\alpha }+T^{\delta }+T^{\alpha
+\delta }+T^{2\alpha }\right) +C_{4}\left( 1+T^{\alpha }\right) \left\Vert
A^{\alpha }F\left( u_{0}\right) \right\Vert _{op}\text{.}\end{aligned}$$
There are three terms appearing in the definition of $\mathbb{L}$, namely $%
P_{t}u_{0}$, the ordinary integral $$J_{t}=\int_{0}^{t}A^{\tau }P_{t-s}Q(u_{s})ds$$and the Young integral $$U_{t}\equiv \int_{0}^{t}P_{t-s}F(u_{s})dw_{s}\text{.}$$Let us estimate their $\mathbb{H}_{T}$ norms one by one. Firstly, $$||A^{\delta }P_{t}u_{0}||\leq ||A^{\delta }u_{0}||\leq \beta$$and $$\begin{aligned}
||P_{t}u_{0}-P_{s}u_{0}|| &=&||\left( P_{t-s}-I\right) P_{s}u_{0}|| \\
&\leq &C_{\alpha }(t-s)^{\alpha }||A^{\alpha }P_{s}u_{0}|| \\
&\leq &C_{\alpha }||A^{\alpha }u_{0}||(t-s)^{\alpha }\text{.}\end{aligned}$$Secondly we consider the ordinary integral $J_{t}$. It is elementary that $$\begin{aligned}
||A^{\delta }J_{t}|| &\leq &C_{1}\int_{0}^{t}(t-s)^{-(\delta +\tau
)}||Q(u_{s})||ds \\
&\leq &C_{1}\int_{0}^{t}(t-s)^{-(\delta +\tau )}K_{2}(||A^{\delta }u_{s}||)ds
\\
&\leq &C_{2}K_{2}(\beta )t^{1-\delta -\tau }\text{.}\end{aligned}$$To estimate the Hölder norm we use the following elementary formula $$J_{t}-J_{s}=\int_{s}^{t}A^{\tau }P_{t-r}Q(u_{r})dr+\left( P_{t-s}-I\right)
\int_{0}^{s}A^{\tau }P_{s-r}Q(u_{r})dr\text{. }$$While it is easy to see that $$\begin{aligned}
\left\Vert \int_{s}^{t}A^{\tau }P_{t-r}Q(u_{r})dr\right\Vert &\leq &\frac{%
C_{4}K_{2}(\beta )}{1-\tau }\left( t-s\right) ^{1-\tau } \\
&\leq &\frac{C_{4}K_{2}(\beta )}{1-\tau }\left( t-s\right) ^{\alpha }\left(
t-s\right) ^{1-\tau -\alpha } \\
&\leq &\frac{C_{4}K_{2}(\beta )}{1-\tau }T^{1-\tau -\alpha }\left(
t-s\right) ^{\alpha }.\end{aligned}$$
Using the bound $$||x-P_{t}x||\leq \frac{C(\delta )}{\delta }t^{\delta }||A^{\delta }x||$$we deduce that$$\begin{aligned}
\left\vert \left\vert \left( P_{t-s}-I\right) \int_{0}^{s}A^{\tau
}P_{s-r}Q\left( u_{r}\right) dr\right\vert \right\vert &\leq &\frac{C\left(
\alpha \right) }{\alpha }\left( t-s\right) ^{\alpha }\left\vert \left\vert
A^{\alpha }\int_{0}^{s}A^{\tau }P_{s-r}Q\left( u_{r}\right) dr\right\vert
\right\vert \\
&\leq &\frac{C\left( \alpha \right) }{\alpha \left( 1-\tau -\alpha \right) }%
K_{2}\left( \beta \right) s^{1-\tau -\alpha }\left( t-s\right) ^{\alpha } \\
&\leq &\frac{C\left( \alpha \right) }{\alpha \left( 1-\tau -\alpha \right) }%
K_{2}\left( \beta \right) T^{1-\tau -\alpha }\left( t-s\right) ^{\alpha }.\end{aligned}$$
Finally we handle the Young integral. Let $[s,t]\subset \lbrack
0,T]$ then $$\begin{aligned}
U_{t}-U_{s}
&=&\int_{0}^{t}P_{t-r}F(u_{r})dw_{r}-\int_{0}^{s}P_{s-r}F(u_{r})dw_{r} \\
&=&\int_{s}^{t}P_{t-r}F(u_{r})dw_{r}+(P_{t-s}-I)%
\int_{0}^{s}P_{s-r}F(u_{r})dw_{r},\end{aligned}$$so that $$\begin{aligned}
||U_{t}-U_{s}|| &\leq &\left\Vert
\int_{s}^{t}P_{t-r}F(u_{r})dw_{r}\right\Vert +\left\Vert
(P_{t-s}-I)\int_{0}^{s}P_{s-r}F(u_{r})dw_{r}\right\Vert \label{difference}
\\
&\leq &\left\Vert \int_{s}^{t}P_{t-r}F(u_{r})dw_{r}\right\Vert +\frac{%
C\left( \alpha \right) }{\alpha }\left( t-s\right) ^{\alpha }\left\Vert
\int_{0}^{s}A^{\alpha }P_{s-r}F(u_{r})dw_{r}\right\Vert \notag\end{aligned}$$To handle the two integrals, let us consider $$R_{s,t}^{\epsilon }\equiv A^{\epsilon }\int_{s}^{t}P_{t-r}F(u_{r})dw_{r}$$for $\varepsilon \in \lbrack 0,\alpha ]$. From Young’s inequality we have that $$\begin{aligned}
\left\vert \left\vert R_{s,t}^{\epsilon }\right\vert \right\vert
&=&\left\Vert \int_{s}^{t}A^{\varepsilon }P_{t-r}F(u_{r})dw_{r}\right\Vert \\
&\leq &C\left\vert \left\vert A^{\varepsilon }P_{t-\cdot }F(u_{\cdot
})\right\vert \right\vert _{\delta -\text{H\"{o}l}}\left\vert \left\vert
w\right\vert \right\vert _{\alpha -\text{H\"{o}l}}\left( t-s\right) ^{\delta
+\alpha }+\left\vert \left\vert A^{\varepsilon }P_{t-s}F(u_{s})\left(
w_{t}-w_{s}\right) \right\vert \right\vert \\
&\leq &C\left( \beta +1+T^{\alpha -\delta }\right) \left( t-s\right)
^{\delta +\alpha }+\left\vert \left\vert A^{\varepsilon
}P_{t-s}(F(u_{s})-F\left( u_{0}\right) )\left( w_{t}-w_{s}\right)
\right\vert \right\vert \\
&&+\left\vert \left\vert A^{\varepsilon }P_{t-s}F(u_{0})\left(
w_{t}-w_{s}\right) \right\vert \right\vert\end{aligned}$$Hence, using condition \[Fcond\] and lemma \[Flemma\] we can deduce $$\left\vert \left\vert R_{s,t}^{\epsilon }\right\vert \right\vert \leq
C\left( \beta +1+T^{\alpha -\delta }\right) \left( t-s\right) ^{\delta
+\alpha }+C\beta s^{\alpha }\left( t-s\right) ^{\alpha }+C\left\vert
\left\vert A^{\alpha }F\left( u_{0}\right) \right\vert \right\vert
_{op}\left( t-s\right) ^{\alpha }. \label{integral bound}$$
An application of this with $\epsilon =\delta $ yields $$\begin{aligned}
\left\Vert A^{\delta }U_{t}\right\Vert &\leq &\left\Vert
\int_{0}^{t}A^{\delta }P_{t-r}F(u_{r})dw_{r}\right\Vert \\
&\leq &C\left( \beta +1+T^{\alpha -\delta }\right) T^{\delta +\alpha
}+C\beta T^{2\alpha }+C\left\vert \left\vert A^{\alpha }F\left( u_{0}\right)
\right\vert \right\vert _{op}T^{\alpha }.\end{aligned}$$Then, two further applications of (\[integral bound\]) with $%
\epsilon =0$ and $\epsilon =\alpha $ show$$\begin{aligned}
\left\Vert \int_{s}^{t}P_{t-r}F(u_{r})dw_{r}\right\Vert &\leq &C\left( \beta
+1+T^{\alpha -\delta }\right) \left( t-s\right) ^{\delta +\alpha }+C\beta
T^{\alpha }\left( t-s\right) ^{\alpha } \label{difference2} \\
&&+C\left\vert \left\vert A^{\alpha }F\left( u_{0}\right) \right\vert
\right\vert _{op}\left( t-s\right) ^{\alpha } \notag \\
\left\Vert \int_{0}^{s}A^{\alpha }P_{s-r}F(u_{r})dw_{r}\right\Vert &\leq
&C\left( \beta +1+T^{\alpha -\delta }\right) T^{\delta +\alpha }+C\beta
T^{2\alpha } \notag \\
&&+C\left\vert \left\vert A^{\alpha }F\left( u_{0}\right) \right\vert
\right\vert _{op}T^{\alpha }.\end{aligned}$$
Assembling (\[difference\]) and (\[difference2\]) provides$$\frac{||U_{t}-U_{s}||}{\left( t-s\right) ^{\alpha }}\leq C\left( \beta
+1+T^{\alpha -\delta }\right) \left( T^{\delta }+T^{\alpha +\delta }\right)
+C\beta T^{\alpha }\left( 1+T^{\alpha }\right) +C\left\vert \left\vert
A^{\alpha }F\left( u_{0}\right) \right\vert \right\vert _{op}\left(
1+T^{\alpha }\right) .$$
The result then follows by collecting together the appropriate terms.
Next we prove that $\mathbb{L}$ is locally Lipschitz.
\[contract\]Let $u,v\in \mathbb{H}_{T}$ such that $||u||_{\mathbb{H}%
_{T}}\leq \beta $, $||v||_{\mathbb{H}_{T}}\leq \beta $, $u_{0}=v_{0}\in
D(A^{\alpha })$, then $$\begin{aligned}
&&||\mathbb{L}u-\mathbb{L}v||_{\mathbb{H}_{T}} \\
&\leq &C\left[ \left( T^{1-\delta -\tau }+T^{1-\alpha -\tau }\right)
K_{1}\left( 2\beta \right) +\left( T^{\alpha }+T^{2\alpha }\right) \left(
\sup_{x\in X}\left\vert \left\vert DA^{\alpha }F\left( x\right) \right\vert
\right\vert +\frac{3}{2}\beta +1\right) \right] ||u-v||_{\mathbb{H}_{T}}.\end{aligned}$$
Let $u,v\in \mathbb{H}_{T}$ such that $||u||_{\mathbb{H}_{T}}\leq \beta $ and $||v||_{\mathbb{H}_{T}}\leq \beta $. Let $\theta _{t}=u_{t}-v_{t}$ and consider $$D_{t}=\int_{0}^{t}A^{\tau }P_{t-r}\left( Q(u_{r})-Q(v_{r})\right) dr\text{.}$$If $t>s$ then $$\begin{aligned}
D_{t}-D_{s} &=&\int_{s}^{t}A^{\tau }P_{t-r}\left[ Q(u_{r})-Q(v_{r})\right] dr
\\
&&+\left( P_{t-s}-I\right) \int_{0}^{s}A^{\tau }P_{s-r}\left[
Q(u_{r})-Q(v_{r})\right] dr\text{. }\end{aligned}$$It follows that $$\begin{aligned}
||D_{t}-D_{s}|| &\leq &C\int_{s}^{t}\left( t-r\right) ^{-\tau }\left\Vert
Q(u_{r})-Q(v_{r})\right\Vert dr \\
&&+C|t-s|^{\alpha }\int_{0}^{s}\left( s-r\right) ^{-\tau -\alpha }\left\Vert
Q(u_{r})-Q(v_{r})\right\Vert dr\end{aligned}$$together with the estimate $$||Q(u_{r})-Q(v_{r})||\leq K_{1}(2\beta )||A^{\delta }\theta _{r}||
\label{cr-02}$$we obtain $$\begin{aligned}
||D_{t}-D_{s}|| &\leq &CK_{1}(2\beta )\int_{s}^{t}\left( t-r\right) ^{-\tau
}||A^{\delta }\theta _{r}||dr \\
&&+CK_{1}(2\beta )|t-s|^{\alpha }\int_{0}^{s}\left( s-r\right) ^{-\tau
-\alpha }||A^{\delta }\theta _{r}||dr \\
&\leq &\frac{C}{1-\tau }K_{1}(2\beta )T^{1-\tau -\alpha }|t-s|^{\alpha
}||\theta ||_{\mathbb{H}_{T}} \\
&&+\frac{C}{1-\tau -\alpha }K_{1}(2\beta )T^{1-\tau -\alpha }||\theta ||_{%
\mathbb{H}_{T}}|t-s|^{\alpha }\text{.}\end{aligned}$$
Using the same argument we can obtain for all $t\in \left[ 0,T%
\right] $ $$\left\vert \left\vert A^{\delta }D_{t}\right\vert \right\vert \leq \frac{%
T^{1-\delta -\tau }}{1-\delta -\tau }K_{1}\left( 2\beta \right) ||\theta ||_{%
\mathbb{H}_{T}}.$$
We now define $$\Delta _{t}\equiv \int_{0}^{t}P_{t-r}\left[ F\left( u_{r}\right) -F\left(
v_{r}\right) \right] dw_{r}$$
and consider $$\begin{aligned}
\Delta _{t}-\Delta _{s} &=&\int_{s}^{t}P_{t-r}\left[ F\left( u_{r}\right)
-F\left( v_{r}\right) \right] dw_{r} \\
&&+\left( P_{t-s}-I\right) \int_{0}^{s}P_{s-r}\left[ F\left( u_{r}\right)
-F\left( v_{r}\right) \right] dw_{r},\end{aligned}$$
which satisfies $$\begin{aligned}
\left\vert \left\vert \Delta _{t}-\Delta _{s}\right\vert \right\vert &\leq
&\left\vert \left\vert \int_{s}^{t}P_{t-r}\left[ F\left( u_{r}\right)
-F\left( v_{r}\right) \right] dw_{r}\right\vert \right\vert \\
&&+C\left( t-s\right) ^{\alpha }\left\vert \left\vert \int_{0}^{s}A^{\alpha
}P_{s-r}\left[ F\left( u_{r}\right) -F\left( v_{r}\right) \right]
dw_{r}\right\vert \right\vert .\end{aligned}$$
As before, we consider for $\epsilon \in \left[
0,\alpha \right] $ $$R_{s,t}^{\epsilon }\equiv \int_{s}^{t}A^{\varepsilon }P_{t-r}\left[ F\left(
u_{r}\right) -F\left( v_{r}\right) \right] dw_{r},$$
and note that Young’s inequality gives $$\begin{aligned}
\left\vert \left\vert R_{s,t}^{\epsilon }\right\vert \right\vert &\leq
&C\left\vert \left\vert A^{\alpha }P_{t-\cdot }\left[ F\left( u_{\cdot
}\right) -F\left( v_{\cdot }\right) \right] \right\vert \right\vert _{\alpha
-\text{H\"{o}l}}\left\vert \left\vert w\right\vert \right\vert _{\alpha -%
\text{H\"{o}l}}\left( t-s\right) ^{2\alpha } \notag \\
&&+\left\vert \left\vert A^{\alpha }P_{t-s}(F\left( u_{s}\right) -F\left(
v_{s}\right) )\left( w_{t}-w_{s}\right) \right\vert \right\vert \notag \\
&\leq &C\left\vert \left\vert A^{\alpha }P_{t-\cdot }\left[ F\left( u_{\cdot
}\right) -F\left( v_{\cdot }\right) \right] \right\vert \right\vert _{\alpha
-\text{H\"{o}l}}\left\vert \left\vert w\right\vert \right\vert _{\alpha -%
\text{H\"{o}l}}\left( t-s\right) ^{2\alpha } \notag \\
&&+C\left\vert \left\vert (u_{s}-u_{0})-(v_{s}-v_{0})\right\vert \right\vert
\left( t-s\right) ^{\alpha }. \label{ineq1}\end{aligned}$$
From lemma we have $$\begin{aligned}
&&\left\vert \left\vert A^{\alpha }\left[ F\left( u_{t}\right) -F\left(
v_{s}\right) \right] -A^{\alpha }\left[ F\left( u_{t}\right) -F\left(
v_{s}\right) \right] \right\vert \right\vert _{L_{2}\left( Z,X\right) }
\label{ineq2} \\
&\leq &\left( \sup_{x}\left\vert \left\vert DA^{\alpha }F\left( x\right)
\right\vert \right\vert +\frac{3}{2}\left\vert \left\vert v\right\vert
\right\vert _{\mathbb{H}_{T}}\right) ||\theta ||_{\mathbb{H}%
_{T}}(t-s)^{\alpha } \notag\end{aligned}$$
Combining (\[ineq1\]) and (\[ineq2\]) we can deduce $$\frac{\left\vert \left\vert \Delta _{t}-\Delta _{s}\right\vert \right\vert }{%
\left( t-s\right) ^{\alpha }}\leq C\left( T^{\alpha }+T^{2\alpha }\right)
\left( \sup_{x}\left\vert \left\vert DA^{\alpha }F\left( x\right)
\right\vert \right\vert +\frac{3}{2}\beta +1\right) ||\theta ||_{\mathbb{H}%
_{T}},$$
and hence $$\begin{aligned}
||Lu-Lv||_{\mathbb{H}_{T}} &\leq &\sup_{t\in \left[ 0,T\right] }\left\vert
\left\vert A^{\delta }D_{t}\right\vert \right\vert +\sup_{t\in \left[ 0,T%
\right] }\left\vert \left\vert A^{\delta }\Delta _{t}\right\vert \right\vert
\\
&&+\sup_{0\leq s<t\leq T}\frac{\left\vert \left\vert D_{t}-D_{s}\right\vert
\right\vert }{\left( t-s\right) ^{\alpha }}+\sup_{0\leq s<t\leq T}\frac{%
\left\vert \left\vert \Delta _{t}-\Delta _{s}\right\vert \right\vert }{%
\left( t-s\right) ^{\alpha }} \\
&\leq &C\left( T^{1-\delta -\tau }+T^{1-\alpha -\tau }\right) K_{1}\left(
2\beta \right) ||\theta ||_{\mathbb{H}_{T}} \\
&&+C\left( T^{\alpha }+T^{2\alpha }\right) \left( \sup_{x}\left\vert
\left\vert DA^{\alpha }F\left( x\right) \right\vert \right\vert +\frac{3}{2}%
\beta +1\right) ||\theta ||_{\mathbb{H}_{T}}.\end{aligned}$$
Let $u_{0}\in D(A^{\alpha })$ then for some $T^{\ast }>0$ depending only on $%
||A^{\alpha }u_{0}||$ and $\left\Vert A^{\alpha }F\left( u_{0}\right)
\right\Vert _{op}$ there is a unique $u\in \mathbb{H}_{T^{\ast }}$ such that $$u_{t}=P_{t}u_{0}-\int_{0}^{t}A^{\tau
}P_{t-s}Q(u_{s})ds+\int_{0}^{t}P_{t-s}F(u_{s})dw_{s}\text{.}$$
By choosing $\beta =\beta \left( ||A^{\alpha }u_{0}||,\left\Vert A^{\alpha
}F\left( u_{0}\right) \right\Vert _{op}\right) $ sufficiently large Theorem \[invariance\] allows us to ensure that
$\mathbb{L}\left( \mathbb{H}_{T}\cap \left\{ u:\left\vert
\left\vert u\right\vert \right\vert _{\mathbb{H}_{T}}\leq \beta \right\}
\right) \mathbb{\subseteq }\mathbb{H}_{T}\cap \left\{ u:\left\vert
\left\vert u\right\vert \right\vert _{\mathbb{H}_{T}}\leq \beta \right\} $ for all $T\in \left[ 0,T_{1}\right] $ and some $T_{1}>0.$ From Theorem [contract]{} we see that $\mathbb{L}$ is a contraction on $\mathbb{H}_{T}\cap
\left\{ u:\left\vert \left\vert u\right\vert \right\vert _{\mathbb{H}%
_{T}}\leq \beta \right\} $ for all $T\in \lbrack 0,T_{2}]$ some $T_{2}>0$. By taking $T^{\ast }=T_{1}\wedge T_{2}$ the result follows by a standard contraction-mapping fixed point argument.
Randomly forced Navier-Stokes equations
=======================================
In particular this theorem applies to Navier-Stokes equation driven by fractional Brownian motion with $h>\frac{1}{2}$. In this case, the equation we want to solve is given together with the Dirichlet boundary conditions and no-slip condition by$$\frac{\partial }{\partial t}u+u.\nabla u=\Delta u-\nabla p+F(u)\dot{w},\text{
}\nabla .u=0,\text{ }u|_{\Gamma }=0$$
in some bounded domain $\Omega $ with compact, smooth boundary $%
\Gamma $. As is well known (see for instance [@ladyzhenskaya]) the orthogonal complement of the set $K^{\infty }\left( \Omega \right) =\left\{
u\in C^{\infty }\left( \Omega \right) :\nabla .u=0\right\} $ in $L^{2}\left(
\Omega \right) $ is given by $G\left( \Omega \right) =\left\{ u:u=\nabla p,%
\text{ }p,\nabla p\in L^{2}\left( \Omega \right) \right\} $, and if we denote the closure of $K^{\infty }\left( \Omega \right) $ in $L^{2}\left(
\Omega \right) $ by $K\left( \Omega \right) $ then we have the decomposition $L^{2}\left( \Omega \right) =K\left( \Omega \right) \oplus G\left( \Omega
\right) $. Letting $P_{\infty }:L^{2}\left( \Omega \right) \rightarrow
K\left( \Omega \right) $ be the orthogonal projection onto $K\left( \Omega
\right) $ the above equation can be written as $$\frac{\partial }{\partial t}u+Au+P_{\infty }(u.\nabla u)=P_{\infty }F(u)\dot{%
w},$$where $A=-P_{\infty }\circ \Delta $ is the Stokes operator. More precisely we can define the usual Stokes operator (i.e. with Dirichlet boundary conditions) in the following manner.
The Stokes operator acting in $K\left( \Omega \right) $ is given by the self-adjoint linear operator $-P_{\infty }\circ \Delta $ with domain $%
D\left( -P_{\infty }\circ \Delta \right) =\left\{ u:u\in W^{2,2}\left(
\Omega \right) ,\left. u\right\vert _{\partial D}=0\right\} $ where $%
\triangle $ is the usual trace Laplacian on functions (or vector fields) where the derivative is taken in the generalized sense.
The following result is well known and we refer the reader to [constantin]{} for the details.
The Stokes operator with Dirichlet boundary conditions is self-adjoint and its inverse is a compact operator in $K\left( \Omega \right) $.
Hence, for the Dirichlet boundary condition case, the Stokes operator satisfies the conditions required by our fixed point theorem. It is also clear that if $w$ is a fractional Brownian motion with Hurst parameter $h>%
\frac{1}{2}$ then it satisfies the conditions required by the fixed point theorem. It remains to verify the condition on $Q:=P_{\infty }\left( \left(
u\cdot \nabla \right) u\right) $.
The operator $Q\left( u\right) =A^{\frac{1}{4}}P_{\infty }\left( \left(
u\cdot \nabla \right) u\right) $ is weakly Gateaux differentiable on $%
D\left( A^{\delta }\right) $ where $A$ is either the Stokes operator associated with either the Dirichlet or Navier and kinematic boundary conditions. Its derivative is given by$$DQ\left( u\right) \xi =A^{\frac{1}{4}}P_{\infty }\left( \left( u\cdot \nabla
\right) \xi \right) +A^{\frac{1}{4}}P_{\infty }\left( \left( \xi \cdot
\nabla \right) u\right)$$
Let $z\in K\left( \Omega \right) ^{\ast }=K\left( \Omega \right) $, then it is clear that $$\begin{aligned}
\frac{d}{d\varepsilon }\left\langle A^{\frac{1}{4}}P_{\infty }^{{}}\left(
\left( \left[ u+\varepsilon \xi \right] \cdot \nabla \right) \left[
u+\varepsilon \xi \right] \right) ,z\right\rangle =& \frac{d}{d\varepsilon }%
\left( \left\langle A^{\frac{1}{4}}P_{\infty }\left( \left( u\cdot \nabla
\right) u\right) ,z\right\rangle \right. \\
& +\varepsilon \left\langle A^{\frac{1}{4}}P_{\infty }\left( \left( u\cdot
\nabla \right) \xi \right) ,z\right\rangle \\
& +\varepsilon \left\langle A^{\frac{1}{4}}P_{\infty }\left( \left( \xi
\cdot \nabla \right) u\right) ,z\right\rangle \\
& \left. +\varepsilon ^{2}\left\langle A^{\frac{1}{4}}P_{\infty }\left(
\left( \xi \cdot \nabla \right) \xi \right) ,z\right\rangle \right)\end{aligned}$$which implies $$\left. \frac{d}{d\varepsilon }\left\langle A^{\frac{1}{4}}P_{\infty }\left(
\left( \left[ u+\varepsilon \xi \right] \cdot \nabla \right) \left[
u+\varepsilon \xi \right] \right) ,z\right\rangle \right\vert _{\varepsilon
=0}=\left\langle A^{\frac{1}{4}}P_{\infty }\left( \left( u\cdot \nabla
\right) \xi \right) +A^{\frac{1}{4}}P_{\infty }\left( \left( \xi \cdot
\nabla \right) u\right) ,z\right\rangle .$$
Then the Sobolev inequality (\[eq:sobiq\]) implies the relative boundedness conditions on $Q$ hold in the case of Dirichlet boundary conditions and so we can apply our fixed point theorem to this case.
This research was gratefully supported in part by EPSRC grants EP/F029578/1 and EP/P502667/1. The first named author would like to thank Daniel Babarto for related discussions.
[99]{} Caruana, M; Lévy, T; Lyons, T.J., Differential Equations Driven by Rough Paths. Lecture Notes in Mathematics, Springer 2007.
Chen, G., Qian, Z., A Study of the Navier-Stokes Equations with the Kinematic and Navier Boundary Conditions (Preprint)
Constantin, P., Foias, C., Navier-Stokes Equations. Chicago Lecture Notes in Mathematics, The University of Chicago Press 1988.
Fujita, H., Kato, T., On the Navier-Stokes Initial Value Problem I. *Arch. Rat. Mech. Anal.* **16** (1964) No. 4, 269-315.
Gubinelli, M., Lejay, A., Tindel, S., Young Integrals and SPDEs. *Potential Analysis* **25** (2006) No. 4, 307-326.
Lyons, T., Differential Equations Driven by Rough Signals (I): An extension of an Inequality of L.C. Young. *Math. Res. Lett.* **1** (1994) 451-464.
Lyons, T., The Interpretation and Solution of Ordinary Differential Equations Driven by Rough Signals. *Proc. Symp. Pure Math.* **57** (1995) 115-128.
Ladyzhenskaya, O.A. Sixth Problem of the Millennium : Navier-Stokes Equations, Existence and Smoothness, *Russian Math. Surveys,* **58** No.2, 251-286
Maslowski, NB., Nualart, D., Evolution Equations Driven by a Fractional Brownian Motion. *J. Funct. Anal.* **202** (2003) 277-305.
Kato, T. Note on Fractional Powers of Linear Operators. *Proc. Japan Acad.* **36** (1960) 94–96
Kato, T. Fractional Powers of Dissipative Operators. *J. Math. Soc. Japan* **13** (1961) 246–274.
Komatsu, H. Fractional Powers of Operators. *Pacific J. Math*. **19** (1966) 285–346
Pazy, A. Semigroups of Linear Opeators and Applications to Partial Differential Equations, Springer 1983
Phillips, R. S. On the Generation of Semigroups of Linear Operators.* Pacific J. Math.* **2** (1952).
Sobolevskii, P.E., On Non-Stationary Equations of Hydrodynamics for Viscous Fluid. *Doklady Akad. Nauk USSR* **128** (1959) 45-48.
Tindel, S., Tudor, C.A., Viens, F., Stochastic Evolution Equations with Fractional Brownian Motion. *Prob. Theory Rel. Fields* **127** (2003) 186-204.
Young, L.C., An Inequality of Holder Type, Connected with Stieltjes Integration. *Acta Math.* **67** (1936) No. 1, 251-282.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We calculate the electromagnetic (EM) form factors of the pseudoscalar mesons in the light-front framework. Specifically, these form factors are extracted from the relevant matrix elements directly, instead of choosing the Breit frame. The results show that the charge radius of the meson is related to both the first and second longitudinal momentum square derivatives of the momentum distribution function. The static properties of the EM form factors and the heavy quark symmetry of the charge radii are checked analytically in the heavy quark limit. In addition, we use the Gaussia-type wavefunction to obtain the numerical results.'
---
6.5in
-2truecm 0truecm 0truecm 16truecm 23truecm
\
\
Introduction
============
The understanding of the electromagnetic (EM) properties of hadrons is an important topic, and the EM form factors which are calculated using non-perturbative methods are the useful tool for this purpose. There have been numerous experimental \[1-7\] and theoretical studies \[8-13\] of the EM form factors of the light pseudoscalar mesons ($\pi$ and $K$). However, due to difficulties in the experiments, the EM form factors of light vector mesons ($\rho$ and $K^*$) have fewer investigations than their pseudoscalar counterparts [@HP; @Simula], even though they could provide much information about the bound-state dynamics. As for the EM form factors of heavy mesons (which containing one heavy quark), there are much fewer studies than the light ones. In the heavy hadron investigation, however, the heavy quark symmetry (HQS) [@HQS] is a fundamental and model-independent property. In this work, we will study the EM form factors of the light and heavy pseudoscalar mesons in the light-front framework. We will also check whether HQS is satisified or not among these EM properties of the heavy mesons.
The light front quark model (LFQM) is the only relativistic quark model in which a consistent and fully relativistic treatment of quark spins and the center-of-mass motion can be carried out. Thus it has been applied in the past to calculate various form factors \[16-22\]. This model has many advantages. For example, the light-front wavefunction is manifestly boost invariant as it is expressed in terms of the momentum fraction variables (in “+" component) in analog to the parton distributions in the infinite momentum frame. Moreover, hadron spin can also be relativistically constructed by using the so-called Melosh rotation [@Melosh]. The kinematic subgroup of the light-front formalism has the maximum number of interaction-free generators including the boost operator which describes the center-of-mass motion of the bound state (for a review of the light-front dynamics and light-front QCD, see [@Zhang]).
The paper is organized as follows. In Sec. 2, the basic theoretical formalism is given and the decay constant and the EM form factors are derived for pseudoscalar mesons. In Sec. 3, we take the heavy quark limit to check whether HQS is satisfied or not. In Sec. 4, the numerical result are obtained by choosing the Gaussian-type wavefunction. Finally, a conclusion is given in Sec. 5.
Framework
=========
A meson bound state consisting of a quark $q_1$ and an antiquark $\bar q_2$ with a total momentum $P$ and spin $S$ can be written as $$\begin{aligned}
|M(P, S, S_z)\rangle
&=&\int \{d^3p_1\}\{d^3p_2\} ~2(2\pi)^3 \delta^3(\tilde
P-\tilde p_1-\tilde p_2)~\nonumber\\
&&\times \sum_{\lambda_1,\lambda_2}
\Psi^{SS_z}(\tilde p_1,\tilde p_2,\lambda_1,\lambda_2)~
|q_1(p_1,\lambda_1) \bar q_2(p_2,\lambda_2)\rangle,\end{aligned}$$ where $p_1$ and $p_2$ are the on-mass-shell light-front momenta, $$\tilde p=(p^+, p_\bot)~, \quad p_\bot = (p^1, p^2)~,
\quad p^- = {m^2+p_\bot^2\over p^+},$$ and $$\begin{aligned}
&&\{d^3p\} \equiv {dp^+d^2p_\bot\over 2(2\pi)^3}, \nonumber \\
&&|q(p_1,\lambda_1)\bar q(p_2,\lambda_2)\rangle
= b^\dagger_{\lambda_1}(p_1)d^\dagger_{\lambda_2}(p_2)|0\rangle,\\
&&\{b_{\lambda'}(p'),b_{\lambda}^\dagger(p)\} =
\{d_{\lambda'}(p'),d_{\lambda}^\dagger(p)\} =
2(2\pi)^3~\delta^3(\tilde p'-\tilde p)~\delta_{\lambda'\lambda}.
\nonumber\end{aligned}$$ In terms of the light-front relative momentum variables $(x, k_\bot)$ defined by $$\begin{aligned}
&& p^+_1=(1-x) P^+, \quad p^+_2=x P^+, \nonumber \\
&& p_{1\bot}=(1-x) P_\bot+k_\bot, \quad p_{2\bot}=x P_\bot-k_\bot,\end{aligned}$$ the momentum-space wavefunction $\Psi^{SS_z}$ can be expressed as $$\Psi^{SS_z}(\tilde p_1,\tilde p_2,\lambda_1,\lambda_2)
= R^{SS_z}_{\lambda_1\lambda_2}(x,k_\bot)~ \phi(x, k_\bot), \label{momentumspace}$$ where $\phi(x,k_\bot)$ describes the momentum distribution of the constituents in the bound state, and $R^{SS_z}_{\lambda_1\lambda_2}$ constructs a state of definite spin ($S,S_z$) out of light-front helicity ($\lambda_1,\lambda_2$) eigenstates. Explicitly, $$R^{SS_z}_{\lambda_1 \lambda_2}(x,k_\bot)
=\sum_{s_1,s_2} \langle \lambda_1|
{\cal R}_M^\dagger(1-x,k_\bot, m_1)|s_1\rangle
\langle \lambda_2|{\cal R}_M^\dagger(x,-k_\bot, m_2)
|s_2\rangle
\langle {1\over2}s_1;
{1\over2}s_2|S,S_z\rangle, \label{RR}$$ where $|s_i\rangle$ are the usual Pauli spinors, and ${\cal R}_M$ is the Melosh transformation operator [@Melosh]: $${\cal R}_M (x,k_\bot,m_i) =
{m_i+x M_0+i\vec \sigma\cdot\vec k_\bot \times \vec n
\over \sqrt{(m_i+x M_0)^2 + k_\bot^2}}, \label{Melosh2}$$ with $\vec n = (0,0,1)$, a unit vector in the $z$-direction, and $$M_0^2={ m_1^2+k_\bot^2\over (1-x)}+{ m_2^2+k_\bot^2\over x}.
\label{M0}$$ In practice, it is more convenient to use the covariant form for $R^{SS_z}_{\lambda_1\lambda_2}$ [@Jaus]: $$R^{SS_z}_{\lambda_1\lambda_2}(x,k_\bot)
={\sqrt{p_1^+p_2^+}\over \sqrt{2} ~{\widetilde M_0}}
~\bar u(p_1,\lambda_1)\Gamma v(p_2,\lambda_2), \label{covariant}$$ where $$\begin{aligned}
&&{\widetilde M_0} \equiv \sqrt{M_0^2-(m_1-m_2)^2}, \nonumber\\
&&\Gamma=\gamma_5 \qquad ({\rm pseudoscalar}, S=0). \non \end{aligned}$$ We normalize the meson state as $$\langle M(P',S',S'_z)|M(P,S,S_z)\rangle = 2(2\pi)^3 P^+
\delta^3(\tilde P'- \tilde P)\delta_{S'S}\delta_{S'_zS_z}~,
\label{wavenor}$$ so that the normalization condition of the momentum distribution function can be obtained $$\int \{dx\}~|\phi(x,k_\bot)|^2 = 1,
\label{momnor}$$ where {dx}In principle, the momentum distribution amplitude $\phi(x,k_\bot)$ can be obtained by solving the light-front QCD bound state equation [@Zhang]. However, before such first-principles solutions are available, we would have to be contented with phenomenological amplitudes. One example that has been often used in the literature for heavy mesons is the Gaussian-type wavefunction, $$\phi(x,k_\bot)_{\rm G}={\cal N} \sqrt{{dk_z\over dx}}
~{\rm exp}\left(-{\vec k^2\over 2\omega^2}\right),
\label{gauss}$$ where ${\cal N}=4(\pi/\omega^2)^{3/4}$ and $k_z$ is of the internal momentum $\vec k=(\vec{k}_\bot, k_z)$, defined through $$1-x = {e_1-k_z\over e_1 + e_2}, \qquad
x = {e_2+k_z \over e_1 + e_2},$$ with $e_i = \sqrt{m_i^2 + \vec k^2}$. We then have M\_0=e\_1 + e\_2, k\_z = [xM\_02]{}-[m\_2\^2+k\_\^2 2 xM\_0]{}, \[kz\] and $${{dk_z\over dx}} = \,{e_1 e_2\over x(1-x)M_0},$$ which is the Jacobian of transformation from $(x, k_\bot)$ to $\vec k$.
Decay Constants
---------------
The decay constant of a pseudoscalar meson $P(q_1\bar{q}_2)$ is defined by 0|A\_|P(p)=if\_Pp\_, \[dcd18\] where $A_\mu$ is the axial-vector current. It can be evaluated using the light-front wavefunction given by (\[gauss\]) 0||[q]{}\_2\^+\_5q\_1|P&=& {d\^3p\_1}{d\^3p\_2} 2(2)\^3\^3(p-p\_1-p\_2)\_P(x,k\_)R\^[00]{}\_[\_1\_2]{} (x,k\_)\
&& 0||[q]{}\_2\^+\_5q\_1|q\_1|[q]{}\_2. Since $\widetilde{M}_0\sqrt{x(1-x)}=\sqrt{{\cal A}^2+k^2_\perp}$, it is straightforward to show that f\_P=4{dx}[\_P(x, k\_)]{}[A]{}, \[fp\] where ${\cal A}=m_1x+m_2(1-x)$. Note that the factor $\sqrt{3}$ in (\[fp\]) arises from the color factor implicitly in the meson wavefunction.
Electromagnetic Form Factors
----------------------------
The EM form factor of a pseudoscalar meson $P$, $F_P(Q^2)$, is determined by the scattering of one virtual photon and one meson. It describes the deviation from the point-like structure of the mesonch, and is a function of the square of the photon momentum $Q$. Here we consider the momentum of the virtual photon in space-like region, so it is always possible to orient the axes in such a manner that $Q^+=0$. Thus the EM form factor is determined by the matrix element P(P’)|J\^+|P(P)= e F\_P(Q\^2) (P+P’)\^+, \[FPdef\] where $J^\mu = \bar q e_q e\gamma^\mu q$ is the vector current, $e_q$ is the charge of quark $q$ in $e$ unit, and $Q^2=-(P'-P)^2\geq 0$. With LFQM, $F_P$ can be extracted by Eq. (\[FPdef\]) F\_P(Q\^2)&=&e\_[q\_1]{}{dx} [\_P(x,k\_)]{}[\_[P’]{}(x,k’\_)]{}\
&+&e\_[|[q\_2]{}]{}{dx}[\_P(x,k\_)]{}[\_[P’]{}(x,k”\_)]{}, \[FPgeneral\] where $k'_\perp=k_\perp+x Q_\perp$, $k''_\perp=k_\perp-(1-x) Q_\perp$. From Eqs. (\[RR\]), (\[Melosh2\]), and (\[covariant\]), it is understandable that the term $\sqrt{{\cal A}^2+k_\perp^2}$ comes from the Melosh transformation. After fixing the parameters which appear in the wavefunction, Eq. (\[FPgeneral\]) can be used to fit the experimental data. But this is not the whole story. We consider the term $\widetilde {\phi}_{P} \equiv {\phi_{P}(x,k_\perp)/{\sqrt{{\cal A}^2+k^2_\perp}}}$ and take the Tayor expansion around $k^2_\perp$ \_[P’]{}(k’\^2\_)=\_[P’]{}(k\^2\_)+[d\_[P’]{}]{}|\_[Q\_=0]{}(k’\^2\_-k\^2\_)+[d\^2\_[P’]{}]{}|\_[Q\_=0]{}(k’\^2\_-k\^2\_)\^2+..... Then, by using the idenity d\^2k\_ (k\_A\_)(k\_B\_)=[1]{}d\^2k\_ k\^2\_ A\_B\_, \[QQQ2\] we can rewrite (\[FPgeneral\]) as F\_P(Q\^2)&=&(e\_[q\_1]{}+e\_[|q\_2]{})\
&+&Q\^2 {dx}\^2\_P(x,k\_)\[x\^2 e\_[q\_1]{}+(1-x)\^2e\_[|q\_2]{}\]( \_[P]{}[[A]{}\^2+2k\^2\_]{}+\_[P]{}k\^2\_)\
&+&[O]{}(Q\^4), \[FPQQ\] where \_[M]{}=[1]{}([d\_[M]{}]{}), \_[M]{}=[1]{}([d\^2\_[M]{}]{}). From Eq. (\[FPQQ\]), the static property $F_P(0)=e_P$ is quite easily checked. The mean square radius of the meson $P$ is determined from the slope of $F_P$ at $Q^2=0$: r\^2 \_P-6[dF\_P(Q\^2)]{}|\_[Q\^2=0]{}. It should be realized that the size and the density of a hadron depend on the probe. For an electromagnetic probe, it is the electric charge radius that is obtained. From the experimental view, $\la r^2 \ra_P$ cannot be measured directly and is obtained by fitting the data on $F_P$ to a pole or dipole form. Here we easily obtained the equation of $\la r^2 \ra_P$ r\^2 \_P &=& r\^2 \_[q\_1]{}+r\^2 \_[[|q]{}\_2]{}\
&=&e\_[q\_1]{}{-6{dx}x\^2\_[P]{}\_[P]{}},\
&+&e\_[|[q]{}\_2]{}{-6{dx}(1-x)\^2\_[P]{}\_[P]{}}.\[MSRP\] From Eq. (\[MSRP\]), it is worthwhile to mention that, first, the mean square radius of a meson is the sum of the contributions of the valence quarks. Second, $\la r^2 \ra$ is related to the first and second longitudinal momentum square derivatives of $\widetilde {\phi}$ which contain the Melosh transformation effect.
Heavy Quark Limit
=================
In this section, we will check the HQS among the charge radii by taking the heavy quark limit. To proceed, we first investigate the heavy-quark-limit behavior of the wavefunction. Since the $x$ in the normalization condition (\[wavenor\]) is the longitudinal momentum fraction carried by the light antiquark, the meson wavefunction should be sharply peaked near $x\sim \Lambda_{\rm QCD}/m_Q$. It is thus clear that only terms of the form “$m_Qx$" survive in the wavefunction as $m_Q\to\infty$; that is, $m_Qx$ is independent of $m_Q$ in the $m_Q\to\infty$ limit. In the $m_Q\to\infty$ limit, we must rewrite Eq. (\[wavenor\]) in the $m_Q$-independent form \^\_0dX|(X,k\_) |\^2=1, \[isnom\] where $X\equiv m_Qx$ and [@CZL] (X,k\_)=[\_[Q\_[|[q]{}]{}]{}(x,k\_)]{}. \[hqamp\] The scaling behavior of Eq. (\[hqamp\]) is the constraint of the light-front wavefunction when we consider the infinite quark mass limit. For the Gaussian-type wavefunction (\[gauss\]), it satisfies an asymptotic form (X,k\_)\_[G]{} = 4([\^2]{})\^[3/4]{}(-[k\^2\_2\^2]{})(-[([X2]{}-[m\_[|[q]{}]{}\^2+ k\^2\_2X]{})\^22\^2]{}). \[GHQ\] Thus we can use this wavefunction when the heavy quark limit is considered.
In the $m_M,m_Q \to \infty$ limit it is appropriate to describe the meson state with the meson velocity $v$ [@HQS] |M(v)= m\_M\^[-1/2]{} |M(P), where $v=P/m_M$. For the decay constant, the definition (\[dcd18\]) becomes 0||[q]{}\_\_5 Q|P(v)&=& i|[f\_P]{}v\_, \[dcHV\] and in the $m_Q \to \infty$ limit it is |[f\_P]{}= 4(X,k\_), \[fPV\] where $\widetilde {\cal A} \equiv X+m_{{\bar q}_2}$. Comparing Eq. (\[fPV\]) with Eq. (\[fp\]), we obtain the HQS scaling law for the decay constant: |[f\_P]{} = f\_P. \[scale\] For the mean square radius Eqs. (\[MSRP\]), when the heavy quark limit is considered, we obtain r\^2\_P = r\^2\_Q+r\^2\_[|[q\_2]{}]{}, \[rPV\] where r\^2 \_Q &=&e\_Q{[-6]{} X\^2}\
&& 0, \[HQL1\] r\^2 \_[|[q]{}\_2]{} = e\_[|[q]{}\_2]{}{-6 },\[HQL2\] and $\widetilde {\Phi} ={\Phi}/\sqrt{\widetilde {\cal A}^2+k^2_\perp}$. Eq. (\[HQL1\]) means that the mean square radius $\la r^2 \ra_P$ is blind to the flavor of $Q$. This is the so-called flavor symmetry. We find that the light degrees of freedom are blind to the flavor of the heavy quark. In addition, Ref. [@plbhwcw] finds the mean square radius also satisfied the spin symmetry. These are the so-called HQS. Up to now, we have not used the wavefunction yet, this also satisfies the well-known property that HQS is model-independent. Reviewing the processes, we can realize that, in this approach, the static properties of the EM form factors and the heavy quark symmetry of the mean square radii can be checked much more easily than in the Breit frame. This is the major reason why we calculate the $Q^2$ dependence of those form factors order by order.
We must emphasized here that, in the $m_Q \to \infty$ limit, the vanishing of the heavy quark sector in the form factor is true only for the $Q^2 \to 0$ region. In the time-like region, near the threshold for the meson pair production the heavy quark sector is dominant and described by the Isgur-Wise function [@MR].
Numerical Results
=================
In this section, we will use the Gaussian-type wavefunction (\[gauss\]) to calculate the EM form factors and the mean square radius. The parameters appearing in the wavefunction, the quark mass $m_q$ and the scale parameter $\omega$, are constrained by the decay constants.
The decay constants of the pseudoscalar mesons $\pi$ and $K$ come from experiments [@PDG00] f\_= 130.7 [MeV]{}, f\_K=159.8 [MeV]{}, the others are obtained by lattice and constituent quark model: f\_D &=& 192 [MeV]{} [@lattprl], f\_[D\_s]{}=210 [MeV]{}[@lattprl], f\_B=157 [MeV]{} [@lattprl],\
f\_[B\_s]{} &=& 171 [MeV]{} [@lattprl], f\_[B\_c]{}=360 [MeV]{} [@QCDSR]. Combining with the quark masses m\_[u,d]{}=0.24 [GeV]{}, m\_s-m\_[u,d]{}=0.18 [GeV]{}, m\_c=1.6 [GeV]{}, m\_b=4.8 [GeV]{}, \[qmass\] we fit the scale parameters &&\_=0.333 [GeV]{}, \_K=0.379 [GeV]{}, \_D=0.443 [GeV]{}, \_[D\_s]{}=0.450 [GeV]{},\
&&\_B=0.477 [GeV]{}, \_[B\_s]{}=0.485 [GeV]{}, \_[B\_c]{}=0.813 [GeV]{}. \[omega\] There are differences between these parameters and the ones in [@QCDSR] because the wavefunctions in the two cases are not the same. However, they have a common tendency such that $\omega_{M_i} < \omega_{M_j}$ if $M_i < M_j$. This corresponds to the ordering law for the size of heavy-light bound states.
The $Q^2$-dependences of $F_\pi$ and $F_K$ can be obtained by Eq. (\[FPgeneral\]), and we compare the results with the data in Fig. 1 and 2, respectively. In addition, the mean square radii of the pseudoscalar meson can be obtained by Eq. (\[MSRP\]). We list the results of the $\la r^2 \ra_{\pi^+,K^+,K^0}$ and the experimental data in Table 1. (the unit is $\rm{fm}^2$)
0.5cm
$\la r^2 \ra$ $\pi^+$ $K^+$ $K^0$
--------------- ------------------------ ---------------------- ------------------------
this work $0.443$ $0.349$ $-0.0676$
$0.314$ $0.240$ $-0.020$
$0.452$ $0.38$ $0.057$
experiment $0.439\pm 0.008$ \[1\] $0.34\pm 0.05$ \[5\] $-0.054\pm0.026$ \[7\]
[Table 1. The mean square radii of the $\pi^+$, $K^+$, and $K^0$ mesons.]{}
The negative signs in Table 1 are interesting, and may be interpreted as the preponderance of negative electric charge in the tail of the distribution. We find these values are all consistent with the data. Comparing with Ref. [@The0], they also used the light-front approach. There were various parameter combinations to fit the data of $F_\pi$ for both small and large momentum transfers.
According to the vector meson dominance (VMD) model [@The4], there is a physical explanation: the pion form factor is determined by a $\rho$-meson pole. Generally speaking, this simple picture fits the data well. A detailed study [@The5] obtained a better fit when one considers the $\rho$-$\omega$ mixing and three vector meson ($\rho$, $\omega$, and $\phi$) poles to the pion and kaon form factors, respectively.
On the other hand, the mean square radii of the heavy pseudoscalar meson have not been measured yet. For comparsion, here we define and calculate them as $\la r^2 \ra_{FM}$ for the finite quark masses and as $\la r^2 \ra_{IM}$ for the infinite quark masses. In the case of the infinite quark masses, the decay constant $\bar{f_P}$ cannot be measured in the true world, so we obtain it approximately by using the values $f_B=157~{\rm MeV}$ and $m_B=5.28~{\rm GeV}$ in Eq. (\[scale\]). The results are listed in Table 2. 0.5cm
$D^+$ $D^0$ $D_s^+$ $B^+$ $B^0$ $B_s^0$ $B_c^+$
-------------------- --------- ---------- --------- --------- ---------- ---------- ----------
$\la r^2 \ra_{FM}$ $0.184$ $-0.304$ $0.124$ $0.378$ $-0.187$ $-0.119$ $0.0433$
$\la r^2 \ra_{IM}$ $0.248$ $-0.496$ $0.181$ $0.496$ $-0.248$ $-0.181$
[Table 2. The mean square radii of the heavy pseudoscalar mesons for the finite\
quark masses $\la r^2 \ra_{FM}$ and for the infinite quark masses $\la r^2 \ra_{IM}$.]{}
From Table 2, we cannot obviously find the situation that, comparing with the values in $D_q$ system, the ones in the $B_q$ system are closer to those in the infinite-quark-mass system. The reason is that the $\la r^2 \ra$ is sensitive to the $f_P$, but the uncertainty of the decay constant is not small. In fact, if we use the most recent value $f_{D_s}=280~{\rm MeV}$ [@chada], the result $\la r^2 \ra_{D^+_s}=0.083~{\rm fm}^2$ is quite different from the one in Table 2. For the $B_c$ meson, the $\la r^2 \ra_{IM}$ are not given here because both $b$ and $c$ quarks are heavy. The HQS must be reconsidered in this case.
Conclusion
==========
We have calculated the EM form factors of the pseudoscalar mesons. The EM form factors are extracted from the relevant matrix elements directly, instead of choosing the Breit frame. We found that the charge radius is related to both the first and second longitudinal momentum square derivatives of the momentum distribution function. We also found that the static properties of the EM form factors and the heavy flavor symmetry of the mean square radii are checked analytically by evaluating the $Q^2$ dependence of those form factors order by order. Therefore, in the heavy quark limit, the charge radii of pseudoscalar have flavor symmetries, and these properties are model-independent. In addition, The $Q^2$-dependences of the form factors $F_{\pi,K}$ and the mean square radius of light and heavy mesons have been calculated by using the Gaussian-type wavefunction. The form factors $F_\pi$ and $F_K$ in small momentum transfer and the values of $\la r^2 \ra_{\pi^+,K^+,K^0}$ are all consistent with the current experimental data. 1.0cm
**ACKNOWLEDGMENTS**
I would like to thank Kingman Cheung for helpful comments. This work was supported in part by the National Science Council of ROC under Contract No. NSC90-2112-M-007-040.
[99]{} [Exp0]{} S.R. Amendolia [*et al.*]{} (NA7 Coll.), , 168 (1986).
[Exp1]{} C.J. Bebek [*et al.*]{}, , 1693 (1978).
[Exp2]{} J. Volmer [*et al.*]{}, , 1713 (2001).
[Exp3]{} P. Brauel [*et al.*]{}, , 101 (1979).
[Exp4]{} S.R. Amendolia [*et al.*]{}, , 435 (1986).
[Exp5]{} E.B. Dally [*et al.*]{}, , 232 (1980).
[Exp6]{} W.R. Molzon [*et al.*]{}, , 1213 (1978).
[The0]{} P.L. Chung, F. Coester, and W.N. Polyzou, , 545 (1988).
[The1]{} C.W. Hwang, Eur. Phys. J. [**C 19**]{}, 105 (2001).
[The2]{} Y.C. Shin [*et al.*]{}, Eur. Phys. J. [**A 9**]{}, 269 (2000).
[The3]{} C.J. Burden, C.D. Roberts, and M.J. Thomson, , 163 (1996).
[The4]{} G.J. Gounaris and J.J. Sakurai, , 244 (1968).
[The5]{} Jun Gao and Bing An Li, , 113006 (2000).
[HP]{} F.T. Hawes and M.A. Pichowsky, , 1743 (1999).
[Simula]{} F. Cardarelli [*et al.*]{}, , 393 (1995).
[HQS]{} N. Isgur amd M.B. Wise, , 113 (1989); N. Isgur amd M.B. Wise, [**B 237**]{}, 527 (1990); H. Georgi, , 447 (1990). For a review, see M. Neubert, Phys. Rept. [**245**]{}, 259 (1994).
F. Cardarelli [*et al.*]{}, , 6682 (1996).
W. Jaus, , 3394 (1990); [*ibid.*]{} [**D 44**]{},2851 (1991); , 611 (1992).
P.J. O’Donnell [*et al.*]{}, , 219 (1994); , 113 (1994); , 3966 (1995).
N.B. Demchuk [*et al.*]{}, Phys. Atom. Nucl [**59**]{}, 2152 (1996).
C.Y. Cheung, C.W. Hwang, and W.M. Zhang, , 657 (1997).
H.Y. Cheng, C.Y. Cheung, and C.W. Hwang, , 1559 (1997).
[GHLZ]{} C.Q. Geng, C.W. Hwang, C.C. Lih, and W.M. Zhang, , 114024 (2001).
[Melosh]{} H.J. Melosh, , 1095 (1974).
W.M. Zhang, Chin. J. Phys. [**31**]{}, 717 (1994); , 1528 (1997).
[CZL]{} C.Y. Cheung, W.M. Zhang, and G.L. Lin, , 2915 (1995).
[plbhwcw]{} C.W. Hwang, , 65 (2001).
[MR]{} T. Mannel and Z. Ryzak, , 412 (1990).
[PDG00]{} D.E. Groom [*et al.*]{} (Particle Data Group), Eur. Phys. J. [**C 15**]{}, 1 (2000).
[lattprl]{} C. Bernard [*et al.*]{}, Phys. Rev. Lett. [**81**]{}, 4812 (1998).
[QCDSR]{} M.A. Ivanov, J.G. K$\ddot{\rm o}$rner, and P. Santorelli, Phys. Rev. [**D 63**]{}, 074010 (2001).
[chada]{} M. Chada [*et al.*]{}, Phys. Rev. [**D 58**]{}, 032002 (1998).
=0 cm
**FIGURE CAPTIONS**
0.5 true cm
[**Fig. 1** ]{} The charge form factor of the pion in small momentum transfer. Data are taken from [@Exp0]. 0.25 true cm
[**Fig. 1** ]{} The charge form factor of the Kaon in small momentum transfer. Data are taken from [@Exp4]. 0.25 true cm
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A program for computing pure angular momentum coefficients in relativistic atomic structure for any scalar one– and two–particle operator is presented. The program, written in Fortran 90/95 and based on techniques of second quantization, irreducible tensorial operators, quasispin and the theory of angular momentum, is intended to replace existing angular coefficient modules from GRASP92. The new module uses a different decomposition of the coefficients as sums of products of pure angular momentum coefficients, which depend only on the tensor rank of the interaction but not on its details, with effective interaction strengths of specific interactions. This saves memory and reduces the computational cost of big calculations signficantly.'
author:
- |
Gediminas Gaigalas$^{\, a,b}$, Stephan Fritzsche$^{a}$ and Ian P. Grant$^{c}$\
\
\
$^a$ Fachbereich Physik, Universität Kassel,\
Heinrich–Plett–Str. 40, D–34132 Kassel, Germany.\
$^b$ Institute of Theoretical Physics and Astronomy,\
A. Goštauto 12, Vilnius 2600, Lithuania.\
$^c$ Mathematical Institute, University of Oxford,\
24/29 St. Giles’, Oxford OX1 3LB, UK.\
title: 'Program to calculate pure angular momentum coefficients in $jj$–coupling'
---
[**PROGRAM SUMMARY**]{}
[*Title of program:*]{} ANCO
[*Catalogue number:*]{} ADOO
[*Program obtainable from:*]{} CPC Program Library, Queen’s University of Belfast, N. Ireland. Users may obtain the program also by down–loading the tar–file `ratip-anco.tar` from our home page at the University of Kassel (http://www.physik.uni-kassel.de/fritzsche/programs.html).
[*Program Summary URL:*]{} http://cpc.cs.qub.ac.uk/summaries/ADOO
[*Licensing provisions:*]{} None.
[*Computer for which the program is designed and has been tested:*]{} IBM RS 6000, PC Pentium II. University of Kassel (Germany). IBM AIX 4.1.2+, Linux 6.1+.
[*Program language used in the new version:*]{} ANSI standard Fortran 90/95.
[*Memory required to execute with typical data:*]{} 100 kB.
[*No. of bits in a word:*]{} All real variables are parametrized by a `selected kind parameter`. Currently this is set to double precision for consistency with other components of the RATIP package \[1\].
[*Distribution format:*]{} Compressed tar file. On a UNIX (or compatible) workstation, the command `uncompress` uncompress this file and the command `tar -xvf ratip_anco.tar` reconstructs the file structure.
[*Keywords:*]{} atomic many–body perturbation theory, complex atom, configuration interaction, effective Hamiltonian, energy level, Racah algebra, reduced coefficients of fractional parentage, reduced matrix element, relativistic, second quantization, standard unit tensors, tensor operators, $9/2$–subshell.
[*Nature of the physical problem:*]{}
The matrix elements of a one–electron tensor operator $\widehat{A}^k$ of rank $k$ with respect to a set of configuration state functions $|\gamma_iJ_i\rangle$ can be written $\sum_{ab} t^{k}_{ij}(ab)\,
(a|\widehat{A}^k |b)$ where the angular coefficients $t^{k}_{ij}(ab)$ are independent of the operator $\widehat{A}^k$, $i,j$ are CSF labels and $a,b$ run over the relevant interacting orbital labels. Similarly, the matrix elements of the Dirac–Coulomb Hamiltonian can be written in the form $\sum_{ab} t^{0}_{ij}(ab)\,
(a|\widehat{H}_D |b) +\sum_k \sum_{abcd} v^k_{ij}(abcd)\,
X^k(abcd)$, where $\widehat{H}_D$ is the one-electron Dirac Hamiltonian operator, with tensor rank zero, $v^k_{ij}(abcd)$ are pure angular momentum coefficients for two–electron interactions, and $X^k(abcd)$ denotes an effective interaction strength for the two electron interaction. The effective interaction strengths for Coulomb and Breit interaction have different selection rules and make use of subsets of the full set of coefficients $v^k_{ij}(abcd)$.
Such matrix elements are required for the theoretical determination of atomic energy levels, orbitals and radiative transition data in relativistic atomic structure theory. The code is intended for use in configuration interaction or multiconfiguration Dirac–Fock calculations \[2\], or for calculation of matrix elements of the effective Hamiltonian in many–body perturbation theory \[3\].
[*Method of solution:*]{} A combination of second quantization and quasispin methods with the theory of angular momentum and irreducible tensor operators leads to a more efficient evaluation of (many–particle) matrix elements and to faster execution of the code \[4\].
[*Restrictions on the complexity of the problem:*]{} Tables of reduced matrix elements of the tensor operators $a^{(q~j)}$ and $W^{(k_q k_j)}$ are provided for ($nj$) with $j = 1/2, 3/2, 5/2, 7/2$, and $9/2$. Users wishing to extend the tables must provide the necessary data.
[*Typical running time:*]{} 3.5 seconds for both examples on a 450 MHz Pentum III processor.
[*Unusual features of the program:*]{} The program is designed for large–scale atomic structure calculations and its computational cost is less than that of the corresponding angular modules of GRASP92. The present version of the program generates pure angular momentum coefficients $t^{0}_{ij}(ab)$ and $v^k_{ij}(abcd)$, but coefficients $t^{k}_{ij}(ab)$ with $k>0$ have not been enabled. An option is provided for generating coefficients compatible with existing GRASP92.
Configurational states with any distribution of electrons in shells with $j \leq 9/2$ are allowed. This pemit user to take into account the single, double, triple excitations form open $d$– and $f$– shells for the systematic MCDF studies of heavy and superheavy elementds (Z $>$ 95).
[*References:*]{} \[1\] S. Fritzsche, C. F. Fischer, and C. Z. Dong, Comput. Phys. Commun. 124 (1999) 240. \[2\] I. P. Grant, Methods of Computational Chemistry, Vol 2. (ed. S. Wilson) pp. 1-71 (New York, Plenum Press, 1988); K. G. Dyall, I. P. Grant, C. T. Johnson, F. A. Parpia and E. P. Plummer, Comput. Phys. Commun. 55 (1989) 425; F. A. Parpia, C. Froese Fischer and I. P. Grant, Comput. Phys. Commun. 92 (1996) 249. \[3\] G. Merkelis, G. Gaigalas, J. Kaniauskas, and Z. Rudzikas, Izvest. Acad. Nauk SSSR, Phys. Series 50 (1986) 1403. \[4\] G. Gaigalas, Lithuanian Journal of Physics 39 (1999) 80.
[**LONG WRITE–UP**]{}
Introduction
============
The improved accuracy of modern experiments challenges theorists to match or exceed experimental precision. Models of many–electron atoms and ions require both relativistic and correlation effects to be taken into account; this can be done, for example, by using various versions of perturbation theory, the configuration interaction method, the multiconfiguration Hartree–Fock method [@kn:fa] or the multiconfiguration Dirac–Fock method [@kn:gk].
The evaluation of matrices of one– and two–electron operators for many–electron states in $jj$–coupling is customarily done by expressing each matrix element as a sum of products of angular coefficients and radial integrals. This strategy, based on earlier work by Fano, was adopted for the MCP program [@MCP; @MCP75] for evaluating angular coefficients for the Coulomb interaction and the related MCBP program [@MCBP] for evaluating angular coefficients for the Breit interaction. Whilst this strategy was adequate on the relatively low–powered computers of the 1970s, it is now possible and desirable to use very large configuration sets which require a more efficient strategy. The new program attains this objective by building up the angular coefficients for one– and two–electron interactions from a relatively small number of common spin-angular parts in the manner of [@method2; @method6; @method7].
The theoretical background is presented in Section 2, program organization in section 3, testing and timing studies in section 4 and simple test problems in Section 5.
Theoretical background
======================
Dirac–Coulomb Hamiltonian
-------------------------
As usual, multiconfiguration self–consistent–field calculations in relativistic atomic theory are based on the Dirac–Coulomb hamiltonian: $$\begin{aligned}
\label{eq:DC-a}
{\displaystyle}\widehat{H}_{DC}=
{\sum_{i = 1}^{N}}
\widehat{H}_{D}(i) +
\frac{1}{2} {\sum_{i,j = 1}^{N}}
\frac{1}{r_{ij}}.\end{aligned}$$ $\widehat{H}_{D}(i)$ is the Dirac one–particle operator, $$\begin{aligned}
\label{eq:DC-b}
{\displaystyle}\widehat{H}_{D}=
c {{\mbox{\boldmath{$\alpha $}}}}(i) \cdot {{\mbox{\boldmath{$p $}}}}_i + [\beta(i) -1] c^2
-\frac{Z}{r_i},\end{aligned}$$ and the second term in (\[eq:DC-a\]) represents the Coulomb interaction of pairs of electrons. In equation (\[eq:DC-b\]) $c$ denotes the speed of light; ${{\mbox{\boldmath{$\alpha $}}}}(i)$ and $\beta(i)$ are 4 $\times$ 4 Dirac matrices for the $i$th electron; $r_i$ and ${{\mbox{\boldmath{$p $}}}}_i$ are the radial coordinate of the $i$th electron and its (3–) momentum, respectively. The first two terms of (\[eq:DC-b\]) comprise the Dirac kinetic energy operator.
Both configuration interaction and multiconfiguration Dirac–Hartree–Fock calculations require the matrix elements of $\widehat{H}_{DC}$ with respect to a basis of $n$–electron configurational states labelled $|\gamma_r J_r\rangle, r =
1,2,3,\ldots$ The present program computes the pure angular coefficients $t_{rs}^k(ab)$ (only for $k=0$ in the present version) and $v_{rs}^k(abcd))$ in the expression $$\label{eq:DC-c1}
{\displaystyle}\langle \gamma_r J_r|\widehat{H}_{DC}|\gamma_s J_s\rangle =
\sum_{ab}\left\{ t^0_{rs}(ab)\,(a|\widehat{H}_D|b) + {\sum_{k}}
\sum_{cd} v^{(k)}_{rs}(abcd)\,X^{(k)}(abcd)\right\},$$ where $X^{(k)}(abcd)$ is the effective interaction strength of the two–electron interaction with respect to the orbitals concerned. Here in (\[eq:DC-c1\]), the interaction is the Coulomb potential only, where $$\begin{aligned}
\label{eis-c}
X^{(k)}(abcd)\\
& & = (-1)^k \langle n_a l_a j_a \| C^{(k)} \| n_{c} l_{c}
j_{c} \rangle \langle n_b l_b j_b \| C^{(k)} \|
n_{d} l_{d} j_{d} \rangle R^k(n_a l_a j_a
n_b l_b j_b n_{c} l_{c} j_{c}
n_{d} l_{d} j_{d} ) , \nonumber\end{aligned}$$ but the formalism can also be used for the Breit interaction, with a different definition of $X^{(k)}(abcd)$ and different selection rules. Equation (\[eq:DC-c1\]) is a rearrangement of the formula used in GRASP92 [@GRASP92 eqn. (3.10)], $$\begin{aligned}
\label{eq:DC-c}
{\displaystyle}\langle \gamma_r J_r|\widehat{H}_{DC}|\gamma_s J_s\rangle =
\sum_{ab}\left\{ T_{rs}(ab)I(ab) + {\sum_{k}}
\sum_{cd}
V^{(k)}_{rs}(abcd)\,R^{(k)}(abcd)\right\},\end{aligned}$$ where $$I(ab) = \left(a|\widehat{H}_D|b \right)$$ and $$R^{(k)}(abcd) = \left(ab \left|{r_<}^k/{r_>}^{k+1}\right|cd\right)$$ is a relativistic Slater integral over orbitals $a,b,c,d$ (in the usual notation). Whilst the one–electron part is the same, equation (\[eq:DC-c\]) expands the effective interaction strength as a traditional sum of Slater integrals, and $V^{(k)}_{rs}(abcd)$ therefore differs from $v^{(k)}_{rs}(abcd)$. The latter can be used for *any* two–electron interaction, and it is no longer necessary to treat Coulomb and Breit interactions on a different footing in the manner of GRASP92. Other intractions such as the lowest–order normal mass shift and the specific mass shift [@GRASP92 eqns. (3.12–3.14)] can be handled within the same scheme. The way in which the coefficients are built up is described below.
One–particle operators
----------------------
The matrix elements of a one–particle scalar operator $\widehat{F}^{(0)}$ between configuration state functions with $u$ open shells can be expressed as a sum over one–electron contributions $$\begin{aligned}
\label{eq:one-a}
{\displaystyle}\left (\psi _u^{bra}( J)\left \| \widehat{F}^{(0)}\right\|
\psi _u^{ket}( J^{\prime }) \right ) =
{\sum_{n_i\kappa _i,n_j\kappa _j}}
\left (\psi _u^{bra}( J) \left\| \widehat{F}
( n_i \kappa _i, n_j \kappa _j )
\right\| \psi _u^{ket}( J^{\prime }) \right)\end{aligned}$$ where $$\begin{aligned}
\label{eq:one-b}
\lefteqn{{\displaystyle}\left (\psi _u^{bra}( J) \left\| \widehat{F}
( n_i \kappa _i, n_j \kappa _j )
\right\| \psi _u^{ket}( J^{\prime }) \right )}
\nonumber \\[1ex]
& & =
{\displaystyle}( -1)^{\Delta +1} \sqrt{2j_i +1}\; R\left( j_i, j_j,\Lambda
^{bra},\Lambda ^{ket} \right) \, \delta( \kappa _i , \kappa _j)\,
\left( n_i\kappa _i\left\|f^{(0) } \right\|n_j\kappa _j \right)
\nonumber \\[1ex]
& & \times \left\{ \delta ( n_i , n_j )
\left( j_{i}^{N_i}\;\alpha _i Q_i J_i \left\|
\left[ a^{\left( q \; \; j_i\right)}_{1/2}
\times a^{\left( q \; \; j _i
\right)} _{-1/2}\right] ^{\left( 0 \right) }
\right\|j_{i}^{N_{i}} \;\alpha _i Q_i J_i\right)
\nonumber \right. \\[1ex]
& & \left. + (1-\delta ( n_i , n_j ))
\left( j_{i}^{N_i}\,\alpha _i Q_i J_i \left\|
a^{( q \,j _i)}_{1/2}
\right\|j_i^{N_{i}^{\prime }} \,\alpha _i Q_i J_i \right)
\left( j_j^{N_j}\;\alpha _j Q_j J_j \left\|
a^{( q \, j _j)}_{-1/2}
\right\|j_j^{N_{j}^{\prime }} \,\alpha_j Q_j J_j \right)
\right\}.
\nonumber\end{aligned}$$ All states are defined in $jj$–coupling. $(\psi _u^{bra}\left(
J\right) ||$ and $||\psi _u^{ket}\left( J^{\prime }
\right) )$ are respectively bra and ket functions with $u$ open subshells, $ \kappa \equiv (2j+1)(l-j)$, $\left(n_i\kappa _i
\left\|f^{( 0 ) }\right\|n_j\kappa _j\right) $ is the one–electron reduced matrix element of the operator $\widehat{F}^{\left( 0 \right)}$, $\Lambda^{bra} \equiv
\left( J_i,J_j,J_{i^{\prime }},J_{j^{\prime }}\right)^{bra}$ and $\Lambda^{ket}\equiv
\left( J_i,J_j,J_{i^{\prime }},J_{j^{\prime }}\right)^{ket}$ denote the respective sets of active subshell angular momenta. The operators $a^{( q \, j)}_{m_q}$ are second quantization operators in quasispin space of rank $q = 1/2$. The operator $a^{( q \, j)}_{1/2 \; m_j} =
a^{\left( j \right) }_{m_j}$ creates electrons with angular momentum quantum numbers $j,m_j$ and its conjugate $a^{( q \, j)}_{-1/2 \, m_j} =
\tilde{a}_{m_j }^{( j )} =
( -1)^{j-m_j }a_{-m_j }^{( j ) +}$ annihilates electrons with the same quantum numbers $j,m_j$ in a given subshell.
### Recoupling matrix
The recoupling matrix $R\left( j_i, j_j,\Lambda ^{bra},\Lambda ^{ket} \right)$ in (\[eq:one-a\]) is particularly simple. It is either a product of delta functions [@method2 eqn. (18)] when $n_i\kappa _i = n_j \kappa _j$ or a combination of delta functions and $6j-$ coefficients [@method2 eqn. (22)] when $n_i\kappa _i \neq n_j\kappa _j $.
### Matrix elements of irreducible tensor operators
By applying the Wigner–Eckart theorem in quasispin space we obtain the submatrix elements of operators of type $a_{m_q}^{( q j ) }$ in the form [@method1] $$\begin{aligned}
\label{eq:theo-g}
\lefteqn{
\left( j^N\,\alpha QJ\left\|a_{m_q}^{( q\;j) }\right\|j^{N^{\prime }}
\,\alpha ^{\prime }Q^{\prime }J^{\prime }\right)
\nonumber} \\
& & {\displaystyle}= -[Q] ^{-1/2}\left[
\begin{array}{ccc}
Q^{\prime } & 1/2 & Q \\
M_Q^{\prime } & m_q & M_Q
\end{array}
\right]
\left( j^N \,\alpha QJ {\left|\!\left|\!\left|}a^{( q \; j)} {\right|\!\right|\!\right|}j\,\alpha^{\prime} Q^{\prime }J^{\prime}\right),\end{aligned}$$ where we have used the conventional shorthand notation $(2k+1)\cdot
...\equiv [k,...]$, and the last factor is a reduced coefficient of fractional parentage. The submatrix elements of the simplest compound tensor operator of type ${\displaystyle}\left[ a^{( q \; j)}_{m_{q2}} \times a^{( q \; j)} _{m_{q2}}
\right]^{( k_j)}$ uses $$\begin{aligned}
\label{eq:theo-gb}
\lefteqn{
\left( nj^N\,\alpha QJ \left\|
\left[ a_{m_{q1}}^{( q \; j) } \times
a_{m_{q2}}^{( q \; j) } \right]^{( k_j) }
\right\|nj^{N^{\prime }}\;\alpha ^{\prime}Q^{\prime}J^{\prime} \right)}
\nonumber \\[1ex]
& & =
{\displaystyle}{\sum_{ k_q, m_q }}\left[ Q\right]^{-1/2}
\left[ \begin{array}{ccc}
q & q & k_q \\
m_{q1} & m_{q2} & m_q \end{array} \right]
\left[ \begin{array}{ccc}
Q^{\prime } & k_q & Q \\
M_Q^{\prime } & m_q & M_Q \end{array} \right]
\nonumber \\[1ex]
& & {\displaystyle}\times
\left( nj\;\alpha QJ {\left|\!\left|\!\left|}W^{\left( k_q k_j \right)} {\right|\!\right|\!\right|}nj\;\alpha ^{\prime }Q^{\prime }J^{\prime}\right).\end{aligned}$$ where $\left( nj\;\alpha QJ {\left|\!\left|\!\left|}W^{\left( k_q k_j \right) }
{\right|\!\right|\!\right|}nj\;\alpha^{\prime }Q^{\prime }J^{\prime }\right) $ denotes the reduced matrix element of the tensor operator $W^{\left( k_q k_j
\right) }\left( nj,nj\right) = \left[
a^{( q\,j) }\times a^{( q\,j) }\right] ^{(k_q k_j) }$ in quasi–spin space. In terms of the fully reduced coefficients of fractional parentage $\left(j\,\alpha QJ{\left|\!\left|\!\left|}a^{\left(
qj\right)}{\right|\!\right|\!\right|}j\,\alpha^{\prime}Q^{\prime } J^{\prime }\right)$, we find $$\begin{aligned}
\label{eq:theo-gc}
\lefteqn{
\left( n j\,\alpha Q J {\left|\!\left|\!\left|}W^{( k_q k_j)} {\right|\!\right|\!\right|}n j\,\alpha^{\prime } Q^{\prime } J^{\prime } \right)}
\nonumber \\[1ex]
& & {\displaystyle}=
(-1)^{Q+J+Q^{\prime }+J^{\prime}+k_q+k_j}
\left[k_q,k_j\right]^{1/2}
\sum_{\alpha^{\prime \prime } Q^{\prime \prime } J^{\prime \prime }}
\left\{
\begin{array}{ccc}
q & q & k_q \\
Q^{\prime } & Q & Q^{\prime \prime }
\end{array} \right\}
\left\{
\begin{array}{ccc}
j & j & k_j \\
J^{\prime } & J & J^{\prime \prime } \end{array} \right\}
\nonumber \\[1ex]
& &{\displaystyle}\mbox{\hspace{7em}} \times
\left( j\,\alpha Q J {\left|\!\left|\!\left|}a^{(q\,j)} {\right|\!\right|\!\right|}j\,\alpha^{\prime\prime} Q^{\prime \prime} J^{\prime \prime} \right)
\left( j\,\alpha^{\prime \prime }
Q^{\prime \prime } J^{\prime \prime }{\left|\!\left|\!\left|}a^{(q\,j)} {\right|\!\right|\!\right|}j\,\alpha^{\prime } Q^{\prime} J^{\prime } \right).
\end{aligned}$$ This construction has the advantage that the completely reduced matrix elements on the right hand side of (\[eq:theo-g\]) and (\[eq:theo-gb\]) are independent of the occupation number of the shell, and so reduces requirements of storage in comparison with earlier work. These formulae are evaluated in the module `rabs_rcfp` [@rabs_rcfp]. The phase factor arises from the reordering needed to match the recoupled creation and annihilation operators in the bra and ket vectors. We have $$\label{eq:one-ti}
\Delta = 0.$$ when $n_i\kappa _i = n_j\kappa _j$; otherwise $$\label{eq:one-tj}
\Delta = 1+ \sum_{r=a}^{b-1}
N_r,$$ where $N_r$ is the occupation number of subshell $r$, $a = \min \{ i,j \} $, and $b = \max \{ i,j \}$.
### The one–electron submatrix element
It only remains to define the one–electron interaction matrix element $$\left( n_i\kappa _i \left\|f^{(0)}\right \|n_j\kappa _j\right)$$ in (\[eq:one-a\]). The only operator required in this implementation is the matrix element of the Dirac operator, a tensor operator of rank zero, $$\begin{aligned}
\label{eq:t-coeff}
\left( n_i\kappa _i
\left\|\widehat{H}_D \right\| n_j\kappa _j\right)
= I(n_il_ij_i,n_il_ij_i)\delta(\kappa_i,\kappa_j),\end{aligned}$$ where $I(n_il_ij_i,n_il_ij_i)$ is defined by [@Grant-a eqn. (22)]. The Dirac kinetic energy operator, denoted by $T$ in [@GRASP92 eqn. (3.13)], can be obtained from this by setting the nuclear charge $Z=0$. The coefficients $T_{rs}(ab)$ in (\[eq:DC-c1\]) can now be identified by inserting (\[eq:t-coeff\]) in (\[eq:one-a\]). Meanwhile the pure angular coefficients $t_{rs}^0(ab)$ for one–electron operators can be identified by inserting $$\left( n_i\kappa _i \left\|f^{(0)}\right \|n_j\kappa _j\right) = 1$$ in (\[eq:one-a\]).
Two–particle operators
----------------------
According to [@method2], the matrix element of any two–particle scalar operator $\widehat{G}^{(kk0)}$ between configuration state functions with $u$ open shells, can be written $$\begin{aligned}
\label{eq:theo-a}
\lefteqn{
\left(\psi _u^{bra}( J) \left\|\widehat{G}^{(kk0)} \right\|
\psi _u^{ket}( J^{\prime }) \right)}\nonumber \\[1ex]
& = & {\displaystyle}{\sum_{n_i \kappa _i,n_j \kappa _j,
n_{i^{\prime }} \kappa _{i^{\prime }},n_{j^{\prime}}
\kappa _{j^{\prime }}}}
\left(\psi _u^{bra}( J) \left\| \widehat{G}
( n_i \kappa _i,n_j \kappa _j,n_{i^{\prime }} \kappa _{i^{\prime }},
n_{j^{\prime }} \kappa _{j^{\prime }} )
\right\|\psi _u^{ket}( J^{\prime })
\right)\end{aligned}$$ where $$\begin{aligned}
\label{eq:theo-a1}
\lefteqn{
(\psi _u^{bra}\left( J\right) || \widehat{G}
\left( n_i \kappa _i,n_j \kappa _j,n_{i^{\prime }}
\kappa _{i^{\prime }},
n_{j^{\prime }} \kappa _{j^{\prime }} \right) ||
\psi _u^{ket}\left( J^{\prime } \right) ) \nonumber } \\[1ex]
& & =
{\displaystyle}{\sum_{k_{12}}} (-1)^\Delta \Theta^{\prime }
( n_i l_i j_i,n_j l_j j_j,n_{i^{\prime }} l_{i^{\prime }}
j_{i^{\prime}},n_{j^{\prime }} l_{j^{\prime }} j_{j^{\prime }},\Xi)
\nonumber \\[1ex]
& & \times T\left(
n_i j_i,n_j j_j,n_{i^{\prime }} j_{i^{\prime }},n_{j^{\prime }}
j_{j^{\prime }}, \Lambda^{bra}, \Lambda^{ket}, \Xi ,\Gamma \right)
R\left( j_i, j_j, j_{i^{\prime }}, j_{j^{\prime }},
\Lambda^{bra}, \Lambda ^{ket}, \Gamma \right),
\nonumber\end{aligned}$$ $\Gamma$ specifies the recoupling scheme required for each matrix element and $\Xi$, when required, specifies the coupling scheme of the tensor operators defining each matrix element. The operator $\widehat{G}^{(kk0)}$ couples tensor operators of rank $k$ for each electron to give an overall scalar operator.
\[israis\]
[**No.**]{} $i$ $j$ $i^{\prime }$ $j^{\prime }$ $\Gamma$ $\Xi$ $\alpha$ $\beta$ $\gamma$ $\delta$ $\varphi$
------------- ---------- ---------- --------------- --------------- ------------------------------------------------- ------------------------- ---------- --------- ---------- ---------- ----------------------------------------------
[**1.**]{} $\alpha$ $\alpha$ $\alpha$ $\alpha$ – $j_{\alpha}, k$ WW – – – –
– $j_{\alpha}$ W – – – –
[**2.**]{} $\alpha$ $\beta$ $\alpha$ $\beta$ $k$ $j_{\alpha}, j_{\beta}$ W W – – $0$
[**3.**]{} $\beta$ $\alpha$ $\beta$ $\alpha$ $k$ $j_{\alpha}, j_{\beta}$ W W – – $0$
[**4.**]{} $\alpha$ $\beta$ $\beta$ $\alpha$ $k_{12}$ $j_{\alpha}, j_{\beta}$ W W – – $0$
[**5.**]{} $\beta$ $\alpha$ $\alpha$ $\beta$ $k_{12}$ $j_{\alpha}, j_{\beta}$ W W – – $0$
[**6.**]{} $\alpha$ $\alpha$ $\beta$ $\beta$ $k_{12}$ $j_{\alpha}, j_{\beta}$ W W – – $j_{\alpha}+j_{\beta}+k_{12}$
[**7.**]{} $\beta$ $\alpha$ $\alpha$ $\alpha$ $j_{\beta}$ $j_{\alpha}, k_{12}$ aW a – – $k+k_{12}$
[**8.**]{} $\alpha$ $\beta$ $\alpha$ $\alpha$ $j_{\beta}$ $j_{\alpha}, k_{12}$ aW a – – $k$
[**9.**]{} $\beta$ $\beta$ $\beta$ $\alpha$ $j_{\alpha}$ $j_{\beta}, k_{12}$ a Wa – – $k+k_{12}$
[**10.**]{} $\beta$ $\beta$ $\alpha$ $\beta$ $j_{\alpha}$ $j_{\beta}, k_{12}$ a Wa – – $k$
[**11.**]{} $\beta$ $\gamma$ $\alpha$ $\gamma$ $j_{\alpha}, j_{\beta}, k$ – a a W – $1+j_{\alpha}+j_{\beta}-k$
[**12.**]{} $\gamma$ $\beta$ $\gamma$ $\alpha$ $j_{\alpha}, j_{\beta}, k$ – a a W – $1+j_{\alpha}+j_{\beta}-k$
[**13.**]{} $\gamma$ $\beta$ $\alpha$ $\gamma$ $j_{\alpha}, j_{\beta}, k_{12}$ – a a W – $1+j_{\alpha}+j_{\beta}-k_{12}$
[**14.**]{} $\beta$ $\gamma$ $\gamma$ $\alpha$ $j_{\alpha}, j_{\beta}, k_{12}$ – a a W – $1+j_{\alpha}+j_{\beta}-k_{12}$
[**15.**]{} $\gamma$ $\gamma$ $\alpha$ $\beta$ $j_{\alpha}, j_{\beta}, k_{12}$ – a a W – $j_{\alpha}+j_{\gamma}+k_{12}$
[**16.**]{} $\gamma$ $\gamma$ $\beta$ $\alpha$ $j_{\alpha}, j_{\beta}, k_{12}$ – a a W – $j_{\alpha}+j_{\gamma}$
[**17.**]{} $\alpha$ $\beta$ $\gamma$ $\gamma$ $j_{\alpha}, j_{\beta}, k_{12}$ – a a W – $j_{\alpha}+j_{\beta}+k_{12}$
[**18.**]{} $\beta$ $\alpha$ $\gamma$ $\gamma$ $j_{\alpha}, j_{\beta}, k_{12}$ – a a W – $j_{\beta}+j_{\gamma}$
[**19.**]{} $\alpha$ $\beta$ $\gamma$ $\delta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\beta}+k_{12}$
[**20.**]{} $\beta$ $\alpha$ $\gamma$ $\delta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\beta}+j_{\gamma}+k_{12}$
[**21.**]{} $\alpha$ $\beta$ $\delta$ $\gamma$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\beta}+j_{\gamma}$
[**22.**]{} $\beta$ $\alpha$ $\delta$ $\gamma$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\beta}+j_{\gamma}$
[**23.**]{} $\gamma$ $\delta$ $\alpha$ $\beta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\delta}+k_{12}$
[**24.**]{} $\gamma$ $\delta$ $\beta$ $\alpha$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\delta}+k_{12}$
[**25.**]{} $\delta$ $\gamma$ $\alpha$ $\beta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\delta}$
[**26.**]{} $\delta$ $\gamma$ $\beta$ $\alpha$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\delta}$
[**27.**]{} $\alpha$ $\gamma$ $\beta$ $\delta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $0$
[**28.**]{} $\alpha$ $\gamma$ $\delta$ $\beta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $0$
[**29.**]{} $\gamma$ $\alpha$ $\delta$ $\beta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $0$
[**30.**]{} $\gamma$ $\alpha$ $\beta$ $\delta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $0$
[**31.**]{} $\beta$ $\delta$ $\alpha$ $\gamma$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\beta}+j_{\gamma}+j_{\delta}$
[**32.**]{} $\delta$ $\beta$ $\gamma$ $\alpha$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\beta}+j_{\gamma}+j_{\delta}$
[**33.**]{} $\beta$ $\delta$ $\gamma$ $\alpha$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\beta}+j_{\gamma}+j_{\delta}$
[**34.**]{} $\delta$ $\beta$ $\alpha$ $\gamma$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $j_{\alpha}+j_{\beta}+j_{\gamma}+j_{\delta}$
[**35.**]{} $\alpha$ $\delta$ $\beta$ $\gamma$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\gamma}+j_{\delta}-k$
[**36.**]{} $\delta$ $\alpha$ $\gamma$ $\beta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\gamma}+j_{\delta}-k$
[**37.**]{} $\alpha$ $\delta$ $\gamma$ $\beta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\gamma}+j_{\delta}-k_{12}$
[**38.**]{} $\delta$ $\alpha$ $\beta$ $\gamma$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\gamma}+j_{\delta}-k_{12}$
[**39.**]{} $\beta$ $\gamma$ $\alpha$ $\delta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\alpha}+j_{\beta}-k$
[**40.**]{} $\gamma$ $\beta$ $\delta$ $\alpha$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\alpha}+j_{\beta}-k$
[**41.**]{} $\beta$ $\gamma$ $\delta$ $\alpha$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\alpha}+j_{\beta}-k_{12}$
[**42.**]{} $\gamma$ $\beta$ $\alpha$ $\delta$ $j_{\alpha}, j_{\beta}, j_{\gamma}, j_{\delta}$ – a a a a $1+j_{\alpha}+j_{\beta}-k_{12}$
: Scheme of the definitions for matrix elements of any two–particle operator. The operators a, W, aW and Wa defined in (\[eq:theo-b\]) – (\[eq:theo-f\]) act on the indicated subshells.
From (\[eq:theo-a\]) we see that the matrix element of any two–particle operator can be written as a sum over all possible sets of active shell quantum numbers $n_i \kappa _i$, $n_j \kappa _j$, $n_{i^{\prime }} \kappa _{i^{\prime }}$, $n_{j^{\prime }} \kappa
_{j^{\prime }}$. The systematic analysis of [@method2] aims to minimize the computation needed in this expansion. The parameter distributions are presented in Table \[israis\]. Note that for distributions 2–5 and 19–42 the subshell labels are ordered so that $\alpha < \beta <
\gamma < \delta$, while for distributions 6–18 no conditions upon the ordering are imposed. We discuss these structures in more detail below.
### Recoupling matrix
The recoupling coefficients defined in [@MCP; @MCP75; @MCBP] did not reduce the recoupling coefficients to their simplest forms but relied on the analysis module of the NJSYM package (later NJGRAF) to perform the reduction mechanically. The analysis [@method2] leads to simpler forms denoted by $R\left( j_i, j_j, j_{i^{\prime }},
j_{j^{\prime }},\Lambda^{bra},\Lambda ^{ket},\Gamma \right)$. In the case of one interacting shell $R\left( j_i, j_j, j_{i^{\prime }}, j_{j^{\prime }},\Lambda
^{bra},\Lambda ^{ket},\Gamma \right) $ reduces to delta functions [@method2 eqn. (18)]. For two, three and four interacting shells, the recoupling coefficients are given by [@method2 eqns. (22), (23) and (24)], replacing $l,L$ by $j,J$ respectively. The recoupling parameters $\Gamma$ for each distribution can be found in Table \[israis\].
### Matrix elements of irreducible tensor operators
The expressions $T\left( n_i j_i,n_j j_j,n_{i^{\prime }} j_{i^{\prime}},
n_{j^{\prime }} j_{j^{\prime }}, \Lambda ^{bra}, \Lambda ^{ket},\Xi,
\Gamma \right)$ are matrix elements of standard subshell creation/annihilation operators $$\begin{aligned}
\label{eq:theo-b}
a & = & a_{m_q}^{\left( q j \right) }, \\
\label{eq:theo-c}
W & = & \left[ a_{m_{q1}}^{\left( q j \right) }\times
a_{m_{q2}}^{\left( q j \right) }\right] ^{\left( k_{12} \right) },
\\
\label{eq:theo-d}
aW & = & \left[ a_{m_{q1}}^{\left( q j \right) }\times
\left[
a_{m_{q2}}^{\left( q j \right) }\times a_{m_{q3}}^{\left( q j \right) }
\right] ^{\left( k_{12} \right) }\right] ^{\left( k_2 \right) },
\\
\label{eq:theo-e}
Wa & = & \left[ \left[ a_{m_{q1}}^{\left( q j \right)}
\times
a_{m_{q2}}^{\left( q j \right) }\right] ^{\left( k_{12} \right) }
\times a_{m_{q3}}^{\left( q j \right) }\right] ^{\left( k_2\right) },
\\
\label{eq:theo-f}
WW & = & \left[ \left[ a_{m_{q1}}^{\left( q j \right)}
\times
a_{m_{q2}}^{\left( q j \right) }\right] ^{\left( k
\right) }\times \left[ a_{m_{q3}}^{\left( q j \right) }\times
a_{m_{q4}}^{\left( q j \right) }\right] ^{\left( k
\right) }\right] ^{\left( 0 \right)} .\end{aligned}$$
The creation and annihilation operators in (\[eq:theo-b\])–(\[eq:theo-f\]) refer to a single subshell. The evaluation of the submatrix elements of operators of type $a$ (\[eq:theo-b\]) and the simplest compound tensor operator of type $W$ (\[eq:theo-c\]) was explained in section 2.1.2. For types (\[eq:theo-d\])–(\[eq:theo-f\]), we use the formula $$\begin{aligned}
\label{eq:theo-h}
\lefteqn{
\left(nj^N\,\alpha QJ \left\|
\left[ U^{( k_1)}(nj) \times V^{( k_2 )}(nj) \right]^{(k)}
\right\| nj^{N^{\prime }}\,\alpha^{\prime }
Q^{\prime } J^{\prime }\right)}
\\[1ex]
& & = (-1)^{J+J^{\prime }+k}[k]^{1/2}
{\displaystyle}{\sum_{\alpha ^{\prime \prime } Q^{\prime \prime }
J^{\prime\prime }}}
\left\{
\begin{array}{ccc}
k_1 & k_2 & k \\
J^{\prime } & J & J^{\prime \prime }
\end{array} \right\},
\nonumber \\
& & \times\,
\left(nj^N\,\alpha Q J \left\| U^{(k_1)}(nj) \right\|
nj^{N^{\prime \prime }}\, \alpha^{\prime \prime } Q^{\prime\prime}
J^{\prime \prime }\right)
\left(nj^{N^{\prime \prime }}\,\alpha ^{\prime \prime }
Q^{\prime \prime } J^{\prime \prime }
\left\| V^{( k_2 )}(nj) \right\|
nj^{N^{\prime }}\,\alpha^{\prime } Q^{\prime} J^{\prime } \right)
\nonumber\end{aligned}$$ where $U^{\left( k_1\right) }\left( n j \right),
V^{( k_2 )}\left( n j \right)$ are either of type (\[eq:theo-b\]) or type (\[eq:theo-c\]). The occupation number $N^{\prime \prime }$ is defined by second quantization operators occurring in $U^{\left( k_1
\right) }\left( n j \right) $ and $V^{( k_2)}\left(n j \right) $. The module `rabs_rcfp` [@rabs_rcfp] performs the evaluation of these formulae.
### Phase factors $\Delta $
These arise from the reordering necessary to match the recoupled creation and annihilation operators in bra and ket vectors contributing to the matrix element. For each of the cases considered in Table \[israis\] we find\
*Cases 1–6*: $$\label{eq:ti}
\Delta = 0.$$ *Cases 7–18*: $$\label{eq:tj}
\Delta = 1+\sum_{r=
i}^{j-1}N_r,$$ where $N_r$ is the occupation number of subshell $r$. If $\alpha
<\beta $, then $i= \alpha ,$ $j= \beta $, and if $\alpha >\beta $, then $i= \beta ,$ $j= \alpha$.\
*Cases 19–42*: $$\label{eq:tl}
\Delta = \sum_{k=
\alpha }^{\beta -1}N_k+
\sum_{k= \gamma}^{\delta
-1}N_k.$$
### The coefficients $\Theta ^{\prime }\left( n_i l_i j_i,
n_j l_j j_j,n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }},
n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}, \Xi \right) $
The effective interaction strength of order $k$ of a two–electron operator, is denoted by $$\left( n_i l_i j_i n_j l_j
j_j\left\| g^{( k k )} \right\| n_{i^{\prime }} l_{i^{\prime }}
j_{i^{\prime }} n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}
\right)$$ in [@method6] and by $$X^k\left( n_i l_i j_i, n_j l_j j_j,
n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }}, n_{j^{\prime}}
l_{j^{\prime }} j_{j^{\prime }} \right)$$ in [@Grant-a]. The coefficients $\Theta ^{\prime }\left( n_i l_i j_i,
n_j l_j j_j, n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }},
n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}, \Xi \right) $ for the different cases tabulated in Table \[israis\]) have different multiplicative factors defined as follows:
*Case 1*: Single subshell ($\alpha \alpha \alpha
\alpha$) $$\begin{aligned}
\label{eq:theta-a}
\lefteqn{
\Theta ^{\prime}_{IIa}\left( n_\alpha l_\alpha j_\alpha ,n_\alpha
l_\alpha
j_\alpha , n_\alpha l_\alpha j_\alpha , n_\alpha l_\alpha j_\alpha ,\Xi
\right)}
\nonumber \\[1ex]
& & = \frac{1}{2}[k]^{-1/2}
\left( n_\alpha l_\alpha j_\alpha n_\alpha l_\alpha j_\alpha \left\|
g^{\left(k k\right) }
\right\| n_\alpha l_\alpha j_\alpha n_\alpha l_\alpha j_\alpha \right)
\delta \left( k_{12},k \right)\end{aligned}$$ and $$\begin{aligned}
\label{eq:theta-b}
\lefteqn{
\Theta ^{\prime} _{IIb}\left( n_\alpha l_\alpha j_\alpha ,n_\alpha
l_\alpha
j_\alpha , n_\alpha l_\alpha j_\alpha , n_\alpha l_\alpha j_\alpha ,\Xi
\right)}
\nonumber \\[1ex]
& & = \left( -1\right) ^{k} \left[ j_{\alpha } \right]
^{-1/2}
\left( n_\alpha l_\alpha j_\alpha n_\alpha l_\alpha j_\alpha \left\|
g^{\left(k k\right) }
\right\| n_\alpha l_\alpha j_\alpha n_\alpha l_\alpha j_\alpha \right)
\delta \left( k_{12},0 \right).\end{aligned}$$
*Cases 2, 3, 11, 12, 27, 29, 31, 32, 35, 36, 39, 40*: Subshell assignments $\alpha\beta\alpha\beta,
\beta\alpha\beta\alpha$, $\beta \gamma \alpha \gamma $, $\gamma \beta \gamma \alpha $, $\alpha \gamma \beta \delta $, $\gamma \alpha \delta \beta $, $\beta \delta \alpha \gamma $, $\delta \beta \gamma \alpha $, $\alpha \delta \beta \gamma $, $\delta \alpha \gamma \beta $, $\beta \gamma \alpha \delta $, $\gamma \beta \delta \alpha $ $$\begin{aligned}
\label{eq:theta-c}
\lefteqn{
\Theta ^{\prime }\left(n_i l_i j_i,n_j l_j j_j,
n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }},
n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}, \Xi \right)}
\nonumber \\[1ex]
& = & \left(-1\right) ^{\varphi} \frac 12\left[
k\right]
^{-1/2} \left( n_i l_i j_i n_j l_j j_j\left\| g^{( k k )} \right\|
n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }} n_{j^{\prime}}
l_{j^{\prime }} j_{j^{\prime }} \right) \delta \left( k_{12},k
\right).\end{aligned}$$ *Cases 6, 15–26*: Subshell assignments $\alpha \alpha \beta
\beta $, $\gamma \gamma \alpha \beta $, $\gamma \gamma \beta \alpha $, $\alpha \beta \gamma \gamma $, $\beta \alpha \gamma \gamma $, $\alpha \beta \gamma \delta $, $\beta \alpha \delta \gamma $, $\alpha \beta \delta \gamma $, $\beta \alpha \gamma \delta $, $\gamma \delta \alpha \beta $, $\delta \gamma \beta \alpha $, $\gamma \delta \beta \alpha $, $\delta \gamma \alpha \beta $ $$\begin{aligned}
\label{eq:theta-d}
\lefteqn{
\Theta ^{\prime }\left(n_i l_i j_i,n_j l_j j_j,
n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }},
n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}, \Xi \right)}
\nonumber \\[1ex]
& & = \left(-1\right) ^{1+k+\varphi} \frac 12\left[
k_{12} \right]
^{1/2}
\left\{
\begin{array}{ccc}
j_i & j_{i^{\prime }} & k \\
j_{j^{\prime }} & j_j & k_{12}
\end{array}
\right\}
\left( n_i l_i j_i n_j l_j j_j\left\| g^{( k k )} \right\|
n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }} n_{j^{\prime}}
l_{j^{\prime }} j_{j^{\prime }} \right).\end{aligned}$$ *Cases 4, 5, 7, 8, 9, 10, 13, 14, 28, 30, 33, 34, 37, 38, 41, 42*: Subshell arrangements $\alpha \beta \beta \alpha $, $\beta
\alpha\alpha \beta $, $\beta \alpha \alpha \alpha $, $\alpha \beta \alpha \alpha $, $\beta \beta \beta \alpha $, $\beta \beta \alpha \beta $, $\gamma \beta \alpha \gamma $, $\beta \gamma \gamma \alpha $, $\alpha \gamma \delta \beta $, $\gamma \alpha \beta \delta $, $\beta \delta \gamma \alpha $, $\delta \beta \alpha \gamma $, $\alpha \delta \gamma \beta $, $\delta \alpha \beta \gamma $, $\beta \gamma \delta \alpha $, $\gamma \beta \alpha \delta $ $$\begin{aligned}
\label{eq:theta-e}
\lefteqn{
\Theta ^{\prime }\left(n_i l_i j_i,n_j l_j j_j,
n_{i^{\prime }} l_{i^{\prime }}j_{i^{\prime }},
n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}, \Xi \right)}
\nonumber \\[1ex]
& & = \left(-1\right) ^{\varphi} \frac 12\left[ k_{12}
\right]
^{1/2}
\left\{
\begin{array}{ccc}
j_i & j_{i^{\prime }} & k \\
j_j & j_{j^{\prime }} & k_{12}
\end{array}
\right\}
\left( n_i l_i j_i n_j l_j j_j\left\| g^{( k k )} \right\|
n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }} n_{j^{\prime}}
l_{j^{\prime }} j_{j^{\prime }} \right).\end{aligned}$$ The phase factors $\varphi$ in expressions (\[eq:theta-c\]) – (\[eq:theta-e\]) are defined in column $\varphi$ of Table \[israis\]. This construction exploits the common tensorial structure of any scalar two–electron operators as the Coulomb, Breit and Gaunt interactions [@Grant-a]) and exploits this similarity to simplify the calculation of spin–angular coefficients. The relativistic $jj$–coupling expressions for the effective interaction strength of the Coulomb interaction [@Grant-a eqn. (86)] is $$\begin{aligned}
\label{eis-c}
\lefteqn{\left( n_i l_i j_i n_j l_j
j_j\left\| g^{( k k )} \right\| n_{i^{\prime }} l_{i^{\prime }}
j_{i^{\prime }} n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}
\right)} \\
& & = (-1)^k \langle n_i l_i j_i \| C^{(k)} \|
n_{i^{\prime }}
l_{i^{\prime }}
j_{i^{\prime }} \rangle \langle n_j l_j j_j \| C^{(k)} \|
n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }} \rangle R^k(n_i l_i j_i
n_j l_j j_j n_{i^{\prime }} l_{i^{\prime }} j_{i^{\prime }}
n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }} ) .
\nonumber\end{aligned}$$
We can now identify the coefficients $V_{rs}^{k}(abcd)$ of (\[eq:DC-c1\]) by substituting the results above in (\[eq:theo-a\]). The same construction can be used for the Gaunt interaction (the leading part of the magnetic Breit interaction) [@Grant-a eqn. (91)] and for the full transverse Breit interaction [@Grant-a eqn. (101)], although the selection rules and the effective interaction strengths corresponding to (\[eq:theta-e\]) are, of course, different. The pure angular coefficients $v_{rs}^{k}(abcd)$ for two–electron operators are the same for all these operators since the $v_{rs}^{k}(abcd)$ can be identified by inserting $$\left( n_i l_i j_i n_j l_j
j_j\left\| g^{( k k )} \right\| n_{i^{\prime }} l_{i^{\prime }}
j_{i^{\prime }} n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime }}
\right) = 1$$ in (\[eq:theo-a\]).
Program organization
====================
Overview of the program
-----------------------
The program ANCO constructs the pure angular coefficients $t^0_{rs}(ab)$ for one–electron operators and the $v^k_{rs}(abcd)$ coefficients contributing to matrix elements of the Dirac–Coulomb–Breit Hamiltonian. The coefficients $T_{rs}(ab)$ and $V^k_{rs}(abcd)$ used in GRASP92 and earlier version of the system are available as an option. The new format generates what we have called “pure” angular momentum coefficients which can be used unchanged with any one–particle tensor operator of rank 0, and any two–particle interaction. The Coulomb and Breit interactions use different subsets of the complete set of $v^k_{rs}(abcd)$ coefficients, which are selected automatically when multiplying by the relevant effective interaction strengths to complete the matrix element calculation. The MCP and MCBP modules of GRASP92 calculated the full matrix elements for each of these subsets, so that the new formulation reduces the computational overheads and the memory requirements, which renders ANCO more suitable for large scale problems.
There are two new modules `rabs_recoupling` and `rabs_anco` for extracting spin–angular coefficients relating to formula (\[eq:theo-a\]). The module `rabs_recoupling` evaluates recoupling coefficients $R\left( j_i, j_j,\Lambda ^{bra},\Lambda ^{ket} \right)$ and $R\left( j_i, j_j, j_{i^{\prime }}, j_{j^{\prime }},\Lambda
^{bra},\Lambda ^{ket},\Gamma \right)$ as described in [@method2], module `rabs_rcfp` [@rabs_rcfp] evaluates the $T\left( n_i j_i,n_j j_j,n_{i^{\prime }} j_{i^{\prime
}},n_{j^{\prime }} j_{j^{\prime }},\Lambda^{bra},\Lambda^{ket},\Xi
,\Gamma \right)$ whilst `rabs_anco` evaluates all the contributions to (\[eq:theo-a\]) for both scalar one– and two–particle operators.
The program ANCO can be run in two modes. The interactive mode is intended for testing the program and for performing short calculations involving a small number of configurations. Normally the modules `rabs_anco` and `rabs_recoupling` will be interfaced to GRASP92 or to some other program with compatible data structure to perform multiconfiguration or configuration–interaction relativistic calculations. The program can handle any combination of open subshells with $j \leq 9/2$, but subshells with $j > 9/2$ are only allowed if they contain not more than two electrons.
ANCO is written in Fortran 90/95 and is designed as an addition to the RATIP package [@Fritzsche/CFF/Dong:99]. The new Fortran 90/95 standard enables us to define new derived data types which will enable us to incorporate this module in our continuing development of large–scale computations for open–shell atoms and ions. The full definition of the various derived structures can be found in the module header of `rabs_anco`. Here we need only those types which concern the program output.
All information about pure spin–angular coefficients of scalar one– and two–particle operators ($t_{rs}(ab)$ and $v_{rs}^{k}(abcd)$ coefficients) of the Hamiltonian matrix or of some part of it is summarized in the derived data `anco_pair_list`.
type(anco_csf_pair), dimension(:), pointer :: anco_pair_list
which is defined by
type :: anco_csf_pair
integer :: r, s
integer :: no_t_coeff, no_v_coeff
type(anco_t_coeff), dimension(:), pointer :: t_coeff
type(anco_v_coeff), dimension(:), pointer :: v_coeff
end type anco_csf_pair
The integers `r` and `s` respectively index the bra– and the ket– configuration state functions (CSF) for the current matrix element. The variable `no_t_coeff` is the number of pure spin–angular coefficients of one–particle operators that can be constructed for the pair `r,s`, and the variable `no_v_coeff` gives corresponding data for two–particle matrix elements. The array `anco_v_coeff` contains pure spin–angular coefficients for two–particle scalar operators. It is defined by
type :: anco_v_coeff
integer :: k
type(nkappa) :: a, b, c, d
real(kind=dp) :: v
end type anco_v_coeff
where `k` is the rank $k$ of the effective interaction strength, `a, b, c, d` point to the relevant subshells $n_i l_i j_i$, $n_j l_j j_j$, $n_{i^{\prime }} l_{i^{\prime
}} j_{i^{\prime }}$, $n_{j^{\prime}} l_{j^{\prime }} j_{j^{\prime
}}$, and the pure spin–angular coefficient itself is given in `v`. The array `anco_t_coeff` is defined in the same way.
Memory for the array `anco_pair_list` is allocated dynamically using the
allocate( anco_pair_list(1:number_of_pair_list_max))
instruction, and can be deallocated subsequently.
Interactive calculations
------------------------
A typical interactive dialog for calculating spin–angular coefficients is shown in Fig. 1.
`Enter a file name for the anco.sum file:`
After this prompt, the user should insert the output file name to which the main output data will be written. This must be followed by the name of the input file listing CSF in the GRASP92 format. The next question
`Generate only not trivial angular coefficients which include (at least one) open shells ?`
The response `y` or `Y` will cause the program to calculate coefficients for peel shells only; the reponse `n` or `N` the program calculate will yield data for all shells (open and closed).
The question
`Generate one-electron angular coefficients for scalar interactions ?`
This requires answer `n` or `N` if one–electron coefficients are not needed. If the user responds `y` or `Y`, then the prompt
`Generate GRASP92-like T coefficients for scalar interactions ?`.
appears. The response `y` or `Y` causes GRASP92–like $T_{rs}(ab)$ coefficients to be generated, whereas the alternative `n` or `N` yields $t^0_{rs}(ab)$ coefficients.
A similar dialog follows for two–electron angular coefficients. A number of examples illustrate the usage of ANCO in this mode in section 4 below.
The prompt
`Enter a file name for the anco.vnu file:`
permits the user to specify where the spin–angular coefficients should be stored.
ANCO: Calculation of angular coefficients for symmetry-adapted CSF functions
from the GRASP92 structure program (Fortran 90 version)
(C) Copyright by G Gediminas and others, Kassel (2000).
Enter a file name for the anco.sum file:
test.sum
Enter the name of the configuration symmetry list file:
argon-sd.inp
Loading configuration symmetry list file ...
There are 16 relativistic subshells;
there are 761 relativistic CSFs;
... load complete.
Generate only non-trivial angular coefficients which include
(at least one) open shells ?
y
Generate one-electron angular coefficients for scalar interactions ?
y
Generate GRASP92-like T coefficients for scalar interactions ?
y
Generate two-electron angular coefficients for scalar interactions ?
y
Generate GRASP92-like V^k coefficients for scalar interactions ?
y
Enter a file name for the anco.vnu file:
test.vnu
[**Figure 1:**]{}
The module ANCO needs one input file `.csl` containing the CSF list output by GRASP92 [@GRASP92], which is gnerated by the program GENCSL.
The program creates two output files. The file `.sum` contains the summary of the problem. The other file `.vnu` contains the angular momentum coefficients and their characteristics. Its format for $T_{rs}(ab)$ and $V_{rs}^{k}(abcd)$ coefficients is the same as that of file `genmcp.dbg` generated by the module `genmcp` of GRASP92 [@GRASP92]. The format for $t^0_{rs}(ab)$ and $v_{rs}^{k}(abcd)$ is very similar but without the sorting process which is used in GRASP92.
Distribution and installation of the program
--------------------------------------------
As a new component of the RATIP package [@Fritzsche/CFF/Dong:99] similar to the module RCFP [@rabs_rcfp] ANCO will be distributed as a `tar` archive file of the directory `ratip_anco`. On a UNIX (or compatible) workstation, the command `tar -xvf ratip_anco.tar` reconstructs the file structure. The directory `ratip_anco` then contains the Fortran 90/95 modules `rabs_anco.f` and `rabs_recoupling.f`, the program `xanco.f` (the main program for interactive work) as well as the makefile `make-anco`. It also includes a number of examples in the subdirectory `test-anco` and a short `Read.me` which explains further details about the installation. Since the same file structure is preserved in both cases, the combination of ANCO with RATIP is simply achieved by running the command `cp -r ratip_anco/. ratip/.` inside the RATIP root directory; then `make -f make-anco` will generate the executable `xanco`, together with the other two components `xcesd99` [@Fritzsche/Anton:99] and `xreos99` [@Fritzsche/CFF/Dong:99] of the RATIP package. The name of the (Fortran 90/95) compiler and special compiler flags can be overwritten in the header of the makefile as necessary. Although ANCO uses six other modules which are part already of RATIP, no further adaptation of the program is needed. At present, the ANCO program has been installed and tested under the Linux and AIX operating systems but, owing to the compliance of the Fortran 90/95 standard, no difficulties should arise on any other platform.
The subdirectory `test-anco` lists a number of examples which demonstrate the usage of the program.
Timing and verification of `ANCO`
=================================
Tests and timing studies using the Dirac-Coulomb Hamiltonian only were performed for the $3s^{2}3p^{6}$ $^{1}S$ state of Ar I with the common closed shells $1s^{2}2s^{2}2p^{6}$ for different values of final orbital momentum $J$. The wave function expansions used were:
1. 3SD: Single and double excitations from $3s^{2}3p^{6}$ to the active set $ \{ 3s,3p,3d \} $ contains 14 configuration state functions (CSF) for $J= 0$ and 34 CSF (the maximum) for $J= 2$.
2. 3SDT: Single, double and triple excitations from $3s^{2}3p^{6}$ to the active set $\{ 3s,3p,3d \} $. The maximum number of CSF is 145 for $J= 2$.
3. 4SD: Single and double excitations from $3s^{2}3p^{6}$ to the active set\
$ \{ 3s,3p,3d,4s,4p,4d,4f\} $. The maximum number of CSF is 465 for $J= 2$.
4. 4SDT: Single, double and triple excitations from $3s^{2}3p^{6}$ to the active set\
$\{ 3s,3p,3d,4s,4p,4d,4f \} $.
5. 5SD; Single and double excitations from $3s^{2}3p^{6}$ to the active set\
$\{ 3s,3p,3d,4s,4p,4d,4f,5s,5p,5d,5f,5g \}$.
We first considered simple cases with a small number of CSF (3SD, 3SDT, 4SD with $J$ = 0, 1, 2, 3, 4, 5, 6, 7, 8, 9). Although ANCO generates the full set of “pure” coefficients for both one– and two–particle operators, the calculation runs from 1.4 – 2.3 times faster than an equivalent calculation with `GRASP92` because of the lower computational overheads. Table \[run\_time\] demonstrates similar enhanced performance for the much larger 4SDT and 5SD eamples, showing the improvement expected for large–scale calculations.
From the results presented in the Table \[run\_time\] we conclude that the new program is not much faster in simple cases, but does better in more complicated cases. The fact that `ANCO` calculates approximately twice the number of angular coefficients as `GRASP92` reduces the effective cost per coefficient by a further factor of two.
Although the program is completely new, we have verified that the results presented agree completely with those obtained from GRASP92. We have also verified that the Breit interaction is treated correctly, although no data are presented here.
\[run\_time\]
------------------------ ------------- ------------------------------- ------------------------------ ------------------------------ --------------- ------------ -------------
\[-0.2cm\] [**ASF**]{} [**Speed**]{}
\[0.3cm\] [**-up**]{}
\[-0.2cm\] [**CSF**]{} [**$T_{rs} ^{k} (ab)$ or**]{} [**$V_{rs} ^{k} (abcd)$**]{} [**$v_{rs} ^{k} (abcd)$**]{} [`GRASP92`]{} [`ANCO`]{}
[**$t_{rs} ^{k} (ab)$** ]{}
\[0.3cm\]
\[-0.2cm\] 4SDT (J= 0) 2 149 3 606 756 023 1 530 086 00:08:11 00:03:01 2.7
4SDT (J= 1) 5 786 14 017 4 070 156 8 188 130 00:59:01 00:15:19 3.9
4SDT (J= 2) 8 016 21 356 7 018 885 14 077 044 01:42:47 00:26:28 3.9
4SDT (J= 3) 8 378 21 342 7 634 136 15 290 955 01:53:55 00:30:37 3.7
4SDT (J= 4) 7 284 15 971 6 111 074 12 260 139 01:33:17 00:23:01 4.1
4SDT (J= 5) 5 349 9 435 3 810 165 7 656 054 00:50:27 00:14:18 3.6
4SDT (J= 6) 3 370 4 556 1 836 602 3 706 544 00:21:52 00:06:40 3.3
4SDT (J= 7) 1 788 1 789 693 761 1 412 443 00:07:26 00:02:29 3.0
5SD (J= 0) 468 621 75 192 150 455 00:00:32 00:00:17 1.9
5SD (J= 1) 1 134 2 324 395 450 792 560 00:03:10 00:01:29 2.1
5SD (J= 2) 1 609 3 704 697 651 1 395 839 00:06:27 00:02:44 2.4
5SD (J= 3) 1 584 3 441 721 907 1 444 095 00:06:43 00:02:59 2.3
5SD (J= 4) 1 361 2 500 558 223 1 117 681 00:05:15 00:02:17 2.3
5SD (J= 5) 920 1 361 314 909 632 306 00:02:30 00:01:22 1.8
5SD (J= 6) 559 644 141 328 284 102 00:01:02 00:00:36 1.7
5SD (J= 7) 259 226 44 137 89 398 00:00:15 00:00:12 1.3
------------------------ ------------- ------------------------------- ------------------------------ ------------------------------ --------------- ------------ -------------
: Timing Comparison for [`GRASP92`]{} and [`ANCO`]{} codes. Times are given in hours, minutes, seconds
Examples
========
To illustrate the use of ANCO in its interactive mode, we studied Ar II. The program `genmcp` of GRASP92 [@GRASP92] is used first to generate the CSF file `argon-sd.inp` before running the program `xanco`.
The first example does a GRASP92–style calculation. After checking all triangular conditions for $X_{Coulomb}^{k}(\alpha \beta \gamma \delta)$ (see expression (88) in [@Grant-a]), it multiplies each pure–two particle coefficient by the factor $X_{Coulomb}^{k}(\alpha \beta \gamma \delta)$ and prints all non zero $V_{rs}^{k}(abcd)$ coefficients. With corrsponding $T_{rs}(ab)$ coefficients, this generates a total of 11279 coefficients. The second example, with the same input, calculates a total of 433911 pure non–trivial spin–angular coefficients at one go, as is more convenient for large scale calculations. However, only 11279 coefficients from this set are required in the first example.
The Test Run Output displays the `.sum` files and the first 10 lines of `.vnu` files for both examples.
[99]{} C.F. Fischer, *The Hartree-Fock Method for Atoms* (Wiley, New York, 1977). I.P. Grant and M. Quiney, Advances in Atomic and Molecular Physics 23 (1987) 37. I.P. Grant, Comput. Phys. Commun. 5 (1973) 264. I.P. Grant, Comput. Phys. Commun. 11 (1976) 397. N. Beatham, I.P. Grant and B.J. McKenzie, Comput. Phys. Commun. 18 (1979) 245. G. Gaigalas, Z. Rudzikas and C. Froese Fischer, J. Phys. B: At. Mol. Opt. Phys. 30 (1997) 3747. G. Gaigalas, Lithuanian Journal of Physics 39 (1999) 80. G. Gaigalas, S. Fritzsche and Z. Rudzikas, At. Data and Nucl. Data Tables (accepted). F.A. Parpia, C. Froese Fischer and I.P. Grant, Comput. Phys. Commun. 94 (1996) 249. G. Gaigalas and Z. Rudzikas, J. Phys. B: At. Mol. Phys. 29 (1996) 3303. G. Gaigalas and S. Fritzsche, Comput. Phys. Commun. (2000) in print. I.P. Grant, in: *Methods in Computational Chemistry* [**2**]{}, S. Wilson, ed. (Plenum, New York, 1988) p. 1. S. Fritzsche, C. F. Fischer, and C. Z. Dong, Comput. Phys. Commun. 124 (2000) 340. S. Fritzsche and J. Anton, Comput. Phys. Commun. 124 (1999) 353.
TEST RUN OUTPUT {#test-run-output .unnumbered}
===============
...........................Example 1...........................
>>type example1.sum
ANCO run at 15:50:39 on Feb 14 2000.
There are 17 electrons in the cloud
in 761 relativistic CSFs
based on 16 relativistic subshells.
Total number of pair is: 289941
Generate only not trivial angular coefficients
there are 716 Grasp92-like T coefficients
there are 112078 Grasp92-like Vk coefficients
the total number of coefficients is 112794
>> type example1.vnu (first 10 lines of example1.vnu file)
ANCO
761 16 4
V^[( 2)]_[ 1, 1] ( 3p , 3p ; 3p , 3p ) = -1.200000000000E-01
V^[( 1)]_[ 1, 1] ( 3p , 3d-; 3d-, 3p ) = 6.666666666667E-02
V^[( 3)]_[ 1, 1] ( 3p , 3d-; 3d-, 3p ) = -2.571428571429E-01
V^[( 1)]_[ 2, 1] ( 3p , 3d ; 3d-, 3p ) = -1.632993161855E-01
V^[( 2)]_[ 3, 1] ( 3p , 4s ; 3p , 3d-) = 2.529822128135E-01
V^[( 1)]_[ 3, 1] ( 3p , 4s ; 3d-, 3p ) = -2.108185106779E-01
T^[ 6, 1] ( 4d-, 3d-) = 1.000000000000E+00
V^[( 0)]_[ 6, 1] ( 1s , 4d-; 1s , 3d-) = 2.000000000000E+00
...........................Example 2...........................
>>type example2.sum
ANCO run at 16:35:02 on Feb 14 2000.
There are 17 electrons in the cloud
in 761 relativistic CSFs
based on 16 relativistic subshells.
Total number of pair is: 289941
Generate only not trivial angular coefficients
there are 716 pure one-particle angular coefficients
there are 433195 pure two-particle angular coefficients
the total number of coefficients is 433911
>> type example2.vnu (first 10 lines of example2.vnu file)
ANCO
761 16 4
pure two-particle [( 1)]_[ 1, 1] ( 3p , 3p ; 3p , 3p ) = 5.000000000000E-02
pure two-particle [( 2)]_[ 1, 1] ( 3p , 3p ; 3p , 3p ) = -1.500000000000E-01
pure two-particle [( 3)]_[ 1, 1] ( 3p , 3p ; 3p , 3p ) = 5.000000000000E-02
pure two-particle [( 1)]_[ 1, 1] ( 3p , 3d-; 3p , 3d-) = 3.000000000000E-01
pure two-particle [( 3)]_[ 1, 1] ( 3p , 3d-; 3p , 3d-) = -2.000000000000E-01
pure two-particle [( 0)]_[ 1, 1] ( 3p , 3d-; 3d-, 3p ) = -2.500000000000E-01
pure two-particle [( 1)]_[ 1, 1] ( 3p , 3d-; 3d-, 3p ) = -2.500000000000E-01
pure two-particle [( 2)]_[ 1, 1] ( 3p , 3d-; 3d-, 3p ) = -1.500000000000E-01
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A study of searching for doubly charged Higgs $(\delta^{\pm\pm})$ is performed in two-Higgs-doublet extension of the type-II seesaw model. We find that a significant mixing effect between singly charged Higgs of Higgs doublet and of triplet is arisen from the scalar potential. The mixing leads to new production processes and decay patterns of doubly charged Higgs. With luminosity of 40 fb$^{-1}$ and collision energy of 13 TeV at the LHC, we show that $\delta^{\pm\pm}$ with mass below $330$ GeV could be observed at the $5\sigma$ level. Moreover, for the luminosity of 300 fb$^{-1}$, the observed mass of $\delta^{\pm\pm}$ could reach up to 450 GeV.'
author:
- Takaaki Nomura
title: 'Type-II seesaw model with two-Higgs doublets'
---
Introduction
============
The origin of neutrino mass is one of the unsolved issue in the standard model (SM). This mystery could be solved by a Higgs triplet extension of the SM called Type-II seesaw model [@Magg:1980ut; @Konetschny:1977bn]. The novel feature of a Type-II seesaw model is the existence of a doubly charged Higgs $\delta^{\pm \pm}$ in the Higgs triplet $\Delta$. In the original Type-II seesaw model, $\delta^{\pm \pm}$ decays into two same-sign charged leptons $\ell^\pm \ell'^\pm$ or W bosons $W^\pm W^\pm$ according to the triplet vacuum expectation value (VEV) when masses in the multiplet are degenerate. Thus the experimental searches of doubly charged Higgs bosons have focused on the signals of same-sign dilepton or W bosons [@Chatrchyan:2012ya; @ATLAS:2012hi; @Khachatryan:2014sta].
The extension of Type-II seesaw model by including second Higgs doublet can lead new effect in doubly charged Higgs production and decays [@Chen:2014xva; @Chen:2014qda]. The most attractive new effect is the dimension-3 terms in scalar potential, read by $\mu_x H^T_j i\tau_2 \Delta^\dagger H_k$ (j,k =1,2; $x$=1,2,3), where $H_j$ is a Higgs doublet. Since the coefficients $\mu_x$ can be of order of electroweak (EW) scale, the new terms lead a large mixing angle between the singly charged Higgs of doublet ($H^\pm$) and of triplet ($\delta^{\pm}$). Consequently the doubly charged Higgs can be produced via QCD process owing to the the mixing effect.
In this study we focus on the new production and decay patterns of $\delta^{\pm \pm}$. We then carry out a simulation study for the signal of $\delta^{\pm \pm}$ at the LHC and estimate the significance of discovering the doubly charged Higgs.
The interactions of doubly charged Higgs boson
==============================================
To investigate the production and decay of doubly charged Higgs boson $\delta^{\pm \pm}$, we first discuss the relevant interactions of charged Higgs bosons in the two Higgs doublet(THD)-Type-II seesaw model. The gauge interactions and Yukawa couplings of $\Delta$ and $H_{1,2}$ are same as original Type-II seesaw model and THD model (Type-II) respectively. We thus do not explicitly write down them here; more details are given in Refs [@Chen:2014xva; @Chen:2014qda]. The Higgs potential is written such that $$\begin{aligned}
V(H_1, H_2, \Delta) &= V_{H_1H_2} + V_\Delta + V_{H_1 H_2 \Delta}\,, \nonumber \\
V_{H_1H_2} &= m^2_1 H^\dagger_1 H_1 + m^2_2 H^\dagger_2 H_2 - m^2_{12} ( H^\dagger_1 H_2 + h.c.)+ \lambda_1 ( H^\dagger_1 H_1)^2 \nonumber \\
&+ \lambda_2 (H^\dagger_2 H_2)^2 + \lambda_3 H^\dagger_1 H_1 H^\dagger_2 H_2+ \lambda_4 H^\dagger_1 H_2 H^\dagger_2 H_1 + \frac{\lambda_5}{2} [(H^\dagger_1 H_2)^2+h.c. ] \,, \nonumber \\
V_\Delta &= m^2_\Delta Tr \Delta^\dagger \Delta + \lambda_{9} (Tr \Delta^\dagger \Delta)^2 + \lambda_{10} Tr (\Delta^\dagger \Delta)^2\,, \nonumber \\
V_{H_1H_2\Delta} &= \left( \mu_1 H^T_1 i\tau_2 \Delta^{\dagger} H_1 + \mu_2 H^T_2 i \tau_2 \Delta^\dagger H_2 + \mu_3 H^T_1 i\tau_2 \Delta^\dagger H_2 + h.c. \right) \nonumber \\
&+ ( \lambda_6 H^\dagger_1 H_1 + \bar \lambda_6 H^\dagger_2 H_2 ) Tr \Delta^\dagger \Delta + H^\dagger_1 ( \lambda_7 \Delta \Delta^\dagger + \lambda_8 \Delta^\dagger \Delta ) H_1
+ H^\dagger_2 \left( \bar \lambda_7 \Delta \Delta^\dagger + \bar \lambda_8 \Delta^\dagger \Delta \right) H_2\,, \label{eq:v}
\end{aligned}$$ where $V_{H_1 H_2}$ and $V_\Delta$ denote the scalar potential of THD and of pure triplet, and $V_{H_1 H_2 \Delta}$ is the part involving $H_1$, $H_2$ and $\Delta$. The VEV of the triplet $v_\Delta$ is required to be $$v_\Delta \approx \frac{1}{\sqrt{2}} \frac{\mu_1 v^2_1 + \mu_2 v^2_2 + \mu_3 v_1 v_2}{m^2_\Delta + (\lambda_6+\lambda_7) v^2_1/2 + (\bar\lambda_6+\bar\lambda_7)v^2_2 /2}\,, \label{eq:v_d}$$ where $v_{1,2}$ is the VEV of $H_{1,2}$. We only keep the leading power for $v_\Delta$ in Eq. (\[eq:v\_d\]) since it should satisfy $v_\Delta << v=\sqrt{v_1^2 + v_2^2}$ from $\rho$ parameter measurement. Interestingly, the parameters $\mu_{1,2,3}$ as large as electroweak scale can provide $v_\Delta << v$ if the parameters satisfy the relation $\mu_3 \sim - (\mu_1 v_1^2 + \mu_2 v_2^2)/(v_1 v_2)$ in contrast to original Type-II seesaw model. In our analysis, we adopt the relation for simplicity.
Two physical singly charged Higgs bosons are provided by both THD sector and Higgs triplet which are denoted by $H^\pm$ and $\delta^\pm$ respectively. The mass eigenstates are obtained as a combination of $H^\pm$ and $\delta^\pm$ as $$\left( \begin{array}{c}
H^\pm_1\\
H^\pm_2 \\
\end{array}\right) = \left(\begin{array}{cc}
\cos\theta_\pm & \sin\theta_\pm \\
-\sin\theta_\pm & \cos\theta_\pm\\
\end{array}\right) \left( \begin{array}{c}
H^\pm \\
\delta^{\pm} \\
\end{array}\right)\,.\label{eq:ma}$$ The masses and their mixing angles are written by $$\begin{aligned}
\left( m_{H_{1,2}^{\pm}}\right)^2 = \frac{1}{2}\left(m^2_{\delta^\pm} + m^2_{H^\pm}\right) \mp \frac{1}{2} \left[ \left( m^2_{\delta^\pm} - m^2_{H^\pm}\right)^2 + 4 m^4_{H^-\delta^+}\right]^{1/2}\,, \quad
\tan2\theta_\pm = - \frac{2 m^2_{H^- \delta^+}}{m^2_{\delta^\pm} - m^2_{H^\pm}}\,,
\label{eq:mass_mixing}
\end{aligned}$$ where $H_1^\pm$ is identified as the lighter charged Higgs, $m^2_{H^\pm} = (m_{12}^2 - (\lambda_4 + \lambda_5)v_1 v_2/2)/(\sin \beta \cos \beta)$, $m^2_{H^- \delta^+}=v[\mu_1 \cos^4 \beta -\mu_2 \sin^4 \beta +(\mu_1 -\mu_2)\sin^2 \beta \cos^2 \beta]/(2 \sin \beta \cos \beta) $, and $m^2_{\delta^\pm} = m_\Delta^2 +v_1^2(2 \lambda_6 + \lambda_7 + \lambda_8)/2 + v_2^2(2 \bar \lambda_6 + \bar \lambda_7 \bar \lambda_8)/2$. The magnitude of the mixing angle $\theta_\pm$ depends on the massive parameter $\mu_{1,2}$. We are going to explore the influence of a large $\theta_\pm$ on the search of doubly charged Higgs.
In our model, two singly charged Higgses can interact with quarks owing to the large mixing effect. The $H_{1,2}^\pm$-quark interactions are given by $$\begin{aligned}
\frac{\sqrt{2}}{v} \left[\bar u\left( \tan\beta V_{CKM} {\bf m_D} P_R + \cot\beta {\bf m_U} V^\dagger_{CKM} P_L \right) d \right]
(\cos\theta_\pm H^+_1 - \sin\theta_\pm H^+_2) + h.c. \,, \label{eq:hff}
\end{aligned}$$ where we suppress all flavor indices, $u^T=( u, c, t)$ and $d^T = ( d, s, b)$ denote the up and down type quarks, $V_{CKM}$ is the Cabibbo-Kobayashi-Maskawa (CKM) matrix, ${\bf m_{D(U)}}$ is the diagonalized mass matrix of down (up) type quarks, and $P_{R, L} = (1 \pm \gamma_5)/2$.
Production and decays of doubly charged Higgs boson
===================================================
The doubly charged Higgs could be produced by EW interactions via the s channel processes: $$\begin{aligned}
& p p \rightarrow Z/\gamma \rightarrow \delta^{++} \delta^{--} \,, \label{eq:EW1}\\
& p p \rightarrow W^\pm \rightarrow \delta^{\pm \pm} H_{1,2}^{\mp} \,. \label{eq:EW2}\end{aligned}$$ Except for the new mixing effect $\theta_+$, the production channels are similar to those in the original Type-II seesaw model. Moreover, with the sizable mixing angle $\theta_+$, the on-shell $\delta^{\pm \pm}$ could be produced through QCD interactions: $$\begin{aligned}
\label{eq:QCDI}
& p p \rightarrow H_2^+ \bar{t} b \ (H_2^- t \bar b) \rightarrow \delta^{++} W^- \bar{t} b \ (\delta^{--} W^+ t \bar b) \,, \\
\label{eq:QCDII}
& p p \rightarrow H_2^+ \bar{t} \ (H_2^- t) \rightarrow \delta^{++} W^- \bar{t} \ (\delta^{--} W^+ t)\,.
\end{aligned}$$ The on-shell $\delta^{\pm \pm}$ in Eq. (\[eq:QCDI\]) and (\[eq:QCDII\]) is generated by the decay $H^\pm_2 \to \delta^{\pm\pm} W^{\mp}$ since we adopt mass relation of $m_{H^\pm_1} < m_{\delta^{\pm \pm}} < m_{H^\pm_2}$. The production of $\delta^{\pm\pm}$ through lighter charged Higgs $H^\pm_1$ is off-shell effects and small, we therefore ignore its contributions. For the processes in Eq. (\[eq:QCDII\]), the main QCD reaction is associated with the interactions of b-quark and gluons, e.g. $\bar b(b) g \to H^+_2 \bar t ( H^-_2 t)$.
In our analysis, we apply following conditions for the parameters in the model: $$\begin{aligned}
m_{\delta^\pm} = m_{\delta^{\pm \pm}} + 100 \ {\rm GeV} \,, \quad
m_{H^\pm} = \frac{4}{5} m_{\delta^{\pm}} \,, \quad
\mu_1 = -\mu_2 = m_{\delta^\pm} \sin \beta \cos \beta \,, \label{eq:setting}\end{aligned}$$ which give sizable mixing angle $\theta_+$ and the mass relation of $m_{H^\pm_1} < m_{\delta^{\pm \pm}} < m_{H^\pm_2}$. The masses of singly charged Higgs and mixing angle are then obtained from Eq. (\[eq:mass\_mixing\]). We then calculate the cross sections of $\delta^{\pm \pm}$ production processes using [*CalcHEP 3.6.15*]{} code [@CalcHEP] by implementing the model. Applying the settings of Eq. (\[eq:setting\]) and $\tan\beta =1$, we show the production cross sections for the processes in Eqs. (\[eq:EW1\])-(\[eq:QCDII\]) as a function of $m_{\delta^{\pm\pm}}$ in Fig. \[fig:XS\], where the collision energy at LHC is 13 TeV and [CTEQ6L]{} PDF is used; the dotted, dash-dotted and dash-dot-dotted lines denote the EW processes while the solid and dashed lines stand for QCD processes, respectively. We therefore find that QCD production cross sections are significantly larger than that of EW production in our setting.
The produced $\delta^{\pm \pm}$ can decay into $\ell^\pm \ell^\pm$, $W^\pm W^\pm$, $H^\pm_i H^\pm_j$, $W^\pm H^\pm_j$, etc. With our parameter setting, the dominant decay modes are $W^\pm H^\pm_{1} (H^{*\pm}_2)$ where $H^{*\pm}_2$ is off-shell, since we assume $v_\Delta << v$ and small Yukawa couplings for triplet suppressing $W^\pm W^\pm$ and $\ell^+ \ell^+$ modes. Then the lighter singly charged Higgs $H^{+(-)}_{1}$ dominantly decays into $t \bar b(\bar t b)$; the off-shell $H^{*+(-)}_2$ also coverts to $t \bar b(\bar t b)$.
![Product cross sections of the doubly charged Higgs as a function of mass for a collision energy of 13 TeV at the LHC; the dotted, dash-dotted, and dash-dot-dotted lines denote the EW processes, while the solid and dashed lines stand for the QCD processes. The parameter setting of Eq. (\[eq:setting\]) and $\tan \beta =1$ are used. (reproduced from [@Chen:2014qda].)[]{data-label="fig:XS"}](XS_Plot.eps){width="70mm"}
Simulation study
================
In this section, we discuss the simulation study of possible signal and background events. We then estimate the significance of discovering the doubly charged Higgs after applying relevant kinematic cuts (see also [@Chen:2014qda] for details). In order to generate the simulation events, we employ the [MADGRAPH/MADEVENT5]{} [@Ref:MG] and use [PYTHIA6]{} [@Ref:Pythia] to deal with the fragmentation of hadronic effects, the initial-state radiation (ISR) and final-state radiation (FSR) effects, and the decays of SM particles e.g. $W$-boson, $t$-quark, etc. In addition, the generated events are also run though the [PGS4]{} detector simulation [@Ref:PGS].
The $\delta^{\pm \pm}$ is dominantly produced by the QCD production processes Eqs. (\[eq:QCDI\]) and (\[eq:QCDII\]), and decays as $\delta^{++(--)} \to W^{+(-)} H^{+(-)}_{1,2} \to W^{+(-)} t \bar b (\bar t b) \to W^{+(-)} W^{+(-)} b \bar b$. We thus require the signal events as $$\ell^\pm \ell^\pm + {\rm four \, or \, more \, jets}\,. \label{eq:sig}$$ For background processes we consider following final states at the LHC: $\ell^+ \ell^-$ (+ISR/FSR), $W^\pm W^\pm j j (\alpha^4)$, $W^\pm W^\pm j j (\alpha^2 \alpha_s^2)$, $W^\pm t \bar{t}$, $W^\pm t \bar{t} j$, $W^\pm Z +n j$ and $Z Z+ nj$, where the number of jets $n$ for VV backgrounds is taken as $n \leq 2$. $W^\pm W^\pm +nj$ events in $VV$ background have been included in EW and QCD background, therefore they should be excluded.
For enhancing the significance of $\delta^{\pm\pm}$ signals by reducing the possible backgrounds, we need to propose some strategies of kinematical cuts. For excluding the soft leptons and jets, when we generate the events by event generator, we set the preselection conditions for leptons and jets to be $$\begin{aligned}
\label{eq:cuts_basic}
p_T (\ell) > 10 \ {\rm GeV}, \quad \eta(\ell) < 2.5, \quad p_T (j) > 20 \ {\rm GeV}, \quad \eta(j) < 5.0, \end{aligned}$$ where $p_T$ is the transverse momentum and $\eta = 1/2 \ln (\tan \theta/2)$ is pseudo-rapidity with $\theta$ being the scattering angle in the laboratory frame. Furthermore we apply the selection cuts $$\begin{aligned}
N_{\rm b-jet} \geq 1, \quad p_T(\ell_2) < 60 \ {\rm GeV}\,, \quad
M_{\ell^\pm \ell^\pm} < \frac{m_{\delta^{\pm \pm}}}{4}. \label{eq:Mll}\end{aligned}$$ where $N_{\rm b-jet}$ denotes the number of b-jet, $\ell_2$ stands for the second highest $p_T$ charged lepton.
Finally we estimate significance which is defined as [@Ball:2007zza] $S = \sqrt{2[(n_s+n_b)\ln (1+n_s/n_b)-n_s]}$, where $n_s$ and $n_b$ denote the number of signal and background events, respectively. The left (right) panel of Fig. \[fig:significance\] is the estimated significance (luminosity to get 5$\sigma$ discovery) as a function of $m_{\delta^{\pm\pm}}$. By the figure, one can find that the doubly charged Higgs with a mass lower than 330 GeV can be discovered at the LHC with an integrated luminosity of 40 fb$^{-1}$. Additionally, the doubly charged Higgs with a mass of 450 GeV can be discovered at the LHC with an integrated luminosity of 300 fb$^{-1}$.
![ Significance with 40 fb$^{-1}$ (left) and luminosity for $S > 5$ (right) as a function of $m_{\delta^{\pm\pm}}$. The collision energy of 13 TeV is applied for both plots. (reproduced from [@Chen:2014qda].) \[fig:significance\]](Significance.eps "fig:") ![ Significance with 40 fb$^{-1}$ (left) and luminosity for $S > 5$ (right) as a function of $m_{\delta^{\pm\pm}}$. The collision energy of 13 TeV is applied for both plots. (reproduced from [@Chen:2014qda].) \[fig:significance\]](Luminosity.eps "fig:")
Summary
=======
We have discussed the Type-II seesaw model with two Higgs doublet. The new interaction in the model can lead the significant mixing effect between singly charged Higgs of the Higgs doublet and triplet. The mixing results in new production processes and decay patterns of the doubly charged Higgs.
We have shown that QCD processes are the predominant effects to produce the $\delta^{\pm\pm}$, read as $pp \to H_2^+ \bar t b (H_2^- t \bar b) \to \delta^{++} W^- \bar t b (\delta^{--} W^+ t \bar b)$ and $\bar b(b) g \to H_2^+ \bar t (H_2^- t) \to \delta^{++}W^- \bar t (\delta^{--} W^+ t)$ owing to the mixing effect, while other Higgs triplet models are arisen from EW processes. Subsequently the doubly charged Higgs decays as $\delta^{++(--)} \to W^{+(-)} H^{+(-)}_{1,2} \to W^{+(-)} t \bar b (\bar t b) \to W^{+(-)} W^{+(-)} b \bar b$.
We then have investigated the significance of doubly charged Higgs signal by numerical simulation including relevant kinematical cuts. We find that with luminosity of 40 fb$^{-1}$ and collision energy of 13 TeV, $\delta^{\pm\pm}$ with mass below $330$ GeV could be observed at the $5\sigma$ level. Additionally, the observed mass of $\delta^{\pm\pm}$ could be up to 450 GeV when the luminosity approaches 300 fb$^{-1}$.
This work is supported by the Ministry of Science and Technology of R.O.C. under Grant \#: MOST-103-2811-M-006-030. We also thank the National Center for Theoretical Sciences (NCTS) for supporting the useful facilities.
[99]{} M. Magg and C. Wetterich, Phys. Lett. B [**94**]{}, 61 (1980); G. Lazarides, Q. Shafi and C. Wetterich, Nucl. Phys. B [**181**]{}, 287 (1981); R. N. Mohapatra and G. Senjanovic, Phys. Rev. D [**23**]{}, 165 (1981); E. Ma and U. Sarkar, Phys. Rev. Lett. [**80**]{}, 5716 (1998) \[hep-ph/9802445\]. W. Konetschny and W. Kummer, Phys. Lett. B [**70**]{}, 433 (1977); J. Schechter and J. W. F. Valle, Phys. Rev. D [**22**]{}, 2227 (1980); T. P. Cheng and L. -F. Li, Phys. Rev. D [**22**]{}, 2860 (1980); S. M. Bilenky, J. Hosek and S. T. Petcov, Phys. Lett. B [**94**]{}, 495 (1980). S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], Eur. Phys. J. C [**72**]{}, 2189 (2012) \[arXiv:1207.2666 \[hep-ex\]\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Eur. Phys. J. C [**72**]{}, 2244 (2012) \[arXiv:1210.5070 \[hep-ex\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], arXiv:1410.6315 \[hep-ex\]. C. H. Chen and T. Nomura, Phys. Rev. D [**90**]{}, 075008 (2014) \[arXiv:1406.6814 \[hep-ph\]\]. C. H. Chen and T. Nomura, Phys. Rev. D [**91**]{}, no. 3, 035023 (2015) \[arXiv:1411.6412 \[hep-ph\]\].
A. Pukhov, \[hep-ph/0412191\].
J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H.-S. Shao and T. Stelzer [*et al.*]{}, JHEP [**1407**]{}, 079 (2014) \[arXiv:1405.0301 \[hep-ph\]\]. T. Sjostrand, S. Mrenna, P. Z. Skands, JHEP [**0605** ]{}, 026 (2006). <http://www.physics.ucdavis.edu/conway/research/software/pgs/pgs4-general.htm>.
G. L. Bayatian [*et al.*]{} \[CMS Collaboration\], J. Phys. G [**34**]{}, 995 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We prove a bijection between the triangulations of the 3-dimensional cyclic polytope $C(n,3)$ and persistent graphs with $n-2$ vertices. We show that under this bijection, the Stasheff-Tamari orders naturally translate to subgraph inclusion. Moreover, we describe a connection to the second higher Bruhat order $B(n-2,2)$. We also give an algorithm to efficiently enumerate all persistent graphs on $n-2$ vertices and thus all triangulations of $C(n,3)$.
**Keywords:** computational geometry, persistent graphs, time series visibility, Bruhat order, counting and enumeration
author:
- Vincent Froese
- 'Malte Renken[^1]'
bibliography:
- 'bibliography.bib'
title: Terrain Visibility Graphs and Cyclic Polytope Triangulations
---
Introduction
============
In this work we reveal a new connection between cyclic polytopes and certain geometric graphs by proving a one-to-one correspondence between the triangulations of the 3-dimensional cyclic polytope and persistent graphs. Cyclic polytopes are natural generalizations of convex polygons to higher dimensions and are among the most studied classes of polytopes. They are neighborly and achieve the maximum number of faces according to the upper bound theorem of @mcmullen_1970. Triangulations of the $d$-dimensional cyclic polytope $C(n,d)$ are well-studied [@RV12]. It is known that a triangulation is fully determined by the set of its $\lfloor d/2 \rfloor$-dimensional faces [@Dey93]. For even dimension $d$, a combinatorial description of this set is known [@OT12]. For odd dimension, however, no characterization is known so far.
We give a characterization for $d=3$ by proving a bijection between the set of triangulations of $C(n,3)$ and the class of so-called *[persistent graph]{}s*. These (vertex-ordered) graphs are best known for their conjectured equality to the set of *[terrain visibility graph]{}s*. A graph $G$ is a *[terrain visibility graph]{}* (sometimes referred to as 1.5-dimensional [terrain visibility graph]{}), if there exists a sequence of points $p_i \in \RR^2$ with ascending $x$-coordinates (the vertices of $G$) such that there is an edge between $p_i$ and $p_j$ if and only if the line segment connecting $p_i$ and $p_j$ does not pass below any other point in between. [Terrain visibility graph]{}s are known to be persistent ( they satisfy certain combinatorial properties). It is an open question, whether every [persistent graph]{} is also a [terrain visibility graph]{}. Our result opens new directions to answer this question and simultaneously provides new insights into the combinatorial structure cyclic polytope triangulations.
Related Work
------------
For a general overview on triangulations, see the monograph by @DRS10. The triangulations of cyclic polytopes and their poset structures have received considerable interest [@KV91; @ER96; @Rambau97; @RV12; @OT12]. Also, efforts have been made to determine the number of triangulations [@AS02; @RV12; @JK18]. @Thomas02 gave a bijection between the triangulations of the cyclic polytope $C(n, d)$ and so-called *snug partitions* of the set $[n-1]^d$.
[Terrain visibility graph]{}s are closely related to so-called orthogonal staircase polygons [@Colley92]. In this context, they were studied by @AbelloStaircaseI, who proved that they are persistent (and claimed the converse implication, which is still open). A simplified proof of this results was more recently published by @Saeedi2015, who also showed a connection to certain restricted 3-signotopes. Some graph-theoretic results regarding (forbidden) induced subgraphs of [terrain visibility graph]{}s and relation to other graph classes are known [@FR19]. Interestingly, in the context of time series data, [terrain visibility graph]{}s (there called time series visibility graphs) have received a lot of attention as an analytical tool [@Lacasa4972] (see also references in [@FR19]). Also related classes such as [terrain visibility graph]{}s with uniform step length [@AbelloStaircaseUniform] and *horizontal visibility graphs* [@GUTIN20112421] have been individually studied (the latter are shown to be exactly the outerplanar graphs containing a Hamilton path).
Preliminaries
-------------
We introduce some notation, basic definitions and preliminary results.
#### Notation.
We define $[n] := \{1,\ldots,n\}$ and denote the set of all size-2 subsets of $[n]$ by $\binom{[n]}{2}$. The convex hull of a set $S$ of points is denoted $\conv(S)$. We assume the reader to be familiar with the basics of the theory of polytopes (see e.g. @Ziegler95). For a polytopal complex $C$, we denote the set of $i$-dimensional faces of $C$ as $F_i(C)$ and we write $f_i(C) := \abs{F_i(C)}$. The $i$-skeleton of $C$ is defined as $\operatorname{skel}_i(C)=\bigcup_{j=0}^iF_i(C)$. Note that the 1-skeleton defines a graph with vertices $F_0(C)$ and edges $F_1(C)$. Throughout this work, we always consider combinatorial faces and simplices, that is, we only consider the corresponding vertex sets.
#### Cyclic Polytopes.
\(1) at (-3,9) ; (2) at (-2,4) ; (3) at (-1,1) ; (4) at (0,0) ; (5) at (1,1) ; (6) at (2,4) ; (7) at (3,9) ; (1) – (2) – (3) – (4) – (5) – (6) – (7); in [1,2,...,6]{} [ () – (7); ]{} in [2,3,...,6]{} [ (1) – (); ]{}
For an integer $d \ge 1$, the $d$-dimensional cyclic polytope is defined via the $d$-th *moment curve*: $$\mu_d \colon \mathbb{R} \to \mathbb{R}^d,\; t \mapsto (t,t^2,\ldots,t^d).$$ Let $t_1 < t_2 < \ldots < t_n$ be $n > d$ real numbers. Then, $$C(n,d) \coloneqq \conv\{\mu_d(t_1),\ldots,\mu_d(t_n)\}$$ is the $d$-dimensional cyclic polytope with $n$ vertices. It is well-known that the combinatorics of $C(n,d)$ do not depend on the particular values of $t_1,\ldots,t_n$ but just on the number $n$. In this work, we consider $C(n,3)$ and denote its vertices by $0, 1, \dots, n-1$, ordered by their first coordinate. The faces of $C(n,3)$ are determined by *Gale’s evenness criterion* [@Gale63 Theorem 3] as follows (see for an example): $$\begin{aligned}
F_1(C(n,3)) &= \set{\{0,n-1\}, \{0,i\},\{i,n-1\},\{i,i+1\} ; 0< i < n-1},\\
F_2(C(n,3)) &= \set{\{0,i,i+1\}, \{i-1,i,n-1\} ; 0 < i < n-1}.\end{aligned}$$
A *triangulation* of $C(n,3)$ is a collection $T=\{S_1,\ldots,S_m\}$ of 3-simplices (that is, tetrahedra) $S_i=\{a,b,c,d\}\subseteq\{0,\ldots,n-1\}$, such that $\bigcup_{i=1}^m\conv(S_i) = C(n,3)$ and each pair of 3-simplices intersects in a common (possibly empty) face. We denote the set of all triangulations of $C(n,3)$ by ${\mathcal{T}_{n}}$.
We proceed with some known results about characterizing triangulations of $C(n,3)$. We will use these in order to prove our main result. A *circuit* (also called a *primitive Radon partition*) is a pair $(X,Y)$ of disjoint minimal subsets of vertices of $C(n,3)$ such that $\conv(X) \cap \conv(Y) \neq \emptyset$. The circuits of $C(n,3)$ are easily characterized as follows:
\[thm:cyclic\_circuits\] The circuits of $C(n,3)$ are exactly the pairs $(\{u, v, w\}, \{x, y\})$ with $u < x < v < y < w$.
The above result on circuits allows us to give the following characterization of a triangulation of $C(n,3)$ as a direct consequence of [@Rambau97 Proposition 2.2].
\[thm:triangulation\_conditions\] A set $T$ of 3-simplices with vertices from $C(n,3)$ is a triangulation of $C(n,3)$ if and only if
1. for each $S \in T$ and each facet $F$ of $S$ either $F \in F_2(C)$ or there is another 3-simplex $S' \in T$ of which $F$ is a facet (Union-Property), and
2. there is no pair of 3-simplices $S, S' \in T$ such that $\{x_1, x_3, x_5\} \subset S$ and $\{x_2, x_4\} \subset S'$ for any $x_1 < x_2 < x_3 < x_4 < x_5$ (Intersection-Property).
The next observation states that an internal edge of a triangulation is contained in at least three 3-simplices. It follows directly from the union-property of .
\[thm:internal\_edge\_surrounded\] Let $T\in {\mathcal{T}_{n}}$ and let $\{v, w\}$ be an internal edge, i.e., $\{v, w\} \in F_1(T) \setminus F_1(C(n,3))$. Then, there are $k \geq 3$ vertices $x_1, \ldots, x_k, x_{k+1} = x_1$ such that $\{v, w, x_i, x_{i+1}\} \in T$ for all $i = 1, \dots, k$.
This leads us to the following helpful lemma about internal edges.
at (-1,-1) [$H^-$]{}; at (-0.2,-2) [$H^+$]{}; (1) at (-3,9) ; (2) at (-2,4) ; (3) at (0,0) ; (4) at (0.8,0.64) ; (5) at (1.5,2.25) ; (6) at (2.3,5.29) ; (7) at (3,9) ; (1) – (2) – (3) – (4) – (5) – (6) – (7); in [1,2,...,6]{} [ () – (7); ]{} in [2,3,...,6]{} [ (1) – (); ]{} ($-0.5*(2.3,5.29)$) – ($1.4*(2.3,5.29)$);
at (7,13.5) [$H^+$]{}; at (7,11) [$H^-$]{}; (v) at (0,0) ; (1) at (1,1) ; (2) at (2,2.82) ; (w) at (3.2,5.72) ; (3) at (6,14.69) ; (4) at (2,-2.82) ; (5) at (4,-8) ; (v) – (1) – (2) – (w) – (3); (v) – (4) – (5); (3) – (2); (3) – (1); (3) – (v); (3) – (5); (3) – (4); (3) – (5); (5) – (v); (5) – (1); (5) – (2); (5) – (w); (-1,-1.78) – (7,12.512);
\[thm:super\_bar\_property\] Let $T$ be a triangulation of $C:=C(n,3)$ and let $\{v, w\} \in F_1(T) \setminus F_1(C)$ be an internal edge with $v < w$. Then, there are vertices $a, b, c$ with $a < v < b < w < c$ such that $\{\{v, w, a\},\{v, w, b\},\{v, w, c\}\}\subseteq F_2(T)$.
Note that $0 < v$ and $w < n-1$ since $\{v,w\}$ is an internal edge. Let $x_1, \ldots, x_{k+1}=x_1$ be the $k\ge 3$ vertices given by \[thm:internal\_edge\_surrounded\]. Then $T' := \set{\set{v, w, x_i, x_{i+1}}; i=1,\dots,k}$ is a triangulation of $K := \conv\{v, w, x_i, \dots, x_k\}$. Since the facets of $T'$ that contain $\{v, w\}$ are exactly those of the form $\{v, w, x_i\}$ and since each of these appears in two 3-simplices in $T'$, it follows that $\{v, w\}$ is an internal edge of $T'$. Therefore, every plane $H \subset \RR^3$ containing $v$ and $w$ (and thus $\conv\{v, w\}$) divides $K$ into two nonempty 3-dimensional polytopes. Then, the two corresponding open half-spaces $H^+$ and $H^-$ must each contain a vertex from $\{x_1, \dots, x_k\}$. Now, assume without loss of generality that the coordinates of $v$ are $(0, 0, 0)$.
If we take $H$ as the plane containing $\{v, w, (0, 0, 1)\}$ (that is, containing the $z$-axis and $w$) and $H^+$ as the open half-space not containing the vertex $0$, then $H^+ \cap F_0(C) = \set{u \in F_0(C); v < u < w}$ (see \[fig:plane\] (left)). Since $H^+ \cap \{x_1, \dots, x_k\} \subseteq H^+ \cap F_0(C)$ and $H^+ \cap \{x_1, \dots, x_k\}\neq \emptyset$, there exists a vertex $b$ as claimed.
If we instead take $H$ to be the plane containing $\{v, w, (1, 0, 0)\}$ (that is, containing the $x$-axis and $w$), and $H^+$ as the open half-space containing the vertex $n-1$, then $H^+ \cap F_0(C) = \set{u \in F_0(C); w < u}$ (see (right)), and thus, $H^+ \cap \{x_1, \ldots, x_k\}$ contains a vertex $c$ as claimed. The existence of a vertex $a$ as claimed follows by symmetry.
#### Terrain Visibility and Persistent Graphs.
[Terrain visibility graph]{}s are visibility graphs of 1.5-dimensional *terrains*, that is, $x$-monotone polygonal chains in the plane defined by a set $V\subseteq\mathbb{R}^2$ of *terrain vertices* with pairwise different $x$-coordinates. Two vertices $v_1=(x_1,y_1)$ and $v_2=(x_2,y_2)$ are adjacent if and only if they *see each other*, that is, there is no vertex between them that lies on or above the line segment connecting them. Formally, there exists an edge $\{v_1,v_2\}$, for $x_1< x_2$, if and only if all terrain vertices $(x,y)$ with $x_1 < x < x_2$ satisfy $$\begin{aligned}
y < y_1 + (x - x_1) \frac{y_2 - y_1}{x_2 - x_1}.\end{aligned}$$ depicts an example. We denote the vertices by $1,\ldots,n$ in increasing order of their $x$-coordinates.
[Terrain visibility graph]{}s are known to be persistent [@Saeedi2015] where a graph $G=([n], E)$ is called *persistent* if it satisfies the following three properties.
1. It contains a Hamilton path from $1,\ldots,n$, that is, $\{\{i,i+1\}\mid 1\le i < n\}\subseteq E$
2. **[X-property]{}**: If $\{a,c\}\in E$ and $\{b,d\}\in E$ for some vertices $a < b < c < d$, then $\{a,d\}\in E$.
3. **[bar-property]{}**: For every edge $\{a,b\}\in E$ with $a < b-1$, there exists a vertex $x$ with $a < x < b$ such that $\{a,x\}\in E$ and $\{x,b\}\in E$.
It is still open whether every persistent graph is also a [terrain visibility graph]{}. We denote the set of all persistent graphs with $n$ vertices by ${\mathcal{P}_{n}}$.
\(1) at (0,1.5) ; (2) at (1,0) ; (3) at (2,1) ; (4) at (2.5,-1) ; (5) at (3.5,4) ; (6) at (4.25,2) ; (7) at (5,3) ; (1) – (2) – (3) – (4) – (5) – (6) – (7); (1) – (3); (1) – (5); (2) – (5); (3) – (5); (5) – (7);
\(1) at (1,0) ; (2) at (2,0) ; (3) at (3,0) ; (4) at (4,0) ; (5) at (5,0) ; (6) at (6,0) ; (7) at (7,0) ; (1) – (2) – (3) – (4) – (5) – (6) – (7); (1) to (3); (1) to (5); (2) to (5); (3) to (5); (5) to (7);
We prove the following elementary property about [persistent graph]{}s, which states that consecutive neighbors of a vertex are also neighbors of each other. Here, $N(v)$ denotes the neighborhood of vertex $v$ and $N[v] = N(v)\cup\{v\}$.
\[thm:consecutive\_neighbors\] Let $G=([n],E)$ be a [persistent graph]{} and let $a,b,c$ be vertices such that $b < c$, $\{b,c\}\subseteq N(a)$, and there is no vertex $x \in N[a]$ with $b < x < c$. Then $\{b, c\} \in E$.
We assume that $a < b < c$ (the case $b < c < a$ is fully symmetric). By the [bar-property]{}, there exists a common neighbor $x$ of $a$ and $c$ with $a < x < c$. Note that $b < x$ is not possible by assumption. If $x = b$, then we are done. Otherwise, we have $a < x < b$ and the [bar-property]{} again implies the existence of a common neighbor $x'$ of $x$ and $c$ with $x < x' < c$. Now, if $b < x'$, then the [X-property]{} (applied to $\{a,b\}$ and $\{x,x'\}$) implies that $x'$ is a neighbor of $a$ which contradicts our assumption on $b$ and $c$. Thus, $a < x' \le b$. Note that we can repeat the above argument again on $x'$ if $x' < b$. Since $G$ is finite, we can conclude that $b$ is a neighbor of $c$.
A Bijection Between ${\mathcal{T}_{n+2}}$ and ${\mathcal{P}_{n}}$
=================================================================
In this section, we prove a bijection between triangulations of $C(n+2,3)$ and persistent graphs on $n$ vertices. The central observation is that the 1-skeleton of a triangulation restricted to the vertices $1,\ldots,n$ forms a persistent graph (see for an example). Formally, the bijection is defined as follows.
\(1) at (-3,9) ; (2) at (-2,4) ; (3) at (-1,1) ; (4) at (0,0) ; (5) at (1,1) ; (6) at (2,4) ; (7) at (3,9) ; (1) – (2) – (3) – (4) – (5) – (6) – (7); in [1,2,...,6]{} [ () – (7); ]{} in [2,3,...,6]{} [ (1) – (); ]{} (7) – (2); (7) – (4); (7) – (6); (2) – (6); (2) – (4); (4) – (6);
\(1) at (1,0) ; (2) at (2,0) ; (3) at (3,0) ; (4) at (4,0) ; (5) at (5,0) ; (1) – (2) – (3) – (4) – (5); (1) to (3); (3) to (5); (1) to (5);
For $n\ge 2$, the map $\Gamma \colon {\mathcal{T}_{n+2}} \rightarrow {\mathcal{P}_{n}}$ is defined as $$\Gamma(T) := \left([n],F_1(T)\cap \binom{[n]}{2}\right),$$ that is, two vertices $i$ and $j$ are adjacent in $\Gamma(T)$ if and only if $\{i,j\}\subseteq S$ for some 3-simplex $S\in T$.
First, we show that $\Gamma$ is well-defined, that is, $\Gamma(T)$ is in fact a persistent graph.
For every $T\in{\mathcal{T}_{n+2}}$, it holds $\Gamma(T)\in {\mathcal{P}_{n}}$.
Let $C:=C(n+2,3)$. Clearly, $\Gamma(T)$ contains a Hamilton path from 1 to $n$ since $\{i,i+1\}\in F_1(C)$ and thus $\{i,i+1\}\in F_1(T)$ for each $i\in[n-1]$.
Next, we show that $\Gamma(T)$ satisfies the [bar-property]{}. Let $e = \{v, w\}$ be an edge of $\Gamma(T)$ with $v < w -1$. Then, $e$ is an internal edge, that is, $e \in F_1(T) \setminus F_1(C)$. Hence, by \[thm:super\_bar\_property\], there exists a vertex $b$ with $v < b < w$ such that $\{v,w, b\}\in F_2(T)$. Therefore, $\{v,b\}$ and $\{b, w\}$ are edges of $\Gamma(T)$.
Now, for the [X-property]{}, assume towards a contradiction that $\Gamma(T)$ contains the edges $\{u, w\}$ and $\{v, x\}$ with $u < v < w < x$, but $\{u, x\} \notin E(\Gamma(T))$. Let $(u, v, w, x)$ be lexicographically minimal with this property. Note that $\{u,w\}$ and $\{v,x\}$ are both internal edges. Thus, \[thm:super\_bar\_property\] applied to $\{v,x\}$ implies that there exists a vertex $a < v$ such that $\{a,v,x\}\in F_2(T)$ (and thus $\{\{a,v\},\{a,x\}\}\subseteq E(\Gamma(T))$). By minimality of $v$, it follows $a \leq u$. If $a < u$, then $\{a, v, x\}$ and $\{u, w\}$ are subsets of two different 3-simplices of $T$, contradicting the intersection-property of \[thm:triangulation\_conditions\] (since $(\{a,v,x\},\{u,w\})$ is a circuit). Thus, it follows $a = u$, that is, $\{u,x\}\in E(\Gamma(T))$, which is a contradiction.
In order to show that $\Gamma$ is a bijection, we next define a map that maps a persistent graph to a triangulation. We then prove that this map is the inverse of $\Gamma$. To start with, we define the following auxiliary graph.
\[def:supergraph\] For a [persistent graph]{} $G = ([n], E)$, we define the supergraph $\hat{G}:=(\{0, \ldots, n+1\}, E \cup (\{0,n+1\}\times\{0,\ldots,n+1\}))$, that is, $\hat{G}$ contains two additional vertices that are connected to all other vertices.
It is easy to see that $\hat{G}$ is a [persistent graph]{} since adding a vertex that is adjacent to all others cannot violate the X- or [bar-property]{}. Using , we now introduce the inverse map $\Xi$.
Let $G=([n], E)$ be a [persistent graph]{}. For $e = \{v, w\} \in E$, $v < w$, we define $$\begin{aligned}
\ell_G(e) &:= \max\set{i \in V(\hat{G}) ; i < v \land \{i, w\} \in E(\hat{G})} , \text{ and}\\
r_G(e) &:= \min\set{i \in V(\hat{G}) ; w < i \land \{i, w\} \in E(\hat{G})} .
\end{aligned}$$ Further, we define the 3-simplex $\xi_G(e):=\{\ell_G(e), v, w, r_G(e)\}$ and the map $\Xi\colon {\mathcal{P}_{n}} \rightarrow {\mathcal{T}_{n+2}}$ as $$\Xi(G) := \set{\xi_G(e); e \in E}.$$ We omit the index $G$ whenever it is clear from the context.
Note that, by construction of $\hat{G}$, the vertices $\ell(e)$ and $r(e)$ always exist. Moreover, by \[thm:consecutive\_neighbors\], the vertices in $\xi(e)$ form a clique in $\hat{G}$. We now show that $\Xi$ is well-defined, that is, $\Xi(G)$ is indeed a triangulation. To this end, we show that $\Xi(G)$ satisfies the union-property and the intersection-property according to . We start with the intersection-property.
\[thm:triangle\_disjointness\] Let $G$ be a [persistent graph]{} and let $a < b < c$ be vertices of a 3-simplex $S \in \Xi(G)$. Then, $G$ does not contain any edge $\{x, y\}$ with $a < x < b < y < c$.
Assume towards a contradiction that there exists such an edge $\{x,y\}$. We assume the vertices to be chosen such that $(c - a) + (y - x)$ is minimal. Note that $\hat{G}$ contains the edge $\{x, y\}$ and a clique on $\{a, b, c\}$. By the [X-property]{}, $\hat{G}$ then also contains the edges $\{a, y\}$ and $\{x, c\}$.
Let $e \in E(G)$ be an edge with $\xi(e) = S$. Then, by definition of $\xi$, it follows that $e \nsubseteq \{a, b, c\}$. To see this, note that $e=\{a,c\}$ is not possible since $b\not\in\{\ell(e),r(e)\}$. Also $e = \{a, b\}$ is not possible, since $c\neq r(e)$ (since $y < c$ is also neighbor of $a$). Analogously, $e=\{b,c\}$ is not possible, since $a\neq \ell(e)$.
Therefore, assume that $e = \{b, b'\}$ with $b < b' < c$ and $a = \ell(e)$ and $c = r(e)$. We cannot have $b' > y$, because then we could replace $c$ by $b'$ and decrease $(c-a) + (y-x)$. Also $b' = y$ is impossible since then $a = \ell(e)$ would contradict the fact that $x$ is a neighbor of $y$. Hence, $b' < y$. From the [bar-property]{} and the minimality of $(c-a) + (y-x)$, it follows that $x$ and $y$ have a common neighbor $z$ with $b \leq z \leq b'$ (see ). Then, by the [X-property]{}, there exists the edge $\{b, y\}$ (contradicting $r(e) = c$) or the edge $\{x, b'\}$ (contradicting $\ell(e) = a)$.
\(a) at (0,0) ; (x) at (1.5,0) ; (b) at (3,0) ; (z) at (4,0) ; (b’) at (5,0) ; (y) at (6.5,0) ; (c) at (8,0) ; (a) to (c); (b) to (c); (b) to (b’); (x) to (y); (a) to (b’); (x) to (z); (y) to (z); (x) to (b’); (b) to (y); (a) to (b); (b’) to (c);
Next, we prove that $\Xi(G)$ satisfies the union-property.
\[thm:image\_triangulation\] Let $G\in{\mathcal{P}_{n}}$, $C := C(n+2, 3)$, and let $S \in \Xi(G)$ be a 3-simplex containing vertices $a < b < c$ such that $\{a, b, c\}\not\in F_2(C)$. Then, there exists another 3-simplex $S'\in\Xi(G)$ with $\{a,b,c\}\subset S'$.
Fix an edge $e \in E(G)$ with $S = \xi(e)$ (recall that the vertices $\xi(e)$ form a clique in $\hat{G}$ by ). By definition of $\xi$, it holds $b \in e$. , we can assume that either $e = \{a, b\}$ or $e = \{b, b'\}$ with $b < b' < c$ (the cases $e=\{b,c\}$ or $e = \{b, b'\}$ with $a < b' < b$ are symmetric). The following case distinction yields the existence of a vertex $x$ with $a < x < b$ such that $x$ is a common neighbor of $b$ and $c$ in $\hat{G}$.
**Case 1:** $c = n+1$. Since $\{a, b, c\}$ is not a face of $C$, we have $a + 1 < b$. Clearly, the vertex $b-1$ is a neighbor of $b$ and $c$ in $\hat{G}$ (by construction).
**Case 2:** $c < n+1$. If $\ell(\{b, c\}) = a$, then the 3-simplex $\xi(\{b, c\})$ also contains $\{a, b, c\}$ and we are done. Otherwise, $x=\ell(\{b,c\})$ is a common neighbor of $b$ and $c$ in $\hat{G}$ by .
In the following, we assume $x$ to be chosen minimally. By \[thm:triangle\_disjointness\], $x$ has no neighbor between $b$ and $c$. Thus, $r(\{x, b\}) = c$. Furthermore, $b$ has no neighbor between $a$ and $x$ because, by the [X-property]{}, this would also be a neighbor of $c$, contradicting the minimality of $x$. Therefore, $\ell(\{x, b\}) = a$ and thus, $\xi(\{x, b\})$ contains $\{a, b, c\}$ (note that $\{x,b\}\neq e$).
together with now yield the following.
For every $G\in{\mathcal{P}_{n}}$, it holds $\Xi(G)\in{\mathcal{T}_{n+2}}$.
Finally, we prove that $\Xi$ is the inverse of $\Gamma$.
\[thm:bijection\] The map $\Gamma$ is a bijection with $\Gamma^{-1} = \Xi$ (and thus $\Xi$ is also a bijection).
First, we show that $\Gamma \circ \Xi = {\mathrm{id}}$. Let $G=([n],E)$ be a persistent graph. Note that, by definition, for each edge $e\in E$, the 3-simplex $\xi(e)$ contains $e$, that is, $e\in F_1(\Xi(G))$. Thus, $E\subseteq E(\Gamma(\Xi(G)))$. Moreover, since the vertices in $\xi(e)$ form a clique in $\hat{G}$ (by ), it follows that $(F_1(\xi(e)) \cap \binom{[n]}{2}) \subseteq E$. Thus, $E(\Gamma(\Xi(G))) \subseteq E$. Hence, $\Gamma(\Xi(G)) = G$.
To see that $\Xi \circ \Gamma = {\mathrm{id}}$, let $T$ be any triangulation of $C = C(n+2, 3)$ and let $S \in T$ be a 3-simplex. Let $a < b < c < d$ be the vertices of $S$. We claim that $a = \max\set{i; 0 \leq i < b, \{i,c\}\in F_1(T)}$. Assume towards a contradiction that there exists a vertex $x$ with $a < x < b$ and $\{x,c\}\in F_1(T)$. Then, $\{a, b, d\}$ and $\{x, c\}$ are subsets of two different simplices of $T$, contradicting the intersection-property of . By symmetry, we also obtain that $d = \min\set{i; c < i \leq n+1, \{i,b\}\in F_1(T)}$. Now, since $\Gamma(T)$ contains the edge $\{b, c\}$, it follows that $\Xi(\Gamma(T))$ contains $S$. Thus, $T \subseteq \Xi(\Gamma(T))$. Since $T$ and $\Xi(\Gamma(T))$ are triangulations of $C$ (by \[thm:image\_triangulation\]), this implies $T = \Xi(\Gamma(T))$.
An interesting observation is that, for any $G \in {\mathcal{P}_{n}}$, the map $\xi_G \colon E(G) \rightarrow \Xi(G)$ is a bijection. Its inverse is given by the map $\{a, b, c, d\} \mapsto \{b, c\}$, where $a < b < c < d$. This implies that the number of edges in $G$ equals the number of 3-simplices in $\Xi(G)$.
To close this section, we compare our result for $d=3$ with the characterization for even $d$ by @OT12. They showed that for every triangulation $T$ of $C(n,2k)$, the set of $k$-dimensional faces of $T$ that do not contain $\{i,i+1\}$ for some $i$ contains exactly $\binom{n-k-1}{k}$ *non-intertwining* tuples from $\{0,\ldots,n-1\}^{k+1}$, where $(a_0,\ldots,a_k)$ *intertwines* $(b_0,\ldots,b_k)$ if $a_0 < b_0 < a_1 < b_1 < \dots < a_k < b_k$. Conversely, they also proved that every non-intertwining set of size $\binom{n-k-1}{k}$ (which is maximal) defines a unique triangulation. For $k=1$, this gives a one-to-one correspondence between triangulations of $C(n,2)$ and maximal outerplanar graphs (which are chordal). Now, when moving to $d=3$ dimensions, we lose planarity since edges can intertwine but have to satisfy the [X-property]{}. Also, chordality is lost and replaced by the [bar-property]{}.
Stasheff-Tamari Order on Persistent Graphs
==========================================
A classic tool for the analysis of triangulations of cyclic polytopes are the first and second *Stasheff-Tamari orders*, which are certain partial orders on the set of triangulations. In this section we show how these partial orders translate to partial orders on persistent graphs. It is known that the first and second Stasheff-Tamari order are identical on ${\mathcal{T}_{n}}$ [@ER96]. Hence, we will only define and use the first Stasheff-Tamari order here.
Let $C = C(n, 3)$ and $W := \{v_1 < v_2 < \dots < v_5\}$ be a set of five vertices of $C$. Note that $\conv(W)$ equals $C(5,3)$ and has exactly two triangulations: $$\begin{aligned}
T^* &:= \set{\{v_1, v_2, v_3, v_5\}, \{v_1, v_3, v_4, v_5\}} \text{ and}\\
T_* &:= \set{\{v_1, v_2, v_4, v_5\}, \{v_1, v_2, v_3, v_4\}, \{v_2, v_3, v_4, v_5\}}.\end{aligned}$$
Now, let $T$ be a triangulation of $C$ with $T_* \subseteq T$. Then, we obtain a new triangulation $T'$ of $C$ via $T' := (T \setminus T_*) \cup T^*$. In this case, we say that $T'$ is obtained from $T$ by a *bistellar up-flip*, and conversely, $T$ is obtained from $T'$ by a *bistellar down-flip*. For any two triangulations $T, T'$ of $C$, we write $T \leq_1 T'$ if $T'$ is obtained from $T$ by a sequence of bistellar up-flips. This defines a partial order called the first Stasheff-Tamari order [@KV91]. Note that $T \leq_1 T'$ implies that $\abs{T} \geq \abs{T'}$.
The following theorem shows that a bistellar up-flip corresponds to removing a certain edge from the corresponding persistent graph.
\[thm:stasheff-tamari\] Let $T, T' \in {\mathcal{T}_{n}}$. Then, $T'$ is obtained from $T$ by a bistellar up-flip if and only if $E(\Gamma(T)) = E(\Gamma(T')) \cup \{e\}$ for some edge $e\in E(\Gamma(T))$. In particular, $T \leq_1 T'$ if and only if $\Gamma(T) \supseteq \Gamma(T')$.
Let $T \leq_1 T'$ be related by a bistellar up-flip on the vertices $v_1 < \dots < v_5$, that is, $$\begin{aligned}
T &\supseteq \set{
\{v_1, v_2, v_3, v_4\},
\{v_1, v_2, v_4, v_5\},
\{v_2, v_3, v_4, v_5\}} \text{ and}\\
T' &\supseteq \set{
\{v_1, v_2, v_3, v_5\},
\{v_1, v_3, v_4, v_5\}}.
\end{aligned}$$ Then, $E(\Gamma(T)) = E(\Gamma(T')) \cup \set{\{v_2, v_4\}}$.
Conversely, let $G,G' \in {\mathcal{P}_{n-2}}$ with $E(G) = E(G') \cup \{v, w\}$ with $v < w$. Then, clearly $v+1 < w$. Thus, by the [bar-property]{}, there exists $v < y < w$ with $\set{\{v, y\}, \{y, w\}} \subseteq E(G')$. Moreover, $y$ is unique because otherwise the [X-property]{} would imply that $\{v, w\} \in E(G')$. In fact, the [X-property]{} even implies that $\ell_G(\{y, w\}) = v$ and $r_G(\{v, y\}) = w$. Let $x := \ell_G(\{v, w\})$ and $z := r_G(\{v, w\})$.
We claim that $\xi_{G'}(\{v, y\}) = \{x, v, y, z\}$, that is, $\ell_{G'}(\{v, w\})=x$ and $r_{G'}(\{v, w\})=z$. First, note that, since $r_G(\{v, y\}) = w$, we must have $r_{G'}(\{v, y\}) > w$ and thus $r_{G'}(\{v, y\}) = r_G(\{v, w\}) = z$. Now, if $\ell_{G'}(\{v, y\})> x$, then, by the [X-property]{}, $\ell_{G'}(\{v, y\})$ would also be a neighbor of $w$ in $G$, contradicting $x = \ell_G\{v, w\}$. To see that $\hat{G'}$ contains the edge $\{x, y\}$, note that otherwise $x$ would have two consecutive neighbors $a, a'$ with $v \leq a < y < a' \leq w$. By \[thm:consecutive\_neighbors\], this implies that $G'$ contains the edge $\{a, a'\}$ (thus, $\{a, a'\} \neq \{v, w\}$). The [X-property]{} then implies that $G'$ also contains the edges $\{v, a'\}$ and $\{a, w\}$. If $a \neq v$, then this contradicts $\ell_G(\{y, w\}) = v < a$, and if $a' \neq w$, then this contradicts $r_G(\{v, y\}) = w > a'$. Thus, we have $\{x,y\}\in E(\hat{G'})$ implying $\ell_{G'}(\{v, y\})=x$ (and thus also $\ell_G(\{v, y\})=x$). This proves the claim $\xi_{G'}(\{v,y\}) = \{x, v, y, z\}$. Moreover, we clearly have $\xi_G(\{v, y\}) = \{x, v, y, w\}$. From symmetric arguments it follows that $\xi_{G'}(\{y, w\}) = \{x, y, w, z\}$ and $\xi_G(\{y, w\}) = \{v, y, w, z\}$.
Finally, it is not difficult to check that $\xi_G(e) = \xi_{G'}(e)$ for any edge $e\in E(G')$ with $e\nsubseteq \{v,y,w\}$. This is clear if $e\cap\{v,w\}=\emptyset$. It is also clear if $e\in\{v,w\}\times\{u \mid (u < v) \vee (w < u)\}$. For an edge $e=\{v,u\}$ with $v < u < w$, note that $u < y$ holds since otherwise $G'$ would contain the edge $\{v,w\}$ (by the [X-property]{}). Since $r_{G}(\{v,u\})\le y$, it follows $r_{G'}(\{v,u\}) = r_{G}(\{v,u\})$, and thus $\xi_G(e) = \xi_{G'}(e)$. Similarly, for an edge $e=\{w,u\}$ with $y < u < w$, we have $\ell_{G'}(\{w,u\}) = \ell_{G}(\{w,u\}) \ge y$, and thus $\xi_G(e) = \xi_{G'}(e)$.
To sum up, we obtain $$\Xi(G') = (\Xi(G)\setminus \{\{x,v,y,w\},\{x,v,w,z\},\{v,y,w,z\}\}) \cup \{\{x,v,y,z\},\{x,y,w,z\}\},$$ that is, $\Xi(G')$ is obtained from $\Xi(G)$ by a bistellar up-flip on $x < v < y < w < z$.
We close with observing a connection to higher Bruhat orders. @Saeedi2015 [Theorem 3] described a map $\alpha: {\mathcal{P}_{n}} \to B(n, 2)$, where $B(n,2)$ is the second higher Bruhat order (which is isomorphic to the set of 3-signotopes [@FelsnerSignotopesBruhat]).
Moreover, @Rambau97 showed an order-preserving map $f_d \colon B(n,d) \to \text{HST}_1(n+2,d+1)$ from the higher Bruhat order to the first higher Stasheff-Tamari order (see also [@RV12 Theorem 8.9]). It is open whether this map is surjective. It can be observed that our bijection $\Xi$ equals $f_2\circ\alpha$, which implies that $f_2$ is surjective.
Enumerating Triangulations
==========================
The bijection between ${\mathcal{T}_{n+2}}$ and ${\mathcal{P}_{n}}$ has the practical implication that in order to enumerate all triangulations of $C(n,3)$, one can instead enumerate all persistent graphs on $n-2$ vertices. Since these graphs are combinatorially simpler structures, we can thus improve upon previous enumeration efforts [@JK18]. We present a simple and efficient algorithm for the enumeration of [persistent graph]{}s.
For given $n$, let $\Ee := \binom{[n]}{2} \setminus \set{\{i, i+1\}; i\in[n-1]}$ be the set of all potential edges that are not on the obligatory Hamilton path of a [persistent graph]{}. Further, we define a lexicographical order $\preceq$ on $\Ee$ by setting, for any $x_1 < y_1$ and $x_2 < y_2$, $$\{x_1, y_1\} \preceq \{x_2, y_2\} \iff (y_1 < y_2) \lor (y_1 = y_2 \land x_1 \leq x_2).$$
Starting from a path $P_n$, \[alg:count\_persistent\] processes the potential edges $\Ee$ in ascending order and recurses on each edge, either adding or not adding it to the graph. Its efficiency arises mainly from the fact that we can quickly identify and skip edges whose addition would violate the X- or [bar-property]{}. We remark that, while the listing of \[alg:count\_persistent\] assumes that all inputs are copied upon invocation, it is easy to modify the algorithm such that no copying of $G$ is necessary.
output $G$\[line:k=n\] \[line:k<n\] \[line:case1\] $y \gets $ rightmost neighbor of $x$ \[line:case1recursion\] add $\{x, k\}$ to $E$\[line:add\_e\] \[line:case2loop\] add $\{y, k\}$ to $E$ remove $\{y, k\}$ from $E$ (unless $y = k-1$) $y \gets$ rightmost neighbor of $y$
The following proposition states the correctness.
Let $G=([n],E)$ be a graph containing a path on $1,2,\ldots,n$ and let $e=\{x,k\}$, $1 \le x < k \le n$, be such that the following properties hold:
- $E \cap \{e'\in\Ee \mid e' \succ e\} = \emptyset$.
- If $e\not\in E$, then $G$ is persistent.
- If $e\in E$, then either $G$ is persistent or $G$ satisfies the [X-property]{} and $e$ is the only edge violating the [bar-property]{}.
Then outputs exactly all graphs in the set $$\mathcal{P}_G^e := \{G'=([n], E')\in {\mathcal{P}_{n}} \mid (E\subseteq E') \wedge ((E' \setminus E) \subseteq \set{e' \in \Ee; e' \succeq e})\}.$$
If $x+1=k$, then clearly $e\in E$ and $G$ is persistent. If now $k=n$, then clearly $\mathcal{P}_G^e=\{G\}$, that is, \[line:k=n\] is correct. If $k < n$, then $\mathcal{P}_G^e = \mathcal{P}_G^{\{1,k+1\}}$. Thus, \[line:k<n\] is correct.
Now assume that $x + 1 < k$. If $e \notin E$ (\[line:case1\]), then $G$ is persistent. The set $\mathcal{P}_G^e$ can be partitioned into two sets: $$\begin{aligned}
A&:=\{G'=([n],E')\in \mathcal{P}_G^e\mid e \not\in E'\} \text{ and}\\
B&:=\{G'=([n],E')\in \mathcal{P}_G^e\mid e \in E'\}.
\end{aligned}$$ Consider a graph $G'=([n],E')\in A$. Let $y$ be the rightmost neighbor of $x$ in $G$ and note that $y < k$ since $E$ does not contain any edge $e'$ with $e' \succ e$. Due to the [X-property]{}, $E'$ does not contain any edge $\{x', k\}$ with $x < x' < y$. Thus, $A = \mathcal{P}_G^{\{y,k\}}$ and all these graphs are produced by the recursive call in \[line:case1recursion\]. As regards the set $B$, note that $B = \mathcal{P}_{G+e}^e$, where $G+e := ([n],E\cup\{e\})$. Thus, we add $e$ to $E$ in \[line:add\_e\] and then handle this case in \[line:case2loop\].
If $e\in E$, then, for every $G'=([n],E')\in \mathcal{P}_G^e$, there must be a minimal vertex $y$ with $x < y < k$ and $\set{\{x, y\}, \{y, k\}} \subseteq E'$ (by the [bar-property]{}). Since $\{x, y\} \prec e$, it follows that $\{x,y\}\in E$. Hence, $y$ has to be a neighbor of $x$ in $G$ with $x < y < k$. Furthermore, any neighbor $y'$ of $x$ with $x < y' < y$ cannot have a neighbor to the right of $y$, because the [X-property]{} would otherwise imply that also $\{y', k\} \in E'$, contradicting the minimality of $y$. That is, $y$ can only be neighbor of $x$ such that no other neighbor $y'$ of $x$ with $x < y' < y$ has a neighbor to the right of $y$. Let $Y$ denote the vertex set containing all these possible candidates. The for-loop in \[line:case2loop\] iterates exactly over the candidates in $Y$. For a given $y\in Y$, let $A_y\subseteq\mathcal{P}_G^e$ be the subset of graphs $G'=([n],E')$, where $y$ is the minimal vertex with $x<y<k$ and $\set{\{x, y\}, \{y, k\}} \subseteq E'$, and note that $A_y = \mathcal{P}_{G+\{y,k\}}^{\{y,k\}}$. Moreover, $\{A_y\mid y\in Y\}$ is clearly a partition of $\mathcal{P}_G^e$. Hence, calling for each possible $y$, outputs exactly the graphs in $\mathcal{P}_G^e$.
\[thm:alg\_cor\] Let $G=([n], E)$ be the path on $1,2,\ldots,n$. Then outputs exactly $\mathcal{P}_G^{\{1,2\}}={\mathcal{P}_{n}}$.
By \[thm:alg\_cor\], we can use \[alg:count\_persistent\] to efficiently count the number of elements of ${\mathcal{P}_{n}}$ and thus of ${\mathcal{T}_{n+2}}$. The results for $n \le 16$ are listed in \[tab:counts\]. The computations[^2] were performed using an Intel Xeon W-2125 CPU.
$n$ $\abs{{\mathcal{P}_{n}}} = \abs{{\mathcal{T}_{n+2}}}$ computation time
----- ------------------------------------------------------- ------------------
1 1 $<$0.1s
2 1 $<$0.1s
3 2 $<$0.1s
4 6 $<$0.1s
5 25 $<$0.1s
6 138 $<$0.1s
7 972 $<$0.1s
8 8477 $<$0.1s
9 89405 $<$0.1s
10 1119280 $<$0.1s
11 16384508 0.15s
12 276961252 2s
13 5349351298 30s
14 116985744912 12m
15 2873993336097 4h 30m
16 78768494976617 4d 23h 2m
: Number of [persistent graph]{}s with $n\le 16$ vertices. The values for $n\leq13$ were already known [@JK18].[]{data-label="tab:counts"}
Conclusion
==========
Our results yield further insights into the structure of the triangulations of the 3-dimensional cyclic polytope by relating their 1-skeleton to persistent graphs. It remains open to characterize the structure of the $\lfloor d/2\rfloor$-skeleton for arbitrary odd dimension $d$. It is also open whether a closed formula for the number of triangulations of $C(n,3)$ can be given [@RV12 Open Problem 9.2].
On the other side, the bijection might also lead to new insights about persistent graphs. Can the bijection be of help in resolving the conjecture that every [persistent graph]{} is a [terrain visibility graph]{}?
#### Acknowledgment.
We thank Lito Goldmann for his work on enumerating persistent graphs, which led us to the discovery of the bijection. We further thank André Nichterlein for helpful initial discussions.
[^1]: Supported by the DFG project NI 369/17-1.
[^2]: Implementation available at <https://www.akt.tu-berlin.de/menue/software>.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A method is presented for the systematic derivation of a hierarchy of coupled equations for the computation of two-time correlation functions of operators for open many-body quantum systems. We show how these systems of equations can be closed in mean-field and beyond approximations. Results for the specific example of the spectral weight functions are discussed. Our method allows one to access the full temporal evolution, not just the stationary solution, of non-equilibrium open quantum problems described by a Markovian master equation.'
author:
- Zakari Denis
- Sandro Wimberger
title: 'Two-Time Correlation Functions in Dissipative and Interacting Bose–Hubbard Chains'
---
Introduction {#sec:1}
============
A very controlled way of introducing non-trivial dynamics to a Bose–Einstein condensate is to deplete a narrow region of the condensate by localised loss and watch the subsequent evolution of the many-body quantum system. For the experimental situation, as realised, e.g., in Herwig Otts’s group at Kaiserslautern, atoms are ionised in a controlled way by an electronic beam and the produced ions and electrons are quickly extracted [@Gericke2008; @Wuertz2009; @Barontini2013; @Ott2015; @Ott2016]. Consequently, there is scarcely any backaction onto the remaining atoms in the Bose condensate, provided that the filling factors (particle numbers per site) along the lattice are large. For such a setup, we can assume a Markovian coupling of the system to the environment. A corresponding Markovian master equation was used for such systems, taking into account localised loss in the lattice and phase noise, arising from interactions with the background gas or other experimental imperfections [@Anglin2001; @Witthaut2011].
For small systems, typically two to four lattice sites, yet with reasonably large filling factors, we can unravel the Master equation exactly, using quantum jump Monte Carlo simulations [@Witthaut2008; @Kordas2012; @Witthaut2011]. [For larger system sizes, approximative stochastic methods such as the truncated Wigner have been used to compute the single-particle density matrix (SPDM) [@Chianca2011; @Martinet2017] and normally ordered two-time correlation functions can be calculated in the Glauber–Sudarshan and positive-P representations [@Olsen2009; @Olsen2017]. Another]{} beyond-mean-field method was successfully applied to propagate specifically chosen initial conditions, typically fully coherent Bose condensed states. This so-called Bogoliubov Back-reaction (BBR) method is based on a Bogoliubov-Born-Green-Kirkwood-Yvon (BBGKY) hierarchy expansion, consisting of dynamical equations for two-, four-, etc., point [*equal-time*]{} correlation functions, i.e., the expectation values of the two-, four-, etc., body reduced density matrices [@Anglin2001; @Tikhonenkov2007]. The interaction term in the Hubbard–Hamiltonian then induces the coupling between these dynamical equations toward higher orders. Typically, one truncates at the second order, approximating the six-point correlator by products of two- and four-point correlation functions, in the way the moments of Gaussian variables would exactly split. This allows one to arrive at a closed system of coupled equations, which we can subsequently solve [@Witthaut2011].
In this paper, we present a method to compute [*non-equal time*]{} correlation functions of operators, much in the spirit of the BBGKY hierarchy truncation for equal-time observables, which allows one to take into account higher orders in the fluctuations in a systematic way. To do so, we adapt the quantum regression theorem, extensively used in quantum optics, see e.g., [@Carmichael1993; @GardinerZoller2000; @Breuer2002], to interacting ultracold atoms in an open system’s setting. Here, atom–atom interactions play an important role and result in hierarchies of dynamical equations that have to be truncated. More precisely, we want to compute two-time Green functions that are heavily used in solid-state physics, typically for fermonic transport problems, see e.g., [@MWL1993; @Komnik2008; @Mahan]. In contrast to the latter applications, our open quantum systems are not time-translational invariant, which discards working in Fourier (frequency or energy) space and mapping the equations of motions into purely alegabric equations, as done e.g., in Ref. [@Gong2006].
The paper is organised as follows: Section \[sec:2\] introduces our open many-body boson system. Section \[sec:3\] presents the computation scheme for two-time correlation functions, which is then applied in Section \[sec:4\] in mean-field approximation. Section \[sec:5\] discusses the next order beyond mean-field, with similar equations presented in Section \[sec:6\] for the density–density correlation functions. The last Section \[concl\] concludes the paper.
Dissipative Finite Bose–Hubbard Chain {#sec:2}
=====================================
We model ultracold bosons in sufficiently deep optical lattices by a tight-binding approximation, using a single-band Bose–Hubbard model. The geometry is assumed to be quasi one-dimensional, corresponding to a cigar-shaped confinement of the atoms (which is much stronger in the radial direction). Then, the dynamics of coherent interacting ultracold atoms tunnelling through an [$M$]{}-site lattice is described by the Bose–Hubbard Hamiltonian [@BDZ2008]: $$\hat{H}_{\textrm{BH}} = -J \sum^{M-1}_{j=1} (\hat{a}^\dagger_{j+1}\hat{a}^{\mathstrut}_j+\hat{a}^\dagger_j\hat{a}^{\mathstrut}_{j+1})+\frac{U}{2}\sum^M_{j=1}\hat{a}^\dagger_j\hat{a}^\dagger_j\hat{a}^{\mathstrut}_j\hat{a}^{\mathstrut}_j\label{eq:1}\,.$$
Here, $\hat{a}^\dagger_j$ and $\hat{a}^{\mathstrut}_j$ denote respectively the creation and annihilation operators at site $j$, $J$ is the tunnelling rate and $U$ is the interaction strength. The reduced Planck’s constant $\hbar$ is set to one, which corresponds to measuring all energies in frequency units. The natural unit of time is then $J^{-1}$.
Dissipation processes are accounted for by introducing the master equation in Lindblad form [@Breuer2002] together with a suitable Liouville superoperator: $$\begin{aligned}
\partial_t\hat{\rho}(t) &= \mathcal{L}\hat{\rho}(t) = -i\big[\hat{H}_{\textrm{BH}},\hat{\rho}(t)\big]+\widetilde{\mathcal{L}}\hat{\rho}(t),\label{eq:2}\\
\widetilde{\mathcal{L}}\hat{\rho} &= -\sum_{j=1}^{M}\frac{\gamma_j}{2}\big(\hat{L}_j^\dagger\hat{L}_j^{\mathstrut}\hat{\rho}+\hat{\rho}\hat{L}_j^\dagger\hat{L}_j^{\mathstrut}-2\hat{L}_j^{\mathstrut}\hat{\rho}\hat{L}_j^\dagger\big).
\label{eq:3}\end{aligned}$$
The Lindblad operators $\hat{L}_j$ are chosen depending on the type of relaxation or decoherence process relevant for the system under study. In the following, we restrict to local single-body dissipation as motivated in the introduction, i.e., $\hat{L}_j=\hat{a}_j$, and $\gamma_j$ is the dissipation rate at site $j$. [This choice of the Lindblad operator, as shown in [@Witthaut2011], leads to equations of motion for the SPDM equivalent to the heuristic non-Hermitian discrete nonlinear Schrödinger equation introduced in [@Ott2009] and successfully applied to the description of localized single-body dissipation processes in Bose–Hubbard chains in good agreement with experimental realizations [@Barontini2013; @Ott2016]. Further sets of Lindblad operators modelling other processes in Bose–Hubbard chains can be found in the review [@Kordas2015].]{}
The expectation value of some system operator $\hat{A}$ is then provided by the trace $\langle\hat{A}(t)\rangle = \Tr\lbrace\hat{A}\hat{\rho}(t)\rbrace = \Tr\lbrace\hat{A}_H(t)\hat{\rho}(0)\rbrace$. From this, the equivalent Heisenberg representation is defined as $\hat{A}_H(t) = V^\dagger(t,0)\hat{A}$, with the propagator $V(t,t_0) = \exp((t-t_0)\mathcal{L})$, for a time-independent Liouvillian. Henceforth, the Heisenberg representation is not specified by an index but simply indicated by the presence of a time argument. In this representation, the time evolution of the Heisenberg operator $\hat{A}(t)$ is carried out by the adjoint master equation: $$\begin{aligned}
\partial_t\hat{A}(t) &= V^\dagger(t,0)(\mathcal{L}^\dagger \hat{A}) = i\big[\hat{H}_{\textrm{BH}},\hat{A}\big](t)+\big(\widetilde{\mathcal{L}}^\dagger\hat{A}\big)(t),\label{eq:4}\\
\widetilde{\mathcal{L}}^\dagger\hat{A} &= -\sum_{j=1}^{M}\frac{\gamma_j}{2}\big(\hat{L}_j^\dagger\hat{L}_j^{\mathstrut}\hat{A}+\hat{A}\hat{L}_j^\dagger\hat{L}_j^{\mathstrut}-2\hat{L}_j^\dagger\hat{A}\hat{L}_j^{\mathstrut}\big) = -\sum_{j=1}^{M}\frac{\gamma_j}{2}\big(\hat{L}_j^\dagger[\hat{L}_j^{\mathstrut},\hat{A}]+[\hat{A},\hat{L}_j^\dagger]\hat{L}_j^{\mathstrut}\big).\label{eq:5}\end{aligned}$$
The main observable in the introduced system is the single-particle density matrix (SPDM) . Its diagonal matrix elements give the local populations of the chain, while its off-diagonal elements provide information about the coherence of the state [@Witthaut2011]. The quartic interaction term leads to a coupling of the equations of motion verified by the SPDM and higher-order moments. Consequently, computing the SPDM requires a prior truncation of the hierarchy into a closed set of differential equations. The mean-field (MF) approximation keeps only the first order of the hierarchy by neglecting the covariances , whereas the BBR close to mean-field method evolves simultaneously the SPDM and the covariance, truncating higher-order moments [@Anglin2001; @Tikhonenkov2007; @Witthaut2011].
The aim of this article is to provide a method for obtaining multi-time correlation functions for quantum many-body systems from the knowledge of the equal-time correlation functions. Our method works for systems with relaxation channels according to the master Equation (\[eq:3\]) and gives access to the transient dynamics of the functions, not just to the stationary-state solutions.
Computation Scheme for Two-Point Correlation Functions {#sec:3}
======================================================
Two-point Green functions (GF) provide useful information about many-body systems such as temporal and spatial correlations or information about the response of the system to an external perturbation [@Mahan]. As a first example, if one is interested in particular in evaluating the [average]{} probability of a particle propagating from some site [$k$]{} at time $t$ to any other site [$j$]{} at time $t^\prime=t+\tau$, the retarded Green function provides relevant information. In this case, this correlation function reads: $$G_{j,k}^{R}(t+\tau;t) = \theta_{\mathrm{H}}(\tau)(G_{j,k}^{>}(t+\tau;t)-G_{j,k}^{<}(t+\tau;t)),
\label{eq:6}$$ where the *lesser* and *greater* bosonic GFs are defined as follows: $$G_{j,k}^{<}(t^\prime;t) = -i\langle \hat{a}^\dagger_k(t)\hat{a}^{\mathstrut}_j(t^\prime) \rangle\,,\qquad
G_{j,k}^{>}(t^\prime;t) = -i\langle \hat{a}^{\mathstrut}_j(t^\prime)\hat{a}^\dagger_k(t) \rangle,
\label{eq:7}$$ where the ladder operators are expressed in the above-recalled Heisenberg representation $V^\dagger(t,0)\hat{a}^{(\dagger)}_j$. In order to compute these last two GFs, we derive a closed set of equations of motion for a dissipative setting by means of a modified version of the quantum regression theorem.
Quantum Regression Hierarchy {#sec:3-1}
----------------------------
Let $\lbrace\hat{A}_i\rbrace$ be a set of arbitrary system operators, e.g., $\hat{n}_i = \hat{a}^\dagger_i\hat{a}^{\mathstrut}_i$, and $\mathcal{L}$ a time-independent Liouville superoperator, e.g., the one defined in Equations (\[eq:4\]) and (\[eq:5\]), the adjoint master equation reads: $$\partial_t\hat{A}_i(t) = V^\dagger(t,0)(\mathcal{L}^\dagger \hat{A}_i).
\label{eq:8}$$
Let us now make the general assumption that the adjoint Liouville operator acts on $\hat{A}_i$ in such a way that its expectation value can be rewritten as: $$\Tr\lbrace(\mathcal{L}^\dagger\hat{A}_i)\hat{\rho}(t)\rbrace = \Tr \bigg\lbrace\bigg(\sum_\ell T^{(1)}_{i\ell}\hat{A}_\ell+\sum_{\ell,\ell^\prime}K^{(1)}_{i\ell^\prime\ell}\hat{B}_{\ell^\prime}\hat{A}_{\ell}\bigg)\hat{\rho}(t)\bigg\rbrace,
\label{eq:9}$$ where $\lbrace\hat{B}_\ell^\prime\rbrace$ are operators composed of an [*even*]{} number of creation and annihilation operators, e.g., a power of the density operator. This is the typical relation one gets when dealing with non-quadratic Hamiltonians and/or nonlinear Lindbald operators, and in particular the previously defined Liouvillian, which result in a coupling between the equations of motion satisfied by [$n$]{}-point and higher-order correlation functions. This relation is assumed to hold for any initial density matrix $\hat{\rho}$ so that one is able to identify both operators in parentheses: $$\mathcal{L}^\dagger\hat{A}_i = \sum_\ell T^{(1)}_{i\ell}\hat{A}_\ell+\sum_{\ell,\ell^\prime}K^{(1)}_{i\ell^\prime\ell}\hat{B}_{\ell^\prime}\hat{A}_{\ell}.
\label{eq:10}$$
The $K$-term requires an extension of the quantum regression theorem (see [@Lax1963; @Lax1967] or the Sections [5.2.3]{} in [@GardinerZoller2000] and [3.2.3]{} in [@Breuer2002]). Combining Equations (\[eq:8\]) and (\[eq:10\]), the equation of motion of the two-point correlation function reads: $$\begin{aligned}
\partial_\tau\langle\hat{A}_i(t+\tau)\hat{A}_j(t)\rangle &= \Tr\lbrace(\mathcal{L}^\dagger\hat{A}_i)V(t+\tau,t)\hat{A}_j V(t,0)\hat{\rho}(0)\rbrace = \langle(\mathcal{L}^\dagger\hat{A}_i)V(t+\tau,t)\hat{A}_j V(t,0)\rangle\\
&= \sum_\ell T^{(1)}_{i\ell}\langle\hat{A}_\ell(t+\tau)\hat{A}_j(t)\rangle+\sum_{\ell,\ell^\prime}K^{(1)}_{i\ell^\prime\ell}\langle(\hat{B}_{\ell^\prime}\hat{A}_\ell)(t+\tau)\hat{A}_j(t)\rangle\\
&= \sum_\ell T^{\prime(1)}_{i\ell}(t+\tau)\langle\hat{A}_\ell(t+\tau)\hat{A}_j(t)\rangle+\sum_{\ell,\ell^\prime}K^{(1)}_{i\ell^\prime\ell}\langle(\Delta\hat{B}_{\ell^\prime}\hat{A}_\ell)(t+\tau)\hat{A}_j(t)\rangle,
\label{eq:11}
\end{aligned}$$ where $T^{\prime(1)}_{i\ell}(t) = T^{(1)}_{i\ell}+\sum_{\ell^\prime}K^{(1)}_{i\ell^\prime\ell}\langle\hat{B}_{\ell^\prime}(t)\rangle$ and the central moment operator is defined as $\Delta\hat{A}=\hat{A}-\langle\hat{A}\rangle$. Naturally, the same can be done for $\langle\hat{A}_j(t)\hat{A}_i(t+\tau)\rangle$, with the only difference being that $\hat{A}_j$ is then placed at the left end of the moments.
With this proper rewriting, one gets a hierarchy of coupled dynamical equations in the form of the BBGKY hierarchy:
$$\begin{aligned}
\partial_\tau\langle\hat{A}_i(t+\tau)\hat{A}_j(t)\rangle &= \sum_\ell T^{\prime(1)}_{i\ell}(t+\tau)\langle\hat{A}_\ell(t+\tau)\hat{A}_j(t)\rangle+\sum_{\ell,\ell^\prime}K^{(1)}_{i\ell^\prime\ell}\langle(\Delta\hat{B}_{\ell^\prime}\hat{A}_\ell)(t+\tau)\hat{A}_j(t)\rangle,\label{eq:12}\\
\partial_\tau\langle(\Delta\hat{B}_{i^\prime}\hat{A}_i)(t+\tau)\hat{A}_j(t)\rangle &= \sum_\ell T^{\prime(2)}_{i\ell}(t+\tau)\langle(\Delta\hat{A}_{\ell^\prime}\hat{A}_\ell)(t+\tau)\hat{A}_j(t)\rangle\nonumber\\
&\hspace{4.1cm}+\smash{\sum_{\ell,\ell^\prime,k}K^{(2)}_{i^\prime i\ell^\prime\ell k}\langle(\Delta\hat{B}_k\Delta\hat{B}_{\ell^\prime}\hat{A}_\ell)(t+\tau)\hat{A}_j(t)\rangle}.\label{eq:13}\\
&\mathrel{\ooalign{\hss\vdots\hss\cr\phantom{=}}}\nonumber\end{aligned}$$
This system of differential equations can then be closed by truncating the hierarchy to some order in the fluctuations provided that the time-local averages $\langle\hat{B}_{\ell^\prime}(t)\rangle$ can be calculated at any time $t$. In the simplest case, for which $K^{(1)}$ is a zero matrix, one naturally recovers the standard quantum regression theorem expression. Otherwise, the first approximation is to make a mean-field approximation and neglect the covariances of the operators $\hat{B}_{\ell^\prime}$ and $(\hat{A}_{\ell}\hat{A}_{j})$ in Equation (\[eq:12\]). To perform this approximation, it is convenient to first rewrite Equation (\[eq:12\]) in such a way that the order in the fluctuations is explicit: $$\begin{aligned}
\partial_\tau\langle\hat{A}_i(t+\tau)\hat{A}_j(t)\rangle &= \sum_\ell T^{\prime(1)}_{i\ell}(t+\tau)\langle\hat{A}_\ell(t+\tau)\hat{A}_j(t)\rangle\label{eq:14}\\
&+\sum_{\ell,\ell^\prime}K^{(1)}_{i\ell^\prime\ell}\langle\Delta\hat{B}_{\ell^\prime}(t+\tau)\Delta(\hat{A}_\ell(t+\tau)\hat{A}_j(t))\rangle\label{eq:15}.\end{aligned}$$
From this expression, one observes that here the mean-field approximation amounts to keeping only the line (\[eq:14\]), which results in a quantum regression expression with a time-dependent coefficient matrix $T^{\prime(1)}$. The knowledge of $T^{\prime(1)}$ at any time puts then the system into a closed form. Moreover, if one is able to compute $\langle\hat{A}_{\ell^\prime}\rangle$ and $\langle\hat{A}_\ell\hat{A}_j\rangle$ at any time at second order in the fluctuations, for instance by employing the BBR truncation [@Witthaut2011], then one can obtain an approximation of the GFs at this order by including the line (\[eq:15\]) and computing the Equation (\[eq:13\]) of the hierarchy in the BBR approximation.
Mean-Field Approximation {#sec:4}
========================
We study now the Bose–Hubbard Hamiltonian of Equation (\[eq:1\]) with single-body loss putting $\hat L_j=\hat a_j$ in Equation (\[eq:2\]). In the mean-field approximation, the equations of motion of the annihilation operators read in that case $$\begin{aligned}
\partial_t\hat{a}^{\mathstrut}_j(t) &= iJ\big(\hat{a}^{\mathstrut}_{j+1}(t)+\hat{a}^{\mathstrut}_{j-1}(t)\big)-iU\hat{a}^\dagger_j(t)\hat{a}^{\mathstrut}_j(t)\hat{a}^{\mathstrut}_j(t)-\frac{\gamma_j}{2}\hat{a}^{\mathstrut}_j(t)\\
&= \sum_\ell\underbrace{\big(iJ(\delta_{j+1,\ell}+\delta_{j-1,\ell})-\frac{\gamma_\ell}{2}\delta_{j,\ell}\big)}_{T_{j\ell}}\hat{a}^{\mathstrut}_\ell(t)+\sum_{\ell,\ell^\prime}\underbrace{\big({-}iU\delta_{j,\ell^\prime}\delta_{\ell^\prime,\ell}\big)}_{K_{j\ell^\prime\ell}}\hat{n}^{\mathstrut}_{\ell^\prime}(t)\hat{a}_\ell(t)\\
&= \textstyle\sum_\ell T^{\prime}_{j\ell}(t)\hat{a}^{\mathstrut}_\ell(t) +\sum_{\ell,\ell^\prime}K_{j\ell^\prime\ell}\Delta\hat{n}^{\mathstrut}_{\ell^\prime}(t)\hat{a}_\ell(t)\label{eq:16}\,,
\end{aligned}$$ where $$T^{\prime}_{j\ell}(t) = iJ(\delta_{j+1,\ell}+\delta_{j-1,\ell})-iU\delta_{j,\ell}n_\ell(t)-\frac{\gamma_\ell}{2}\delta_{j,\ell}\label{eq:17}.$$
The simplicity of the equation of motion satisfied by the annihilation operator allows one to get directly a relation of the form of Equation (\[eq:10\]), without the need of considering its expectation value and taking advantage of the cyclic permutation invariance of the trace.
Then, the quantum regression theorem yields: $$\partial_t\hat{a}^{\mathstrut}_j(t) {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize MF}}}{\approx}}\textstyle\sum_\ell T^\prime_{j\ell}(t) \hat{a}^{\mathstrut}_\ell(t) \quad\Rightarrow\quad
\begin{cases}
\partial_\tau\langle\hat{a}^{\dagger}_k(t)\hat{a}^{\mathstrut}_j(t+\tau)\rangle {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize MF}}}{\approx}}\textstyle\sum_\ell T^\prime_{j\ell}(t+\tau) \langle\hat{a}^{\dagger}_k(t)\hat{a}^{\mathstrut}_\ell(t+\tau)\rangle,\\
\partial_\tau\langle\hat{a}^{\mathstrut}_j(t+\tau)\hat{a}^{\dagger}_k(t)\rangle {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize MF}}}{\approx}}\textstyle\sum_\ell T^\prime_{j\ell}(t+\tau) \langle\hat{a}^{\mathstrut}_\ell(t+\tau)\hat{a}^{\dagger}_k(t)\rangle.
\end{cases}\label{eq:18}$$
Due to the dependence of $T^\prime$ on the local density these equations of motion have to be evolved along with those of the single particle density matrix. This leads to the following closed set of differential equations: $$\begin{aligned}
i\partial_t\sigma_{j,k} &{\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize MF}}}{\approx}}-J(\sigma_{j,k+1}+\sigma_{j,k-1}-\sigma_{j+1,k}-\sigma_{j-1,k})+U(n_j-n_k)\sigma_{j,k}-\frac{i}{2}(\gamma_j+\gamma_k)\sigma_{j,k},\label{eq:19}\\
i\partial_\tau G_{j,k}^{\lessgtr}(t+\tau;t) &{\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize MF}}}{\approx}}-J\big(G_{j+1,k}^{\lessgtr}(t+\tau;t)+G_{j-1,k}^{\lessgtr}(t+\tau;t)\big)+\big(Un_j(t+\tau)-i\frac{\gamma_j}{2}\big)G_{j,k}^{\lessgtr}(t+\tau;t).\label{eq:20}\end{aligned}$$
These can be evolved in $\tau\geq 0$ for each $t$. In this way, the first time argument is always later than the second one, but this is not restrictive as the opposite situation can be obtained from the relation:
$$G_{j,k}^{\lessgtr}(t;t+\tau) = -(G_{k,j}^{\lessgtr}(t+\tau;t))^*\label{eq:21}\,.$$
Illustrative Results {#sec:4-1}
--------------------
As an example of the performance of the method presented here, we compute the correlation function defined as $A_{j,k}(t^\prime,t) = \langle[\hat{a}^{\mathstrut}_j(t^\prime),\hat{a}^\dagger_k(t)]\rangle = i(G_{j,k}^{>}(t^\prime;t)-G_{j,k}^{<}(t^\prime;t)) = i(G_{j,k}^{R}(t^\prime;t)-G_{j,k}^{A}(t^\prime;t))$ for two different settings. The Fourier transform of $A_{j,k}(t^\prime,t)$ presents the spectral weight function [@Mahan], used e.g., for computing currents in transport setups. This particular correlation function is chosen because its equal-time values and its lower and upper bounds can readily be checked and because its magnitude is equal to that of the retarded Green’s function for $t^\prime > t,$ whereas the magnitude of the advanced Green’s function is given by the $t > t^\prime$ half-quadrant. In addition, in the noninteracting non-dissipative case, its profile simply consists of periodic oscillations extending to both sides away from the diagonal $t = t^\prime$. This harmonic behaviour of $A_{j,k}(t^\prime,t)$ is related to the mentioned fact that its Fourier transform characterises the spectrum without perturbation in the closed system’s case, in the example shown below of a three-level system. Figure \[fig:1\] shows the magnitude of two matrix elements of the spectral weight function of a condensate initially made of $N_0 = 1000$ atoms loaded into a three-well Bose–Hubbard chain with an initial population imbalance between a low populated central well $2$ and its two more populated neighbours $1$ and $3$ in the presence of strong dissipation occurring at the central site. The values taken at the diagonal $t=t^\prime$ correspond to the constant $\lvert A_{j,k} \rvert = \lvert\langle[\hat{a}^{\mathstrut}_j(t),\hat{a}^\dagger_k(t)]\rangle\rvert = \delta_{j,k}$, as expected. In Figure \[fig:1\]a, which depicts the $\lvert A_{2,1}\rvert$ matrix element, the $t^\prime > t$ half-quadrant presents an elevation at early $t^\prime$ that drops as $t^\prime$ becomes larger than a few $J^{-1}$. This indicates that, on average, the flow of particles from the site $1$ to the leaky site $2$ is suppressed at times above this typical value.
This suppression is a signature of the so-called quantum Zeno effect, meaning that particles are blocked on average from flowing into the site with dissipation, see e.g., the results and descriptions in [@Trimborn2011; @Barontini2013; @Kordas2015]. Dissipation can thus induce an effectively self-trapped regime, considerably lowering the average inter-well tunnelling, although the values of the population imbalance and the interaction strength do not suffice to reach this regime in the absence of dissipation. In Figure \[fig:1\]b, which depicts the $\lvert A_{2,1}\rvert$ matrix element, whereas, in the noninteracting non-dissipative case, this function shows periodic oscillations from either side of the diagonal around the value 1/2; in this case, the spectral weight function quickly attains a plateau at this value. The fact that it does not take values below one half after a few $J^{-1}$ is another signature of the quenching of the inter-well tunnelling induced by a quantum Zeno effect.
[0.45]{} ![\[fig:1\] Magnitude of the spectral weight function $A_{j,k}$ for an initially pure Bose–Einstein condensate with $n_1(0) = 500$, $n_2(0) = 50$, $n_3(0) = 450$, and interaction $UN_0/J = 10$; the dissipation rate is set to $\gamma_2/J = 5$. (**a**) interwell correlations $\lvert A_{2,1}(t^\prime,t)\rvert$; (**b**) onsite correlations $\lvert A_{1,1}(t^\prime,t)\rvert$. ](fig1.pdf "fig:"){width="\linewidth"}
[0.45]{} ![\[fig:1\] Magnitude of the spectral weight function $A_{j,k}$ for an initially pure Bose–Einstein condensate with $n_1(0) = 500$, $n_2(0) = 50$, $n_3(0) = 450$, and interaction $UN_0/J = 10$; the dissipation rate is set to $\gamma_2/J = 5$. (**a**) interwell correlations $\lvert A_{2,1}(t^\prime,t)\rvert$; (**b**) onsite correlations $\lvert A_{1,1}(t^\prime,t)\rvert$. ](fig2.pdf "fig:"){width="\linewidth"}
Figure \[fig:2\] represents the same correlation functions for the same three-well setting but in the absence of dissipation. In this case, the profile of the spectral weight function is very rugged due to the interactions. The correlation function at $t\neq t^\prime$ fluctuates around $1/2$ regardless of the matrix element, which indicates that the system is [*not*]{} in a self-trapped regime, as the average inter-site flow of bosons is not suppressed.
From these two examples, one observes that correlations in Bose–Hubbard chains qualitatively differ depending on the presence or absence of dissipation. This as well as the observation of some particular structures in the profiles of the investigated correlation functions can be used to determine and differentiate the dynamical regimes in Bose–Hubbard chain set-ups, where the interplay of interactions and dissipation plays a major role.
[0.45]{} ![\[fig:2\] Magnitude of the spectral weight function $A_{j,k}$ for an initially pure Bose–Einstein condensate with $n_1(0) = 500$, $n_2(0) = 50$, $n_3(0) = 450$, and interaction $UN_0/J = 10$; the dissipation rate is set to zero, $\gamma_2 = 0$. (**a**) interwell correlations $\lvert A_{2,1}(t^\prime,t)\rvert$; (**b**) onsite correlations $\lvert A_{1,1}(t^\prime,t)\rvert$. ](fig3.pdf "fig:"){width="\linewidth"}
[0.45]{} ![\[fig:2\] Magnitude of the spectral weight function $A_{j,k}$ for an initially pure Bose–Einstein condensate with $n_1(0) = 500$, $n_2(0) = 50$, $n_3(0) = 450$, and interaction $UN_0/J = 10$; the dissipation rate is set to zero, $\gamma_2 = 0$. (**a**) interwell correlations $\lvert A_{2,1}(t^\prime,t)\rvert$; (**b**) onsite correlations $\lvert A_{1,1}(t^\prime,t)\rvert$. ](fig4.pdf "fig:"){width="\linewidth"}
Beyond Mean-Field Approximation {#sec:5}
===============================
One of the assets that motivate seeking correlation functions as solutions to a hierarchy of coupled equations of motion is that the contribution of higher order fluctuations can be integrated in a systematic way—for instance, by taking into account the second equation, Equation (\[eq:13\]), of the hierarchy and truncating sextic moments as follows [@Anglin2001; @Tikhonenkov2007]:
@size[9]{}@mathfonts @@@\#1$$\begin{aligned}
\langle\hat{a}_j^\dagger\hat{a}_m^{\mathstrut}\hat{a}_k^\dagger\hat{a}_n^{\mathstrut}\hat{a}_r^\dagger\hat{a}_s^{\mathstrut}\rangle {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize BBR}}}{\approx}}& \langle \hat{a}_j^\dagger\hat{a}_m^{\mathstrut}\hat{a}_k^\dagger\hat{a}_n^{\mathstrut} \rangle\langle \hat{a}_r^\dagger\hat{a}_s^{\mathstrut} \rangle+ \langle \hat{a}_j^\dagger\hat{a}_m^{\mathstrut}\hat{a}_r^\dagger\hat{a}_s^{\mathstrut} \rangle\langle \hat{a}_k^\dagger\hat{a}_n^{\mathstrut} \rangle+\langle \hat{a}_k^\dagger\hat{a}_n^{\mathstrut}\hat{a}_r^\dagger\hat{a}_s^{\mathstrut} \rangle\langle \hat{a}_j^\dagger\hat{a}_m^{\mathstrut} \rangle- 2\langle \hat{a}_j^\dagger\hat{a}_m^{\mathstrut} \rangle\langle \hat{a}_k^\dagger\hat{a}_n^{\mathstrut} \rangle\langle \hat{a}_r^\dagger\hat{a}_s^{\mathstrut} \rangle.\label{eq:22}\end{aligned}$$
One gets a closed set of differential equations for the GFs in the Bogoliubov Back-reaction beyond mean-field approximation. For example, for the greater GF, this reads: $$\begin{aligned}
i\partial_\tau G_{j,k}^{>}(t+\tau;t) {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize BBR}}}{\approx}}& -J\big(G_{j+1,k}^{>}(t+\tau;t)+G_{j-1,k}^{>}(t+\tau;t)\big)+\big(Un_j(t+\tau)-i\frac{\gamma_j}{2}\big)G_{j,k}^{>}(t+\tau;t)\nonumber\\
&+U\,F_{j,j,j;k}^{>}(t+\tau;t),\label{eq:23}\end{aligned}$$ @size[9]{}@mathfonts @@@\#1$$\begin{aligned}
i\partial_\tau F_{j,m,n;k}^{>}(t+\tau;t) {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize BBR}}}{\approx}}& +J\big( \phantom{+}F_{j+1,m,n;k}^{>}(t+\tau;t) + F_{j-1,m,n;k}^{>}(t+\tau;t) - F_{j,m+1,n;k}^{>}(t+\tau;t)\\
&\phantom{+J\big(}-F_{j,m-1,n;k}^{>}(t+\tau;t) - F_{j,m,n+1;k}^{>}(t+\tau;t) - F_{j,m,n-1;k}^{>}(t+\tau;t) \,\big)\\
&-U\big( \phantom{+}F_{j,j,n;k}^{>}(t+\tau;t)\sigma_{jm}(t+\tau) - (\Delta_{j,m,m,m}+\Delta_{j,m,n,n})(t+\tau)iG^{>}_{n,k}(t+\tau;t)\\
&\phantom{-U\big(}+F_{j,m,n;k}^{>}(t+\tau;t)(n_j-n_m-n_n)(t+\tau)\\
&\phantom{-U\big(}+\delta_{m,n}(F_{j,m,n;k}^{>}(t+\tau;t)+\sigma_{j,m}G_{n,k}^{>}(t+\tau;t)) \,\big) \,.
\label{eq:24}
\end{aligned}$$
Here, we defined $$iF_{j,m,n;k}^{>}(t^\prime;t) = \langle\Delta(\hat{a}^\dagger_j\hat{a}^{\mathstrut}_m)(t^\prime)\Delta(\hat{a}^{\mathstrut}_n(t^\prime)\hat{a}^\dagger_k(t))\rangle = \langle(\hat{a}^\dagger_j\hat{a}^{\mathstrut}_m\hat{a}^{\mathstrut}_n)(t^\prime)\hat{a}^\dagger_k(t)\rangle - \sigma_{j,m}(t^\prime)iG_{n,k}^{>}(t^\prime;t)
\label{eq:25}\,.$$
The SPDM and the covariances are given in the BBR approximation by (c.f. [@Witthaut2011]): @size[9]{}@mathfonts @@@\#1$$\begin{aligned}
i&\partial_t\sigma_{jk} {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize BBR}}}{\approx}}-J\big(\sigma_{j,k+1}+\sigma_{j,k-1}-\sigma_{j+1,k}-\sigma_{j-1,k}\big)
+U\big(\Delta_{jkkk}+\sigma_{jk}\sigma_{kk}-\Delta_{jjjk}-\sigma_{jj}\sigma_{jk}\big)-i\frac{\gamma_j+\gamma_k}{2}\sigma_{j,k},\label{eq:26}\end{aligned}$$
However, although being more accurate in the low population regime, no qualitative difference is expected with respect to the mean-field version, at least in the limit of large filling factors $n_j \gg 1$.
Density–Density Correlation Function {#sec:6}
====================================
[Besides improving the mean-field precision, BBR also provides a way to compute non-trivial density–density correlation functions. The latter indicate, for instance, possible temporal bunching or anti-bunching effects [@Witthaut2011; @Carmichael1993]. Indeed, at the mean-field level, one just gets . Instead, by defining the density–density correlation function as follows:]{} $$C_{j,m;k,n}(t^\prime;t) = -i\langle\Delta(\hat{a}_j^\dagger\hat{a}_m^{\mathstrut})(t^\prime)\Delta(\hat{a}_k^\dagger\hat{a}_n^{\mathstrut})(t)\rangle = \langle(\hat{a}_j^\dagger\hat{a}_m^{\mathstrut})(t^\prime)(\hat{a}_k^\dagger\hat{a}_n^{\mathstrut})(t)\rangle - \sigma_{jm}(t^\prime)\sigma_{kn}(t)
\label{eq:28}$$ and truncating sextic moments according to Equation (\[eq:22\]) in the equation of motion similar to Equation (\[eq:12\]) that it satisfies, one gets: $$\begin{aligned}
i\partial_\tau C_{j,m;k,n}(t+\tau;t) {\overset{\makebox[0pt]{\mbox{\normalfont\scriptsize BBR}}}{\approx}}\hspace{-2cm}&\\
&+J\big(C_{j+1,m;k,n}(t+\tau;t) + C_{j-1,m;k,n}(t+\tau;t)-C_{j,m+1;k,n}(t+\tau;t) - C_{j,m-1;k,n}(t+\tau;t)\big)\\
&-U\big((n_j-n_m)(t+\tau)C_{j,m;k,n}(t+\tau;t)+\sigma_{jm}(C_{j,j;k,n}(t+\tau;t)-C_{m,m;k,n}(t+\tau;t))\big)\\
&-i\frac{\gamma_j+\gamma_m}{2}C_{j,m;k,n}(t+\tau;t),\label{eq:29}
\end{aligned}$$ which has to be evolved together with Equations (\[eq:26\]) and (\[eq:27\]). Again, if one wants to compute the density–density correlation function with the first time argument evaluated at an earlier time than the second, one can simply use $$C_{j,m;k,n}(t;t+\tau) = -\big(C_{n,k;m,j}(t;t+\tau)\big)^*.
\label{eq:30}$$
This shows that the systematic expansions in order to approximate the temporal correlation functions of operators proposed here are indeed very useful to compute physical relevant quantities in open quantum many-body systems.
Conclusions {#concl}
===========
Usually for a quantum many-body problem, the Schrödinger equation, or in our case the master equation Equation (\[eq:2\]), is not analytically solvable. One possible way to access the non-equilibrium dynamics of such systems is with the help of Green functions. The two-time correlators, such as the time-dependent first order coherences from Equations (\[eq:6\]) and (\[eq:7\]), or the density–density correlation functions Equation (\[eq:28\]), allow one to characterize the out-of-equilibrium dynamics of many-body interacting quantum systems. We have presented a method on how to systematically compute a hierarchy of approximations for two-time correlation functions, in principle of arbitrary operators, which applies to a wide class of interacting systems whose dissipation is described by a master equation of the form of Equation (\[eq:2\]). Based on our method, we can treat particle interactions on a mean-field level (see Section \[sec:4\]) and also one-order beyond (see Section \[sec:5\]). Higher-order expansions are possible, at the cost of very lengthy formulae, which at some point will become difficult to deal with even numerically (see e.g., [@SchmelcherI; @SchmelcherII] for just time-equal correlators).
With our method, we computed the spectral weight function in Section \[sec:4-1\] as an example of a physical observable. In principle, we can compute the temporal evolution and the correlations of many interesting quantities for experiments, such as the onsite populations or the density–density correlations between two lattice sites. For about ten years, state-of-the-art experiments with ultracold atoms can access information on static correlation functions with high spatial resolution (see e.g., ). Time-dependent two-point correlation functions can be measured too, see e.g., [@Esslinger2007], which would allow for direct applications of our developed theoretical approach.
Our problem is defined by a time-local master equation, which means that the coupling to the environment, in our case a zero-temperature sink, is supposed to be sufficiently weak for justifying the Markovian assumption underlying Equation (\[eq:2\]). Strong coupling to the environment, such as that present in lead-to-lead transport across solid-state samples (see e.g., [@MWL1993; @Komnik2008]) remains an open problem since then we may not apply the quantum regression theorem valid for time-local master equations. For tedious extensions of the theorem to non Markovian setups, the reader may consult e.g., [@Alonso2005; @Paolo2009; @Bassano2014; @Ban2017]. An alternative approach would be then to fully include the leads into the treatment and use a diagrammatic expansion of the Green functions computed for the full lead-system-lead system, such as done for bosons in steady state in Refs. [@Gutman2012; @Ivanov2013]. Needless to say, such an approach as just mentioned is very hard for many-body quantum systems whose main central part, without the leads, is itself non-integrable, such as our many-body Bose–Hubbard model for more than two wells [@BK2003; @Tomadin2007].
for his invitation.
[45]{}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 [****, ()](\doibase 10.1038/nphys1102) [****, ()](\doibase
10.1103/PhysRevLett.103.080404) [****, ()](\doibase
http://dx.doi.org/10.1103/PhysRevLett.110.035302) [****, ()](\doibase
10.1103/PhysRevLett.115.050601) [****, ()](\doibase
10.1103/PhysRevLett.116.235302) [****, ()](\doibase 10.1103/PhysRevLett.86.568) [****, ()](\doibase
http://dx.doi.org/10.1103/PhysRevA.83.063608) [****, ()](\doibase 10.1103/PhysRevA.79.033621) [****, ()](\doibase 10.1209/0295-5075/100/30007) [****, ()](\doibase 10.1103/PhysRevA.84.043636) [****, ()](\doibase 10.1140/epjd/e2016-70663-9) [****, ()](\doibase
10.1103/PhysRevA.80.033624) [****, ()](\doibase 10.1103/PhysRevA.95.063607) [****, ()](\doibase 10.1103/PhysRevA.75.013613) [**](\doibase 10.1007/978-3-540-47620-7) (, ) @noop [**]{}, ed. (, ) @noop [**]{} (, ) [****, ()](\doibase 10.1103/PhysRevLett.70.2601) [****, ()](\doibase 10.1103/PhysRevB.78.235110) @noop [**]{} (, ) [****, ()](\doibase
10.1103/PhysRevB.73.245329) [****, ()](\doibase 10.1103/RevModPhys.80.885) [****, ()](\doibase 10.1103/PhysRevLett.102.144101) [****, ()](\doibase
10.1140/epjst/e2015-02528-2) [****, ()](\doibase 10.1103/PhysRev.129.2342) [****, ()](\doibase 10.1103/PhysRev.157.213) [****, ()](\doibase
10.1140/epjd/e2011-10702-7) @noop [ ()]{} @noop [ ()]{} [****, ()](\doibase 10.1103/PhysRevLett.94.110401) [****, ()](\doibase
10.1103/PhysRevLett.95.090404) [****, ()](http://dx.doi.org/10.1038/nature03500) [****, ()](\doibase 10.1126/science.1192368) [****, ()](http://dx.doi.org/10.1038/nature09378) [****, ()](\doibase
10.1103/PhysRevLett.96.130403) [****, ()](\doibase 10.1103/PhysRevLett.105.230402) [****, ()](\doibase
10.1103/PhysRevLett.98.090402) [****, ()](\doibase 10.1103/PhysRevLett.94.200403) [****, ()](\doibase 10.1103/PhysRevA.80.022103) [****, ()](\doibase 10.1103/PhysRevA.90.022110) [****, ()](\doibase 10.1103/PhysRevA.96.042111) [****, ()](\doibase http://dx.doi.org/10.1103/PhysRevB.85.125102) [****, ()](\doibase
10.1140/epjb/e2013-40417-4) [****, ()](\doibase 10.1103/PhysRevLett.91.253002) [****, ()](\doibase 10.1103/PhysRevLett.98.130402)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Visual tracking can be easily disturbed by similar surrounding objects. Such objects as hard distractors, even though being the minority among negative samples, increase the risk of target drift and model corruption, which deserve additional attention in online tracking and model update. To enhance the tracking robustness, in this paper, we propose a cascaded regression tracker with two sequential stages. In the first stage, we filter out abundant easily-identified negative candidates via an efficient convolutional regression. In the second stage, a discrete sampling based ridge regression is designed to double-check the remaining ambiguous hard samples, which serves as an alternative of fully-connected layers and benefits from the closed-form solver for efficient learning. During the model update, we utilize the hard negative mining technique and an adaptive ridge regression scheme to improve the discrimination capability of the second-stage regressor. Extensive experiments are conducted on 11 challenging tracking benchmarks including OTB-2013, OTB-2015, VOT2018, VOT2019, UAV123, Temple-Color, NfS, TrackingNet, LaSOT, UAV20L, and OxUvA. The proposed method achieves state-of-the-art performance on prevalent benchmarks, while running in a real-time speed.'
author:
- 'Ning Wang, Wengang Zhou, Qi Tian, *Fellow, IEEE*, and Houqiang Li, *Senior Member, IEEE* [^1] [^2] [^3]'
bibliography:
- 'reference\_full.bib'
title: 'Cascaded Regression Tracking: Towards Online Hard Distractor Discrimination'
---
[Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{}
Visual tracking, regression tracking, cascaded framework, hard distractor.
Introduction {#sec:intro}
============
a fundamental task in computer vision, visual object tracking has received lots of attention over the last decades. It plays an important role in many applications such as autonomous driving, robotics, human-computer interaction, etc. In generic visual tracking, the target is arbitrary with only the initial bounding box available. With such limited prior information, the tracker is still highly required to both model the target appearance and distinguish the negative samples on the fly, which is challenging due to the blurry boundary between appearance changes of the target itself and unforeseen similar distractors.
![[**Top**]{}: the tracking results of SiamRPN++ [@siamrpn++], ATOM [@ATOM], and our approach on the *Bolt2* video. Note that the prior motion model is removed in these methods. [**Bottom**]{}: the per-frame overlap between the tracking results and ground-truth bounding box on the *Bolt2*. Without the cosine window, previous methods tend to switch between the target and distractors (e.g., $ 5 $-th and $ 75 $-th frame), while ours steadily tracks the target without drift.[]{data-label="fig:1"}](op_example.pdf "fig:"){width="8.3cm"} ![[**Top**]{}: the tracking results of SiamRPN++ [@siamrpn++], ATOM [@ATOM], and our approach on the *Bolt2* video. Note that the prior motion model is removed in these methods. [**Bottom**]{}: the per-frame overlap between the tracking results and ground-truth bounding box on the *Bolt2*. Without the cosine window, previous methods tend to switch between the target and distractors (e.g., $ 5 $-th and $ 75 $-th frame), while ours steadily tracks the target without drift.[]{data-label="fig:1"}](op_plot.pdf "fig:"){width="8.8cm"}
Recently, thanks to the strong representational power of deep CNN models, a simple two-stream template matching based Siamese pipeline [@SiamFC; @SINT] has been proved effective in visual tracking, even without the online model update. However, as reported in the Visual Object Tracking (VOT) challenge [@VOT2018], the robustness of Siamese trackers still has a margin with the discriminative trackers equipped with an update mechanism. In the latest literature [@DiMP; @SiameseUpdate], substantial attentions have been cast to the updatable deep trackers with superior discrimination capability. Despite the rapid advances, in the tracking and updating stages, how to distinguish similar distractor objects from the target and effectively leverage these hard negative samples to boost the model discrimination capability still leaves exploration space. There exist vast uninformative samples that can be easily distinguished without much effort (i.e., easy sample), while a handful of distracting examples heavily mislead the tracker, enlarging the error accumulation and causing the tracking failures (Figure \[fig:1\]). These unexpectedly emerged distractors, even though being the minority, have a non-trivial effect on degrading the tracking performance, and deserve to be carefully checked *online* for robust tracking.
In this paper, we propose a cascaded regression tracker, which consists of two sequential stages with different regression models for high-performance visual tracking. In the first stage, we employ an efficient convolutional regression [@ATOM] to densely predict all the searching locations, which filters out plentiful easy samples. In the second stage, we only consider the remaining ambiguous candidates and propose a discrete sampling based ridge regression for further discrimination. The ridge regressor performs as an alternative of the fully-connected layer but exhibits superior efficiency thanks to its closed-form solution. These two stages complement each other as follows. The dense prediction with the convolutional regressor in the first stage [@ATOM] covers a large search area, while its model tends to be disturbed by an overwhelming number of easy samples. In contrast, the second-stage regressor trained using the carefully selected hard samples naturally avoids the class-imbalance issue and yields better discrimination on distractors, while its sampling manner fails to perfectly cover the search area and will increase the computational cost when drawing plentiful candidates. By virtue of such a dense-to-discrete and coarse-to-fine two-tier verification, these two stages contribute to a superior robust tracking system. More importantly, both of them allow to update the corresponding models, achieving the *online* adaptability.
During online tracking, to enhance the tracker discrimination, we employ the hard negative mining [@example-based_HNM; @felzenszwalb2009object_HNM] for the second stage. Moreover, we dynamically reweigh the training samples based on their reconstruction errors in an adaptive ridge regression formula, forcing the second-stage regressor to focus more on valuable samples. Benefited from the high robustness, the second-stage regressor is able to re-detect the lost target when the first stage fails to confidently track the target, and search a large region without excessively worrying about the risk of tracking drift. As a consequence, our framework differs from most existing short-term trackers typically focusing on a limited search region with a prior cosine window to penalize the far-away distractors (e.g., Siamese trackers [@SiamFC; @SiamRPN]). It is worth mentioning that our method shows outstanding performance on *both* short-term and long-term tracking datasets without adding additional sophisticated modules thanks to our excellent online discrimination capability.
We summarize the contributions of our work as follows:
- We propose a discrete sampling based ridge regression, which can flexibly absorb the online hard samples and is efficient to learn under a closed-form formula. Furthermore, we propose a cascaded regression tracker, which achieves favorable robustness via a dense-to-discrete large-scale search and a coarse-to-fine two-tire verification.
- To improve the online distractor discrimination, we propose an adaptive ridge regression to further exploit the valuable samples selected by the hard negative mining technique [@example-based_HNM; @felzenszwalb2009object_HNM]. With the merit of promising discrimination, the second-stage regressor also serves as an effective re-detection module to complement the first stage.
- We extensively evaluate the proposed method on 11 short-term and long-term tracking benchmarks including OTB-2013 [@OTB-2013], OTB-2015 [@OTB-2015], Temple-Color [@TempleColor128], UAV123 [@UAV123], VOT2018 [@VOT2018], VOT2019 [@VOT2019], NfS [@NFSdataset], TrackingNet [@2018trackingnet], LaSOT [@LaSOT], UAV20L [@UAV123], and OxUvA [@2018longtermBenchmark]. The proposed approach exhibits state-of-the-art performance on prevalent datasets with a real-time speed.
In the following of the paper, we first survey related works in Section \[sec:related\]. Then, we elaborate the proposed cascaded framework in Section \[sec:method\]. After that, we evaluate our method with extensive experiments in Section \[sec:experiments\]. Finally, we conclude this work in Section \[sec:conclusion\].
Related Work {#sec:related}
============
In recent years, the Siamese network has gained significant popularity in visual tracking, which deals with the tracking task by searching for the image region most similar to the initial template [@SiamFC; @SINT]. The GOTURN algorithm [@GOTURN] adopts a Siamese pipeline to regress the target bounding box. By introducing the RPN module [@SiamRPN; @DaSiamRPN], ensemble learning [@SASiam], attention mechanism [@RASNet], and target-aware formulation [@TADT], the Siamese trackers gain substantial improvements. Besides visual tracking, similar ideas such as one-shot learning [@caelles2017one-shotSegmentation] and online adaptation scheme [@voigtlaender2017onlineSegmentation] are widely explored in the video object segmentation task. Without video annotations, the unsupervised deep tracking framework is explored in UDT [@UDT]. In [@siamrpn++], SiamRPN++ adopts a deeper backbone network to achieve superior performance. By switching multiple Siamese trackers using an agent network, POST tracker [@POST] achieves a good balance of accuracy and efficiency. Recently, model update mechanisms are incorporated with the Siamese network [@GCT; @Dsiam; @MemTrack; @SiameseUpdate]. However, these approaches mainly focus on the template adaptation and still fail to exploit the background context. Since most Siamese trackers ignore the informative negative samples for discrimination enhancement, they tend to drift when similar distractors appear. Recently, the cascaded framework has been investigated within the Siamese tracking framework [@CRPN; @SPM]. SPM [@SPM] combines the SiamRPN with a relation network to further classify the candidates. C-RPN [@CRPN] utilizes cascaded region proposal networks for accurate target localization. Nevertheless, they do not involve the online model update. The overlook of online emerged samples heavily limits the performance. In other words, how to take advantage of the hard negative samples to distinguish potential distractors in future frames is ignored in the recent cascaded frameworks. Compared with them, the main distinction of this work is that our cascaded framework is built on two complementary regression models, both of which are able to absorb the online samples for the persistent model update.
{width="18.1cm"}
Another popular tracking family is the regression based approach, which generally regresses a large Region of Interest (RoI) to a response map for target localization. The Correlation Filter (CF) solves the ridge regression in the Fourier domain, showing extremely attractive efficiency [@KCF; @DSST; @HCF; @CSR-DCF; @target_response_CF; @SCT; @ACFN; @Context-AwareCorrelationFilter; @MCCT; @C-COT]. To alleviate the unwanted boundary effect, regularization terms [@ASRCF; @SRDCF; @STRCF] and background-aware formulation [@BACF] are proposed. ECO tracker [@ECO] introduces a factorized convolution operator, a generative sample space model, and the sparse update strategy to further boost the efficiency of correlation tracking. Recently, by jointly compressing and transferring the heavyweight feature extractors in deep CF trackers, CPU real-time efficiency is also feasible [@CF-VGG]. Besides CF, with the recent astonishing development of deep learning, convolutional regression gains an increasing attention in visual tracking [@CREST; @DSLT; @ATOM; @DiMP]. In these approaches, a CNN kernel is learned to convolve with the RoI feature for response generation, which effectively avoids the boundary effect in CF. Unfortunately, this convolutional formulation does not have a closed-form solution, and needs the gradient back-propagation to learn the filter. Besides, the large RoI size in the regression approach brings in the class-imbalance issue. In CREST [@CREST], residual terms are incorporated into the convolutional regression to cope with the target appearance changes. DSLT [@DSLT] introduces shrinkage loss to balance the training samples in the convolutional regression. To accelerate the kernel learning process, ATOM [@ATOM] exploits the conjugate gradient in the deep learning framework. The recent DiMP approach [@DiMP] proposes an iteratively optimized discriminative model for classification and trains the whole framework in an end-to-end manner. Despite the recent progress, the discrimination capability in regression trackers, especially for hard distractors, still leaves room for improvement.
In contrast to the aforementioned regression methods that generate a dense prediction, previous discriminative trackers learn a binary classifier to classify the discretely sampled candidates for tracking (e.g., MDNet [@MDNet]). In spite of their shallow backbone networks and limited discrete samples (e.g., 256 candidates per frame), by an effective model update with hard negative mining, these approaches [@MDNet; @RTMDNet; @VITAL] still exhibit impressive robustness on various tracking benchmarks, suggesting the importance of online learning.
Our proposed approach is partially inspired by the above observations to retain both the dense and discrete predictions in a coarse-to-fine manner. Hard negative mining, as a powerful technique in object detection [@example-based_HNM; @felzenszwalb2009object_HNM], has been successfully equipped into some discrete sampling based visual trackers such as MDNet [@MDNet]. However, existing regression based trackers fail to effectively explore the hard negative samples since they train the regression model using densely sampled candidates and generally equally weigh them. The recent ATOM tracker [@ATOM] reduces the training weights of easy samples to focus on the valuable negative samples to some extent, but we observe that it still struggles to distinguish hard distractors. In this work, our first stage densely searches a large RoI to generate high-quality proposals, while the second stage is more flexible in the model update and hard negative mining to better distinguish the hard negative samples. Even though aiming at predicting discrete samples, unlike [@MDNet; @RTMDNet; @VITAL] that leverage fully-connected layers for classification, we learn an efficient closed-form solver in the feed-forward pass without back-propagation, potentially alleviating the overfitting issue due to much fewer parameters to be optimized online. By design, we absorb the strength of both regression trackers and discrete sampling based tracking-by-detection approaches to form a unified cascaded tracking framework. Our method is also motivated by the two-stage framework in object detection (e.g., faster RCNN [@FasterRCNN]), which has witnessed tremendous success in recent years. Differently, we exploit two regression models specially designed for the online tracking task with an incremental model update.
Methodology {#sec:method}
===========
In Figure \[fig:main\] (top), we show an overview of the proposed cascaded tracker. In the first stage, a convolutional regressor densely predicts the target location over a large RoI. Then, the ambiguous proposals are fed to the second regression stage for further discrimination. Under such a dense-to-discrete and coarse-to-fine verification, the proposed tracking framework achieves favorable tracking robustness. In Figure \[fig:main\] (bottom), we exhibit the online update process of the second-stage regression model. By virtue of the hard negative mining and an adaptive ridge regression formulation, the learned regressor is readily ready for distinguishing hard distractors.
In the following, we first review the regression based tracking in Section \[review regression tracking\] for the sake of completeness. In Section \[discrete ridge regression\], we present our discrete sampling based ridge regression and provide a detailed analysis in comparison with the previous methods. Then, in Section \[online tracking\], we depict the cascaded regression tracking and re-detection mechanism. Finally, we introduce the details of the online model update in Section \[model update\].
Revisiting Regression Tracking {#review regression tracking}
------------------------------
In this subsection, we briefly review the correlation filter and convolutional regression.
[**Correlation Filter.**]{} The correlation filter (CF) [@DSST; @KCF] tackles visual tracking by solving the following regression problem: $$\label{eq1}
\min_{{\bf W}_{\text{CF}}}\|{\bf X}\star{{\bf W}_{\text{CF}}}-{{\bf Y}_{\text{G}}}\|^{2}_{2} + \lambda\|{{\bf W}_{\text{CF}}}\|^{2}_{2},$$ where $ \star $ denotes the circular correlation, $ \lambda $ is a regularization parameter that controls overfitting, $ {\bf X}\in\mathbb{R}^{M\times N\times C} $ is the feature map of the RoI patch, $ {\bf Y}_{\text{G}}\in\mathbb{R}^{M\times N} $ is the Gaussian-shaped label, and $ {{\bf W}_{\text{CF}}}\in\mathbb{R}^{M\times N\times C} $ is the desired correlation filter.
Let $ \bf A $ denote the data matrix that contains all the circulant shifts of the base feature representation $ \bf X $. Then, the circular correlation $ {\bf X}\star{{\bf W}_{\text{CF}}} $ is equal to $ {\bf A}{{\bf W}_{\text{CF}}} $, and the filter $ {\bf W}_{\text{CF}} $ has the following closed-form solution [@rifkin2003regularized; @KCF; @bertinetto2018meta]: $$\label{eq2}
{{\bf W}_{\text{CF}}} = ({\bf A}^{\mathrm{T}}{\bf A}+\lambda {\bf I})^{-1}{\bf A}^{\mathrm{T}}{{\bf Y}_{\text{G}}},$$ where $ \bf I $ is the identity matrix. Due to the circulant structure of $ \bf A $, it can be diagonalized via $ {\bf A}={\bf F}~\text{diag}(\hat{\bf X})~{\bf F}^{\mathrm{H}} $, where $ \hat{\bf X} $ is the Discrete Fourier Transform (DFT) of $ \bf X $, $\bf F $ is the DFT matrix and $ {\bf F}^{\mathrm{H}} $ is the Hermitian transpose of $ \bf F $. Therefore, Eq. \[eq2\] results in a very efficient element-wise multiplication solution in the Fourier domain without matrix inversion. Please refer to [@KCF] for more details.
[**Convolutional Regression.**]{} The convolutional regression [@CREST; @DSLT; @ATOM] considers the following minimization problem: $$\label{eq3}
\min_{{\bf W}_{\text{Conv}}} \|{\bf X}\ast {{\bf W}_{\text{Conv}}} - {{\bf Y}_{\text{G}}}\|^{2}_{2} + \lambda\|{{\bf W}_{\text{Conv}}}\|^{2}_{2}.$$ Different from the circular correlation in Eq. \[eq1\], the $ \ast $ operation in Eq. \[eq3\] denotes the standard multi-channel convolution, which is the core component in CNNs.
Without a closed-form formula, the solution of Eq. \[eq3\] can be optimized via the standard gradient descent as follows: $$\label{solution of eq3}
{\bf W}_{\text{Conv}}^{i+1} = {\bf W}_{\text{Conv}}^{i} - \alpha \nabla {\cal L}({\bf W}_{\text{Conv}}^{i}),$$ where $ \alpha $ is the learning rate of the gradient descent and $ {\cal L}(\cdot) $ denotes regression error presented in Eq. \[eq3\]. Given the feature map $ {\bf X}\in\mathbb{R}^{M\times N\times C} $, the learned filter (or convolutional kernel) $ {{\bf W}_{\text{Conv}}}\in\mathbb{R}^{m\times n\times C} $ regresses the feature map $ \bf X $ to the desired Gaussian label $ {\bf Y}_{\text{G}} $. Note that the correlation filter $ {{\bf W}_{\text{CF}}} $ in Eq. \[eq1\] has the same spatial size with $ \bf X $, while the convolutional filter $ {{\bf W}_{\text{Conv}}} $ requires to be smaller than $ \bf X $, i.e., $ m<M,~n<N $, as shown in Figure \[fig:comparison\] (b).
Discrete Sampling based Ridge Regression {#discrete ridge regression}
----------------------------------------
In the CF and convolutional regression, the learned filters regress the RoI to a dense response map. This continuous prediction generally brings in the class-imbalance issue [@DSLT], where plentiful uninformative samples will overwhelm the valuable ones in the filter training. Actually, there is no need to limit ourselves to the dense prediction in a regression scheme. To focus on the hard samples, we propose a simple, flexible yet effective Discrete Sampling based Ridge Regression (DSRR). The *discrete* lies in two aspects: (1) The training data are sampled discretely (Figure \[fig:comparison\] (c)), which is similar to the classic classification based tracking approach [@MDNet]. By carefully selecting the training samples, the learned filter pays more attention to the hard negative samples and naturally avoids the class-imbalance issue. (2) The label is discrete (binary) instead of the soft Gaussian shape, which introduces the label margin between positive and hard negative samples. As shown in Figure \[fig:comparison\] (c), the learned discrete ridge regressor can be interpreted as a fully-connected layer with a single node, but provides a fast solution in a single pass to learn the model instead of learning with time-consuming back-propagation.
![Illustration of correlation filter, convolutional regression, and our discrete sampling based ridge regression.[]{data-label="fig:comparison"}](comparison.pdf){width="8.7cm"}
To train this regressor, we represent each sample by a high-dimensional feature embedding via a CNN mapping function $ \varphi({\bf X}, {\bf B}_i) $, whose inputs consist of the base feature map $ \bf X $ and the $ i $-th sample’s bounding box $ {\bf B}_i\in\mathbb{R}^{4} $. These training samples are discretely sampled with binary labels, representing the target or background. As shown in Figure \[fig:main\], the mapping function $ \varphi(\cdot) $ first refines the backbone feature $ \bf X $ through two convolutional layers, and further generates the feature embedding via an RoI pooling operation followed by a fully-connected layer. Then we assemble these feature embeddings to form the data matrix $ {\bf D}=[\varphi({\bf X},{\bf B}_1), \cdots,\varphi({\bf X},{\bf B}_P)]^{\mathrm{T}} \in\mathbb{R}^{P\times L} $, which contains $ P $ embeddings and each of them is $ L$-dimensional. Based on the overlap ratios between candidates’ boxes $ {\bf B} $ and ground-truth box, these feature embeddings are assigned by positive or negative labels. Leveraging data matrix $ \bf D $ and its binary label $ {\bf Y}_{\text{B}} $, the discrete sampling based ridge regressor $ {\bf W}_{\text{DSRR}} $ can be obtained by solving the following minimization problem: $$\label{eq4}
\min_{{\bf W}_{\text{DSRR}}}\|{\bf D} {{\bf W}_{\text{DSRR}}} -{{\bf Y}_{\text{B}}}\|^{2}_{2} + \lambda\|{{\bf W}_{\text{DSRR}}}\|^{2}_{2},$$ where $ {\bf Y}_{\text{B}}\in\mathbb{R}^{P} $ is the binary label.
[**Primal Domain.**]{} Since Eq. \[eq4\] still follows the standard ridge regression, similar to Eq. \[eq2\], it has the closed-form solution $ {\bf W}_\text{DSRR} = ({\bf D}^{\mathrm{T}}{\bf D}+\lambda {\bf I})^{-1}{\bf D}^{\mathrm{T}}{{\bf Y}_{\text{B}}} $. Compared with the solution to CF, the main advantage is that this data matrix $ \bf D $ no longer contains fake (cyclically shifted) samples, while the tradeoff is that the Fourier domain solution becomes unfeasible. In the above solution, the main computational burdern lies in the matrix inverse, whose time complexity is $ O(L^3) $ for the matrix $ ({\bf D}^{\mathrm{T}}{\bf D}+\lambda {\bf I})\in \mathbb{R}^{L\times L}$.
[**Dual Domain.**]{} Eq. \[eq4\] can also be solved in the dual domain, where the regressor $ {\bf W}_\text{DSRR} $ is expressed by a linear combination of the samples, i.e., $ {\bf W}_\text{DSRR} ={\bf D}^{\mathrm{T}}\bm{\alpha}$. The variables under optimization are thus $ \bm{\alpha} $ instead of $ {\bf W}_{\text{DSRR}} $. The dual variables $ \bm{\alpha} $ can be solved by $ \bm{\alpha} = ({\bf D}{\bf D}^{\mathrm{T}}+\lambda {\bf I})^{-1}{{\bf Y}_{\text{B}}} $ [@rifkin2003regularized]. Therefore, the ridge regressor can be computed in the dual domain as follows: $$\label{Eq5}
{\bf W}_\text{DSRR} = {\bf D}^{\mathrm{T}}\bm{\alpha} = {\bf D}^{\mathrm{T}}({\bf D}{\bf D}^{\mathrm{T}}+\lambda {\bf I})^{-1}{{\bf Y}_{\text{B}}}.$$ Since $ ({\bf D}{\bf D}^{\mathrm{T}} + \lambda {\bf I})\in \mathbb{R}^{P\times P}$, the matrix inverse in Eq. \[Eq5\] has the time complexity of $ O(P^3) $ instead of $ O(L^3) $ in primal domain, which relates to the sample number $ P $ instead of feature dimension $ L $. Thanks to the limited number of hard examples, a small $ K $ is generally practicable. While in case of a low feature dimension $ L $, the primal domain solution will be more efficient. Overall, depending on the sizes of $ L $ and $ P $, we can always find a good efficiency balance between the primal and dual solutions.
[**Offline Training.**]{} In the training stage, we aim to learn a CNN function $ \varphi(\cdot) $ to ensure the learned feature representation suitable for the designed ridge regression. To this end, we adopt a Siamese-like pipeline in the training stage, where the template branch is utilized to learn the ridge regressor while the search branch is used to generate plentiful test candidates for loss computation. In the large Region of Interest (RoI), we randomly draw plentiful samples. The positive and negative samples are collected following the ratio of 1 $:$ 3, which have $ \geqslant$ 0.7 and $ \leqslant $ 0.5 overlap ratios with ground-truth bounding boxes, respectively. In our experiment, the total sample number is 400 in each frame, i.e., 100 positive samples and 300 negative samples.
Instead of using the prototype $ \varphi(\cdot) $ in Figure \[fig:main\] for simplicity, to achieve better performance, we exploit the multi-scale feature representations from both $ \mathrm{Block3} $ and $ \mathrm{Block4} $ of the ResNet-18 [@ResNet] as the inputs of two individual $ \varphi(\cdot) $ networks. The Precise RoI Pooling ($ \mathrm{PrPool} $) [@IoUNet] is utilized in $ \varphi(\cdot) $ to crop the $ \mathrm{Block3} $ and $ \mathrm{Block4} $ features, whose output sizes are 5$\times$5 and 3$\times$3, respectively. The following fully-connected layer maps the pooled features to a 256-dimensional feature vector. Finally, the $ \mathrm{Block3} $ and $ \mathrm{Block4} $ feature vectors are concatenated along the channel dimension as the 512-dimensional output feature embedding.
Thanks to the closed-form solution of ridge regression, it can be embedded as a differentiable layer for end-to-end training. Leveraging the regressor $ {\bf W}_\text{DSRR} $ learned via template branch, the regression scores of the test candidates in the search branch can be calculated by $ {\bf Y}_\text{Test} = {\bf D}_\text{Test}{\bf W}_\text{DSRR} $. To train the network $ \varphi(\cdot) $, we adopt the standard $ L_2 $ loss as the training objective: $ \ell = \|{\bf Y}_\text{Test}- {\bf Y}_\text{GT}\|^{2}_{2} $, where $ {\bf Y}_\text{GT} $ is the ground-truth binary label of the test samples. After offline training, $ \varphi(\cdot) $ is fixed in the tracking stage.
![Tracking examples of the proposed cascaded framework. In the first stage, we select top three peaks as the high-quality proposals, which are further checked via the second stage. If the first stage fails to predict confidently, we draw plentiful candidates in the second stage for target re-detection.[]{data-label="fig:tracking instance"}](cascade_example.pdf){width="7.3cm"}
[**Connection with Related Methods.**]{} We compare the CF, convolutional regression, and our discrete sampling based ridge regression in the following 4 aspects. (1) [**Efficiency.**]{} Convolutional regression typically requires gradient back-propagation to learn the filter. CF exploits the closed-form solution in the Fourier domain, showing extremely attractive efficiency. The proposed DSRR also has a closed-form solution, yielding satisfactory efficiency. (2) [**Label.**]{} Both CF and convolutional regression predict dense response scores. In contrast, our approach considers discrete proposals, which is flexible to focus on the hard examples and eliminate the class-imbalance issue. (3) [**Effectiveness.**]{} The performance of CF is heavily limited by the boundary effect, i.e., the data matrix $\bf A $ consists of plentiful fake samples. In contrast, the convolutional regression and our DSRR are learned using *real* samples. (4) [**Flexibility.**]{} The CF can only detect the RoI with a fixed size (Figure \[fig:comparison\]). In contrast, the convolutional regression and DSRR are more flexible, which can be applied to the RoI of any size and explore a larger area when necessary (e.g., target out-of-view). Considering the above characteristics, we choose the convolutional regression and discrete ridge regression as the first and second stages in our approach, respectively.
Our discrete ridge regression also shares partial similarity with the classification based approach (MDNet [@RTMDNet]). The main distinction is that we learn a closed-form solver to *regress* the proposals instead of leveraging several fully-connected (FC) layers to *classify* them, which is much more efficient via a feed-forward computation without back-propagation to update the FC parameters.
Online Tracking {#online tracking}
---------------
[**Cascaded Regression Tracking.**]{} Before tracking, we first learn the aforementioned two regressors using the initial frame. For the first stage, instead of adopting the stochastic gradient descent (SGD) to learn the convolutional filter, we follow Danelljan *et al.* [@ATOM] to use Newton-Gaussian descent as the optimization strategy for fast convergence, and learn a 4$ \times$4 kernel to regress the Gaussian response map. To learn the second-stage regressor, based on the initial ground-truth label, we crop the positive and negative samples following a ratio of 1 $:$ 3 to form the data matrix. Then, the discrete ridge regressor $ {\bf W}_\text{DSRR} $ is obtained by the primal or dual solution, depending on the sample number and feature dimension. In the initial few frames, the sample number is smaller than the feature dimension (i.e., $P < L$), and we choose the dual domain. With the arrival of new frames, if $ P > L$, we switch to the primal domain.
![An illustration of the hard negative mining and sample reweigh for the proposed adaptive ridge regression learning.[]{data-label="fig:hard_negative"}](hard_negative_mining.pdf){width="8.6cm"}
During online tracking, in each frame, top-3 peaks in the first stage’s score map are selected, and the corresponding proposals are fed to the network $ \varphi(\cdot) $ to generate the feature embeddings. These ambiguous proposals are further checked by the second-stage regressor, as shown in Figure \[fig:tracking instance\]. Finally, we equally combine the prediction scores of these two stages, and select the highest proposal as the current target. After target localization, we utilize the IoU predictor proposed in [@ATOM] to further refine the target scale.
[**Cascaded Re-detection.**]{} As a common strategy in many visual trackers [@DiMP; @ATOM; @MDNet], we set two reliability thresholds $ \tau_1 $ and $ \tau_2 $ for the two regressors, respectively. In case that the first stage cannot confidently predict the target, i.e., the highest response score is lower than $ \tau_1 $, we sample abundant candidates (512 per frame) and leverage the second-stage regressor for re-detection, as shown in Figure \[fig:tracking instance\]. If the confidence score of the re-detected target exceeds $ \tau_2 $, we regard it as the target. Otherwise, we keep the target position as in the previous frame. Since the backbone features are shared, this re-detection process and the following model update only involve a slight computational burden.
Online Model Update {#model update}
-------------------
[**Hard Negative Mining.**]{} Model update is the core component for discriminating the online distractors. To alleviate the corruption of the memory buffer, we only collect the training samples in reliable frames. Here, a *reliable* frame represents that both two regressors predict confidently, i.e., their estimated scores exceed $ \tau_1 $ and $ \tau_2 $, respectively.
The first stage is incrementally updated by the Gauss-Newton descent using newly collected RoI samples following [@ATOM]. The second stage is expected to distinguish ambiguous samples. To this end, we discretely draw two times of the desired negative samples and select only half of them with a high regression score, as shown in Figure \[fig:hard\_negative\]. These hard training samples are added to the buffer for the model update.
**Input**: Video sequence and initial ground-truth. Initialize two regressors $ {\bf W}_\text{Conv} $ and $ {\bf W}_\text{DSRR} $. Initialize the reliable frame buffer $ {\cal T}=\{1\} $. Conduct cascaded regression tracking;
Collect the current RoI region $ {\cal R}_{t} $; Draw pos/neg samples $ {\cal S}^{+}_{t} $ and $ {\cal S}^{-}_{t} $; Drop $ 50\% $ easy negative samples from $ {\cal S}^{-}_{t} $; $ {\cal T}\leftarrow {\cal T}\cup\{t\} $;
$ {\cal T}\leftarrow {\cal T}\setminus\{\text{min}_{k\in{\cal T}}k\} $;
Update $ {\bf W}_\text{Conv} $ using $ {\cal R}_{k\in{\cal T}} $; Update $ {\bf W}_\text{DSRR} $ using $ {\cal S}^{+/-}_{k\in{\cal T}} $;
\[code1\]
[**Adaptive Ridge Regression.**]{} For the second stage, under consistent model update, the ambiguity degrees of different training samples dynamically change. Therefore, we further assign a weight $ \beta_i $ to each training sample $ \varphi({\bf X},{\bf B}_i) $ in the memory buffer. As a result, the discrete ridge regression is re-formulated as follows: $$\label{eq6}
\min_{{\bf W}_{\text{DSRR}}} \sum_i \beta_i\|{\varphi({{\bf X}, {\bf B}_i})} {{\bf W}_{\text{DSRR}}} -{y_i}\|^{2}_{2} + \lambda\|{{\bf W}_{\text{DSRR}}}\|^{2}_{2}.$$ By defining a weight matrix $ {\bf M} = [\sqrt{\beta_1},\cdots,\sqrt{\beta_P}]^{\mathrm{T}} $, Eq. \[eq6\] can be converted into the matrix form as follows: $$\label{eq6_new}
\min_{{\bf W}_{\text{DSRR}}} \| {\bf M} \odot {\bf D} {{\bf W}_{\text{DSRR}}} -{\bf M} \odot {\bf Y}_{\text{B}}\|^{2}_{2} + \lambda\|{{\bf W}_{\text{DSRR}}}\|^{2}_{2}.$$ As a result, the the solution to Eq. \[eq6\] can be computed by $$\label{eq7}
{\bf W}_\text{DSRR} = (\widetilde{\bf D}^{\mathrm{T}}\widetilde{\bf D}+\lambda {\bf I})^{-1}\widetilde{\bf D}^{\mathrm{T}}{\widetilde{{\bf Y}_{\text{B}}}} = \widetilde{\bf D}^{\mathrm{T}}(\widetilde{\bf D}\widetilde{\bf D}^{\mathrm{T}}+\lambda I)^{-1}{\widetilde{{\bf Y}_\text{B}}},$$ where $ \widetilde{\bf D} = {\bf M}\odot {\bf D} $, $ \widetilde{{\bf Y}_{\text{B}}} = {\bf M}\odot {\bf Y}_\text{B} $, and $ \odot $ is the element-wise product. We empirically define the weight matrix $ \bf M $ as the reconstruction error of the sample label by previous ridge regressor, as follows: $$\label{eq10}
{\bf M} = \text{norm}\left({\bf Y}_\text{B} - {\bf D}{\bf W}^{t-1}_\text{DSRR}\right) \cdot P,$$ where $ {\bf W}^{t-1}_\text{DSRR} $ is the ridge regressor in the previous frame, $ \text{norm}({\bf x}) = {\bf x} / \|{\bf x}\|_1 $ denotes $L_1$ normlization, and $ P $ is the total sample number in the data matrix. Intuitively, Eq. \[eq10\] normalizes the reconstruction errors of different samples and then rescales the weights to ensure the summation of $ \bf M $ equals to $ P $. A large prediction error means the corresponding sample performs as a hard one for the previously learned $ {\bf W}^{t-1}_\text{DSRR} $, which deserves more attention in the current learning.
In our experiments, a new discrete regressor is learned every 10 frames, and is updated to the previous model in a moving average manner: $ {\bf W}^{t}_\text{DSRR} = (1-\eta){\bf W}^{t-1}_\text{DSRR} + \eta{\bf W}_\text{DSRR} $. An overview of the above model update process is presented in Algorithm \[code1\].
Experiments {#sec:experiments}
===========
Implementation Details
----------------------
In offline training, we freeze all the weights of the backbone network (ResNet-18 [@ResNet]) and adopt a multi-task training strategy to train the network $ \varphi(\cdot) $ and IoU predictor. Note that the inputs of IoU predictor and ridge regression are different. Following ATOM [@ATOM], the IoU predictor leverages the samples that have a certain overlap with the ground-truth box (at least 0.1). In contrast, our ridge regression branch utilizes the aforementioned positive and negative samples to learn the discriminative model. The input RoI region is 5 times of the target size and is further resized to 288$ \times $288. We utilize the training splits of LaSOT [@LaSOT], TrackingNet [@2018trackingnet], GOT-10k [@GOT10k], and COCO [@COCO] for offline training. The model is trained for 50 epochs with 1000 iterations per epoch and 36 image pairs per batch. The ADAM optimizer [@ADAM] is employed with an initial learning rate of 0.01, and use a decay factor 0.2 for every 15 epochs. The first-stage regressor uses ResNet-18 $ \mathrm{Block3} $ features as in ATOM, while the second-state regressor and the IoU predictor takes both $ \mathrm{Block3} $ and $ \mathrm{Block4} $ backbone features as input. In online tracking, to update the second-stage regressor, we collect 30 positive and 90 hard negative samples per reliable frame, and maintain a buffer for the last 30 frames. The learning rate $ \eta $ of the second stage is 0.2. The reliability thresholds $ \tau_1 $ and $ \tau_2 $ are set to 0.25 and 0.4, respectively.
We denote our scaded gression method as [CARE]{} in the following experiments. Our tracker is implemented in Python using PyTorch, and operates about 25 frames per second (FPS) on a single Nvidia GTX 1080Ti GPU. We evaluate our method on each benchmark 3 times and report the average performance.
Ablation Experiments
--------------------
We utilize the OTB-2015 [@OTB-2015], UAV123 [@UAV123], and LaSOT testing set [@LaSOT], with total 503 videos, to comprehensively verify the effectiveness of our framework.
[**Cascaded Framework.**]{} In Table \[table:ablation study\], we compare the performance of each single stage and their cascaded combination. Note that we draw 512 samples per frame if the second stage is tested alone, aiming to obtain satisfactory performance. From Table \[table:ablation study\], we can observe that the first and second stages almost perform identically. The main reason is that the convolutional regression is not discriminative enough, while the discrete sampling strategy fails to well cover a large search region. By combining them in a cascaded manner, superior performance can be obtained. For example, on OTB-2015, our final cascaded tracker outperforms the first and second stages by 3.0% and 3.1%, respectively. On the recent large-scale dataset LaSOT, our final framework surpasses the first stage by 3.0% in AUC. Note that the first stage in our framework is adopted from the ATOM, which already achieves a high performance level on various challenging datasets. Under the same backbone network and bounding box regression manner (i.e., IoUNet), our performance gains can be attributed to the superior discrimination capability of our cascaded framework. As for the tracking speed, with the abundant candidates (512 samples per frame), the second-stage regressor is less efficient than the first stage. In contrast, our cascaded framework achieves a balanced speed and outstanding performance, which only slightly reduces the first-stage efficiency but notably outperforms it in tracking accuracy.
[8.5 cm]{} [@lcccc|ccc|c]{} &First &Second & Re-detection & ADRR & OTB-2015 & UAV123 &LaSOT & Speed\
&Stage &Stage & & & [@OTB-2015] & [@UAV123] &[@LaSOT] & FPS\
&$\checkmark$ & & & &67.5 &63.3 &51.7 &[**30**]{}\
& &$\checkmark$ & & &68.0 &62.1 &49.3 &20\
&$\checkmark$ &$\checkmark$ & &&69.2 &64.4 &53.7 &27\
&$\checkmark$ &$\checkmark$ &$\checkmark$ &&69.5 &65.0 &54.1 &25\
&$\checkmark$ &$\checkmark$ &$\checkmark$ &$\checkmark$&[**70.5**]{} &[**65.4**]{} &[**54.7**]{} &25\
[**Target Re-detection.**]{} As discussed in Section \[online tracking\], our second-stage regressor also acts as a re-detection module due to its high discrimination. As shown in Table \[table:ablation study\], with additional performance improvements, the re-detection mechanism further exploits the potential of the second stage.
[**Adaptive Ridge Regression.**]{} Online model update plays a vital role in our framework. Based on the collected hard samples in the memory buffer, to better concentrate on the valuable ones, we propose an adaptive ridge regression that dynamically reweighs the training samples. As illustrated in Table \[table:ablation study\], our adaptive ridge regression (ADRR) steadily improves the tracking accuracy. Besides, it is worth mentioning that our ADRR is extremely efficient with a negligible computational cost.
Comparison with State-of-the-art Methods
----------------------------------------
We compare our proposed CARE tracker with the recent state-of-the-art trackers on 11 challenging tracking benchmarks including OTB-2013 [@OTB-2013], OTB-2015 [@OTB-2015], UAV123 [@UAV123], LaSOT [@LaSOT], VOT2018 [@VOT2018], VOT2019 [@VOT2019], TrackingNet [@2018trackingnet], Temple-Color [@TempleColor128], UAV20L [@UAV123], Need for Speed [@NFSdataset], and OxUvA [@2018longtermBenchmark].
[**OTB-2013 [@OTB-2013].**]{} OTB-2013 is a widely evaluated tracking dataset with 50 videos. Figure \[fig:otb-50-100\] (left) shows the success plot on the OTB-2013. On this dataset, our method achieves an AUC score of 72.0%, outperforming all previous state-of-the-art trackers such as VITAL [@VITAL] and ECO [@ECO]. Note that the top-performing trackers on this benchmark cannot operate at a real-time speed, e.g., the speeds of VITAL and MDNet are only 1 FPS by using fully-connected layers for candidate classification, while ours is real-time since our closed-form regressor is free of gradient back-propagation. Compared with other state-of-the-art trackers with the same ResNet backbone (e.g., ATOM), our approach exhibits competitive efficiency with a speed of about 25 FPS.
![Success plots on the OTB-2013 [@OTB-2013] (left) and OTB-2015 [@OTB-2015] (right) datasets. The legend shows the AUC score. The proposed CARE method outperforms all the comparison trackers.[]{data-label="fig:otb-50-100"}](CARE_OTB50.pdf "fig:"){width="4.3cm"} ![Success plots on the OTB-2013 [@OTB-2013] (left) and OTB-2015 [@OTB-2015] (right) datasets. The legend shows the AUC score. The proposed CARE method outperforms all the comparison trackers.[]{data-label="fig:otb-50-100"}](CARE_OTB100.pdf "fig:"){width="4.3cm"}
[**OTB-2015 [@OTB-2015].**]{} OTB-2015 benchmark extends OTB-2013 with additional 50 videos, resulting in 100 videos in total. Figure \[fig:otb-50-100\] (right) shows the success plot over 100 videos on the OTB-2015. Our method achieves an AUC score of 70.5% on this benchmark, surpassing the recently proposed SiamRPN++ [@siamrpn++], ECO [@ECO], and VITAL [@VITAL] trackers. Compared with the recent single-stage regression trackers such as ATOM [@ATOM] and DiMP-18 [@DiMP], our CARE method outperforms them by 3.4% and 4.3% in terms of AUC score, respectively. Note that DiMP-18 is the recently proposed regression method with discriminative model learning, which represents the state-of-the-art performance on several datasets.
[**UAV123 [@UAV123].**]{} This dataset includes 123 aerial videos collected by a low-attitude UAV platform. Therefore, UAV123 focuses on evaluating visual trackers in the UAV scenarios with small and fast-moving targets. Figure \[fig:uav-lasot\] (left) illustrates the success plot of the state-of-the-art trackers including SiamRPN++, ATOM, and DiMP-18. Compared with the recent remarkable approaches, our method achieves the best result. Especially, our approach shows an AUC score of 65.4%, outperforming SiamRPN++, ATOM, and DiMP-18 by 4.1%, 1.9%, and 2.0% AUC score, respectively. Since our approach is equipped with the same backbone network and IoU predictor compared with ATOM and DiMP-18, our performance advantage verifies the superiority of the proposed cascaded tracking framework.
![Success plots on the UAV123 [@UAV123] (left) and LaSOT [@LaSOT] (right) datasets. The legend shows the AUC score. The proposed CARE method outperforms all the comparison trackers.[]{data-label="fig:uav-lasot"}](CARE_UAV.pdf "fig:"){width="4.3cm"} ![Success plots on the UAV123 [@UAV123] (left) and LaSOT [@LaSOT] (right) datasets. The legend shows the AUC score. The proposed CARE method outperforms all the comparison trackers.[]{data-label="fig:uav-lasot"}](CARE_LaSOT.pdf "fig:"){width="4.3cm"}
[**LaSOT [@LaSOT].**]{} LaSOT is a recent large-scale tracking dataset including 1200 videos, which is more challenging than the previous short-term benchmarks with an average of 2500 frames per video. We evaluate our approach on the test set of 280 videos. Except for the top-performing trackers like MDNet and VITAL on this dataset, we also include the recent C-RPN [@CRPN], SiamRPN++, ATOM, and DiMP-18 for comparison. The success plot on LaSOT is shown in Figure \[fig:uav-lasot\] (right). On this dataset, our approach achieves an AUC score of 54.7%, outperforming the previous best method on this benchmark (i.e., MDNet) by a considerable margin of 15.0% AUC score. Compared with the recent C-RPN, SiamRPN++, ATOM, and DiMP-18, our CARE surpasses them by 9.2%, 5.1%, 3.3%, and 1.2% in AUC, respectively.
![Expected average overlap (EAO) graph with trackers ranked from right to left. Our method obviously outperforms all the participants on the VOT2018 [@VOT2018].[]{data-label="fig:vot"}](vot.pdf){width="8.6cm"}
[8.6 cm]{} [@lcccccccc]{} & SPM &C-RPN &DWSiam &SiamMask & SiamRPN++ &ATOM &DiMP-18 &[**CARE**]{}\
& [@SPM] &[@CRPN] &[@deeperwiderSiamFC] &[@SiamMask] &[@siamrpn++] &[@ATOM] &[@DiMP] &\
R &0.30 &- &- &0.276 &0.234 &0.204 &[ **0.182**]{} &[ **0.201**]{}\
A &0.58 &- &- &[ **0.609**]{} &[ **0.600**]{} &0.590 &0.594 &0.597\
EAO &0.338 &0.289 &0.301 &0.380 &[**0.414**]{} &0.401 &0.402 &[ **0.410**]{}\
FPS &110 &32 &150 &55 &35 &30 &46 &25\
[8.5 cm]{} [@lccccccc]{} & SPM &SiamMask &SiamMask-E & SiamRPN++ &ATOM &SiamDW &[**CARE**]{}\
& [@SPM] &[@SiamMask] &[@chen2019SiamMaskRotated] &[@siamrpn++] &[@ATOM] &[@deeperwiderSiamFC] &\
R &0.507 &0.461 &0.487 &0.482 &[ **0.411**]{} &0.467 &[ **0.343**]{}\
A &0.577 &0.594 &[ **0.652**]{} &0.599 &[ **0.603**]{} &0.600 &0.601\
EAO &0.275 &0.287 &0.309 &0.285 &0.292 &[ **0.299**]{} &[ **0.323**]{}\
FPS &110 &55 &50 &35 &30 &- &25\
{width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"} {width="4.4cm"}
In Figure \[fig:lasot-attribute\], we further provide the attribute evaluation on the LaSOT benchmark [@LaSOT]. On this large-scale dataset, our approach shows good results on fast motion, out-of-view, and viewpoint change. On the above attributes, our method even surpasses the recently remarkable DiMP-18 tracker [@DiMP] by a large margin, which can be attributed to the strong discrimination of our second stage. Our second-stage regressor further checks the ambiguous candidates and serves as a re-detection module, which significantly improves the tracking performance on the challenging scenarios such as fast motion, out-of-view, and target occlusion. Besides, our method outperforms its baseline method ATOM [@ATOM] in all attributes on the LaSOT dataset. In particular, our method significantly outperforms ATOM in viewpoint change, low resolution, and partial occlusion by 8.2%, 4.7%, and 4.3%, respectively, which demonstrates the effectiveness of our second stage for cascaded verification. Since our framework mainly focuses on the target re-identification and re-detection, in the attributes such as aspect ratio change and scale variation, our method is less effective and slightly improves the baseline.
[18.0 cm]{} [@lcccccccccccccccc]{} &BACF &Staple &Staple-CA &CSR-DCF &ECOhc &ECO &SiamFC &CFNet &MDNet &UPDT &DaSiamRPN &SPM &C-RPN & ATOM &DiMP-18 &[**CARE**]{}\
&[@BACF] &[@Staple] &[@Context-AwareCorrelationFilter] &[@CSR-DCF] &[@ECO] &[@ECO] &[@SiamFC] &[@CFNet] &[@MDNet] &[@UPDT] &[@DaSiamRPN] &[@SPM] &[@CRPN] &[@ATOM] &[@DiMP] &\
Precision &46.1 &47.0 &46.8 &48.0 & 47.6 &49.2 &53.3 &53.3 &56.5 &55.7 &59.1 &66.1 &61.9 &64.8 &[ **66.6**]{} &[ **66.7**]{}\
Norm. Prec. &58.0 &60.3 &60.5 &62.2 &60.8 &61.8 &66.3 &65.4 &73.3 &70.2 &73.3 &77.8 &74.6 &77.1 &[ **78.5**]{} &[ **79.0**]{}\
Success &52.3 &52.8 & 52.9 &53.4 &54.1 &55.4 &57.1 &57.8 &63.8 &61.1 &63.8 &71.2 &66.9 &70.3 &[ **72.3**]{} &[ **71.8**]{}\
Speed (FPS) &35 &70 &55 &18 &45 &8 &86 &55 &1 &- &[ **160**]{} &[ **110**]{} &32 &30 &46 &25\
[18.0 cm]{} [@lccccccccccccccc]{} &KCF &DSST &SRDCF &HCF &SiamFC &CFNet &ECOhc &MDNet &C-COT &ECO &SiamRPN &SiamRPN++ & ATOM &DiMP-18 &[**CARE**]{}\
&[@KCF] &[@DSST] &[@SRDCF] &[@HCF] &[@SiamFC] &[@CFNet] &[@ECO] &[@MDNet] &[@C-COT] &[@ECO] &[@SiamRPN] &[@siamrpn++] &[@ATOM] &[@DiMP] &\
UAV20L [@UAV123] &19.8 &27.0 &34.3 &- &39.9 &34.9 &- &- &- &43.5 &45.4 &56.1 &55.4 &[ **57.1**]{} &[ **60.3**]{}\
TC128 [@TempleColor128] &38.4 &40.6 &50.9 &48.2 &50.5 &45.6 &56.1 &56.3 &58.3 &59.7 &- &56.2 &59.3 &[ **60.6**]{} &[ **61.2**]{}\
NfS [@NFSdataset] &21.7 &28.0 &35.1 &29.5 &- &- &- &42.2 &- &46.6 &- &50.0 &58.4 &[ **61.0**]{} & [ **60.5**]{}\
Speed (FPS) &[ **270**]{} &45 &5 &12 &86 &55 &45 &1 &0.3 &8 &[ **160**]{} &35 &30 &46 &25\
[8.7 cm]{} [@lcccccccccccc]{} &MDNet &LCT &TLD &SiamFC+R &MBMD &SPLT &[**CARE**]{}\
&[@MDNet] &[@LCT] &[@TLD] &[@SiamFC] &[@MBMD] &[@SkimmingPerusalTracking] &\
MaxGM &0.343 &0.396 &0.431 &0.454 &0.544 &[ **0.622**]{} &[ **0.749**]{}\
TPR & 0.472 &0.292 &0.208 &0.427 &[ **0.609**]{} &0.498 &[ **0.609**]{}\
TNR & 0 &0.537 &[ **0.895**]{} &0.481 &0.485 &0.776 &[ **0.922**]{}\
[**VOT2018 [@VOT2018].**]{} VOT2018 dataset contains 60 challenging videos for short-term tracking evaluation, which will reset the tracker to the ground-truth position when tracking failure occurs. On this benchmark, trackers are evaluated by the Expected Average Overlap (EAO), which considers both accuracy (average overlap over successful frames) and robustness (failure rate). From Figure \[fig:vot\], we can observe that our approach outperforms all the participants on the VOT2018. Compared with the recent state-of-the-art approaches, our approach still exhibits satisfactory results. As shown in Table \[table:vot\], our method surpasses the recent regression based methods such as ATOM and DiMP-18 with a relative gain of 2.2% and 2.0% in terms of EAO, respectively. Compared with the cascaded Siamese trackers including SPM [@SPM] and C-RPN [@CRPN], our method significantly outperforms them thanks to our online adaptation capability. Among all the compared trackers, only SiamRPN++ slightly outperforms ours, which adopts a deeper ResNet-50 as the backbone network.
[**VOT2019 [@VOT2019].**]{} VOT2019 is the recently released challenging benchmark, which replaces 12 easy videos in VOT2018 [@VOT2018] by 12 more difficult videos. Therefore, the EAO scores of the state-of-the-art trackers such as SiamRPN++ drop sharply. We compare our approach with the representative approaches in Table \[table:vot19\]. Compared with the SiamRPN++ [@siamrpn++] and SiamDW [@deeperwiderSiamFC] with deeper ResNet-50, our method with a ResNet-18 obviously surpasses them. The ATOM [@deeperwiderSiamFC] is a top-performing single-stage regression tracker, while ours outperforms it with a relative gain of 6.5% in terms of EAO.
[**TrackingNet [@2018trackingnet].**]{} The recent TrackingNet benchmark contains more than 30K videos with more than 14 million dense bounding box annotations. The videos are collected on the YouTube, providing large-scale high-quality data for assessing visual trackers in the wild. We evaluate our method on the test set of the recently released large-scale TrackingNet dataset, which consists of 511 videos. Note that the recent trackers already achieve outstanding AUC scores of more than 70%, which means the improvement room on this dataset is limited. As shown in Table \[table:trackingnet\], the proposed tracker achieves a normalized precision score of 79.0% and a success score of 71.8%, which is comparable or superior to previous state-of-the-art trackers such as ATOM and DiMP-18.
[**Need for Speed [@NFSdataset].**]{} NfS dataset contains 100 challenging videos with fast-moving targets, which aims at evaluating the tracking robustness in object fast-moving scenarios. We evaluate our approach on the 30 FPS version of NfS. The AUC scores of comparison approaches are shown in Table \[table:TC\_NFS\]. Since the search range is limited in SiamRPN++, its performance is relatively unsatisfactory (10.5% lower than ours in AUC). The state-of-the-art DiMP-18 and ATOM represent the top performance on this dataset. Our method is comparable with DiMP-18 and outperforms ATOM by 2.1% AUC.
[**Temple-Color [@TempleColor128].**]{} Temple-Color benchmark is a challenging dataset consisting of 128 color videos. In Table \[table:TC\_NFS\], we show the AUC score of state-of-the-art trackers on this benchmark. Compared with the SiamRPN++ with ResNet-50, our method outperforms it by a large margin of 5.0% AUC score. The recent single-stage regression tracker ATOM and DiMP yield AUC scores of 59.3% and 60.6%, respectively. The proposed approach also outperforms the recent single-stage regression tracker ATOM and DiMP-18 by 1.9% and 0.6% AUC score, respectively.
[17.8 cm]{} [@lcccccccccccc]{} &Backbone Network &OTB2015 [@OTB-2015] &TC128 [@TempleColor128] &UAV123 [@UAV123] &NfS [@NFSdataset] &VOT2018 [@VOT2018] &VOT2019 &LaSOT [@LaSOT] & TrackingNet [@2018trackingnet] &Speed\
& &AUC score &AUC score &AUC score & AUC score &EAO score &EAO score &AUC score & Success score &FPS\
DaSiamRPN [@DaSiamRPN] &AlexNet &65.8 &- &58.6 &- &0.326 &- &41.5 &63.8 &[ **160**]{}\
C-RPN [@CRPN] &AlexNet &66.3 &- &- &- &0.289 &- &45.5 &66.9 &32\
SPM [@SPM] &AlexNet &68.7 &- &- &- &0.338 &0.275 &- &71.2 &[ **110**]{}\
ATOM [@ATOM] &ResNet-18 &67.1 & 59.3 & 63.5 &58.4 & 0.401 &0.292 &51.4 &70.3 &30\
DiMP-18 [@DiMP] &ResNet-18 &66.2 &60.6 & 63.4 &61.0 &0.402 &- & 53.5 &72.3 &46\
[**CARE-18 (Ours)**]{} &ResNet-18 & [ **70.5**]{} &61.2 &[ **65.4**]{} &60.5 &0.410 &0.323 &54.7 &71.8 &25\
SiamRPN++ [@siamrpn++] &ResNet-50 &69.6 &56.2 &61.3 &50.0 &0.414 &0.285 &49.6 &73.3 &35\
DiMP-50 [@DiMP] &ResNet-50 &68.4 &[ **61.5**]{} &64.5 &[ **62.0**]{} &[ **0.440**]{} &[ **0.379**]{} &[ **56.9**]{} & [ **74.0**]{} &40\
[**CARE-50 (Ours)**]{} &ResNet-50 & [ **71.2**]{} &[ **61.7**]{} &[ **64.6**]{} &[ **62.3**]{} &[ **0.427**]{} &[ **0.353**]{} &[ **56.1**]{} & [ **74.2**]{} &21\
[**UAV20L [@UAV123].**]{} This is a long-term tracking benchmark consisting of 20 long UAV videos with an average length of 2934 frames. Our second-stage regressor ensures the tracking robustness and helps re-detect the lost target. As a result, our method significantly surpasses previous methods such as ATOM, DiMP-18, and SiamRPN++ (Table \[table:TC\_NFS\]).
[**OxUvA [@2018longtermBenchmark].**]{} This is a recent large-scale long-term tracking benchmark with 366 videos. The targets in OxUvA undergo frequent partial/full occlusion and out of view. On this dataset, the visual trackers are required to predict the target state (presence or absence) in each frame. We test our method on the test set of 166 videos. The comparison results are shown in Table \[table:OxUvA\]. We do not add any additional mechanisms (e.g., global search) and merely use the reliability thresholds to predict the target presence/absence. Note that the recently proposed Skimming-Perusal method (SPLT) [@SkimmingPerusalTracking] leads the top performance on this dataset, which is specially designed for long-term tracking with a local-global search. Without bells and whistles, our approach outperforms SPLT by a relative gain of 20.4% in terms of MaxGM, showing the importance of online discrimination learning. The prior motion model (e.g., cosine window) in short-term trackers heavily limits their long-term performance. Benefited from strong discrimination, our tracker is free of the motion model (e.g., cosine window) and simultaneously handles short-term and long-term scenarios.
Performance with a Deeper Backbone Network
------------------------------------------
For fair comparison, in this section, we compare our method with state-of-the-art trackers with the same backbone network. In our approach, we follow ATOM [@ATOM] and use a shallow backbone network of ResNet-18 for high efficiency. By adopting the deeper ResNet-50 [@ResNet], our CARE approach obtains further performance improvements and still maintains a near real-time speed of about 21 FPS on a single Nvidia GTX 1080Ti GPU. In Table \[table:res50\], we include the recent SiamRPN++ [@siamrpn++] and DiMP-50 [@DiMP] for comparison, both of which leverage the deep ResNet-50 model. From the results in Table \[table:res50\], we can observe that our CARE-50 steadily outperforms SiamRPN++ and is comparable with the recent DiMP-50. It is worth mentioning that DiMP-50 represents the current state-of-the-art tracker in various tracking benchmarks.
Conclusion {#sec:conclusion}
==========
In this paper, we propose a conceptually simple yet effective discrete sampling based ridge regression, which performs as an alternative of the fully-connected layers to discriminate the candidates, but exhibits promising efficiency under a closed-form solution. Its high flexibility allows the incorporation of hard negative mining as well as our proposed adaptive ridge regression to enhance online discrimination. We further complement it with the convolutional regression to develop a cascaded framework for robust visual tracking. The first stage enables a fast and dense search, while the second stage guarantees distractor discrimination. The proposed method exhibits outstanding results on several challenging benchmarks with a real-time speed.
[^1]: Ning Wang, Wengang Zhou, and Houqiang Li are with the CAS Key Laboratory of Technology in Geo-spatial Information Processing and Application System, Department of Electronic Engineering and Information Science, University of Science and Technology of China, Hefei, China.\
E-mail: [email protected], {zhwg, lihq}@ustc.edu.cn.
[^2]: Qi Tian is with the Huawei Noah’s Ark Laboratory.\
E-mail: [email protected].
[^3]: Corresponding authors: Wengang Zhou and Houqiang Li.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The authenticity of images posted on social media is an issue of growing concern. Many algorithms have been developed to detect manipulated images, but few have investigated the ability of deep neural network based approaches to verify the authenticity of image labels, such as event names. In this paper, we propose several novel methods to predict if an image was captured at one of several noteworthy events. We use a set of images from several recorded events such as storms, marathons, protests, and other large public gatherings. Two strategies of applying pre-trained Imagenet network for event verification are presented, with two modifications for each strategy. The first method uses the features from the last convolutional layer of a pre-trained network as input to a classifier. We also consider the effects of tuning the convolutional weights of the pre-trained network to improve classification. The second method combines many features extracted from smaller scales and uses the output of a pre-trained network as the input to a second classifier. For both methods, we investigated several different classifiers and tested many different pre-trained networks. Our experiments demonstrate both these approaches are effective for event verification and image re-purposing detection. The classification at the global scale tends to marginally outperform our tested local methods and fine tuning the network further improves the results.'
author:
- |
M.Goebel\*; University of California, Santa Barbara; Santa Barbara, CA\
A. Flenner\*; NAVAIR; China Lake, CA\
L. Nataraj, B.S. Manjunath; Mayachitra Inc.; Santa Barbara, CA\
*‘\*’-equal contribution*
bibliography:
- 'event-ver.bib'
title: Deep Learning Methods for Event Verification and Image Repurposing Detection
---
Introduction {#sec:intro}
============
Today social media websites are emerging as a dominant news source, but verifying the validity of the news stories is a difficult problem. In the few months before the 2016 US Presidential elections, the average American saw at least one fake news story, and of those who saw one, half of them believed it to be true [@allcott2017social]. In practice, countering these sources of fake news is a complex problem. There is little entry cost to distributing false information on Facebook, with large potential for ad revenue if the story gains popularity. While in the past people relied on a few well known sources, the website publishing false stories is often removed before being identified as illegitimate [@allcott2017social]. For these reasons, an automated algorithm is needed to identify these fake stories before they reach a large number of users.
A common approach to distribute false information is to select an authentic image from some previously recorded event which convincingly supports their message, and re-brand it as a current story. For example, during the times of storms and hurricanes, images from previous hurricanes are usually re-purposed and uploaded in social media to create a scare. If we have a database of previously recorded events, then we will be able to verify those images that have been re-purposed from older events. In this paper we investigate several methods to automatically identify images for event verification and image re-purposing detection. We explore a transfer learning approach [@transfer] to event verification by using a network pre-trained on the ImageNet dataset [@imagenet_cvpr09], but instead of using the network for image classification we use the network for event verification . After observing many images from different events, certain features may stand out to distinguish between two similar classes. Similar locations, architectures, or identifying symbols may be associated with each. For example, marathon race bibs are generally consistent for all participants in a single event, and different from other races. In this paper, we outline two approaches of applying pre-trained ImageNet network for event verification, one at the global image level and other at the local image level. At the global level, we compare the effect of fine tuning a pre-trained network to a particular dataset to a method that is not tuned to any dataset. At the local level, we explore the effects of spatial context at smaller scales using one method that does averaging and another method without averaging. Figure. \[fig:summary-blk\] provides a summary of the methods. Our experiments on several datasets show that both approaches are promising for event verification and image re-purposing detection.
Related Work {#sec:rel-work}
============
{width="0.95\linewidth"}
While there are several works on image classification, retrieval, event classification, and more, very few works address the problem of event verification. At the time of writing, we are aware of two recent works that are related to this problem [@jaiswal; @Sabir]. The first considers pairs of images and captions[@jaiswal], and then trains a network to decide how consistent the caption is with the associated image. The second work [@Sabir] altered GPS coordinates, captions, and the actual image pixels for image re-purposing detection. In contrast to these methods, our approach operates only at the image pixel level and does not need any metadata, which may not be always available at hand.
We approach the event verification problem as a transfer learning classification problem and distinguish between a finite set of events by using pre-trained networks. A recent paper [@transfer] systematically tested how well features from different architectures transfer to other classification tasks. Their tests showed that when comparing two different architectures, ImageNet performance was only weakly correlated with performance on another task if the network is fixed. However, this effect diminished once the network was fine-tuned [@transfer].
Our methods also do not address the possibility of modification to image pixels. This is a well studied problem, with many viable solutions [@forgery_1; @forgery_2; @forgery_4; @forgery_5] Our models are created under the assumption that they will be used with one of these image manipulation detectors, and that all images given to our detector are unmodified.
[.22]{} {width="0.95\linewidth"}
[.22]{} {width="0.95\linewidth"}
[.22]{} {width="0.95\linewidth"}
[.22]{} {width="0.95\linewidth"}
Event Verification and Image Repurposing Detection
==================================================
In this paper, our goal is to demonstrate an approach to verify if an image was taken at the claimed event. We explored two different approaches where one approach used global image features and another used local image features. More specifically, we treated event verification as a transfer learning problem. A deep learning classifier can be represented as a composition of the two function $f(\cdot) \in R^{N\times K}$ and $g(\cdot) \in R^{K \times M}$ where $f(\cdot)$ consists of the convolutional, non-linear, and pooling layers while $g(\cdot)$ is the fully connected classification layer with $M$ possible classes (see Figure \[fig:summary-blk\]). During the training phase, the parameters of $f(\cdot)$ and $g(\cdot)$ were learned using a large image database such as ImageNet [@imagenet_cvpr09].
We investigated two transfer learning strategies that consisted of using the features from the pre-trained network and learning a new classifier using these features. The first strategy, or global method, extracted features by mapping the entire image using either $f(\cdot)$, $g(f(\cdot))$ or both. The second local method extracted features by mapping image patches using $g(f(\cdot))$, and the output from the patches were then averaged or concatenated into one large feature vector. A classifier was then trained on the extracted features to identify the respective events. Finally, for the global method we also tested allowing the parameters of $f(\cdot)$ to be fine tuned using the event data training set. In total, there were four different strategies as illustrated in Figure $\ref{fig:summary-blk}$.
Global Method
-------------
The global method consists of a two part structure. The first section uses deep convolutional neural networks (CNNs) to extract features from each image. The second is a separate classifier which takes these features as input.
**Untuned model**: The untuned network uses standard models such as ResNet-50 that are trained on ImageNet, and not tuned to the dataset under test. The images are first re-sized to the native resolution used by the CNN (either 224x224 or 299x299). The final feature vector is derived from the outputs from the last convolutional layer and the final output layer. This process replaces a large image with a much smaller feature vector, while hopefully retaining as much relevant information as possible. Using these features, we investigated several classifiers such as extra trees, random forests, nearest neighbors, support vector machines, and convolutional nets. The three variables tested were: the CNN used to extract features, which layers were used as features, and the final classifier algorithm.
**Tuned model**: We also investigated the effect of fine-tuning the CNN. A recent work suggested using at least 800 training images that contained at least 32 images per class [@transfer]. In our case we trained on 800 images with 200 per class, which is close to their minimum for total number of images. The fine tuning was done in Keras, by removing the last fully connected layer of dimension 1000 used for ImageNet and replacing it with a fully connected layer with 4 outputs. An Adam optimizer with learning rate of 0.0001 was run and the training data was iterated through 10 times. Rates slower than 0.0001 did not show any improvement in validation accuracy while taking longer to converge. At our chosen rate, performance generally plateaued after 10 epochs.
Local Method
------------
The local method differs from the global method in two key areas. First, the local method uses a sliding window to extract features from local image patches. Second, the local method derives its feature vector from the output of the fully connected classification layer. The goal with the local method is to capture more image details through the aggregation of information from local image patches.
**Sum Features**: In order to deal with different image sizes, we rescaled each image to have 1120 rows while preserving the aspect ratio of the original image. We then divided the image into overlapping 224x224 patches. Each block was processed through the entire ResNet-50 classifier [@he2016deep]. The output vectors for each patch were summed and the resulting sum was normalized to one, which produced a 1000 dimensional feature vector. This procedure extracted a feature vector of constant length from each image. A classifier was then used to identify the event. Several classifiers were considered including support vector machines, extra trees, random forests, and xgboost [@chen2016xgboost].
The sum feature vectors are similar to the bag-of-words image classification model [@yang2007evaluating]. In particular, the output of each patch can be interpreted as a distribution over classes, which is a 1000 dimensional vector of words or phrases. The words for each patch are then summed, generating a bag of words. The original bag-of-words model would identify one word per patch, however our approach outputs a probability distribution of words for each patch and we sum the different probability distribution to generate our feature vector.
**Full Features**: The sum features average the information gained from all the patches and removes any spatial context. As a simple test to investigate the importance of the spatial knowledge, we removed the sum over the patches. In order to obtain feature vectors that had consistent dimensions, we rescaled each image to have 1120 rows and 1120 columns. ResNet-50 and a sliding window was used to produce a feature vector for each 224x224 block with an overlap of 100 pixels. The final output was the concatenation of the output of all the feature vectors. The same classifiers were tested as in the sum features.
Experiments and Results
=======================
Here we will detail the datasets used in our experiments, the results on the global and local methods and a comparison of both the methods. For each method, a receiver operating curve (ROC) was obtained, and the area under this curve (AUC) is used as the primary metric for comparison.
Event Verification (EV) Dataset
-------------------------------
The classification procedures were tested on the Event Verification dataset generated by NIST as part of the DARPA Media Forensics (MediFor) project. This dataset contained three different subsets, each already divided into training and testing. We refer to them as Ver$_1$, Ver$_2$ and Ver$_{eval}$. The first two versions each had 4 events, with 200 training images for each event, and 100 testing images per event. Ver$_1$ subset contained images associated with the Boston Marathon, Austin Marathon, Occupy Portland, and Occupy Baltimore (see Figure \[fig:my\_label\]). Ver$_2$ had images from Hurricane Sandy, Hurricane Matthew, the Oshkosh Air Show, and Berlin Air Show. The third set was held out for evaluation (Ver$_{eval}$). This subset had 12 events, with 200 images per event, and 600 total test images with held out labels. Testing on Ver$_{eval}$ subset was done by submission of results to a NIST server, which in turn returned the ROC curve. Ver$_1$, Ver$_2$ subsets were used to test across a wide number of models, while Ver$_{eval}$ subset was reserved to see how well these models generalized.
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
Results on Global Method
------------------------
For the global method, many combinations of ImageNet CNNs, feature extraction locations, and standard machine learning classifiers were tested. In general, using only intermediate layer features performed better than only considering output layer features. Using both gave a slight boost for some classifiers though the difference was insignificant for the best classifiers. The results presented for the rest of this section are from the intermediate layer. For most cases, ResNet-50 had the best performance, which is consistent with the results from the recent work on transfer learning [@transfer]. For a network strictly trained on ImageNet, the classifier at the end had a strong impact on performance. Random Forest and Extra Trees classifiers performed the best in most cases. On Ver$_1$, the fine-tuned ResNet universally outperformed the top untuned method. For Ver$_2$ the results were less clear as AUC scores were close. At low false alarm rates the tuned method performs better, while the untuned method has better detection at high false alarm rates. AUC is slightly higher for the untuned case overall and per class ROC curves are shown in Figure \[fig:ver1-global\] for Ver$_1$ subset and Figure \[fig:ver2-global\] for Ver$_2$ subset.
After fine tuning, the classifier at the end seemed to have less of an impact. The best performing of the fine-tuned classifiers were simply the dense output layer generated during the tuning process. Using a different classifier in the end may not have much benefit. However, distance based classifiers do have the ability to provide training images which it found to be similar to the query image, and can be helpful in some applications where justification is needed. More detailed experiments on the effect of different classifiers on various pre-trained networks are presented later in Table. \[tab:global\_results\_array\].
Features ET RF 1-NN SVM XGB
------------ ----------- ------- ------- ------- ------
Local Sum **0.885** .882 0.74 0.855 .872
Local Full **0.857** 0.851 0.612 0.852 .841
: Local method AUC results for different classifiers and different features on the Ver$_{1}$ subset. From left to right the classifiers are extra trees, random forest, 1 nearest neighbor with Euclidean distance, support vector machines, and xgboost.
\[tab:local\]
Dataset Global Untuned Global Tuned Local Sum Local Full
-------------- ---------------- -------------- ----------- ------------
Ver$_1$ 0.889 **0.921** 0.885 0.857
Ver$_2$ **0.857** 0.848 0.831 0.798
Ver$_{eval}$ 0.88 **0.89** 0.85 0.82
: Comparison of AUC between Global and Local Method
\[tab:comparison\]
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
[.3]{} {width="0.8\linewidth" height="0.6\columnwidth"}
Results on Local Method
-----------------------
Based on the results of the global method, only the ResNet-50 network trained using ImageNet was investigated. The classifiers tested were extra trees, random forests, support vector machines, one nearest neighbor, and xgboost which are summarized for the Ver$_{1}$ subset in Table \[tab:local\]. The ensemble methods of extra trees and random forest performed the best for the sum features and were statistically equivalent. For the full features, the extra trees, random forest, and support vector machines had equivalent results. The extra trees and random forest ensemble classifiers performed better on the sum features than the full features, while the support vector machine results was equivalent for the sum and full features. As with the global method, due to random initialization, the results using the same classifier would fluctuate by 1%. The local method was tested on the Ver$_1$, Ver$_2$ and Ver$_{eval}$ subsets of the Event Verification dataset. For all the classifiers except xgboost, the python package scikit-learn version 0.19.1 was used and the default settings obtained the best performance. The python package xgboost was used for the xgboost implementation with a max depth of two and binary logistic objective. The ROC curves for this method are shown in Figures \[fig:ver1-local\] and \[fig:ver2-local\] for the Ver$_{1}$ and Ver$_{2}$ subsets.
Comparison of Global and Local Method
-------------------------------------
The comparison of the global method and the local method on Ver$_1$, Ver$_2$ and Ver$_{eval}$ subsets of the Event Verification dataset are summarized in Table \[tab:comparison\] and Figure \[fig:auc-comp\]. On average, we can see that the global methods perform better than the local methods. Among the global methods, the tuned model obtained the highest AUC for Ver$_1$ and Ver$_{eval}$ subsets while the untuned model performed slightly better than the tuned model for the Ver$_2$ subset. In general, we see that the tuned model would be the most preferable though it would come at the extra cost of tuning the model for every dataset. Next we analyzed the performance of individual events for the four methods. The results are summarized for both the Ver$_1$ and Ver$_2$ subsets in Figure. \[fig:auc-comp\]. The global tuned model performed the best for most events (Austin Marathon, Occupy Baltimore, Boston Marathon, Occupy Portland and Berlin Airshow), while the global untuned model performed the best for two events (Hurricane Sandy, Hurricane Matthew). While the global method tends to outperform the local method in most events, there are cases in which one may be preferred. In the one event where the local method outperformed the global one, Oshkosh Airshow, we hypothesize that there was more detail at the smaller scale. Also, for events such as Occupy Baltimore and Hurricane Sandy, there isn’t any significant difference in the performance of the four methods. Another consideration is that all of the methods besides local sum resize the image to be square. If preserving the aspect ratio is important to a particular task, this may perform relatively better. Testing on several different datasets also showed that performance is highly data dependant. Given our results, we expect that the best out-of-the-box approach is either the global tuned or global untuned.
![Comparison of AUCs on individual events on Ver$_1$ and Ver$_2$ subsets of the Event Verification Dataset.[]{data-label="fig:auc-comp"}](total_bar_chart.pdf){width=".90\linewidth"}
ET RF L1 L2 2NN 4NN
----------------- ------- ------- ------- ------- ------- -------
Xception 0.766 0.753 0.641 0.638 0.675 0.695
VGG16 0.880 0.874 0.751 0.775 0.818 0.845
VGG19 0.883 0.874 0.756 0.768 0.820 0.838
ResNet50 0.889 0.886 0.79 0.781 0.839 0.877
InceptionV3 0.765 0.772 0.648 0.655 0.702 0.714
InceptionResNet 0.682 0.695 0.575 0.573 0.586 0.614
MobileNet 0.842 0.847 0.739 0.735 0.786 0.825
MobileNet2 0.887 0.876 0.766 0.753 0.805 0.843
DenseNet121 0.877 0.871 0.745 0.73 0.785 0.819
DenseNet169 0.869 0.870 0.748 0.745 0.787 0.810
DenseNet201 0.735 0.728 0.636 0.636 0.655 0.690
NASNetMobile 0.764 0.763 0.648 0.646 0.680 0.701
Tuned ResNet 0.909 0.905 0.845 0.831 0.867 0.876
ET RF L1 L2 2NN 4NN
----------------- ------- ------- ------- ------- ------- -------
Xception 0.760 0.756 0.641 0.638 0.675 0.695
VGG16 0.879 0.876 0.751 0.775 0.818 0.845
VGG19 0.884 0.884 0.756 0.768 0.820 0.838
ResNet50 0.894 0.895 0.79 0.781 0.839 0.877
InceptionV3 0.762 0.769 0.648 0.655 0.702 0.714
InceptionResNet 0.681 0.690 0.575 0.573 0.586 0.614
MobileNet 0.843 0.837 0.739 0.735 0.786 0.824
MobileNet2 0.875 0.884 0.766 0.753 0.805 0.843
DenseNet121 0.876 0.872 0.745 0.73 0.785 0.819
DenseNet169 0.877 0.864 0.746 0.745 0.787 0.810
DenseNet201 0.725 0.719 0.636 0.636 0.655 0.690
NASNetMobile 0.770 0.770 0.648 0.646 0.680 0.701
Tuned ResNet 0.912 0.906 0.845 0.835 0.865 0.878
Detailed experiments with Pre-trained models
--------------------------------------------
Given the success of the global method over the local method, we performed a more exhaustive set of tests on the global method, summarized for Ver$_1$ subset in Table \[tab:global\_results\_array\]. Twelve different ImageNet classifiers were tested along with the fine-tuned ResNet50 network. For all networks, the output arrays from the last convolutional layer and the class output layer were saved. The classifiers given only the output layers have been omitted as they did not perform well compared to those given strictly the intermediate layer, or the intermediate and output layers concatenated. Then each combination of the previous two conditions was given to a classifier available in scikit-learn. With this common interface, over a dozen different classifiers with varying parameters were tested. Only a few are shown in Table \[tab:global\_results\_array\]. Based on these results, ResNet seemed to perform the best, which was consistent with the results in [@transfer]. Based on these results, and those from Ver$_2$ subset (not shown due to lack of space), the fine-tuned ResNet50 network had the best performance overall.
Discussion and Future Work
==========================
There are many additional methods which can be used as an add-on to the end-to-end trainable network. For example, face matching was one method we tested with mixed results. On Ver$_1$ subset it gave a 1% boost in AUC, while it had little effect on Ver$_2$ or Ver$_eval$ subset. Given that Ver$_1$ subset contained marathons and protests, while Ver$_2$ subset contained storms and airshows, this result was generally expected. Text extraction may be another area of interest as a supplementary decision. In visually inspecting the data, many images contained text which was unique to the event. This would include street signs, advertisements, and in the case of marathon events, race bibs. In future work we will attempt to accommodate images in their native resolutions. The dataset tested here consisted of images from 0.077 MP to 30MP. Resizing all images into the same height and/or width eliminates much of the additional information that large images contain. A brute force solution may be to sweep the feature extractor across multiple scales of an image pyramid. This will then give a pyramid of features extracted at different scales, leaving the open question of how to combine these. The two approaches represented here are a subset of this, where only the largest scales are used, and all others effectively ignored. Finally, a collection of a larger dataset will facilitate in understanding how our methods will generalize in a real world application that may consist of hundreds of events and several thousand images.
Conclusions
===========
This paper demonstrated the viability of several possible methods for applying pre-trained ImageNet classifiers to the problem of event verification. In particular, we explored analyzing an image at the global level and the local level, and studied the impact of fine-tuning a model for a specific dataset. While the global classification methods out-performed the local methods in our experiments, more data will be needed to confirm this hypothesis. The images included in our tested dataset contained less than a dozen events, and a broader range of data would clarify if there are certain types of events where one method may be superior. More research still needs to be done in this area, as it remains a pressing concern without a clear implementation at full scale.
Acknowledgments
===============
This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA). The views, opinions and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government. The paper is approved for public release, distribution unlimited.
**Michael Goebel** received his B.S. and M.S. degrees in Electrical Engineering from Binghamton University in 2016 and 2017. He is currently a PhD student in Electrical Engineering at University of California Santa Barbara.
**Arjuna Flenner** received his Ph.D. in Physics at the University of Missouri-Columbia located in Columbia MO in the year 2004. His major emphasis was mathematical Physics. After obtaining his Ph.D., Arjuna Flenner obtained a job as a research physicist for NAVAIR at China Lake CA. He won the 2013 Dr. Delores M. Etter Navy Scientist and Engineer award for his work on Machine Learning.
**Lakshmanan Nataraj** received his B.E degree in Electronics and Communications Engineering from Sri Venkateswara College of Engineering, Anna University in 2007, and the Ph.D. degree in the Electrical and Computer Engineering from the University of California, Santa Barbara in 2015. He is currently a Senior Research Staff Member at Mayachitra Inc., Santa Barbara, CA. His research interests include malware analysis and image forensics.
**B. S. Manjunath** (F’05) received the Ph.D. degree in Electrical Engineering from the University of Southern California in 1991. He is currently a Distinguished Professor at the ECE Department at the University of California at Santa Barbara. He has co-authored about 300 peer-reviewed articles. His current research interests include image processing, computer vision and biomedical image analysis.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents MAMoC, a framework which brings together a diverse range of infrastructure types including mobile devices, cloudlets, and remote cloud resources under one unified API. MAMoC allows mobile applications to leverage the power of multiple offloading destinations. MAMoC’s intelligent offloading decision engine adapts to the contextual changes in this heterogeneous environment, in order to reduce the overall runtime for both single-site and multi-site offloading scenarios. MAMoC is evaluated through a set of offloading experiments, which evaluate the performance of our offloading decision engine. The results show that offloading computation using our framework can reduce the overall task completion time for both single-site and multi-site offloading scenarios.'
author:
-
-
bibliography:
- 'IEEEabrv.bib'
- 'bare\_conf.bib'
title: 'MAMoC: Multisite Adaptive Offloading Framework for Mobile Cloud Applications'
---
Introduction
============
With the sheer ubiquity of powerful mobile devices, it is now possible to harness the capabilities of these mobile devices in order to execute, and offload compute-intensive applications. Most of the modern mobile devices on the market have quad-core and octa-core CPUs, and mobile devices of today are more computationally powerful than the PCs of the last decade [@6726211].
The concept of using cloud hosted infrastructure as a means to overcome the resource-constraints of mobile devices is known as *Mobile Cloud Computing (MCC)*, and allows applications to run partially on the device, and partially on a remote cloud instance, thereby overcoming any device-specific resource constraints. However, as smart phones and tablets gain more CPU power and longer battery life, the meaning of MCC gradually changes. Instead of being fully dependent on the cloud, a number of nearby devices can be used to coordinate and distribute content and resources in a decentralized manner; this is known as *Mobile Ad hoc Cloud Computing* [@WCM:WCM2709].
Although MMC has helped many application developers to overcome the limited resources of mobile devices, it has also created a new set of challenges, such as the possibility of high network latencies and low bandwidth availability between the mobile device and the cloud. Other research efforts investigated frameworks which allow the cloud to move closer to the user in the form of cloudlets [@5280678] [@Verbelen:2012:CBC:2307849.2307858]. Cloudlets are trusted, resource-rich computers that are connected to the Internet and available to nearby mobile devices. Other approaches include a group of nearby mobile devices to leverage lower end devices thus the formation of a local mobile cloud [@Shi:2012:SER:2248371.2248394] also referred as *Mobile Device Clouds* [@6753815] and *Mobile Edge Clouds* [@fernando2016computing]. Mobile devices with less computational power and lower battery life can be empowered by the nearby mobile devices to run resource-intensive applications. Therefore, more efficient and reliable methodologies need to be explored for resource-hungry and real-time applications such as face recognition, data-intensive, and augmented reality mobile applications.
The heterogeneous mobile cloud environment contains different types of computing resources such as remote clouds, cloudlets, and mobile devices in the vicinity that can be utilized to offload mobile tasks. Heterogeneity in mobile devices includes software, hardware, and technology variations. In MCC, providing collaboration among various mobile and cloud nodes with different interfaces is a significant matter. Dynamic environmental changes is one of the challenges facing the offloading decision making in mobile cloud applications. Mobile Cloud frameworks need to adapt to these changes for efficient task partitioning and high QoS of the mobile applications running on end user devices. Existing mobile computation offloading frameworks lack the automated transparency feature so that the surrounding devices can be detected and the computation offloading take place in a seamless manner [@sanaei2014heterogeneity].
This paper presents *Multisite Adaptive Mobile Cloud (MAMoC)*, a unified framework which allows each mobile device within the shared environment to intelligently offload its computation to other external platforms. For the individual mobile devices, it is important to make the offloading decision based on network conditions, load of other machines, and mobile device’s own constraints (e.g., mobility and battery). Moreover, to achieve a global optimal task completion time for tasks from all the mobile devices, it is necessary to devise a task scheduling solution that schedules offloaded tasks in real time. To achieve the vision of mobile computing among heterogeneous converged networks and computing devices, designing resource-efficient environment-aware applications is an essential need. The offloading decision engine needs to adapt to the dynamic changes in both the host device and connected nearby and remote devices. Changes in hardware such as CPU workload, available memory, and battery state and level need to be carefully monitored.
The main contributions of this work are:
1. Leveraging mobile applications running on lower end devices with nearby mobile devices, cloudlets, and remote cloud resources using a single framework.
2. An API on top of our framework, which allows application developers to use it as a simple programming model to build mobile cloud applications and abstract complex underlying heterogeneous technologies.
3. Through a real-world evaluation, we demonstrate that it is possible to speedup computation of mobile devices by using the adaptive offloading algorithm in both single-site and multi-site offloading scenarios.
Related Work
============
Mobile Computation offloading transfers processing from the mobile device to other service providers. Mobile application is partitioned and analyzed so that the most computational expensive operations at code level can be identified and offloaded [@Kumar:2013:SCO:2431282.2431342]. The objective is to improve the computation performance, enable advanced functionality, and preserve scarce resources. In the mobile-to-cloud offloading model, most challenges arise from partitioning the mobile application code to remote and local tasks based on the dependencies of each task. As such, the current solutions can be categorized by partitioning technique into static and dynamic. The authors at [@Zhang2011] produce an elastic application model in the form of weblets which can be platform independent or dependent. The decision of offloading is based on contextual components stored in the cloud including device status (CPU load, battery level), performance measures of the application for quality of experience and user preferences. By doing so, the application model supports multiple running modes: power-saving mode, high speed mode, low cost mode or offline mode. Mobile offloading overcomes the resource limitations of lower end devices by splitting resource-intensive tasks and allocating subtasks to other resource-rich devices. Offloading may be performed at different granularities ranging from methods and individual tasks [@Zhang2011] to applications [@Giurgiu:2009] and virtual machines [@Chun:2011:CEE:1966445.1966473].
The mobile cloud framework developed in [@Huerta-Canepa:2010:VCC:1810931.1810937] uses the same interface of existing cloud APIs for the collocated virtual computing providers. This allows a seamless integration with the existing cloud infrastructures. On the other hand, Cuckoo framework [@kemp2010cuckoo] uses the native Android partitioning to separate the user interface from the background computational code. This eases the design and implementation of MCC applications, as mobile application developers do not require any cloud computing knowledge, such as integrating with offloading APIs. Most existing MCC proposals concentrate on single-site offloading [@Cuervo:2010:MMS:1814433.1814441] i.e., offloading application’s parts from the mobile device to a single server. However, as the number of surrounding devices and cloud computing and storage increases, it is more common that an application can be executed on multiple servers [@Sinha2011TechniquesFF]. It is shown that we can obtain better performance from multisite offloading. Therefore, multisite offloading is considered as a generally realistic model in this work. However, making decision for multisite offloading problem is an NP-complete problem, and hence, obtaining the optimal solution is time consuming. Hence, we use a simple near optimal decision algorithm to find the best-possible partitioning for offloading to multisite clouds/servers.
We have shown in our previous work that offloading tasks from lower end devices to nearby devices can save up to 50% in computation performance [@sulaiman2016task]. In this work, we propose the creation of a hybrid mobile cloud framework in which the resources are nearby mobile, immobile (cloudlets) and remote cloud servers. MAMoC focuses on reducing the total task computation time and the overall consumed power of the interconnected devices. Similar to [@kemp2010cuckoo], we rule in favor of injecting the offloading code transparently to the developer. By doing so, we preserve the mobile application development process, as the developer is unaware of the underlying mechanisms and is only left with the possibility of giving hints about code that may be offloaded. Such a development model encourages the decoupling of components and the modularity of mobile applications.
![Heterogeneous Multisite MCC Architecture[]{data-label="MCC-Architecture"}](MEC-Architecture.png){width="\linewidth"}
System Design
=============
Our objectives for MAMoC include improving the running time of the compute-intensive tasks on the mobile devices as well as saving energy and bandwidth by computation offloading. MAMoC is designed to allow mobile devices to discover other surrounding devices over standard Local Area Network in infrastructural Wi-Fi using an access point, peer-to-peer Wi-Fi, and Bluetooth personal area networks as shown in Figure \[MCC-Architecture\]. The main goals of the framework are to allow mobile application developers to achieve a transparent automated offloading to multiple destination clouds (mobile clouds and public clouds) and device dynamic changes over the lifecycle of execution of an application.
Components
----------
### Service Discovery
After the framework is initialized, service discovery is performed. Each mobile device can advertise services and discover what services other nearby devices on the local network are offering. A browser object in a host device searches for peers which have an advertiser object. This can be done using infrastructural Wi-Fi, where the devices are connected to the same Access Point, Peer-to-peer Wi-Fi, or Bluetooth wireless technologies. An offloader device can also scan the local network for cloudlet servers and open a TCP socket communication channel for offloading requests. A standard user interface is developed to be used by the applications for service discovery and managing device connections as shown in Figure \[service\_discovery\].
The mobile ad hoc cloud communication technique over LAN is based on Zero Configuration Network [@cheshire2006zero]. MultiPeer Connectivity (MPC) library [@applempc] uses Zero Configuration Network technology to allow application developers to form short range sessions between nearby devices. The devices do not need a preconfigured network to exchange data with each other. There is no configuration needed because they can discover each other via multicast DNS (mDNS) [@cheshire2013multicast]. mDNS is a service that resolves host names on a local network without the use of a central domain name server. After the connection is established between two devices, data can be exchanged between them in either reliable mode over TCP or unreliable mode over UDP. The reliable mode guarantees the delivery and avoids out of order packages so we have used this mode throughout our framework data transfers.
![Service Discovery[]{data-label="service_discovery"}](MECFrameworkDiscovery.png){width="2.5in"}
### Device Profiler
This component collects real time information of the devices including hardware, software, and networking related context information. This allows the framework to have sufficient information about the connected devices. This eases the process of offloading decision making. Once a new device joins the framework, it goes through two profiling procedures. First, a collection of contextual information of the device is retrieved by the framework including hardware information such as number of CPU cores and processing speed, the type of the network (Wi-Fi or cellular), and battery level and state (whether it is charging or not). Second, a workload is sent over to the device after a successful connection establishment. The makespan result of running the workload is sent back to the framework. The collected information in both phases is then disseminated to all the connected devices in the framework to be used by the decision engine in the offloading process.
### Offloading Decision Engine
The engine is fed with the metrics from device profiler component and it is used by the framework to check whether to offload a task to other devices or execute it locally. This component is only used in the partial offloading execution modes. Unlike full offloading, the framework needs to decide the candidate destinations for offloading the task, partition the task into a number of subtasks, and distribute them.
Offloading Models and Algorithms
--------------------------------
When we model mobile computation offloading problems, we need to consider the time needed to execute computation locally, time needed to execute computation remotely (excluding the cost of data transfer), and time needed to move input data and results back and forth between the local and remote computers. We also need to collect the device context information in our preprocessing phase to model the offloading decision making process including the CPU power, memory, battery state and level of the connected devices.
[ll]{} & [**Description**]{}\
$ B_{man} $ & Benchmark workload result of mandelbrot set\
$ B_{ftt} $ & Benchmark workload result of FFT\
$ B $ & Average Benchmark workload result\
$ B_{mob} $ & Benchmark score of a mobile device\
$ B_{clet} $ & Benchmark score of a cloudlet\
$ B_{c} $ & Benchmark score of a remote cloud\
$ BL_{mob} $ & Battery level of a mobile device\
$ M_{mob} $ & Available memory of a mobile device\
$ M_{clet} $ & Available memory of a cloudlet\
$ M_{c} $ & Available memory of a remote cloud\
$ OS_{mob} $ & Offloading score of a mobile device\
$ OS_{clet} $ & Offloading score of a cloudlet\
$ OS_{c} $ & Offloading score of a remote cloud\
$ P_{mob} $ & Computation speed of a mobile device\
$ P_{clet} $ & Computation speed of a cloudlet\
$ P_{c} $ & Computation speed of a remote cloud\
$ RTT_{mob} $ & Network overhead and data transfer to mobile device\
$ RTT_{clet} $ & Network overhead and data transfer to cloudlet\
$ RTT_{c} $ & Network overhead and data transfer to cloud\
$ TOS $ & Total offloading score of all the connected nodes\
The benchmark score will be calculated after the workloads are sent to the devices and the execution results are received. The quicker the CPU completes the tests, the higher the benchmark score. The workloads measure the instruction performance of the device by performing processor-intensive tasks that make heavy use of integer instructions. Initially, we create two types of workloads: compute-bound and memory-bound. Mandelbrot set [@fernando2016computing] of an 800x800 pixels is used for the first type. The Fast Fourier Transform (FFT) [@katoh2002mafft] is used as a memory heavy workload. The workloads are executed 5 times and the average runtime score in GFlops is returned to MAMoC. Finally, the Benchmarking score of the device is calculated using equation \[equation1\]:
$$B = (B_{man} + B_{ftt}) / 2
\label{equation1}$$
The computation power and memory of connected nearby mobile devices are initially collected. If the mobile device is not currently charging, then the offloading score is deducted by the amount of the used battery level of the device. The offloading score is then calculated using the following equation:
$$OS_{mob} = (B_{mob} + P_{mob} + M_{mob}) - RTT_{mob} - (100 - BL_{mob})$$
Similar to mobile ad hoc cloud model, cloudlet and remote cloud modelling is a summation of their respective benchmark score, computation speed, and available memory subtracted by the data transfer cost.
$$OS_{clet} = (B_{clet} + P_{clet} + M_{clet}) - RTT_{clet}$$
$$OS_{c} = (B_{c} + P_{c} + M_{c}) - RTT_{c}$$
Given the offloading scores for all the offloadee device candidates, we calculate the total offloading score for any particular task execution.
$$TOS = |OS_{mob}| + |OS_{clet}| + |OS_{c}|$$
MAMoC collects the offloading scores of the local device running the mobile application and all the connected service providers. Algorithm 1 shows the process of collecting individual offloading scores calculated and received earlier to generate a dictionary of nodes and their corresponding offloading scores.
connectedNodes nodeScores nodeScores = \[:\] selfNode = getSelfNode() localScore = getScore() nodeScores.add(selfNode, localScore) score = node.getScore() nodeScores.add(node,score)\
nodeScores
The node scores will then be sent to task partitioning algorithm to calculate the final task partitioning percentage for any given task.
nodeScores, totalScore partitioningResult partitioningResult = \[:\] partition = (score / totalScore) \* 100 partitioningResult.add(node,partition)\
partitioningResult
Development Phase
-----------------
We have used Swift [^1] to implement both the client (iOS) and server (Linux) components of MAMoC. Swift is open sourced by Apple on December 2016. This has enabled developers to write Swift applications on broader range of platforms. To the best of our knowledge, this is the first complete Swift and iOS based framework in the mobile cloud computing research literature.
Containers are utilized to host the server components of MAMoC. Linux Containers (LXC) is a virtualization method for running multiple isolated Linux systems on a single machine. Docker[^2] extends LXC to automate the deployment of applications inside software containers. We have used an already developed Swift Docker image to implement a server-side swift application to accept incoming requests from the mobile devices. The developed container provides an environment which is ready to be customised for other mobile applications. It provides a feature-rich yet lightweight execution environment for offloaded tasks.
![Container vs. native benchmarking[]{data-label="containerBenchmarking"}](ContainerBenchmarking.png){width="3.5in"}
We have run a couple of compute-intensive workloads to experiment the speed of computation on Docker containers as well as on native platforms. As the results show in \[containerBenchmarking\], the computation speed of containers are close to native platform performance within 1% region [@7095802] [@7484184]. To ease the creation of the container, we provide a Dockerfile that can automate the process of creating a container for the server-side part of MAMoC. The container is preconfigured with the necessary build environment for handling the client requests.
MAMoC allows other mobile application developers to adapt it for their own compute-intensive mobile applications. A developer must follow the following steps to configure an application to use our framework:
1. Add MAMoC to the mobile application: This process can be done manually by adding the files to the application directory or automatically using CocoaPods[^3] which is a dependency manager for Swift and Objective-C projects.
2. Define the task and its parameters: In this step, the developer has to override *initTask* method to provide any user input or additional data to be processed by the framework. In our text search example, the parameters are the text file and the search keyword which is entered by the user. In a face recognition application, the parameter would be an image of the face and so on.
3. Define the offloading sites: Even though MAMoC is designed to offload tasks to multiple sites by default, the developer can define where to execute the task. There are four choices to choose from: nearby mobile devices, cloudlet servers (local immobile), remote cloud servers, or automatic. The automatic option depends on the outcome of the models in the device profiler component and how the tasks are partitioned to target offloadees by the offloading decision engine.
4. Execute the job: the subtasks will be distributed accordingly to be processed in parallel on the target offloadees. The results are returned and merged together in the offloader device and presented to the user.
5. Set the timeout for the task reprocessing (optional): For the tasks which are not returned successfully by the resource providers, the task will be reprocessed locally in the device once the timeout duration is reached. The default value for this is 10 seconds. This adds fault tolerance feature to the applications developed on top of MAMoC.
The source code of MAMoC and a short documentation for setting up the different components in the framework is publicly available online at *https://github.com/dawand/MAMoC*
Experimentation
===============
[|c|c|c|c|]{} & [**Benchmark**]{} & [**CPU**]{} & [**RAM**]{}\
& [**(in GFlops)**]{} & [**(in GHz)**]{} & [**(in GB)**]{}\
Mobile (small) & 1.09 & 1.3 (Dual) & 1\
Mobile (medium) & 1.24 & 1.4 (Dual) & 1\
Cloudlet & 2.56 & 2.5 (Quad) & 16\
Cloud (small) & 2.32 & 2.4 (Single) & 1\
Cloud (medium) & 2.94 & 2.8 (Quad) & 7.5\
Cloud (large) & 3.02 & 2.8 (Octa) & 15\
We have developed a mobile application to test the performance of the framework and showcase the different execution scenarios. We measure total completion time for the application in different settings. The application is executed in four different modes: local execution, nearby mobile devices, cloudlet, and remote cloud (with three different servers). We use four offloading scenarios: a full offloading and three types of partial offloading (workload sharing in a parallel manner) with different configurations. The application contains a Knuth-Morris-Pratt searching algorithm [@knuth1977fast] to be performed on three different size text files. The large text file consists of 1,095,649 words, the medium text file contains 316,323 words, and the small text file contains 39,799 words. The files are stored in the mobile application running on the offloader device. The files need to be distributed through a wireless medium to the offloadees before the keyword search is conducted in the destination. Each execution was repeated ten times, such that averages could be calculated for more accurate results.
Our experimental testbed consists of two mobile devices (an iPhone 5 and an iPhone 6), one cloudlet, and three remote cloud instance types. The hardware specifications are shown in Table \[testbed\]. For the remote cloud instances, we have used three different Amazon Web Services instance types: small (t2.micro), medium (c4.xLarge), and large (c3.2xlarge). The two phones will be running the mobile application while cloudlet and remote cloud instances will be running the container described earlier.
In all our experiments, mobile (small) is the offloader and the rest of the nodes are service providers (offloadees). The offloading scenarios are:
- Full offloading: In full offloading mode, the execution is performed in the offloadee and the final result is returned to the offloader. On the other hand, partial offloading mode only sends a part of the execution over to the offloadee and performs the rest of the execution itself as we will observe in the other two scenarios. After the results are received, they are merged and stored in the offloader.
\[fulloffloading\] {width="\linewidth" height="2in"}
Our framework is specifically designed to support partial offloading. Nonetheless, we wanted to observe the completion times of full offloading executions and compare them to the decision engine results. We performed a full offloading experimentation on the offloadees separately using the three text files. Figure 4 shows the total completion time of running the application with different text file sizes. As we will see later, most of the completion time is the communication overhead that occurs during the transfer of the necessary data (the text file content) from the mobile device to the offloadees.
- Partial offloading: we perform our first set of partial offloading experiments with no help from MAMoC. The tasks are equally distributed among the connected nodes. In other words, if there is only one available offloadee, the workload is divided into two equivalent halves and distributed to them. Both devices then execute the workload in a parallel fashion. The local result and the result returned from the offloadee are then merged and stored locally. The results of running the same set of workloads as previous experiment are displayed in Figure 5.
\[partialoffloading5050\] {width="\linewidth" height="2in"}
- Partial offloading with decision engine enabled: the decision engine uses the models presented earlier to calculate the percentage of the task which should be offloaded to any particular offloadee. The offloading scores of the offloader and the offloadees are investigated by the decision engine for the task partitioning process. In the case of offloading score being less than zero, no computation is offloaded to that offloadee. We wanted to observe the performance of our decision engine in both single-site and multi-site offloading scenarios. Figure 6 shows the single-site offloading scenario where the task is meant to be executed locally and a single offloadee. The task partitioning percentage and completion times of partitioned tasks for multi-site offloading scenario are displayed in Figure 7. It is worth noting that we have only used the large instance type of the remote cloud along with a cloudlet and a nearby mobile device as offloadees in multi-site offloading scenario. The complete set of results of all the experiments are shown in Table III.
\[partialoffloadingsingle\] {width="\linewidth" height="2in"}
\[partialoffloadingmulti\]
[2]{} {width="\linewidth"}
{width="\linewidth"}
\[alltests\]
[2]{} {width="\textwidth"}
Results and discussion
======================
It is already shown in the literature that offloading does not always benefit the lower end devices [@5445167]. In our first execution scenario, we send the text file to the destination and perform the search operation in the corresponding computation device. Despite few millisecond performance gains in the case of small text file, local execution is preferred to full offloading for medium and large text files. The partial offloading with equal partitioning of tasks among the local and external devices perform better in terms of reducing the overall network overhead occurrence in the previous execution scenario.
The single-site partial offloading produces better results than equal task distribution scenario in all of the offloading modes. It even produces a lower overall completion time for the cloudlet than the multi-site offloading scenario. This is due to the limited number of nearby mobile devices and cloudlets that are used in our testbed. Even though cloudlet-only full offloading performed better in both small and medium text file size execution scenarios, large text file size offloading was more efficient in MAMoC (9.074 vs. 5.667 seconds) as a larger portion of execution is performed locally. This is mainly due to the delay caused by data transfer in full offloading execution. For the large text file size offloading scenarios, the network overhead occupies a much larger portion of the total completion time than the time taken for the task execution. This shows a clear advantage of local execution over offloading when the needed computation is not as much. In short, the more computation and less data transfer is needed, offloading has more advantage.
Since we use Wi-Fi in the evaluation, the time of sending files and receiving results has an impact on the overall performance, but if cellular networks or Bluetooth are used, the offloading time will surely increase.
Conclusions and Future Work
===========================
In this paper, we presented an adaptive multisite offloading framework that takes into consideration dynamic context changes in MCC environment and offloads computation to multiple offloadees including nearby mobile devices, cloudlets, and remote cloud servers. We evaluated the proposed framework, and results showed that it can provide suitable offloading decisions based on the current context of the local device and the external platforms. We developed a text search application and conducted experiments with three different text file sizes on different offloading scenarios. Our results present different insights into the factors that affect the offloading decision. Our future work includes increasing context parameters and the number of nearby devices and cloudlets in our experimental testbed. Our future experiments will not be confined to Wi-Fi and would explore other protocols such as Bluetooth Low Energy for nearby device-to-device communications and 4G for the mobile device and remote cloud communications. Moreover, we have built our framework with the assumption that the mobile user (offloader) stays in the same zone within the computation offloading duration. This cannot be guaranteed if the user moves away before the result from the destination is returned. The experiments in this paper were performed in a controlled setting. Enabling user mobility by live migrating the host mobile cloud containers is another future endeavor. An active research challenge is how to implement live migration in mobile cloud systems.
References
==========
[^1]: https://swift.org
[^2]: https://www.docker.com/
[^3]: https://cocoapods.org/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present diagrammatic approximations to the spin dynamics of the 2d Heisenberg antiferromagnet for all temperatures, employing an auxiliary-fermion representation. The projection onto the physical subspace is effected by introducing an imaginary-valued chemical potential as proposed by Popov and Fedotov. The method requires that the fermion number at any lattice site is strictly conserved. We compare results obtained within a self-consistent approximation using two different auxiliary-particle projection schemes, (1) exact and (2) on average. Significant differences between the two are found at higher temperatures, whereas in the limit of zero temperature (approaching the magnetically ordered ground state) identical results emerge from (1) and (2), providing the qualitatively correct dynamical scaling behavior. An interpretation of these findings is given. We also present in some detail the derivation of the approximation, which goes far beyond mean-field theory and is formulated in terms of complex-valued spectral functions of auxiliary fermions.'
author:
- Jan Brinckmann
- 'Peter W[ö]{}lfle'
date: 22 March 2008
title: 'Diagrammatic approximations for the 2d quantum antiferromagnet: exact projection of auxiliary fermions'
---
Introduction and review {#sec-intro}
=======================
Auxiliary particles are a widely used tool in the theory of correlated electron systems. The principal difficulty in the treatment of these systems is the strong Coulomb repulsion $U$ for two electrons on the same localized orbital, usually of $d$ or $f$ character. In effective model Hamiltonians like the Heisenberg, $t$-$J$, or Kondo model the large $U$ leads to a Gutzwiller projection onto the quantum-mechanical subspace, where none of the $d$- or $f$-orbitals may contain more than one electron at a time.
In this paper we will focus on the spin-1/2 quantum Heisenberg antiferromagnet in two spatial dimensions (2D) on a square lattice[@man91], $$\label{eqn-hbergham}
H= \sum_{<i, j>}J\,\mathbf{S}_i \mathbf{S}_j
\;\;,\;\;\;
J> 0$$ The sum covers all nearest-neighbor pairs . The model (\[eqn-hbergham\]) may be obtained from the single-band Hubbard model for large $U$ with nearest-neighbor hopping amplitude $t$ (leading to $J= 2t^2/U$) in the limit of a half-filled band with one electron per site[@grosjoyntrice87]. It represents the simplest low-energy model for two-dimensional Mott insulators, in particular the CuO-planes in the undoped parent compounds for high-temperature superconductors[@hay90; @keim92; @kim01].
The restriction on states with no doubly occupied sites is reflected in the non-canonical commutation relations of spin operators, $$\label{eqn-comm}
{}[\,S^x\,, S^y\,{}]
= i S^z \ne \textrm{c-number}.$$ For an analytical approach to the model, Eq.(\[eqn-comm\]) poses a severe difficulty, since the standard Wick’s theorem and many-body techniques cannot be applied[@negorl; @grekei81; @bic87; @russenbuch]. A convenient way of circumventing this difficulty is to represent the spin operators in terms of canonical auxiliary-particle operators, of either fermionic[@abr65] or bosonic[@aaa88] character. The cost of this concept is the extension of the Hilbert space into unphysical sectors. These unphysical states have to be removed by imposing a constraint. In this work we use auxiliary fermions, $$\label{eqn-auxferm}
S^\mu_i
= \frac{1}{2}
\sum_{\alpha, \bar{\alpha}}
f^\dagger_{i\alpha}
(\sigma^\mu)^{\alpha \bar{\alpha}}
f_{i\bar{\alpha}}
\;\;,\;\;\;
Q_i= \sum_{\alpha}
f^\dagger_{i\alpha} f_{i\alpha}
= 1 \;\;,$$ $\sigma^\mu$, $\mu= x, y, z$ are the Pauli matrices, and $\alpha=
\uparrow, \downarrow$ is the fermion-spin index. Here and in the following we let $\hbar\equiv 1$. The representation (\[eqn-auxferm\]) fulfills the commutation relations (\[eqn-comm\]). The Fock space of the auxiliary fermions $f_{i\alpha}$ is spanned by the states
\[eqn-states\] $$\begin{aligned}
\label{eqn-physstates}
\textrm{physical:} & \ &
|\uparrow\rangle= f^\dagger_{i\uparrow}|0\rangle \;\;,\;\;\;
|\downarrow\rangle= f^\dagger_{i\downarrow}|0\rangle
\\
\label{eqn-unphysstates}
\textrm{\emph{un}physical:} & \ &
|0\rangle \;\;,\;\;\;
|2\rangle= f^\dagger_{i\uparrow} f^\dagger_{i\downarrow}|0\rangle
\\ \nonumber\end{aligned}$$
where $|0\rangle$ denotes the vacuum, $f_{i\alpha}|0\rangle\equiv
0$.
The projection onto the physical subspace can be performed in several ways[@bic87], and for impurity models like the Kondo model or the single-impurity Anderson model the projection of auxiliary particles is a standard technique[@abrmig70; @biccoxwil87]. (For the latter model, auxiliary bosons have to be added to the scheme[@col84].) However, in lattice models like Eq.(\[eqn-hbergham\]) the constraint $Q_i= 1$ has to be enforced on each site $i$ independently. This leads to the so-called excluded-volume problem[@bro61; @grekei81] and prohibits an infinite-order resummation of the perturbation series in $J$. In the limit of high spatial dimension the problem can be evaded within the (extended) dynamical mean-field theory. Here the infinite lattice system is approximately mapped onto a single site[@kur85; @metvol89; @sismith96] or a small cluster[@note-dca] coupled to a bath.
An alternative path starts from a mean-field-like treatment of the constraint, where the auxiliary charge $Q_i= 1$ is fixed merely on the thermal average[@newrea87], $$\label{eqn-meancons}
Q_i \to \langle Q_i\rangle
= \langle Q_1\rangle
= \sum_{\alpha}\langle f^\dagger_{1\alpha} f_{1\alpha}\rangle
= 1\;.$$ This condition is introduced into the Hamiltonian (\[eqn-hbergham\]), (\[eqn-auxferm\]) through a chemical potential $\mu^f$ for the auxiliary fermions. Due to the particle–hole symmetry of (\[eqn-hbergham\]) we have $\mu^f= 0$. The approximation (\[eqn-meancons\]) is of great advantage, since now the perturbation theory in $J$ starts from non-interacting fermions, and we can make use of the standard Feynman-diagram techniques. Eq.(\[eqn-meancons\]) is also the starting point for numerous mean-field theories of correlated electron systems. An improvement of the mean-field-like constraint (\[eqn-meancons\]) in a perturbative fashion has frequently been made by generalizing $\mu^f$ to a fluctuating Lagrange multiplier (see, e.g., Ref.).
*The Popov–Fedotov approach:* The method proposed by Popov and Fedotov[@popfed88] enables one to enforce the auxiliary-particle constraint *exactly* within an analytical calculation for the infinite system[@grosjohn90; @boukis99; @kisopp00; @dilric06]. The approach starts from a “grand-canonical ensemble”, $$\label{eqn-grand}
H\to H^\ppv=
H - i\frac{\pi}{2}T\sum_{i}Q_i \;\;,\;\;\;
Q_i= \sum_{\alpha}f^\dagger_{i\alpha} f_{i\alpha} \;\;,\;\;\;$$ with a homogeneous, *imaginary-valued* chemical potential $$\label{eqn-ppv}
\mu^f= i\frac{\pi}{2}k_BT\equiv i\frac{\pi}{2}T\;.$$ $H$ is the spin Hamiltonian (\[eqn-hbergham\]), written in terms of the auxiliary-fermion operators (\[eqn-auxferm\]).
There are two main requirements for the Popov–Fedotov method to work: The first is the conservation of the auxiliary charge $Q_i$ on each lattice site, $$\label{eqn-qcons}
{}[\,Q_i\,, H\,{}]= 0 \;\;,\;\;\;
i= 1, 2, \ldots, N_L \;\;,$$ where $N_L$ denotes the number of sites (i.e., spins) in the system (\[eqn-hbergham\]). Since also ${}[\,Q_i\,, Q_j\,{}]= 0$, the eigenstates of $H$ and $H^\ppv$ in the Fock space of the fermions can be specified by some auxiliary-charge configuration $$\label{eqn-config}
c_Q= (Q_1, Q_2, \ldots, Q_{N_L}) \;\;,\;\;\;
Q_i\in \{0, 1, 2\}\;.$$ Physical states belong to the subspace with the configuration $$\label{eqn-physconfig}
c_Q^{phys}= (1, 1, \ldots, 1)\;.$$ For a given configuration $c_Q$ the Hamiltonian (\[eqn-hbergham\]) has a specific set of eigenstates with quantum numbers $n_Q$ and energies $E(c_Q, n_Q)$, i.e., Schr[ö]{}dinger’s equation reads $$\label{eqn-sgl}
H|c_Q, n_Q\rangle
= E(c_Q, n_Q)|c_Q, n_Q\rangle\;.$$ Consider now the partition function $Z^\ppv$ for the grand-canonical Hamiltonian (\[eqn-grand\]), $$\begin{aligned}
Z^\ppv
& = & \label{eqn-zppdef}
\textrm{Tr}^f{}[\,e^{-\beta H^\ppv}\,{}]
\\ \nonumber
& = & \sum_{c_Q}\sum_{n_Q}
\langle c_Q, n_Q|\,
e^{-\beta H^\ppv}\,
|c_Q, n_Q\rangle\;\;,\end{aligned}$$ $\textrm{Tr}^f$ denotes the trace in the enlarged Hilbert space of the auxiliary fermions, and $\beta= 1/k_B T\equiv 1/T$. With Eqs.(\[eqn-sgl\]) and (\[eqn-grand\]) it becomes $$\begin{aligned}
Z^\ppv
& = & \sum_{Q_1, \ldots, Q_{N_L}= 0}^2\;
\sum_{n_Q}\cdot
\\
& & \mbox{}\cdot
\langle c_Q, n_Q|\,
e^{-\beta E(c_Q, n_Q)}\,
e^{i\frac{\pi}{2}Q_1}\cdots
e^{i\frac{\pi}{2}Q_{N_L}}\,
|c_Q, n_Q\rangle\end{aligned}$$
In addition to the $Q_i$-conservation, the Popov–Fedotov method requires that the Hamiltonian and the operators appearing in physical (i.e., observable) correlation functions destruct the unphysical states $|0\rangle$, $|2\rangle$. In the present case, Hamiltonian and correlation functions are composites of spin operators, and we have, using Eqs.(\[eqn-auxferm\]) and (\[eqn-unphysstates\]), $$\label{eqn-destruct}
S_i^\mu|0\rangle= 0 \;\;,\;\;\;
S_i^\mu|2\rangle= 0\;.$$ Consider an arbitrary site $l$ with an unphysical auxiliary charge $Q_l\ne 1$. From Eq.(\[eqn-destruct\]) it follows $$\label{eqn-equalen}
Q_l\ne 1\,: \quad
\left.E(c_Q, n_Q)\right|_{Q_l= 0}
= \left.E(c_Q, n_Q)\right|_{Q_l= 2}\;\;,$$ that is, the spin at site $l$ seems to be removed from the Hamiltonian for all states $|c_Q, n_Q\rangle$ with $Q_l= 0$ or $2$. Accordingly, the contribution from $Q_l= 0, 2$ to $Z^\ppv$ is proportional to $$\label{eqn-cancel}
\sum_{Q_l= 0, 2}\,e^{i\frac{\pi}{2}Q_l}
= (1 + e^{i\pi})
= 0\;.$$ In that way, the unphysical contributions from all sites $l= 1,
\ldots, N_L$ cancel in the grand-canonical partition function, i.e., only the physical charge configuration $c_Q^{phys}$, Eq.(\[eqn-physconfig\]), survives in $Z^\ppv$: $$Z^\ppv
= (i)^{N_L}
\sum_n
\langle n|\,
e^{-\beta E_n}\,
| n\rangle\;.$$ Here $|n\rangle$ and $E_n$ denote the eigenstates and -energies of the Hamiltonian (\[eqn-hbergham\]) in the physical subspace, $$\label{eqn-physen}
\begin{array}[c]{c}
\displaystyle
|n\rangle= |c_Q^{phys}, n_Q\rangle \;\;,\;\;\;
E_n= E(c_Q^{phys}, n_Q)\;\;,
\\[2ex] \displaystyle
H|n\rangle = E_n|n\rangle\;.
\end{array}$$ Thus we end up with $$\label{eqn-physpart}
Z^\ppv = (i)^{N_L}\,Z\;\;,$$ i.e., up to a constant prefactor, the (canonical, physical) partition function $Z$ for the Heisenberg model (\[eqn-hbergham\]) is given by $Z^\ppv$.
The above argument, originally presented by Popov and Fedotov in Ref., is extended to Green’s functions in Appendix \[sec-app-greens\]. It is found that any correlation function of spin operators may be calculated from the grand-canonical Hamiltonian (\[eqn-grand\]). In particular, the imaginary time-ordered spin susceptibility $$\label{eqn-sus-def}
\chi^{\mu\bar{\mu}}_{ij}(\tau - \tau')
= \frac{1}{Z}
\mbox{Tr}{}[\,
e^{-\beta H}
\mathcal{T}_\tau\{S^\mu_i(\tau)
S^{\bar{\mu}}_j(\tau')\}\,{}]$$ can be obtained from $$\label{eqn-sus-ppv}
\chi^{\mu\bar{\mu}}_{ij}(\tau - \tau')
= \langle \mathcal{T}_\tau\{S^\mu_i(\tau)
S^{\bar{\mu}}_j(\tau')\}\rangle^\ppv$$ with $\mu, \bar{\mu}\in \{x, y, z\}$. The expectation value is calculated in the enlarged Fock space, $$\label{eqn-expect-ppv}
\langle \ldots\rangle^\ppv
= \frac{1}{Z^\ppv}
\textrm{Tr}^f{}[\,e^{-\beta H^\ppv}\,\ldots\,{}]\;\;,$$ with $Z^\ppv$ as defined in Eq.(\[eqn-zppdef\]) above. The “modified Heisenberg” picture[@negorl] for spin operators reads $$\label{eqn-taudep}
S^\mu_i(\tau)
= e^{\tau H}S^\mu_i e^{-\tau H}
= e^{\tau H^\ppv}S^\mu_i e^{-\tau H^\ppv}\;\;,$$ using again Eq.(\[eqn-destruct\]). Similarly, the local magnetization is given by $$\label{eqn-magn-ppv}
\langle S^\mu_i\rangle
= \frac{1}{Z}\textrm{Tr}{}[\,
e^{-\beta H} S^\mu_i\,{}]
= \langle S^\mu_i\rangle^\ppv\;.$$ It should be emphasized that expectation values of *un*physical operators are meaningless within the Popov–Fedotov scheme: E.g., the auxiliary-fermion charge $Q_i$ introduced in Eq.(\[eqn-auxferm\]) does not destruct the unphysical states (\[eqn-unphysstates\]), and one has $$\label{eqn-qexpect-ppv}
\langle Q_i\rangle
\boldsymbol{\ne} \langle Q_i\rangle^\ppv\;.$$ For the l.h.s. we know that $Q_i= 1$ in the physical Hilbert space. For the r.h.s., however, we obtain $\langle Q_i\rangle^\ppv= (1 +
i)$. The calculation is given in Appendix \[sec-app-greens\].
*Average projection:* For comparison, we also want to use the mean-field-like treatment of the auxiliary-fermion constraint mentioned below Eq.(\[eqn-meancons\]). Since the (real-valued) chemical potential added to the Hamiltonian (\[eqn-hbergham\]), (\[eqn-auxferm\]) turns out to be zero, due to the particle–hole symmetry of (\[eqn-hbergham\]), the calculation of a spin-correlation function or magnetization amounts to just enlarging the Hilbert space into the Fock space of the auxiliary fermions. The equivalents of the Eqs.(\[eqn-sus-ppv\]), (\[eqn-magn-ppv\]), and (\[eqn-qexpect-ppv\]) then read $$\label{eqn-sus-av}
\chi^{\mu\bar{\mu}}_{ij}(\tau - \tau')
\simeq \langle \mathcal{T}_\tau\{S^\mu_i(\tau)
S^{\bar{\mu}}_j(\tau')\}\rangle^\ave$$ and $$\label{eqn-magn-av}
\langle S^\mu_i\rangle
\simeq
\langle S^\mu_i\rangle^\ave\;\;,$$ whereas $$\label{eqn-qexpect-av}
\langle Q_i\rangle
= \langle Q_i\rangle^\ave
= 1$$ with $$\label{eqn-expect-av}
\langle\ldots\rangle^\av
= \frac{1}{Z^\av}
\textrm{Tr}^f{}[\,
e^{-\beta H}\ldots\,{}]
\;\;,\;\;\;
Z^\av
= \textrm{Tr}^f{}[\,e^{-\beta H}\,{}]\;.$$ The $\simeq$ sign in (\[eqn-sus-av\]), (\[eqn-magn-av\]) stands for the error introduced by the uncontrolled fluctuations of the fermion occupation numbers $Q_i$ into unphysical states. These fluctuations are absent in the Popov–Fedotov scheme.
In the following sections, results from the exact and the averaged constraint will be obtained within the same diagrammatic approximations. The effect of the constraint on physical quantities like the dynamical structure factor and magnetic transition temperature is going to be studied. It will turn out that at sufficiently low temperature the results for averaged and exactly treated constraint become equal. In addition we will show in some detail how a self-consistent approximation that goes far beyond mean-field theory, can be worked out within the Popov–Fedotov approach.
Effect of the constraint: simple approximations {#sec-simple}
===============================================
In order to compare results for the Heisenberg model (\[eqn-hbergham\]) from average projection, Eq.(\[eqn-meancons\]), and exact projection using the Popov–Fedotov scheme, Eq.(\[eqn-grand\]), we consider a more general grand-canonical Hamiltonian of auxiliary fermions, $\widetilde{H}= H - \mu^f\sum_i
Q_i$. With the model Hamiltonian (\[eqn-hbergham\]) written according to (\[eqn-auxferm\]), it reads $$\begin{aligned}
\label{eqn-allgrand}
\widetilde{H}
& = & -\mu^f\sum_{i}
f^\dagger_{i\alpha} f_{i\alpha} \;+\;
\\ \nonumber
& & \mbox{} +\;
\frac{1}{2}\sum_{i,j}J_{ij}
\frac{1}{4}
\boldsymbol{\sigma}^{\alpha\bar{\alpha}}
\boldsymbol{\sigma}^{\beta\bar{\beta}}
f^\dagger_{i\alpha} f^\dagger_{j\beta}
f_{j\bar{\beta}} f_{i\bar{\alpha}}\;.\end{aligned}$$ Sums over spin indices $\alpha, \bar{\alpha}, \beta, \bar{\beta}$ are implied. The antiferromagnetic coupling $J_{ij}$ is nonzero and equal to $J>0$, if $i,j$ are nearest neighbors. Depending on the projection method, the chemical potential $\mu^f$ takes the value $$\label{eqn-chem}
\begin{array}[c]{lcrcl}
\textrm{average projection:} & &
\mu^f & = & 0\;\;, \\[1ex]
\textrm{exact projection:} & &
\mu^f & = & \displaystyle
i\frac{\pi}{2}T\;.
\end{array}$$ In the case of exact projection, the Hamiltonian (\[eqn-allgrand\]) is no longer Hermitian. Nevertheless, physical quantities like the dynamical structure factor for spin excitations or the magnetization will come out real-valued.
Eq.(\[eqn-allgrand\]) represents a system of canonical fermions with a two-particle interaction $\sim J$ and may therefore be treated using standard Feynman-diagram techniques[@negorl]. The bare fermion-Green’s function, written as a matrix in spin space, reads $$\label{eqn-baregf}
\overline{G}^0_{ij}(i\omega)
= \delta_{ij}\frac{1}{i\omega + \mu^f}\;\;,$$ where $\omega= (2n + 1)\pi T$, $n\in\mathbb{Z}$ is a fermionic Matsubara frequency. For the case of exact projection, $\mu^f= i\pi
T/2$, there is some common practice[@popfed88; @boukis99; @kisopp00] to absorb $\mu^f$ into $i\omega$ and to re-define $\omega$ accordingly. Here we do not follow this line, but keep $\omega$ as introduced above (i.e., fermionic).
*Free spins ($J= 0$):* The simplest case is given by setting $J=
0$. Using Eq.(\[eqn-auxferm\]), the susceptibility (\[eqn-sus-def\]), to be calculated from Eq.(\[eqn-sus-ppv\]) or (\[eqn-sus-av\]), is then given by a simple bubble, $$\begin{aligned}
\chi_{ij}^{\mu\bar{\mu}}(i\nu)
& = & -\frac{1}{4}T\sum_{i\omega}
\textrm{Tr}^\sigma{}[\,\sigma^\mu
\overline{G}^0_{ij}(i\omega + i\nu)\,\sigma^{\bar{\mu}}\,
\overline{G}^0_{ji}(i\omega)\,{}]
\\
& = & \delta_{i,j}\,\delta_{\mu,\bar{\mu}}\,\delta_{\nu, 0}\,
\chi^0\end{aligned}$$ with, using Eq.(\[eqn-baregf\]), $$\chi^0
= \frac{1}{2T}f(-\mu^f) f(\mu^f)\;.$$ $\nu= 2n\pi T$ is a bosonic Matsubara frequency, $\textrm{Tr}^\sigma$ denotes a trace in spin space, and $f(x)= 1/(e^{x/T} + 1)$ is the Fermi function. Depending on the constraint method, the result is $$\label{eqn-chinul}
\begin{array}[c]{lclcl}
\textrm{average:} & & \displaystyle
f(0)= \frac{1}{2} & \;\Rightarrow\; & \displaystyle
\chi^0= \frac{1}{8T}\;\;,
\\[2ex]
\textrm{exact:} & & \displaystyle
f(\pm i\frac{\pi}{2}T)
= \frac{1}{1 \pm i} & \;\Rightarrow\; & \displaystyle
\chi^0= \frac{1}{4T}\;.
\end{array}$$ Two observations are in order: The imaginary chemical potential cancels out in the physical quantity $\chi^0$, and the result with and without use of the exact auxiliary-particle projection is qualitatively the same (Curie law). With merely average projection in effect ($\mu^f= 0$), the spin moment $S(S + 1)$ extracted from the Curie law $\chi^0(T)\propto S(S + 1)/T$ is reduced by a factor of 2 compared to the exact result. This is due to fluctuations of the fermion charge $Q_i$.
*Mean-field approximation:* The simplest approach to the interacting system $J> 0$ is the Hartree approximation, i.e., magnetic mean-field theory. This approximation does locally conserve the auxiliary-fermion charge $Q_i$. Dyson’s equation for the auxiliary fermion reads $$\label{eqn-fdyson}
\overline{G}_i(i\omega)
= {}[\,i\omega + \mu^f - \overline{\Sigma}_i(i\omega)\,{}]^{-1}\;\;,$$ and in Hartree approximation the self energy is independent of $i\omega$ and given by $$\label{eqn-hartree}
\overline{\Sigma}_i
= \sum_j J_{ij}\frac{1}{2}\boldsymbol{\sigma}
\underbrace{
\textrm{Tr}^\sigma{}[\,
\frac{1}{2}\boldsymbol{\sigma}
\sum_{i\bar{\omega}}
\overline{G}_j(i\bar{\omega})
e^{i\bar{\omega} 0_+}\,{}]
}_{\displaystyle = \langle \mathbf{S}_j\rangle}\;.$$ Here the mean magnetization $\langle \mathbf{S}_j\rangle$ on the site $j$ has been identified. For a square lattice with coordination number $z= 4$ and only nearest-neighbor interaction $J>0$ we assume a N[é]{}el state on the two sublattices $A$, $B$, $$\langle \mathbf{S}_A\rangle
= - \langle \mathbf{S}_B\rangle
= - \langle S^z_B\rangle\,\mathbf{e}_z\;.$$ The fermion Green’s function for any site on $A$ becomes $$\overline{G}_A(i\omega)
= {}[\,i\omega + \mu^f + \sigma^z h\,{}]^{-1}
\;\;,\;\;\;
h= \frac{zJ}{2}\langle S^z_A\rangle\;,$$ which leads to the self-consistent equation $$\label{eqn-selfhartree}
\langle S^z_A\rangle
= \frac{1}{2}{}[\,f(-h - \mu^f) - f(h - \mu^f)\,{}]\;.$$ For average projection, with $\mu^f= 0$, we find $$\label{eqn-selfhartree-av}
\textrm{average projection:} \quad\quad
h= \frac{z}{4}J\,\tanh\big(\frac{h}{2T}\big)\;.$$ Within the Popov–Fedotov scheme, using $\mu^f= i\frac{\pi}{2}T$, one has $$\label{eqn-selfhartree-ex}
\textrm{exact projection:} \quad\quad
h= \frac{z}{4}J\,\tanh\big(\frac{h}{T}\big)\;,$$ where the following expression for the Fermi function has been utilized, $$\label{eqn-fermicmplx}
f(x - i\frac{\pi}{2}T)
= f(2x) + \frac{i}{2\cosh(x/T)}\;\;,$$ for a real-valued $x$. In the physical observable (\[eqn-selfhartree\]) the imaginary part again cancels. Both projection schemes lead to the same self-consistent equation for the effective magnetic (Weiss) field $h$, except for a factor of 2 in the temperature. Accordingly the equations result in different N[é]{}el temperatures, $$\begin{array}[c]{lcrcl}
\textrm{average projection:} & & \displaystyle
T_N & = & \displaystyle \frac{z}{8}J\;\;,
\\[2ex]
\textrm{exact projection:} & & \displaystyle
T_N & = & \displaystyle \frac{z}{4}J\;.
\end{array}$$ However, at zero temperature both projection methods lead to the same result, $$\label{eqn-hartreesatur}
\textrm{average \emph{and} exact projection:} \quad
\lim_{T\to 0}\langle S^z_A\rangle
= \frac{1}{2}\;.$$ That is, the unphysical reduction of the spin moment observed for the free spin, is completely restored in the magnetically ordered ground state. Apparently, the unphysical charge fluctuations are suppressed at $T\to 0$.
Effect of the constraint: self-consistent theory {#sec-advanced}
================================================
The purpose of this section is to demonstrate the application of the Popov–Fedotov scheme within a self-consistent approximation that goes far beyond mean field. To our knowledge, the Popov–Fedotov approach has at present been applied in mean-field-like calculations with perturbative corrections[@popfed88; @boukis99; @kisopp00; @dilric06], but a self-consistent re-summation of the diagram series has not been attempted.
When choosing an approximation scheme, it has to be kept in mind that the automatic cancellation of unphysical states requires the auxiliary-fermion charge $Q_i$ to be conserved (this has been discussed in Section \[sec-intro\] above). In particular, the local gauge symmetry of the Hamitonian (\[eqn-allgrand\]) under $f_{i\alpha}\to e^{i \varphi_i}f_{i\alpha}$ must not be broken. Accordingly, approximations leading to finite expectation values like $\langle f^\dagger_{i\alpha} f_{j\alpha}\rangle\ne 0$ or $\langle
f_{i\uparrow} f_{j\downarrow}\rangle\ne 0$ cannot be used, while it is safe to consider so-called $\Phi$-derivable approximations[@lutwar60; @bay62]. Spontaneous breaking of physical symmetries (e.g., spin rotation, lattice translation) may be included, since the respective order parameters are gauge invariant. As a consequence of the local gauge symmetry, the fermion Green’s function is always local, $$\label{eqn-local}
\overline{G}_{ij}(i\omega)
= \delta_{i,j}\,\overline{G}_{i}(i\omega)\;.$$
Here we focus on the physics of the Heisenberg model (\[eqn-hbergham\]) in strictly two spatial dimensions at finite temperature $T> 0$. This system has been studied extensively in the past using a variety of numerical and analytical methods, in particular in view of experiments on cuprate superconductors in the undoped limit[@note-oldref]. The Hartree approximation discussed in Section \[sec-simple\] above is, of course, not appropriate for the 2D system, although the N[é]{}el-ordered ground state at $T= 0$ appears to be qualitatively correct[@man91; @briwol04]. At finite $T$ the theorem of Mermin and Wagner requires the magnetization $\langle \mathbf{S}_i\rangle= 0$ to vanish. Therefore we seek an approximation, where the susceptibility $\chi$ is self-consistently coupled back onto itself. Such an approximation has originally been proposed for the Hubbard model[@bicscawhi89], and is commonly referred to as FLEX.
![The self-consistent approximation (SCA) discussed in Section \[sec-advanced\], Eqs.(\[eqn-flex\]). Continuous lines denote the fermion Green’s function $\overline{G}$, Eq.(\[eqn-local\]), the thin dashed line represents $J$, and dots are Pauli matrices $\times\,1/2$.[]{data-label="fig-flex"}](figure1.eps){width="0.9\hsize"}
For the Hamiltonian (\[eqn-allgrand\]) with spin–spin interaction it takes the form shown in Fig. \[fig-flex\]. The fermion self-energy shown in the figure reads
\[eqn-flex\] $$\label{eqn-flex-self}
\overline{\Sigma}_{i}(i\omega)
= \frac{1}{4}
T\sum_{i\nu}
\sum_{\mu, \bar{\mu}}
D_{ii}^{\mu\bar{\mu}}(i\nu)
\,\sigma^\mu\,
\overline{G}_i(i\omega + i\nu)
\,\sigma^{\bar{\mu}}\;\;,$$ $\omega$ and $\nu$ denote a fermionic and bosonic Matsubara frequency, respectively, $\sigma^\mu$ a Pauli matrix, $\mu, \bar{\mu}= x, y,
z$. The renormalized spin–spin interaction on lattice sites $i, j$ is given by $$\label{eqn-flex-effint}
D_{ij}^{\mu\bar{\mu}}(i\nu)
= -J_{ij} \;+\;
\sum_{l,k}
J_{il}\,\chi_{lk}^{\mu\bar{\mu}}(i\nu)\,J_{kj}\;.$$ The susceptibility $\chi$ represents the series of fermion-line bubbles in Fig. \[fig-flex\], $$\label{eqn-flex-chi}
\chi_{ij}^{\mu\bar{\mu}}
= \big(\Pi\,{}[\,1 + J\Pi\,{}]^{-1}\big)_{ij}^{\mu\bar{\mu}}\;\;,$$ with $$\label{eqn-flex-bubble}
\Pi_i^{\mu\bar{\mu}}(i\nu)
= -T\sum_{i\omega}
\frac{1}{4}
\textrm{Tr}^\sigma{}[\,\sigma^\mu\,
\overline{G}_i(i\omega + i\nu)
\,\sigma^{\bar{\mu}}\,
\overline{G}_i(i\omega)\,{}]\;.$$
In the paramagnetic phase with lattice-translational symmetry we have $$\overline{G}_i(i\omega)= \sigma^0\,G(i\omega)
\;\;,\;\;\;
\overline{\Sigma}_i(i\omega)= \sigma^0\,\Sigma(i\omega)\;\;,$$ and therefore $$\Pi_i^{\mu\bar{\mu}}(i\nu)
= \delta_{\mu, \bar{\mu}}\,\Pi(i\nu)\;.$$ The self-consistent equations (\[eqn-flex\]) with Eq.(\[eqn-fdyson\]) now turn into
\[eqn-sflex\] $$\begin{aligned}
\label{eqn-sflex-bubble}
\Pi(i\nu)
& = & -\frac{T}{2}\sum_{i\omega}
G(i\omega + i\nu)\,G(i\omega)\;\;,
\\
\label{eqn-sflex-chi}
\chi(\mathbf{q}, i\nu)
& = & \frac{\Pi(i\nu)}{1 + J(\mathbf{q}) \Pi(i\nu)}\;\;,
\\
\label{eqn-sflex-effint}
D(i\nu)
& = & \frac{1}{N_L}
\sum_{\mathbf{q}}
J^2(\mathbf{q})\,\chi(\mathbf{q}, i\nu)\;\;,
\\
\label{eqn-sflex-self}
\Sigma(i\omega)
& = & \frac{3T}{4}\sum_{i\nu}
D(i\nu)\,G(i\omega + i\nu)\;\;,
\\
\label{eqn-sflex-dyson}
G(i\omega)
& = & \big[\,i\omega + \mu^f - \Sigma(i\omega)\,\big]^{-1}\;.\end{aligned}$$ The bare interaction in wave-vector space reads, for a square lattice in 2D with nearest-neighbor distance $a\equiv 1$, $$\label{eqn-sflex-jj}
J(\mathbf{q})
= 4J\,\gamma(\mathbf{q}) \;\;,\;\;\;
\gamma(\mathbf{q})
= \frac{1}{2}{}[\,\cos(q_x) + \cos(q_y)\,{}]\;.$$
Note that the bare $J$ in Eq.(\[eqn-flex-effint\]) does not contribute to the local $D(i\nu)\equiv D_{ii}(i\nu)$, Eq.(\[eqn-sflex-effint\]), since $J_{ii}= 0$.
It has been emphasized above, that the fermion propagator and as a consequence the irreducible bubble are local, $G_{ij}=
\delta_{i,j}\,G$, $\Pi_{ij}= \delta_{i,j}\,\Pi$. Nevertheless, the interesting measurable[@foot-cons] quantity in the Eqs.(\[eqn-sflex\]) is the susceptibility $\chi(\mathbf{q})$, which is wave-vector dependent through the bare interaction $J(\mathbf{q})$, and therefore may describe even long-range fluctuations.
The SCA shown in Fig. \[fig-flex\] can be derived from a $\Phi$-functional in close analogy[@bicscawhi89] to the FLEX. Therefore, Eqs.(\[eqn-sflex\]) represent a conserving approximation and can be used with the averaged fermion constraint as well as the Popov–Fedotov approach. The case of average projection, $\mu^f= 0$, has been treated in detail in Ref. . The magnetic correlation length $\xi(T)$ and the dynamical structure factor, derived from the susceptibility (\[eqn-sflex-chi\]), came out quite satisfactorily when compared to known results, indicating that the diagrams in Fig. \[fig-flex\] indeed capture the important physics of the 2D system at low $T$. In the following some of the results from Ref. will be re-calculated using the Popov–Fedotov method, i.e., $\mu^f= i\frac{\pi}{2}T$. It will turn out, that the imaginary-valued chemical potential requires the use of complex-valued spectral functions, leading to more involved equations than those derived in Ref. for $\mu^f= 0$. The results calculated with both methods differ, except in the limit of vanishing temperature, where average and exact projection become equal, as will be presented below.
*Equations for exact projection:* Before discussing the numerical solution of Eqs.(\[eqn-sflex\]), we quote some important formal results derived in Appendix \[sec-app-greens\]. For a numerical implementation it is suitable to express all Green’s functions through their respective spectral functions. On account of the Hamitonian being non-Hermitian, the spectral function of the fermion Green’s function $G(i\omega)$ becomes complex-valued. $G$ is given by $$\label{eqn-gfdef}
(\overline{G}_i)^{\alpha\bar{\alpha}}(i\omega)
= \int_0^{1/T}\mathrm{d}\tau\,
e^{i\omega\tau}
\langle\mathcal{T}_\tau\{f_{i\alpha}(\tau)
f^\dagger_{i\bar{\alpha}}(0)\}\rangle^\ff\;\;,$$ with the thermal expectation value and $\tau$-dependence calculated in the enlarged fermion Fock-space with the Hamiltonian (\[eqn-allgrand\]). $G$ has the following spectral representation, $\displaystyle
\overline{G}^{\alpha\bar{\alpha}}_i(i\omega)
= \delta_{\alpha, \bar{\alpha}}
G(i\omega)$, $$\label{eqn-gfspect}
G(i\omega)
= \int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\frac{\widehat{G}(\varepsilon)}{i\omega + \mu^f - \varepsilon}
\;\;,\;\;\;
\mu^f = i\frac{\pi}{2}T\;.$$ The energy variable $\varepsilon$ is a real number. For simplicity a system invariant under lattice translations and spin rotations has been assumed. The spectral function $\widehat{G}$ is complex-valued, $$\label{eqn-fspect}
\widehat{G}(\varepsilon)
= \rho_1(\varepsilon) \;+\;
i\rho_2(\varepsilon)\;.$$ In the Appendix \[sec-app-greens\] we also derive the sum rule $$\label{eqn-fsum}
\int\mathrm{d}\varepsilon\,
\widehat{G}(\varepsilon)= 1\;\;,$$ $$\;\;\;\Rightarrow\;\;\;
\int\mathrm{d}\varepsilon\,
\rho_1(\varepsilon)= 1 \;\;,\;\;\;
\int\mathrm{d}\varepsilon\,
\rho_2(\varepsilon)= 0\;\;,$$ $\varepsilon$ is again real valued. For the special case of spin degeneracy considered here, the fermion spectral-function in addition obeys the “particle–hole” symmetry, $$\label{eqn-phsymm}
\widehat{G}(-\varepsilon)
= \widehat{G}(\varepsilon)^\ast\;\;,$$ $$\;\;\;\Rightarrow\;\;\;
\rho_1(-\varepsilon)= \rho_1(\varepsilon) \;\;,\;\;\;
\rho_2(-\varepsilon)= - \rho_2(\varepsilon)\;.$$ The relations (\[eqn-gfspect\]), (\[eqn-fspect\]), (\[eqn-phsymm\]) hold as well for the fermion self-energy $\Sigma$.
For a numerical solution of the Eqs.(\[eqn-sflex\]), we introduce structure factors for $\Pi(i\nu)$ and $D(i\nu)$ according to $$S^0(\omega)
= {}[\,1 + g(\omega)\,{}]\,\Pi''(\omega)
\;\;,\;\;\;
\Pi''(\omega)
= \textrm{Im}\,\Pi(\omega + i0_+)\;\;,$$ $$U(\omega)
= {}[\,1 + g(\omega)\,{}]\,
\textrm{Im}\,D(\omega + i0_+)\;\;,$$ with analytic continuation to the real axis via $\displaystyle
\Pi(i\nu)\to \Pi(\omega + i0_+)$ and $\displaystyle
D(i\nu)\to D(\omega + i0_+)$, $\displaystyle
\omega\in\mathbb{R}$. The Bose function is denoted by $g(\omega)= 1/(e^{\omega/T} - 1)$. As shown in detail in Appendix \[sec-app-deriv\], Eqs.(\[eqn-sflex\]) now turn into
\[eqn-nflex\] $$\begin{aligned}
U(\omega)
& = & \label{eqn-nflex-effint}
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\frac{S^0(\omega)\,\,\mathcal{N}(\varepsilon)\,\varepsilon^2}
{{}\big(1 + \varepsilon\,\Pi'(\omega){}\big)^2 +
{}\big(\varepsilon\,\Pi''(\omega){}\big)^2}\;\;,
\\
\Pi''(\omega)
& = & \label{eqn-nflex-piim}
S^0(\omega) - S^0(-\omega)\;\;,
\\
\Pi'(\omega)
& = & \label{eqn-nflex-pire}
\textrm{P}\!\!\int_{-\infty}^\infty
\frac{\mathrm{d}\varepsilon}{\pi}
\frac{\Pi''(\varepsilon)}{\varepsilon - \omega}\;\;,
\\
S^0(\omega)
& = & \label{eqn-nflex-snul}
\frac{\pi}{2}
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\widehat{G}^+(\varepsilon)\,
\widehat{G}^-(\varepsilon - \omega)\;\;,\end{aligned}$$ for the real (physical) functions $U, S^0, \Pi'', \Pi'$ with the Kramers–Kroenig transform $\Pi'$, and $$\begin{aligned}
\widehat{G}^+(\omega)
& = & \label{eqn-nflex-gfplus}
{}[\,1 - f(\omega - i\frac{\pi}{2}T)\,{}]
\,\widehat{G}(\omega)\;\;,
\\
\widehat{G}^-(\omega)
& = & \label{eqn-nflex-gfmin}
f(\omega - i\frac{\pi}{2}T)
\,\widehat{G}(\omega)\;\;,
\\
\widehat{G}(\omega)
& = & \label{eqn-nflex-gf}
\frac{\widehat{\Sigma}(\omega)}
{{}\big(\omega - \overline{\Sigma}(\omega)\big)^2 +
{}\big(\pi\widehat{\Sigma}(\omega)\big)^2}\;\;,
\\
\widehat{\Sigma}(\omega)
& = & \label{eqn-nflex-selfspec}
\frac{3}{4\pi}
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
U(\varepsilon)\,
\big[\,\widehat{G}^+(\omega - \varepsilon) +
\\
& & \nonumber\hspace*{2cm}\mbox{}+
\widehat{G}^-(\omega + \varepsilon)\,\big]\;\;,
\\
\overline{\Sigma}(\omega)
& = & \label{eqn-nflex-selfhil}
\textrm{P}\!\!\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\frac{\widehat{\Sigma}(\varepsilon)}
{\omega - \varepsilon}\;\;,\end{aligned}$$
for the complex (unphysical) spectra $\widehat{G}^\pm, \widehat{G},
\widehat{\Sigma}, \overline{\Sigma}$ with the Hilbert transform $\overline{\Sigma}$. Note again that the energy arguments $\omega,
\varepsilon$ are real-valued. As is also shown in the Appendix \[sec-app-deriv\], the Eqs.(\[eqn-nflex\]) can be somewhat simplified by utilizing the symmetry (\[eqn-phsymm\]). The imaginary chemical potential $i\frac{\pi}{2}T$ appears in Eq.(\[eqn-nflex-gfplus\]) and (\[eqn-nflex-gfmin\]), adding an imaginary part to the fermion spectra $\widehat{G}^+$ and $\widehat{G}^-$ via Eq.(\[eqn-fermicmplx\]).
The density-of-states that enters Eq.(\[eqn-nflex-effint\]) is defined as $$\label{eqn-dos}
\mathcal{N}(\varepsilon)
= \frac{1}{N_L}\sum_{\mathbf{k}}
\delta(\varepsilon - J(\mathbf{q}))\;\;,$$ and for the nearest-neighbor interaction (\[eqn-sflex-jj\]) in two dimensions it becomes $$\label{eqn-dosnum}
\mathcal{N}(\varepsilon)
= \frac{K(m)}{2\pi^2 J}
\theta(4J - |\varepsilon|)
\;\;,\;\;\;
m = 1 - \left(\frac{\varepsilon}{4J}\right)^2\;\;,$$ with the complete elliptic integral of the first kind, $K(m)=
\int_0^1\mathrm{d}t\,{}[\,(1 - t^2)(1 - m t^2)\,{}]^{-1/2}$.
The physical output from the numerical iteration of Eqs.(\[eqn-nflex\]), (\[eqn-dosnum\]) is the structure factor $U(\omega)$ of the effective interaction, which is essentially the local (on-site) spin-excitation spectrum, and the wave-vector dependent dynamical spin-structure factor $$\label{eqn-struct}
S(\mathbf{q}, \omega)
= {}[\,1 + g(\omega)\,{}]\,
\textrm{Im}
\left\{\frac{\Pi(\omega)}{1 + J(\mathbf{q})\Pi(\omega)}
\right\}\;\;,$$ with $\displaystyle
\Pi= \Pi' + i\Pi''$. Furthermore, the magnetic correlation length $\xi(T)$ is extracted from the static magnetic susceptibility $\chi(\mathbf{q}, 0)$: Eq.(\[eqn-sflex-chi\]), for $\mathbf{q}$ close to the N[é]{}el-ordering vector $\mathbf{Q}= (\pi,\pi)$, i.e., $$J(\mathbf{q})
\simeq -4J + (\mathbf{q} - \mathbf{Q})^2J\;\;,$$ leads to $$\chi(\mathbf{q}, 0)
= \frac{\Pi'(0)}{1 + J(\mathbf{q})\Pi'(0)}
\simeq \frac{1}{J}
\frac{1}{\xi^{-2} + (\mathbf{q} - \mathbf{Q})^2}\;\;,$$ where $\Pi(0)= \Pi'(0)$ has been used, and the correlation length is identified as $$\label{eqn-xidef}
\xi(T)
= \left(\frac{J\Pi'(0)}{1 - 4J\Pi'(0)}\right)^{1/2}\;.$$ For completeness, in Appendix \[sec-app-simpl\] the self-consistent equations (\[eqn-nflex\]) are re-written in real-valued spectral functions. These Eqs.(\[eqn-simpl-num\]) can directly be compared to the Eqs.(A1) in Ref. : Both sets of equations represent the same diagrammatic approximation shown in Fig.\[fig-flex\], the former derived within the Popov–Fedotov scheme (exact projection), the latter within average projection.
Numerical results {#sec-num}
=================
The numerical results presented below are obtained from an iterative solution of the self-consistent equations (\[eqn-sflex\]), using the identical procedure and parameters for exact projection ($\mu^f=
i\frac{\pi}{2}T$, leading to Eqs.(\[eqn-simpl-num\]) in App.\[sec-app-simpl\]) as well as average projection ($\mu^f= 0$, leading to Eqs.(A1) in Ref. ). The procedure also utilizes Eqs.(\[eqn-dosnum\]) and (\[eqn-xidef\]), which apply to both projection schemes. The data shown in Ref. for the case of average projection have not been re-used in the present study.
*Magnetic correlation length:* The correlation length $\xi(T)$ is shown in Fig. \[fig-correl\]. $\xi(T)$ becomes larger than one lattice spacing for $T\lesssim J$ and reaches values up to $\simeq
1200$ lattice spacings for the lowest temperature $T= 0.048 J$ considered in this work. The numerical data shown in Fig.\[fig-correl\] are well reproduced by
![**Main figure:** Magnetic correlation length $\xi(T)$ from the numerical solution of the self-consistent approximation Fig.(\[fig-flex\]) for the 2D square lattice, in units of the lattice spacing. The continuous line with diamonds corresponds to Eqs.(\[eqn-nflex\]), where the Popov–Fedotov approach has been used. The dashed line with circles belongs to Eqs.(A1) in Ref. , where the auxiliary-particle constraint has been approximated by its thermal average. **Inset:** The correlation length, plotted as $\ln(\xi)$ vs. $J/T$. Shown is the temperature region $0.048\,J\le T\le 0.16\,J$, where a fit to the data has been performed as described in the text.[]{data-label="fig-correl"}](figure2.eps){width="\hsize"}
$$\label{eqn-fit}
\xi(T)
= c\left(\frac{J}{T}\right)^b
\exp\big(a\,\frac{J}{T}\big)\;.$$
The parameters $a, b, c$ are determined by plotting the data as $\ln(\xi)$ vs. $J/T$ (see the inset of Fig. \[fig-correl\]) and a numerical fit of the function $\;\ln(c) + b\,\ln(J/T) + a\,J/T\;$ to the data. We find $$\label{eqn-fitval}
\begin{array}[c]{lrcl}
\textrm{average projection:}
& a & = & 0.296 \\
& b & = & 0.592 \\
& c & = & 0.393 \\ \\
\textrm{exact projection:}
& a & = & 0.304 \\
& b & = & 0.418 \\
& c & = & 0.595
\end{array}$$ At low temperatures the effect of the exact auxiliary-particle projection is only marginal, as is already apparent from inspecting Fig. \[fig-correl\]. The “spin stiffness” $a= 2\pi\rho_s/J$ in the exponent in Eq.(\[eqn-fit\]) is the same in both cases[@note-num], $a\simeq 0.30$, merely the power $b\simeq 0.5$ of the algebraic prefactor is slightly modified in going from average to exact projection.
![**Top:** The structure factor $U(\omega)$, Eq.(\[eqn-nflex-effint\]), of the renormalized local interaction $D(i\nu)$, Eq.(\[eqn-sflex-effint\]), at high temperature $T=
2J$. Dashed and continuous lines as in Fig. \[fig-correl\]. **Bottom, main figure:** $U(\omega)$ at low temperature $T=
0.05J$, corresponding to a correlation length $\xi\simeq 900$. **Bottom, inset:** Same data as in the main figure, but zoomed around the energy $\omega= 0$. For clarity, the peaks for average constraint (dashed) and exact constraint (continuous) are shifted from their original position $\omega= 0$ by $J$ and $J$, respectively. []{data-label="fig-effint"}](figure3top.eps "fig:"){width="\hsize"} ![**Top:** The structure factor $U(\omega)$, Eq.(\[eqn-nflex-effint\]), of the renormalized local interaction $D(i\nu)$, Eq.(\[eqn-sflex-effint\]), at high temperature $T=
2J$. Dashed and continuous lines as in Fig. \[fig-correl\]. **Bottom, main figure:** $U(\omega)$ at low temperature $T=
0.05J$, corresponding to a correlation length $\xi\simeq 900$. **Bottom, inset:** Same data as in the main figure, but zoomed around the energy $\omega= 0$. For clarity, the peaks for average constraint (dashed) and exact constraint (continuous) are shifted from their original position $\omega= 0$ by $J$ and $J$, respectively. []{data-label="fig-effint"}](figure3bottom.eps "fig:"){width="\hsize"}
*Spin spectral-function and energy scale:* The almost vanishing effect of the auxiliary-particle constraint at low temperature is also visible in the spectra: Fig. \[fig-effint\] shows the effective interaction $U(\omega)$, Eq.(\[eqn-nflex-effint\]), which is the structure factor of the $D(i\nu)$ given in Eq.(\[eqn-sflex-effint\]). Since $J(\mathbf{q})^2$ in Eq.(\[eqn-sflex-effint\]) depends only weakly on $\mathbf{q}$, $U(\omega)$ is essentally the local magnetic structure factor or spin-excitation spectrum. Note that Eq.(\[eqn-nflex-effint\]) is the same in the average-projected case, Eq.(A1h) in Ref.. The data for low temperature, shown in the bottom panel of Fig. \[fig-effint\], features a broad shoulder of width $\sim J$, which is reminiscent of the box-like density-of-states for spin waves in 2D. Around zero energy $U(\omega)$ displays a huge peak (see the inset of the figure), which contains the critical fluctuations at $\mathbf{q}\simeq (\pi,\pi)$ close to the antiferromagnetic ordering wave-vector. The curves for average and exactly treated constraint are on top of each other; merely the amplitudes of the peaks at $\omega= 0$ differ by a factor of $\mathcal{O}(1)$. At high temperature $T= 2J$, on the contrary, the curves for average and exact constraint are well separated, in particular the total spectral weight is smaller for average projection. This can be seen from the data in the top panel of Fig.\[fig-effint\]. The reduction of spectral weight in $U(\omega)$ is related to an unphysical lack of local spin moment, which occurs if the constraint is not taken exactly. This will be discussed further below.
![**Main figure:** The scaling function Eq.(\[eqn-scalefun\]) of the renormalized spin–spin interaction, using the energy scale $\omega_0(T)$ from Eq.(\[eqn-enscale\]) and $\Gamma(0)\equiv 1$. The curve shown here is the same for all temperatures $0.048\le T/J\le 0.13$, corresponding to $1000\ge \xi(T)\ge 10$, and independent of the auxiliary-particle constraint being treated exactly or on average. **Inset:** Detailed view of the region near $\omega= 0$.[]{data-label="fig-enscale"}](figure4.eps){width="\hsize"}
![The spectral function $\widehat{G}(\omega)$ of the auxiliary fermion, introduced in Eqs.(\[eqn-gfdef\]), (\[eqn-gfspect\]). For exact auxilary-fermion projection it is complex valued and given by Eq.(\[eqn-nflex-gf\]), for average projection Eq.(A1e) from Ref. holds, with $\widehat{G}(\omega)$ being real. **Top:** high temperature, **Bottom:** low temperature. Parameters as in Fig. \[fig-effint\].[]{data-label="fig-fdos"}](figure5top.eps "fig:"){width="\hsize"} ![The spectral function $\widehat{G}(\omega)$ of the auxiliary fermion, introduced in Eqs.(\[eqn-gfdef\]), (\[eqn-gfspect\]). For exact auxilary-fermion projection it is complex valued and given by Eq.(\[eqn-nflex-gf\]), for average projection Eq.(A1e) from Ref. holds, with $\widehat{G}(\omega)$ being real. **Top:** high temperature, **Bottom:** low temperature. Parameters as in Fig. \[fig-effint\].[]{data-label="fig-fdos"}](figure5bottom.eps "fig:"){width="\hsize"}
The difference of peak amplitudes visible in the inset of Fig.\[fig-effint\] bottom can be traced back to the slightly different correlation length $\xi(T)$, compare Eq.(\[eqn-fitval\]). The influence of the absolute value of $\xi$ vanishes if the scaling behaviour of $U(\omega)$ is considered: We start from the dynamical scaling hypothesis[@chn89], $$\label{eqn-hypo}
S(\mathbf{q}, \omega)
= \frac{1}{\omega_0}
S^{st}(\mathbf{Q})
\,\varphi(k\xi)\,\Phi(k\xi, \omega/\omega_0)\;.$$ Here $S(\mathbf{q}, \omega)$ is the dynamical structure factor Eq.(\[eqn-struct\]), $S^{st}(\textbf{Q})$ denotes the static (equal-time) correlation function $$\label{eqn-statstruct}
S^{st}(\mathbf{q})
= \langle S^x_{\mathbf{q}} S^x_{-\mathbf{q}}\rangle
= \frac{1}{\pi}
\int_{-\infty}^\infty\mathrm{d}\omega\,
S(\mathbf{q}, \omega)
\;\;,\;\;\;$$ taken at the AF ordering wave vector $\mathbf{Q}= (\pi,\pi)$. $\varphi(x)$, $\Phi(x, y)$ are scaling functions, $\mathbf{k}=
\mathbf{q} - \mathbf{Q}$, and $\omega_0$ is the energy scale for critical fluctuations. At small energies $\omega\ll J$ we expect $U(\omega)\propto\int\mathrm{d}^2q\,S(\mathbf{q}, \omega)$, and from an intergration of Eq.(\[eqn-hypo\]) over wave-vector space there follows the scaling property $$\label{eqn-lochypo}
U(\omega)
= \frac{1}{\omega_0}
S^{st}(\mathbf{Q})
\,\Gamma(\omega/\omega_0)\;.$$ $\Gamma(y)$ is the (a priori unknown) scaling function for the local effective spin-spin interaction $U(\omega)$. According to Eq.(\[eqn-hypo\]) the energy scale can be extracted from the numerical data, up to a constant prefactor, using $$\label{eqn-enscale}
\omega_0
= \left.\frac{S^{st}(\mathbf{Q})}{S(\mathbf{Q}, \omega)}
\right|_{\omega= 0}\;.$$ We obtained the energy scale in the temperature range $0.048\le T/J\le
0.13$, which corresponds to correlation lengths $1000\ge \xi\ge
10$, using Eqs.(\[eqn-enscale\]), (\[eqn-statstruct\]), and (\[eqn-struct\]). The scaling function is then determined for each temperature from Eq.(\[eqn-lochypo\]), $$\label{eqn-scalefun}
\Gamma(\omega/\omega_0)
= \frac{U(\omega)}{U(0)}\;.$$ All curves $\Gamma(\omega/\omega_0)$, whether calculated with average or with exact auxiliary-particle constraint, agree to within numerical accuracy. The scaling function is shown in Fig.\[fig-enscale\]. Deviations from scaling behaviour become visible only for higher energies $\omega/\omega_0> 20$. In particular, going from exact to average projection has no effect on the scaling behavior.
A slight difference in the low-temperature properties of the two auxiliary-particle methods shows up in the energy scale itself: From a linear regression on $\omega_0(T)\times\xi^2(T)$, obtained from Eq.(\[eqn-enscale\]) and Eq.(\[eqn-xidef\]), we find $$\label{eqn-numenscale}
\begin{array}[c]{lrcl}
\textrm{average:}
& \omega_0 \xi^2 & = & (0.234 \pm 0.002)\,J\;\;,
\\
\textrm{exact:}
& \omega_0 \xi^2 & = & (0.234 \pm 0.002)\,J \;+\; 0.089\,T\;.
\end{array}$$ Nevertheless, the temperature behavior of the energy scale essentially is the same, $$\label{eqn-enscaleresult}
\omega_0(T)
= 0.23\cdot J \,(\xi)^{-2} \;\;,\;\;\;$$ which corresponds to a dynamical critical exponent $z= 2$.
*Fermion propagator:* Fig. \[fig-fdos\] display the spectral function of the auxiliary-fermion propagator for the two projection methods. As has been discussed below Eq.(\[eqn-gfdef\]), the spectrum $\widehat{G}(\omega)$ of the fermion is complex valued, if the auxiliary-particle constraint is enforced exactly via the imaginary-valued chemical potential $\mu^f= i\frac{\pi}{2}T$. If the constraint is treated on the average using $\mu^f= 0$, $\widehat{G}(\omega)$ remains real. In Fig. \[fig-fdos\] the spectrum $\widehat{G}(\omega)$ is shown for high (top panel) and low temperature (bottom). At high $T$ the spectra for exact and average constraint differ significantly, in particular the $\widehat{G}(\omega)$ from exact projection has a considerable imaginary part. At low temperature, on the other hand, the imaginary part is quite small, while the real part becomes equal to the spectrum of the average projection. In accordance with the results for spin-structure factor and correlation length, the projection onto the physical part of the fermion-Hilbert space has almost no effect at sufficiently low temperature.
![The local spin moment $S^{st}_{loc}$ as defined in Eq.(\[eqn-momdef\]), multiplied by 4. The limiting values for $T\gg J$ given in Eq.(\[eqn-momhigh\]) are indicated as dotted lines. The exact value, given by the sum rule (\[eqn-momrule\]), is $4\,S^{st}_{loc}= 1$. []{data-label="fig-locmom"}](figure6.eps){width="\hsize"}
*Local spin moment:* Another interesting quantity for studying the influence of the auxiliary-fermion constraint is the local spin moment, given by the local equal-time correlation function at an arbitrary site $i$, $$\label{eqn-momdef}
S^{st}_{loc}
= \langle S^x_i S^x_i\rangle
= \frac{1}{N_L}\sum_{\mathbf{q}}S(\mathbf{q})\;.$$ The static structure factor $S(\mathbf{q})$ has been defined in Eq.(\[eqn-statstruct\]). For a spin-1/2 system in the paramagnetic phase $S^{st}_{loc}$ should fulfill the sum rule $$\label{eqn-momrule}
S^{st}_{loc}
= \frac{1}{3}(\mathbf{S}_i)^2
= \frac{1}{3}S(S + 1)
= \frac{1}{4}\;.$$ At very high temperature $T\gg J$, interaction effects can be ignored, and $S^{st}_{loc}$ is given by the simple bubble $\Pi$ shown in Fig. \[fig-flex\], calculated with free auxiliary fermions. With Eq.(\[eqn-sflex-bubble\]) we have $$\begin{aligned}
T\gg J\,: \quad
S^{st}_{loc}
& \simeq & T\sum_{i\nu}\Pi(i\nu)
\big|_{J= 0}
\\
& = & \frac{1}{2}
\langle f^\dagger_{i\uparrow} f_{i\uparrow}\rangle^\ff\,
\langle f_{i\downarrow} f^\dagger_{i\downarrow}\rangle^\ff
\big|_{J= 0}\;.\end{aligned}$$ For both projection methods, the expectation values are to be taken in the generalized grand-canonical ensemble (\[eqn-allgrand\]), (\[eqn-chem\]). For the case of average projection we find $\langle f^\dagger_{i \sigma} f_{i \sigma}\rangle= f(0)= 1/2$ for any spin direction $\sigma= \pm 1$, with exact projection we have $\langle f^\dagger_{i \sigma} f_{i \sigma}\rangle=
f(-i\frac{\pi}{2}T)= \frac{1}{2}(1 + i)$. $f$ denotes the Fermi function. That is, for $T\gg J$, $$\label{eqn-momhigh}
\begin{array}[c]{lrcl}
\textrm{average constraint:} &
4\,S^{st}_{loc} & = &
\frac{1}{2}\;\;,
\\
\textrm{exact constraint:} &
4\,S^{st}_{loc} & = &
2\,\left|\frac{1}{2}(1 + i)\right|^2= 1\;.
\end{array}$$ If the projection onto the physical Hilbert space is performed exactly, the sum rule (\[eqn-momrule\]) is correctly reproduced. With average projection, on the other hand, it is significantly violated. This is due to thermal charge fluctuations into unphysical, spinless states, which reduce the spin moment.
When temperature is lowered, we find a partly unexpected result: Fig.\[fig-locmom\] shows $4\,S^{st}_{loc}$ from the numerical solution as function of $T$. At high temperature the free-spin result is approached, whereas at low temperature the average and exact projection methods lead to the same value for $S^{st}_{loc}$. This observation fits into the line of results obtained so far: at $T\to 0$ it does not matter whether the auxiliary-particle constraint is treated exactly or on the thermal average.
However, the local spin moment at $T\to 0$, $4\,S^{st}_{loc}\simeq
0.75$ is too small. This is not due to an ill-treated constraint, but an artifact of the approximation to the interacting system. The local moment (\[eqn-momdef\]) measures the total spectral weight of spin excitations, averaged over the Brillouin zone. The self-consistent approximation we use, see Fig. \[fig-flex\], apparently lacks some weight in the spin-excitation spectrum. In Ref. we studied an approximation with a somewhat reduced self consistency (called “MSCA”), which delivered a better result for $S^{st}_{loc}$ at low $T$, $4\,S^{st}_{loc}\simeq
0.85$. Moreover, the wave-vector dependence of $S(\mathbf{q},
\omega)$ came out better. However, the MSCA cannot straight-forwardly be extended to the Popov–Fedotov scheme, since we don’t know a $\Phi$-functional for that approximation to guarantee the conservation of the auxiliary-charge $Q_i$, which is a necessary condition for the Popov–Fedotov method (see Section \[sec-intro\] above).
*The limit $T\to 0$:* At first sight it seems trivial that the imaginary-valued chemical potential $\mu^f= i\frac{\pi}{2}T$ has almost no effect at $T\to 0$: $\mu^f$ enters the self-consistent equations through the Fermi function $f(\omega - i\frac{\pi}{2}T)$ in Eqs.(\[eqn-nflex-gfplus\]) and (\[eqn-nflex-gfmin\]). Assuming $T\ll\omega$, Eq.(\[eqn-fermicmplx\]) yields
\[eqn-flimit\] $$\label{eqn-flimit-low}
T\ll\omega\,: \quad
f(\omega - i\frac{\pi}{2}T)
= \Theta(-\omega) + \mathcal{O}(e^{-2|\omega|/T})\;\;,$$ which matches the case $\mu^f= 0$ (average projection) at $T= 0$. On the other hand, the energy scale $\omega_0$ of spin excitations, Eq.(\[eqn-enscaleresult\]), is exponentially small compared to $T$, therefore the opposite limit should apply, $$\label{eqn-flimit-high}
T\gg\omega\,: \quad
f(\omega - i\frac{\pi}{2}T)
= \frac{1}{2}(1 + i)
+ \mathcal{O}(\frac{\omega}{T})\;\;,$$
adding a significant imaginary part to the fermion spectral-function $\widehat{G}(\omega)$. It requires a solution of the set of Eqs.(\[eqn-nflex\]), however, to reveal that $\widehat{G}(\omega)$ has no features[@note-fspecen] at $\omega\lesssim\omega_0$ (see Fig. \[fig-fdos\]). The fermion spectrum is governed by its bandwidth $\sim J$, and therefore the crossover from high-temperatures corresponding to Eq.(\[eqn-flimit-high\]) to low temperatures, where Eq.(\[eqn-flimit-low\]) becomes valid, happens at $T\sim J$. A more physical interpretation is provided in the following Sections \[sec-const\] and \[sec-sum\].
Measuring the constraint {#sec-const}
========================
In order to understand the weak influence of the fermion constraint at low temperature, it is useful to calculate the auxiliary-charge fluctuations in average projection. Starting from Eq.(\[eqn-expect-av\]), we have to calculate $$\label{eqn-qfluct}
\langle \Delta Q_i\, \Delta Q_j\rangle^\av
= \langle Q_i Q_j\rangle^\av -
\langle Q_i\rangle^\av\,\langle Q_j\rangle^\av\;.$$ Since all charges $Q_i$ are conserved, ${}[\,Q_i\,,\, H\,{}]= 0$, ${}[\,Q_i\,,\, Q_j\,{}]= 0$, the correlation function (\[eqn-qfluct\]) may be obtained from the charge propagator in Matsubara space as $$\begin{aligned}
\chi^Q_{ij}(i\nu)
& = & \nonumber
\int_0^{1/T}\mathrm{d}\tau\,
e^{i\nu\tau}
\langle\mathcal{T}_\tau\{\Delta Q_i(\tau)\, \Delta Q_j(0)\}\rangle^\av
\\
& = & \label{eqn-qmatsu}
\frac{1}{T}\,\delta_{\nu, 0}\,
\langle \Delta Q_i\, \Delta Q_j\rangle^\av\;.\end{aligned}$$ $\chi^Q_{ij}(i\nu)$ is conveniently calculated with the Feynman-diagram rules introduced in Sect. \[sec-simple\], using the Hamiltonain (\[eqn-allgrand\]) with $\mu^f= 0$ and the bare fermion Green’s function (\[eqn-baregf\]).
In Section \[sec-simple\] we first discussed the limit $J= 0$. In that case, $\chi^Q$ is given by a simple bubble of bare fermion Green’s functions, $$\begin{aligned}
\chi^Q_{ij}(i\nu)
& = & - T\sum_{i\omega}\textrm{Tr}^\sigma{}[\,
\overline{G}^0_{ij}(i\nu + i\omega)\,
\overline{G}^0_{ji}(i\omega)\,{}]
\\
& = & \frac{1}{2T}\delta_{i,j}\delta_{\nu,0}\;\;,\end{aligned}$$ that is, $$\label{eqn-qflu-free}
J= 0\,: \quad
\langle \Delta Q_i\, \Delta Q_j\rangle^\av
= \delta_{i,j}\frac{1}{2}\;.$$ As expected, the auxiliary-charge fluctuations are finite. Note that $\chi^Q_{ij}(i\nu)$ is local ($\sim\delta_{i,j}$) and static, ($\sim\delta_{\nu,0}$) in accordance with Eq.(\[eqn-qmatsu\]), i.e., the conservation of the $Q_i$.
*Mean-field theory:* The second example presented in Section \[sec-simple\] is the Hartree approximation. $\chi^Q$ is again given by the simple bubble[@note-qfluct-hartree], with $\overline{G}_{ij}^0$ replaced by $$\overline{G}_{ij}(i\omega)
= \delta_{i,j}{}[\,i\omega + (-1)^i \sigma^z\,h\,{}]^{-1}\;\;,$$ with the Weiss field $h$ given in Eq.(\[eqn-selfhartree\]). For $T\ll J$ we obtain $$\label{eqn-qflu-hartree}
\textrm{Hartree:} \quad
\langle \Delta Q_i\, \Delta Q_j\rangle^\av
= \delta_{i,j}2\exp(-h/T)\;\;,$$ with $h= \frac{zJ}{4} + \mathcal{O}(e^{-J/T})$. That is, in the magnetically ordered phase the unphysical charge fluctuations are strongly suppressed, since the fermions develop a charge gap similar to a spin-density-wave state.
*Self-consistent theory:* The approximation discussed in Sections \[sec-advanced\] and \[sec-num\] requires a more elaborate calculation. For the case of average constraint, $\mu^f= 0$, the approximation given by Eqs.(\[eqn-sflex\]) and Fig. \[fig-flex\] has been studied earlier in Ref. . In Section IVC of that paper the conserving-approximation scheme has been applied to the spin susceptibility, leading to a vertex function in the fermion bubble, which is determined by a Bethe–Salpeter equation. The corresponding diagrams are shown in Fig. 6 of Ref.. For the auxiliary-fermion-charge susceptibility we want to calculate, the diagrams are exactly the same, except that the two spin vertices appearing in the bubble $\widehat{\Pi}$ in Fig. 6 are to be replaced by charge vertices $\sigma^0= 1$. The response function (\[eqn-qmatsu\]) then reads in wave-vector space $$\chi^Q(\mathbf{q}, 0)
= -T\sum_{i\omega}\textrm{Tr}^{\sigma}{}[\,
\sigma^0\,\overline{G}(i\omega)\,\overline{\Gamma}(\mathbf{q},
i\omega)\,\overline{G}(i\omega)\,{}]\;.$$ For $\langle \Delta Q\,\Delta Q\rangle^\av$ merely the static limit $\displaystyle\lim_{\nu\to 0}\chi^Q(\mathbf{q}, i\nu)$ is needed. With spin-rotation symmetry one has $\overline{G}= \sigma^0\,G$, and it turns out that $\overline{\Gamma}= \sigma^0\,\Gamma$ (i.e., the charge and spin channels do not mix in the vertex function), leading to $$\label{eqn-flex-qbub}
\chi^Q(\mathbf{q}, 0)
= -2T\sum_{i\omega} G(i\omega)^2\,\Gamma(\mathbf{q}, i\omega)\;.$$ The vertex function is specified through the following Bethe–Salpeter equation, taken from the diagrams in Fig. 6 of Ref., $$\label{eqn-flex-bethe}
\begin{array}[c]{l}
\displaystyle
\Gamma(\mathbf{q}, i\omega) = 1 +
\\[1ex] \displaystyle
\mbox{}+
\frac{3}{4}T\sum_{i\omega_1} G(i\omega_1)^2\,
D(i\omega - i\omega_1)\,\Gamma(\mathbf{q}, i\omega_1) -
\\[3ex] \displaystyle
\mbox{}-
\frac{3}{8}T\sum_{i\nu} G(i\omega + i\nu)
\frac{1}{N_L}\sum_{\mathbf{k}}
D(\mathbf{k} + \mathbf{q}, i\nu)\,D(\mathbf{k}, i\nu) \times
\\[3ex] \displaystyle
\mbox{}\times
T\sum_{i\omega_1} G(i\omega_1)^2\,
{}[\,G(i\omega_1 + i\nu) +
G(i\omega_1 - i\nu)\,{}]\,
\Gamma(\mathbf{q}, i\omega_1)
\end{array}$$ The fermion Green’s function $G(i\omega)$ and the local spin-spin interaction $D(i\nu)$ have to be taken from the solution of the Eqs.(\[eqn-sflex\]) for $\mu^f= 0$. The non-local spin interaction appearing in (\[eqn-flex-bethe\]) reads $$D(\mathbf{q}, i\nu)
= - J(\mathbf{q}) + J(\mathbf{q})^2\,\chi(\mathbf{q}, i\nu)\;\;,$$ with $\chi$ from Eq.(\[eqn-sflex-chi\]).
In Appendix \[sec-app-qbub\] it is shown that the 2nd term in Eq.(\[eqn-flex-bethe\]) actually becomes zero by symmetry arguments, i.e., the Bethe–Salpeter equation simplifies to $$\label{eqn-flex-bethe2}
\Gamma(\mathbf{q}, i\omega)
= 1 + \frac{3}{4}T\sum_{i\omega_1} G(i\omega_1)^2\,
D(i\omega - i\omega_1)\,\Gamma(\mathbf{q}, i\omega_1)$$ Instead of solving the last equation numerically, we find it more instructive to aim at an approximate analytical solution. We employ the static approximation introduced in Ref. , i.e., let $D(i\nu)= D(0)\delta_{\nu, 0}$ in Eq.(\[eqn-flex-bethe2\]) as well as the fermion self-energy (\[eqn-sflex-self\]). The calculation closely follows Sections IVA and C of Ref., leading to $$\Gamma(\mathbf{q}, i\omega)
= \frac{1}{1 - (\omega_f/2)^2 G(i\omega)^2}
\;\;,\;\;\;
\omega_f
= J\frac{16}{3\pi} + \mathcal{O}(T^2)$$ for temperatures $T\ll J$. $\omega_f$ is a typical 1/2 bandwidth of the continuous fermion spectrum, compare the bottom panel of Fig.\[fig-fdos\]. Performing the Matsubara-sum in Eq.(\[eqn-flex-qbub\]) eventually leads to $$\chi^Q(\mathbf{q}, 0)
= \frac{1}{\omega_f}\Phi^Q(T/J)
\;\;,\;\;\;
\Phi^Q(t)
= \frac{4}{\pi} + \mathcal{O}(t^2)\;.$$ If the vertex function is ignored, $\Gamma \to 1$, the result does not change significantly, $\Phi^Q(t) \to \frac{16}{3\pi} +
\mathcal{O}(t^2)$. Note that $\chi^Q$ is independent of $\mathbf{q}$, i.e., local.
From Eq.(\[eqn-qmatsu\]) we thus find the auxiliary-charge fluctuations of our self-consistent approximation in average projection, $$\label{eqn-qflu-flex}
\textrm{self cons.:} \quad
\langle \Delta Q_i \,\Delta Q_j \rangle^\av
= \delta_{i, j}
\frac{3 T}{4 J} + \mathcal{O}(T^3)\;.$$ Since the fermion spectrum is gapless[@note-fspecen] around $\omega= 0$ (see Fig. \[fig-fdos\]), and the vertex function has very little effect, $\chi^Q$ comes out Pauli-like. The explicit $T$ factor in Eq.(\[eqn-qmatsu\]) suppresses the charge fluctuations at low temperature. Compared to the magnetically ordered state described in mean-field (Hartree) theory, where the fermions develop a gap $\sim
J$ (see Eq.(\[eqn-qflu-hartree\])), the suppression of charge fluctuations is much weaker[@note-pregap]. However, at $T\to 0$ the unphysical fluctuations still vanish, and the average projection becomes exact.
Summary and conclusions {#sec-sum}
=======================
Resummed perturbation theory is a powerful tool for calculating dynamical properties of strongly correlated electron systems. In this paper we focused on the spin-1/2 antiferromagnetic quantum Heisenberg model on the two dimensional square lattice. Summing infinite classes of contributions in perturbation theory is most economically done within a quantum-field-theoretic formulation employing canonical fields. We therefore use an auxiliary-particle representation of spin operators, which is a faithful representation in the physical sector of the Hilbert space.
The use of auxiliary fermions requires a projection onto the physical part of the fermion-Fock space, where the fermion charge $Q_i$ equals one for each lattice site.
While for a single lattice site the projection may be done exactly, e.g., by introducing an auxiliary-fermion energy $\lambda$, which is sent to infinity at the end of the calculation[@abr65; @col84], these standard methods cannot directly be generalized to effect the projection at each lattice site independently (this would require handling a large number of independent limiting procedures, an impossible task in practice).
For lattice systems the most simple approach to the projection is an approximative treatment, where a global chemical potential (Lagrange multiplier) $\mu^f$ is introduced, which is sufficient to fulfill the constraint on the thermal average, $\langle Q_i\rangle= 1$.
However, Popov and Fedotov have proposed a rather unusual projection method, where a global *imaginary-valued* chemical potential $\mu^f= i\frac{\pi}{2}T$ leads to an *exact* cancellation of unphysical states, therefore enforcing the operator constraint $Q_i=
1$. Unfortunately, the Popov-Fedotov method may not straight-forwardly be generalized to systems away from particle-hole symmetry[@grosjohn90].
In this paper we explored the usability of this concept by identifying the conditions to be satisfied by any, necessarily self-consistent, approximation scheme. Most important is the conservation of the fermion charge $Q_i$ by the model Hamiltonian, ${}[Q_i, H{}]= 0$. If the approximation under consideration violates this conservation law, results become meaningless (see Sec. \[sec-intro\] and App.\[sec-app-greens\]). Therefore, self-consistent approximations are most safely based on the conserving-approximation principle. Any hopping of auxiliary fermions, for example, is precluded by this requirement: The fermions are strictly local entities. The physically observable momentum dependence of spin correlation functions originates from the momentum dependence of the exchange interaction.
Within the Popov–Fedotov approach the well known Feynman-skeleton-diagram expansion is applicable in conjunction with an exact projection of the auxilary particles onto the physical Hilbert space. We have shown in some detail how a self-consistent approximation, which goes far beyond mean-field theory similar to the “fluctuation-exchange approximation”, can be formulated using complex-valued spectral functions of the (unphysical) renormalized fermion propagator. The resulting equations have been solved by numerical iteration.
We applied the Popov-Fedotov method on several approximation levels: the free spin, the Hartree approximation (magnetic mean-field theory), and the above-mentioned self-consistent approximation, using both *average projection* ($\mu^f= 0$) and *exact projection* ($\mu^f= i\frac{\pi}{2}T$). The results obtained for the latter approximation show the expected suppression of the ordered state down to zero temperature, the exponential divergence of the spin correlation length, and a spin-structure factor consistent to the dynamical scaling hypothesis.
A comparison of the results from average and exact projection reveals that there is a significant effect of the exact projection at higher temperatures. In the limit of low temperature, however, the deviation of the average-constraint results from the exact-constraint results become (numerically) indistinguishable, except for the case $J= 0$ (free spins).
In order to support this observation, we calculated the fluctuations $\langle \Delta Q_i\,\Delta Q_j\rangle$ of the auxiliary-fermion charge within the average-projection scheme. We find (by analytical calculation) $\;\lim_{T\to 0}\langle \Delta Q_i\,\Delta Q_j\rangle=
0$, except for the case of free spins, where $\langle \Delta
Q_i\,\Delta Q_j\rangle$ stays finite as $T\to 0$. That is, as long as the spin–spin interaction $J$ is taken into account, the fermion-charge fluctuations into unphysical Hilbert-space states are quenched at $T= 0$. If temperature is increased from zero, we find that $\langle \Delta Q_i\,\Delta Q_j\rangle$ raises continuously with $T$.
These at first sight surprising results find their explanation in the tendency towards antiferromagnetic order in the interacting system, which helps to suppress the fluctuations in the fermion-occupation number: Starting from the physical (“true”) ground state, which features long-range magnetic order[@man91], a fluctuation of the fermion charge $Q_l=1$ at some site $l$ into an unphysical state[@note-qtransfer] with $Q_l= 0$ or $Q_l= 2$ is equivalent to *removing the spin* $\mathbf{S}_l$ in the Hamiltonian (recall Eqs.(\[eqn-destruct\]) and (\[eqn-equalen\])). The lowest-lying state in this unphysical subspace thus lacks the binding energy of the spin at site $l$, which is of order $J$. Therefore, the ground state in the Fock space of arbitrary fermion occupancy is the “true” ground state in the physical segment, and the lowest-lying unphysical state is separated from the ground state by a gap[@note-qflexpon] $\Delta E_Q\sim J$.
Consequently, at low temperatures $T\ll \Delta E_Q$, to a good approximation the exact projection may be omitted in favour of the technically somewhat simpler average-projection approach. At $T= 0$ the approximate treatment of the constraint even becomes exact. Note that $T\ll \Delta E_Q$ does not impose any restriction on the excitation energy $\omega$, e.g., in the structure factor $S(\mathbf{q}, \omega)$: Since the fermion charge is conserved locally, all excitations at any $\omega$ out of the ground state remain in the physical Hilbert space.
The above argument is quite apparent for magnetically ordered systems. However, it should also apply to systems without magnetic order but strong correlations in the ground state. Examples are the various valence bond states discussed for, e.g., Heisenberg models with frustration[@mislhu05]. In these systems the gap $\Delta E_Q$ to unphysical states is also expected to be $\sim J$. Somewhat different examples are systems with a ground state that is dominated by local Kondo singlets. Here the gap $\Delta E_Q$ is exponentially small in $J$, since the binding energy of a localized spin to the Fermi sea is given by the exp. small Kondo energy $T_K$. For calculations in the important temperature range $T\gtrsim T_K$ a solid treatment of the fermion constraint is therefore desirable.
As far as the low-temperature behavior is concerned, the criticism of the auxiliary-particle approach often expressed in view of the uncontrolled handling of the constraint may be refuted on the basis of the results presented here. However, one has to keep in mind that the above arguments are based on the assumption that the approximation method (whether based on self-consistent diagrams or functional integrals) does conserve the local fermion charge $Q_i$.
The Popov–Fedotov approach opens the way to using resummed perturbation theory in specific strongly correlated systems, on the basis of standard Feynman diagrams, and for all temperatures. It requires identifying and performing the summation of physically relevant terms (diagram classes), which, however, remains a challenge for these systems. The self-consistent approximation presented here, for example, still fails to satisfy the notoriously hard to meet sum rule on the local spin moment. More elaborate resummation schemes are necessary to correct this and other deficiencies, the reward being a detailed description of the spin dynamics not accessible by any other analytical method.
Acknowledgements
================
We acknowledge useful discussions with J. Reuther. This work has partially been supported by the Research Unit “Quantum-Phase Transitions” of the Deutsche Forschungsgemeinschaft.
Properties of Green’s functions in the Popov–Fedotov technique {#sec-app-greens}
==============================================================
In this appendix we consider thermal (Matsubara) Green’s functions in the Popov–Fedotov scheme. For some operators $A$ and $B$, which are both either fermionic ($s= +1$) or bosonic ($s= -1$), the Green’s function is defined as[@negorl], $$G(\tau)
= (-s)\langle \mathcal{T}_\tau\{ A(\tau) B(0)\}\rangle^\ppv
\;\;,\;\;\;
-\beta\le \tau< \beta\;\;,$$ with $\beta= 1/T$, the thermal expectation value as defined in Eq.(\[eqn-expect-ppv\]) and (\[eqn-zppdef\]), the Hamiltonian given by Eqs.(\[eqn-grand\]), (\[eqn-hbergham\]), and the usual “time”-ordering symbol $$\mathcal{T}_\tau\{A(\tau) B(0)\}
= A(\tau) B(0) \Theta(\tau) -s\,
B(0) A(\tau) \Theta(-\tau)\;.$$ The fact that $H^\ppv$, Eq.(\[eqn-grand\]), is non-Hermitian, does not influence the (anti-) symmetry properties resulting from the cyclic invariance of the trace. Therefore it is sufficient to consider $\tau> 0$, i.e., $$G(\tau)
= (-s)\frac{1}{Z^\ppv}
\textrm{Tr}^f{}[\,
e^{-(\beta - \tau) H^\ppv} A
e^{-\tau H^\ppv} B\,{}]\;.$$ Using Eqs.(\[eqn-sgl\]) and (\[eqn-grand\]) this becomes $$\begin{aligned}
G(\tau)
& = & \nonumber
(-s)\frac{1}{Z^\ppv}
\sum_{c_Q, n_Q}\sum_{c'_Q, n'_Q} \cdot
\\
& & \nonumber \cdot
\langle c_Q, n_Q|A|c'_Q, n'_Q\rangle
\langle c'_Q, n'_Q|B|c_Q, n_Q\rangle
\cdot
\\
& & \nonumber \cdot
e^{-\beta E(c_Q, n_Q)}\,
e^{\tau{}[\,E(c_Q, n_Q) - E(c'_Q, n'_Q)\,{}]}
\cdot
\\
& & \cdot
\label{eqn-green-spec1}
\big(\prod_{k= 1}^{N_L}e^{i\frac{\pi}{2}Q_k}\big)\,
\big(\prod_{q= 1}^{N_L}
e^{-i\frac{\pi\tau}{2\beta}{}[\,Q_q - Q'_q\,{}]}\big)\end{aligned}$$ $Q_q$ and $Q'_q$ denote the auxiliary charge on lattice site $q$ as it appears in the charge configurations $c_Q$ and $c'_Q$, respectively.
*Physical propagator:* The simplest physical Green’s function is the dynamical spin susceptibility (\[eqn-sus-def\]), (\[eqn-sus-ppv\]), for two lattice sites $l$ and $m$, $$\chi_{lm}^{\mu\bar{\mu}}(\tau)
= \langle \mathcal{T}_\tau\{
S^\mu_l(\tau) S^{\bar{\mu}}_m(0)
\rangle^\ppv\;.$$ Due to the property (\[eqn-destruct\]) of spin operators, the fermion charge on the sites $l, m$ is automatically constrained to 1 in Eq.(\[eqn-green-spec1\]), i.e., $Q_l= Q_m= Q'_l= Q'_m= 1$. For all other sites $p\ne l, m$, the orthonormal matrix elements in Eq.(\[eqn-green-spec1\]) lead to $Q_p= Q'_p$, thus we have $c'_Q=
c_Q$, and the second factor $(\prod_{q= 1}^{N_L}e^{\cdots})$ becomes 1. Now the term $(\prod_{k= 1}^{N_L}\exp(i\frac{\pi}{2}Q_k))$, in combination with the property (\[eqn-equalen\]) of the energies, leads to a cancellation of all unphysical states with charge $Q_p= 0$ and $Q_p= 2$ on any lattice site $p\ne l, m$. Thus, only physical states with a single fermion per site, $c_Q= c'_Q= c_Q^{phys}= (1, 1,
\ldots, 1)$ remain in Eq.(\[eqn-green-spec1\]). Using the notation $$E_n= E(c_Q^{phys}, n_Q) \;\;,\;\;\;
|n\rangle= |c_Q^{phys}, n_Q\rangle$$ for energies and states in the physical subspace, Eq.(\[eqn-green-spec1\]) reads $$\label{eqn-green-phys}
\chi_{lm}^{\mu\bar{\mu}}(\tau)
= \frac{(i)^{N_L}}{Z^\ppv}
\sum_{n, n'}
\langle n|S^\mu_l|n'\rangle
\langle n'|S^{\bar{\mu}}_m|n\rangle
e^{-\beta E_n}
e^{\tau(E_n - E_{n'})}$$ With the result (\[eqn-physpart\]) for the partition function, this is exactly the expression we would have obtained directly, working in the physical Hilbert space.
*Green’s function of the fermions:* The fermion propagator is not a meaningful physical quantity. However, within a self-consistent diagrammatic expansion of, e.g., the dynamical spin susceptibility, the renormalized fermion Green’s function is of technical importance. Therefore it is useful to derive some exact properties of the Green’s function $$\label{eqn-green-fgfdef}
G_{lm,\sigma}(\tau)
= - \langle\mathcal{T}_\tau\{
f_{l \sigma}(\tau) f^\dagger_{m \sigma}(0)\}
\rangle^\ppv\;.$$ The matrix elements in Eq.(\[eqn-green-spec1\]) now read $$\langle c_Q, n_Q|f_{l \sigma}|c'_Q, n'_Q\rangle
\langle c'_Q, n'_Q|f^\dagger_{m \sigma}|c_Q, n_Q\rangle\;.$$ $f^\dagger_{m \sigma}$ increases the auxiliary charge at lattice site $m$ by 1, which can only be compensated by $f_{l \sigma}$, that is, the exact fermion propagator is local, $l= m$. For all sites $p\ne
l$ the arguments from above hold: The orthonormal wave functions lead to $Q_p= Q'_p$, and the unphysical constributions with $Q_p= 0, 2$ cancel. The states that remain in the trace in Eq.(\[eqn-green-spec1\]) then have $Q_p= 1$ at all sites $p\ne l$ and some charge $Q_l, Q'_l\in \{0, 1, 2\}$ at site $l$. With the notation $$|Q, n\rangle^l= |c_q, n_Q\rangle
\quad\textrm{for}\quad
Q_l= Q \;\;,\;\;\;
Q_p= 1\,, \;p\ne l\;\;,$$ and similarly $E^l(Q, n)$ for the eigenenergies, the Green’s function (\[eqn-green-spec1\]), (\[eqn-green-fgfdef\]) becomes $$\begin{aligned}
\label{eqn-green-gferm}
G_{lm, \sigma}(\tau)
& = & (-1)\delta_{lm}\frac{(i)^{N_L - 1}}{Z^\ppv}
\sum_{Q= 0, 1}
\sum_{n, n'}
\cdot
\\
& & \mbox{} \nonumber
\cdot
|\,{}^l\!\langle Q, n|f_{l \sigma}|Q+1, n'\rangle^l|^2
\cdot
\\
& & \mbox{} \nonumber
\cdot
\exp\big({-\beta E^l(Q, n)}\big)\,
\exp\big({i\frac{\pi}{2}Q}\big)
\cdot
\\
& & \mbox{} \nonumber
\cdot
\exp\big({\tau{}\big[\,E^l(Q, n) - E^l(Q + 1, n')
+ i\frac{\pi}{2\beta}\,{}\big]}\big)\end{aligned}$$ The energies and states that occur in Eq.(\[eqn-green-gferm\]) are now denoted by $$\begin{aligned}
Q= 1\,: & &
E^l(1, n) = E_n\;\;,
\\
Q= 0, 2\,: & &
E^l(0, n') = E^l(2, n') = E^l_{n'}\;\;,\end{aligned}$$ and $$\begin{aligned}
Q= 1\,: & &
|1, n\rangle^l = |n\rangle\;\;,
\\
Q= 0\,: & &
f^\dagger_{l \sigma}|0, n'\rangle^l = |\sigma, n'\rangle^l\;\;,
\\
Q= 2\,: & &
f_{l \sigma}|2, n'\rangle^l = \pm|-\sigma, n'\rangle^l\;.\end{aligned}$$ According to Eqs.(\[eqn-physconfig\]), (\[eqn-sgl\]) the $E_n$ and $|n\rangle$ are the eigenenergies and -states of the model Hamiltonian $H$, Eq.(\[eqn-hbergham\]). Referring to Eqs.(\[eqn-destruct\]) and (\[eqn-equalen\]), the $E_{n'}^l$ can be interpreted as the eigenenergies of $H$ *with a “defect” at site $l$*, i.e., with all couplings $J$ to the spin at site $l$ set to zero. The states $|\sigma, n'\rangle^l$ therefore contain the orientation $\sigma\in\{\uparrow, \downarrow\}$ of the resulting free spin at site $l$ as a good quntum number. The set of quantum numbers $n'$ as well as the $E_{n'}^l$ do not depend on $\sigma$. Note that the number of states is the same, $\#\{|n\rangle\}= \#\{|\sigma,
n'\rangle\}= 2^{N_L}$.
The fermion Green’s function (\[eqn-green-spec1\]) now reads, $$\begin{aligned}
\label{eqn-green-gtau}
G_{l, \sigma}(\tau)
& = & \frac{-1}{Z}
\sum_{n, n'}\cdot
\\
& & \nonumber
\mbox{}\cdot\bigg\{
e^{-\beta E_n}
|\langle n| -\sigma, n'\rangle^l |^2
e^{\tau{}[\,E_n - E^l_{n'} + i\frac{\pi}{2\beta}\,{}]} -
\\
& & \nonumber
\mbox{} -
i\,e^{-\beta E^l_{n'}}
|\langle n| \sigma, n'\rangle^l |^2
e^{\tau{}[\,E^l_{n'} - E_n + i\frac{\pi}{2\beta}\,{}]}
\bigg\}\end{aligned}$$
In frequency space, the fermion propagator is given by $$\label{eqn-green-mats}
G_{lm, \sigma}(i\omega)
= \int_0^\beta\mathrm{d}\tau\,
e^{i\omega \tau}
G_{lm, \sigma}(\tau)\;\;,$$ with the fermionic (odd) Matsubare frequency $\omega= (2n +
1)\frac{\pi}{\beta}$. Inserting Eq.(\[eqn-green-gtau\]) into Eq.(\[eqn-green-mats\]) and utilizing the earlier result Eq.(\[eqn-physpart\]) for $Z^\ppv$, we obtain the fermion Green’s function, $$\label{eqn-green-fgfspec}
G_{lm, \sigma}(i\omega)
= \delta_{lm}
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\frac{\widehat{G}_{l, \sigma}(\varepsilon)}
{i\omega + i\frac{\pi}{2\beta} - \varepsilon}\;\;,$$ with the complex-valued spectral function $$\label{eqn-green-cmplxspec}
\widehat{G}_{l, \sigma}(\varepsilon)
= \rho^1_{l, \sigma}(\varepsilon) +
i\,\rho^2_{l, \sigma}(\varepsilon)\;\;,$$ $$\begin{aligned}
\rho^1_{l, \sigma}(\varepsilon)
& = & \frac{1}{Z}
\sum_{n, n'} e^{-\beta E_n}\cdot
\\
& & \cdot\bigg\{
|\langle n|\sigma, n'\rangle^l|^2\,
\delta\big(\varepsilon - {}[\,E_n - E^l_{n'}{}]\big) +
\\
& & \mbox{} +
|\langle n|-\sigma, n'\rangle^l|^2\,
\delta\big(\varepsilon - {}[\,E^l_{n'} - E_n{}]\big)
\bigg\}\end{aligned}$$ $$\begin{aligned}
\rho^2_{l, \sigma}(\varepsilon)
& = & \frac{1}{Z}
\sum_{n, n'} e^{-\beta E^l_{n'}}\cdot
\\
& & \cdot\bigg\{
|\langle n|-\sigma, n'\rangle^l|^2\,
\delta\big(\varepsilon - {}[\,E^l_{n'} - E_n{}]\big) -
\\
& & \mbox{} -
|\langle n|\sigma, n'\rangle^l|^2\,
\delta\big(\varepsilon - {}[\,E_n - E^l_{n'}{}]\big)
\bigg\}\end{aligned}$$
*Sum rule and symmetry:* The $|n\rangle$ as well as the $|\sigma,
n'\rangle^l$ form a complete normalized basis in the physical Hilbert space, $$\sum_n|n\rangle\langle n|= 1 \;\;,\;\;\;
\sum_{n', \sigma}
{}^l\!|\sigma, n'\rangle\langle \sigma, n'|^l
= 1\;\;,$$ and therefore integrating Eqs.(\[eqn-green-cmplxspec\]) over $\varepsilon$ leads to the sum rule $$\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\rho^1_{l, \sigma}(\varepsilon)
= 1
\;\;,\;\;\;
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\rho^2_{l, \sigma}(\varepsilon)
= 0\;\;,$$ $$\label{eqn-green-sum}
\;\;\;\Rightarrow\;\;\;
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\widehat{G}_{l, \sigma}(\varepsilon)
= 1\;.$$
From Eq.(\[eqn-green-cmplxspec\]) we can also read off a “particle–hole” symmetry, $$\rho^1_{l, \uparrow}(\varepsilon)
= \rho^1_{l, \downarrow}(-\varepsilon)
\;\;,\;\;\;
\rho^2_{l, \uparrow}(\varepsilon)
= - \rho^2_{l, \downarrow}(-\varepsilon)\;\;,$$ $$\label{eqn-green-phsym}
\;\;\;\Rightarrow\;\;\;
\widehat{G}_{l, \uparrow}(\varepsilon)
= \widehat{G}_{l, \downarrow}(- \varepsilon)^\ast\;.$$ In the paramagnetic phase, where the overlaps in Eq.(\[eqn-green-cmplxspec\]) are spin degenerate, $$\langle n|\uparrow, n'\rangle^l =
\langle n|\downarrow, n'\rangle^l\;\;,$$ Eq.(\[eqn-green-phsym\]) simplyfies to the result already quoted in Eq.(\[eqn-phsymm\]).
*The expectation value $\langle Q_l\rangle$:* In order to conclude this appendix, we discuss the average auxiliary charge $\langle Q_l\rangle$ at site $l$.
In the enlarged Hilbert space the expectation value $\langle
Q_l\rangle^\ppv$ can be formally calculated; however, although $Q_l$ is a gauge-invariant operator, it does not fulfill the property (\[eqn-destruct\]) of physical observables, and therefore the result becomes meaningless. This is most easily demonstrated by explicitly calculating $\langle Q_l\rangle^\ppv$:
Using the Green’s function (\[eqn-green-fgfdef\]) it may be written as $$\langle Q_l\rangle^\ppv
= \sum_{\sigma}
G_{l, \sigma}(\tau= 0_-)\;.$$ In Eq.(\[eqn-green-gtau\]) the propagator has been given for $\tau>
0$, which can be utilized by help of the anti-symmetric property of fermionic Green’s functions, $$\tau< 0\,: \quad
G(\tau)= - G(\tau + \beta)\;.$$ Thus we find from Eq.(\[eqn-green-gtau\]), setting $\tau=
\beta$, $$\begin{aligned}
\langle Q_l\rangle^\ppv
& = & \frac{1}{Z}\sum_n e^{- \beta E_n}\langle n|n\rangle +
\\
& & \mbox{} +
i\,\frac{1}{Z}\sum_{n'}\sum_{\sigma}
e^{-\beta E^l_{n'}}
\,{}^l\!\langle \sigma, n'|\sigma, n'\rangle^l
\\
& = & \frac{1}{Z}(Z + i Z^l)\end{aligned}$$ Here $Z$ is the partition function of $H$, Eq.(\[eqn-hbergham\]), in the physical subspace, while $Z^l$ is the partition function of $H$ with the “defect” at site $l$, i.e., with all couplings $J$ to the site $l$ made zero. Since all interactions in $H$ are short ranged, $Z$ and $Z^l$ become equal in the thermodynamic limit, $$\label{eqn-green-qexpect}
N_L\to\infty\,: \quad
\langle Q_l\rangle^\ppv
= (1 + i\frac{Z^l}{Z})
\;\to\; (1 + i)$$ Note that $Z$ and $Z^l$ contain the same number of states, $(2)^{N_L}$.
Derivation of the self-consistent Equations (\[eqn-nflex\]) {#sec-app-deriv}
===========================================================
In this appendix the intermediate steps in going from Eqs.(\[eqn-sflex\]) to Eqs.(\[eqn-nflex\]) are presented.
Starting point is the spectral representation (\[eqn-gfspect\]) or (\[eqn-green-fgfspec\]) of the fermion Green’s function (\[eqn-sflex-dyson\]). The Matsubara frequency $i\omega$ can be analytically continued to the complex plane, $i\omega\to z$, with $G(z)$ showing a branch cut at $\textrm{Im}(z)=
-\frac{\pi}{2\beta}$. Close to this cut, at $z= \omega -
i\frac{\pi}{2\beta} \mp i0_+$, ($0_+$ is a positive infinitesimal) we have $$\label{eqn-deriv-fgfdecomp}
G(\omega - i\frac{\pi}{2\beta} \mp i0_+)
= \pm i\pi \widehat{G}(\omega) + \overline{G}(\omega)\;\;,$$ with the spectral function $\widehat{G}$ and its Hilbert transform $$\overline{G}(\omega)
= \textrm{P}\!\!
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\frac{\widehat{G}(\varepsilon)}{\omega - \varepsilon}\;.$$ For the susceptibilities $\Pi(i\nu)$ and $D(i\nu)$, appearing in Eq.(\[eqn-sflex-bubble\]) and (\[eqn-sflex-effint\]), the usual analytic continuation of the bosonic Matsubara frequency $i\nu$ to the real axis applies, $i\nu\to \omega + i0_+$, $$\label{eqn-deriv-cont}
\begin{array}[c]{rcl} \displaystyle
\Pi(i\nu)
& \to &
\Pi(\omega + i0_+) = \Pi'(\omega) + i\,\Pi''(\omega)\;\;,
\\ \\ \displaystyle
D(i\nu)
& \to &
D(\omega + i0_+) = D'(\omega) + i\,D''(\omega)\;.
\end{array}$$ The imaginary part $D''(\omega)$ represents the spectral function of the effective local interaction, $$\label{eqn-deriv-effint}
D(i\nu)
= \frac{1}{\pi}
\int_{-\infty}^\infty\mathrm{d}\varepsilon\,
\frac{D''(\varepsilon)}{\varepsilon - i\nu}\;.$$ Note that $D''(\omega)$ obeys the symmetry $$\label{eqn-deriv-dsymm}
D''(-\omega)= - D''(\omega)\;\;,$$ which comes from $\chi(\mathbf{q}, \omega)= \chi^\ast(-\mathbf{q},
-\omega)$ in Eq.(\[eqn-sflex-effint\]). For $\Pi(i\nu)$, equations similar to (\[eqn-deriv-effint\]) and (\[eqn-deriv-dsymm\]) hold.
The fermion self-energy (\[eqn-sflex-self\]) is re-written using the spectral representations (\[eqn-green-fgfspec\]) and (\[eqn-deriv-effint\]), $$\begin{aligned}
\Sigma(i\omega)
& = & -\frac{3}{4\pi}
\int\mathrm{d}\varepsilon\,
D''(\varepsilon)\,
\int\mathrm{d}\varepsilon'\,
\widehat{G}(\varepsilon')\,\cdot
\\
& & \mbox{}\quad\cdot
\frac{1}{\beta}
\sum_{i\nu}
\frac{1}{(i\nu - \varepsilon)
(i\nu + i\omega + i\frac{\pi}{2\beta} - \varepsilon')}
\\
& = & \frac{3}{4\pi}
\int\mathrm{d}\varepsilon\,
D''(\varepsilon)\,
\int\mathrm{d}\varepsilon'\,
\widehat{G}(\varepsilon')
\frac{g(\varepsilon) + f(\varepsilon' - i\frac{\pi}{2\beta})}
{i\omega + i\frac{\pi}{2\beta} + \varepsilon - \varepsilon'}\end{aligned}$$ $f$ and $g$ stand for the Fermi and Bose function. Apparently, $\Sigma(i\omega)$ obeys a spectral representation similar to Eq.(\[eqn-green-fgfspec\]), namely, $$\label{eqn-deriv-selfspec}
\Sigma(i\omega)
= \int\mathrm{d}\varepsilon\,
\frac{\widehat{\Sigma}(\varepsilon)}
{i\omega + i\frac{\pi}{2\beta} - \varepsilon}\;\;,$$ with the (complex valued) spectral function $$\widehat{\Sigma}(\omega)
= \frac{3}{4\pi}\int\mathrm{d}\varepsilon\,
D''(\varepsilon)
\widehat{G}(\varepsilon + \omega)
{}[\,g(\varepsilon) + f(\varepsilon + \omega - i\frac{\pi}{2\beta})\,{}]$$ and its Hilbert transform $\overline{\Sigma}$, given in Eq.(\[eqn-nflex-selfhil\]) above.
Introducing the structure factor of the renormalized interaction, $$U(\omega)
= {}[\,1 + g(\omega)\,{}] D''(\omega)
\;\;,\;\;\;$$ which is by Eq.(\[eqn-deriv-dsymm\]) equivalent to $$U(-\omega)
= g(\omega) D''(\omega)
\;\;,\;\;\;$$ and using the relation $$g + f
= g(1 - f) + (1 + g)f
\;\;,\;\;\;$$ the spectrum $\widehat{\Sigma}$ takes the form Eq.(\[eqn-nflex-selfspec\]), with the short hands $\widehat{G}^+$ and $\widehat{G}^-$ defined in Eqs.(\[eqn-nflex-gfplus\]) and (\[eqn-nflex-gfmin\]).
The fermion spectrum $\widehat{G}$ is obtained from the Dyson’s equation (\[eqn-sflex-dyson\]) using Eq.(\[eqn-deriv-fgfdecomp\]), i.e., $$\begin{aligned}
\widehat{G}(\omega)
& = & \frac{1}{2\pi i}\left[\,
\frac{1}{\omega - \Sigma(\omega - i\frac{\pi}{2\beta} - i0_+)}
\right. -
\\
& & \mbox{}- \left.
\frac{1}{\omega - \Sigma(\omega - i\frac{\pi}{2\beta} + i0_+)}
\,\right]\end{aligned}$$ By inserting the decomposition $$\Sigma(\omega - i\frac{\pi}{2\beta} \mp i0_+)
= \pm i\pi \widehat{\Sigma}(\omega) + \overline{\Sigma}(\omega)\;\;,$$ which results from the spectral representation (\[eqn-deriv-selfspec\]) and Eq.(\[eqn-nflex-selfhil\]), we obtain $\widehat{G}$ as given in Eq.(\[eqn-nflex-gf\]).
In the fermion bubble $\Pi(i\nu)$, Eq.(\[eqn-sflex-bubble\]), the spectral representation (\[eqn-green-fgfspec\]) of the fermion Green’s function is inserted, and we arrive at $$\Pi(i\nu)
= \frac{1}{2}
\int\mathrm{d}\varepsilon\,
\widehat{G}(\varepsilon)\,
\int\mathrm{d}\varepsilon\,
\widehat{G}(\varepsilon')\,
\frac{f(\varepsilon - i\frac{\pi}{2\beta}) -
f(\varepsilon' - i\frac{\pi}{2\beta})}
{i\nu - \varepsilon + \varepsilon'}$$ Note that the imaginary-valued chemical potential $\mu^f=
i\frac{\pi}{2\beta}$ cancels in the denominator, since $\Pi$ represents an observable susceptibility. Apparently, $\Pi(i\nu)$ obeys the usual spectral representation, similar to Eq.(\[eqn-deriv-effint\]), with the imaginary part $$\begin{aligned}
\label{eqn-deriv-piimag}
\Pi''(\omega)
& = & \frac{\pi}{2}
\int\mathrm{d}\varepsilon\,
\widehat{G}(\varepsilon)
\widehat{G}(\varepsilon - \omega)\,\cdot
\\ \nonumber
& & \mbox{}\quad\cdot
{}[\,f(\varepsilon - \omega - i\frac{\pi}{2\beta}) -
f(\varepsilon - i\frac{\pi}{2\beta})\,{}]\end{aligned}$$ and the corresponding real part $\Pi'$ is computed via Eq.(\[eqn-nflex-pire\]).
It is convenient to introduce a structure factor for the bubble, $$S^0(\omega)
= {}[\,1 + g(\omega)\,{}]\,
\Pi''(\omega)\;\;,$$ and with the relation $${}[\,f(x - y) - f(x)\,{}]
{}[\,1 + g(y)\,{}]
= {}[\,1 - f(x)\,{}]\,f(x - y)\;\;,$$ which is valid for arbitrary complex numbers $x$, $y$, we have $$S^0(\omega)
= \frac{\pi}{2}
\int\mathrm{d}\varepsilon\,
\widehat{G}(\varepsilon)
\widehat{G}(\varepsilon - \omega)\,
{}[\,1 - f(\varepsilon - i\frac{\pi}{2\beta})\,{}]\,
f(\varepsilon - \omega - i\frac{\pi}{2\beta})$$ Using the notation $\widehat{G}^+$, $\widehat{G}^-$ introduced Eqs.(\[eqn-nflex-gfplus\]), (\[eqn-nflex-gfmin\]), the result stated in Eq.(\[eqn-nflex-snul\]) immediately follows.
In order to compute $\Pi''(\omega)$ from $S^0(\omega)$, Eq.(\[eqn-nflex-piim\]) is used, which is a consequence of the symmetry $\Pi''(-\omega)= -\Pi''(\omega)$.
The last equation to be derived in this appendix is Eq.(\[eqn-nflex-effint\]) for the effective interaction $U(\omega)$. Performing the analytic continuation $i\nu\to (\omega +
i0_+)$ in Eqs.(\[eqn-sflex-effint\]), (\[eqn-sflex-chi\]), and using the decomposition (\[eqn-deriv-cont\]), we find $$D''(\omega)
= \frac{1}{N_L}
\sum_{\mathbf{q}}
J^2(\mathbf{q})
\frac{\Pi''(\omega)}
{\big|1 + J(\mathbf{q})\Pi(\omega + i0_+)\big|^2}\;.$$ Eq.(\[eqn-nflex-effint\]) is now obtained using the definition of $U(\omega)$ and $S^0(\omega)$ given in this appendix and the density-of-states $\mathcal{N}(\varepsilon)$ introduced in Eq.(\[eqn-dos\]).
*Particle–hole symmetry:* In the Appendix \[sec-app-greens\] above, a symmetry for the spectrum $\widehat{G}$ of the fermion Green’s function has been derived in Eq.(\[eqn-green-phsym\]), namely $$\widehat{G}(-\omega)
= \widehat{G}(\omega)^\ast\;.$$ Accordingly, the spectra $\widehat{G}^+$ and $\widehat{G}^-$ introduced in Eqs.(\[eqn-nflex-gfplus\]), (\[eqn-nflex-gfmin\]) obey the relation $$\label{eqn-simpl-phsymm}
\widehat{G}^-(-\omega)
= \widehat{G}^{+}(\omega)^\ast\;.$$ This may be used to simplify the Eqs.(\[eqn-nflex\]) somewhat by eliminating $\widehat{G}^-$: Applying Eq.(\[eqn-simpl-phsymm\]) to Eq.(\[eqn-nflex-selfspec\]) leads to
$$\label{eqn-simpl-selfspec}
\widehat{\Sigma}(\omega)
= \frac{3}{4\pi}
\int\mathrm{d}\varepsilon\,
U(\varepsilon)
\,{}[\,
\widehat{G}^+(\omega - \varepsilon) \;+\;
\widehat{G}^{+}(-\omega - \varepsilon)^\ast
\,{}]\;.$$
For $S^0$, we start from Eq.(\[eqn-nflex-snul\]) by writing the expression twice and using the symmetry (\[eqn-simpl-phsymm\]) in the second term, $$S^0(\omega)
= \frac{\pi}{4}
\int\mathrm{d}\varepsilon\,
{}[\,\widehat{G}^+(\varepsilon)
\widehat{G}^-(\varepsilon - \omega) \;+\;
\widehat{G}^{-}(-\varepsilon)^\ast
\widehat{G}^{+}(\omega -\varepsilon)^\ast
\,{}]$$ By renamimg $\varepsilon\to(\omega - \varepsilon)$ in the second term and applying Eq.(\[eqn-simpl-phsymm\]) once more, it follows $$\label{eqn-simpl-snul}
S^0(\omega)
= \frac{\pi}{2}
\int\mathrm{d}\varepsilon\,
\textrm{Re}\big\{
\widehat{G}^+(\varepsilon)\,
\widehat{G}^{+}(\omega - \varepsilon)^\ast
\big\}\;.$$
The remaining equations in (\[eqn-nflex\]) stay unchanged, except that Eq.(\[eqn-nflex-gfmin\]) becomes obsolete.
The self-consistent equations using real spectra {#sec-app-simpl}
================================================
For a direct comparison of the self-consistent equations with those derived within average projection in Ref. (Eqs.(A1) in that reference), we find it instructive to re-write the Eqs.(\[eqn-nflex\]) entirely in real-valued spectral functions. To that end we decompose all unphysical spectra into real and imaginary parts as follows,
\[eqn-simpl-decomp\] $$\begin{aligned}
\widehat{G}(\omega)
& = & \rho_1(\omega) + i\rho_2(\omega)\;\;,
\\
\widehat{G}^+(\omega)
& = & \rho_1^+(\omega) + i\rho_2^+(\omega)\;\;,
\\
\widehat{\Sigma}(\omega)
& = & \hat{\sigma}_1(\omega) + i\hat{\sigma}_2(\omega)\;\;,
\\
\overline{\Sigma}(\omega)
& = & \bar{\sigma}_1(\omega) + i\bar{\sigma}_2(\omega)\;.\end{aligned}$$
Inserting these definitions in Eqs.(\[eqn-nflex\]), with Eqs.(\[eqn-nflex-selfspec\]) and (\[eqn-nflex-snul\]) replaced by Eqs.(\[eqn-simpl-selfspec\]) and (\[eqn-simpl-snul\]), we find the set of equations (\[eqn-simpl-num\]) stated below. For completeness, in (\[eqn-simpl-num\]) we also quote those equations from (\[eqn-nflex\]), that remain unchanged by using (\[eqn-simpl-decomp\]). Making use of Eq.(\[eqn-fermicmplx\]) the result reads
\[eqn-simpl-num\] $$\begin{aligned}
U(\omega)
& = & S^0(\omega)
\int\mathrm{d}\varepsilon\,
\frac{\mathcal{N}(\varepsilon)\,\varepsilon^2}
{\big|1 + \varepsilon\,\Pi(\omega + i0_+)\big|^2}
\\
\Pi''(\omega)
& = & S^0(\omega) - S^0(-\omega)
\\
\Pi'(\omega)
& = & \frac{1}{\pi}
\textrm{P}\!\!\int\mathrm{d}\varepsilon\,
\frac{\Pi''(\varepsilon)}{\varepsilon - \omega}
\\
S^0(\omega)
& = & \frac{\pi}{2}
\int\mathrm{d}\varepsilon\,
{}\big[\,\rho_1^+(\varepsilon) \rho_1^+(\omega - \varepsilon)
+ \\
& & \nonumber
\mbox{}\quad\quad
+ \rho_2^+(\varepsilon) \rho_2^+(\omega - \varepsilon)
\,{}\big]
\\
\rho_1^+(\omega)
& = & {}[\,1 - f(2\omega)\,{}]\,\rho_1(\omega) \;+\;
\frac{\rho_2(\omega)}{2\cosh(\beta\omega)}
\\
\rho_2^+(\omega)
& = & {}[\,1 - f(2\omega)\,{}]\,\rho_2(\omega) \;-\;
\frac{\rho_1(\omega)}{2\cosh(\beta\omega)}
\\
\rho_1(\omega)
& = & \\ \nonumber
\lefteqn{\textrm{Re}
\frac{\hat{\sigma}_1(\omega) + i\hat{\sigma}_2(\omega)}
{(A_+ A_- - B_+ B_-) - i(B_+ A_- + A_+ B_-)}}
\\
\rho_2(\omega)
& = & \\ \nonumber
\lefteqn{\textrm{Im}
\frac{\hat{\sigma}_1(\omega) + i\hat{\sigma}_2(\omega)}
{(A_+ A_- - B_+ B_-) - i(B_+ A_- + A_+ B_-)}}
\\
\hat{\sigma}_1(\omega)
& = & \frac{3}{4\pi}\int\mathrm{d}\varepsilon\,U(\varepsilon)
{}[\,\rho_1^+(\omega - \varepsilon) +
\\ \nonumber
& & \mbox{}\quad\quad
+ \rho_1^+(-\omega - \varepsilon)\,{}]
\\
\hat{\sigma}_2(\omega)
& = & \frac{3}{4\pi}\int\mathrm{d}\varepsilon\,U(\varepsilon)
{}[\,\rho_2^+(\omega - \varepsilon) -
\\ \nonumber
& & \mbox{}\quad\quad
- \rho_2^+(-\omega - \varepsilon)\,{}]
\\
\bar{\sigma}_1(\omega)
& = & \textrm{P}\!\!\int\mathrm{d}\varepsilon\,
\frac{\hat{\sigma}_1(\varepsilon)}{\omega - \varepsilon}
\\
\bar{\sigma}_2(\omega)
& = & \textrm{P}\!\!\int\mathrm{d}\varepsilon\,
\frac{\hat{\sigma}_2(\varepsilon)}{\omega - \varepsilon}\end{aligned}$$
The short hands $A_{\pm}$, $B_{\pm}$ are defined as $$\begin{aligned}
A_{\pm}
& = & \omega - \bar{\sigma}_1(\omega)
\pm \pi\hat{\sigma}_2(\omega)
\\
B_{\pm}
& = & \bar{\sigma}_2(\omega) \pm \pi\hat{\sigma}_1(\omega)\end{aligned}$$
Calculation of the auxiliary-charge response (\[eqn-flex-qbub\]) {#sec-app-qbub}
================================================================
In this appendix the intermediate steps in going from Eq.(\[eqn-flex-bethe\]) to Eq.(\[eqn-flex-bethe2\]) are explained. We start by writing Eq.(\[eqn-flex-bethe\]) in the form $$\begin{aligned}
\label{eqn-qbub-bethe}
\Gamma(\mathbf{q}, i\omega)
& = & 1 + \sum_{i\omega_1}
\big[\,A(i\omega, i\omega_1) +
\\ \nonumber
& & \quad\mbox{} +
B(\mathbf{q}; i\omega, i\omega_1)\,\big]
\,\Gamma(\mathbf{q}, i\omega_1)\end{aligned}$$ with the matrices $$A(i\omega, i\omega_1)
= \frac{3}{4}T\,G(i\omega_1)^2 D(i\omega - i\omega_1)$$ and $B(\mathbf{q}; i\omega, i\omega_1)$, the latter can be read off Eq.(\[eqn-flex-bethe\]). For the constituents of $A$ and $B$ we observe the symmetries $$G(-i\omega)= -G(i\omega) \;\;,\;\;\;
D(\mathbf{k}, -i\nu)
= D(\mathbf{k}, i\nu) \;\;,\;\;\;$$ which implies
\[eqn-qbub-symm\] $$A(-i\omega, -i\omega_1)
= A(i\omega, i\omega_1)$$ and $$\begin{array}[c]{rcl}
B(-i\omega, i\omega_1)
& = & - B(i\omega, i\omega_1)\;\;,
\\[1ex]
B(i\omega, -i\omega_1)
& = & - B(i\omega, i\omega_1)\;.
\end{array}$$
Here and in the following the wave vector $\mathbf{q}$ is not written. The vertex function is split into components behaving symmetrical $(+)$ or anti-symmetrical $(-)$ under $i\omega\to -i\omega$, $$\Gamma(i\omega)
= \Gamma^{(+)}(i\omega) + \Gamma^{(-)}(i\omega)\;\;,$$ corresponding to $$\Gamma^{(\pm)}(i\omega)
= \frac{1}{2}{}[\,\Gamma(i\omega) \pm \Gamma(-i\omega)\,{}]\;.$$ Using these definitions and the symmetry relations (\[eqn-qbub-symm\]) in Eq.(\[eqn-qbub-bethe\]), it can be seen that $\Gamma^{(+)}$ and $\Gamma^{(-)}$ decouple,
\[eqn-qbub-split\] $$\label{eqn-qbub-splitplus}
\Gamma^{(+)}(i\omega)
= 1 + \sum_{i\omega_1}
A(i\omega, i\omega_1)\,
\Gamma^{(+)}(i\omega_1)\;\;,$$ $$\Gamma^{(-)}(i\omega)
= \sum_{i\omega_1}
\big[\,A(i\omega, i\omega_1) +
B(i\omega, i\omega_1)\,\big]\,
\Gamma^{(-)}(i\omega_1)\;.$$
In the fermion bubble (\[eqn-flex-qbub\]) only the symmetric component contributes (not writing $\mathbf{q}$), $$\chi^Q(0)
= -2T\sum_{i\omega}G(i\omega)^2\,
\Gamma^{(+)}(i\omega)\;.$$ Therefore, by identifying $\Gamma^{(+)}$ with $\Gamma$ in Eq.(\[eqn-qbub-splitplus\]), we arrive at Eq.(\[eqn-flex-bethe2\]) for the vertex function to be used in the fermion-charge response (\[eqn-flex-qbub\]).
[40]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , ****, ().
, ** (, , ).
, ****, ().
, ****, ().
, ** (, , ).
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, in **, edited by (, , ), p. .
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, in **, edited by (, ), .
, , , , ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '*Application and use of deep learning algorithms for different healthcare applications is gaining interest at a steady pace. However, use of such algorithms can prove to be challenging as they require large amounts of training data that capture different possible variations. This makes it difficult to use them in a clinical setting since in most health applications researchers often have to work with limited data. Less data can cause the deep learning model to over-fit. In this paper, we ask how can we use data from a different environment, different use-case, with widely differing data distributions. We exemplify this use case by using single-sensor accelerometer data from healthy subjects performing activities of daily living - ADLs (source dataset), to extract features relevant to multi-sensor accelerometer gait data (target dataset) for Parkinson’s disease classification. We train the pre-trained model using the source dataset and use it as a feature extractor. We show that the features extracted for the target dataset can be used to train an effective classification model. Our pre-trained source model consists of a convolutional autoencoder, and the target classification model is a simple multi-layer perceptron model. We explore two different pre-trained source models, trained using different activity groups, and analyze the influence the choice of pre-trained model has over the task of Parkinson’s disease classification.*'
author:
- 'Anirudh Som^1,2^, Narayanan Krishnamurthi^3^, Matthew Buman^4^, Pavan Turaga^1,2^'
bibliography:
- 'egbib.bib'
date: June 2019
title: '**Unsupervised Pre-trained Models from Healthy ADLs Improve Parkinson’s Disease Classification of Gait Patterns**'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider here pseudo-differential operators whose symbol $\sigma(x,\xi)$ is not infinitely smooth with respect to $x$. Decomposing such symbols into four -sometimes five- components and using tools of paradifferential calculus, we derive sharp estimates on the action of such pseudo-differential operators on Sobolev spaces and give explicit expressions for their operator norm in terms of the symbol $\sigma(x,\xi)$. We also study commutator estimates involving such operators, and generalize or improve the so-called Kato-Ponce and Calderon-Coifman-Meyer estimates in various ways.'
address: |
MAB, Université Bordeaux I et CNRS, UMR 5466\
351 Cours de la Libération,\
33405 Talence Cedex, France.
author:
- David Lannes
title: Sharp estimates for pseudodifferential operators with symbols of limited smoothness and commutators
---
pseudodifferential operators ,paradifferential calculus , commutator estimates
Introduction
============
General setting and description of the results
----------------------------------------------
Among the widely known properties of pseudo-differential operators with symbol in Hormander’s class $S^m_{1,0}$, two are discussed in this paper. The first one concerns their action on Sobolev spaces, and the second one deals with the properties of commutators.
It is a classical result that for all $\sigma^1\in S^{m_1}_{1,0}$, the operator ${\mbox{\textnormal{Op}}}(\sigma^1)$ maps $H^{s+m_1}({{\mathbb R}}^d)$ into $H^s({{\mathbb R}}^d)$ for all $s\in{{\mathbb R}}$. Moreover, the proof shows that $$\label{intro1}
\forall u\in H^{s+m_1}({{\mathbb R}}^d),\qquad
\vert {\mbox{\textnormal{Op}}}(\sigma^1)u\vert_{H^s}\leq \underline{C}(\sigma^1)
\vert u\vert_{H^{s+m_1}}.$$
Concerning the study of commutators, Taylor (following works of Moser [@Moser] and Kato-Ponce [@Kato-Ponce]) proved in [@TaylorM0] that for all $\sigma^1\in S^{m_1}_{1,0}$, with $m_1>0$, and all $\sigma^2\in H^\infty({{\mathbb R}}^d)$, one has for all $s\geq 0$, $$\label{intro2}
\big\vert [{\mbox{\textnormal{Op}}}(\sigma^1),\sigma^2]u\big\vert_{H^s}
\leq \underline{C}(\sigma^1)\big(\vert\sigma^2\vert_{W^{1,\infty}}
\vert u\vert_{H^{s+m_1-1}}
+\vert \sigma^2\vert_{H^{s+m_1}}\vert u\vert_\infty\big);$$ another well-known commutator estimate is the so-called Calderon-Coifman-Meyer estimate: if $m_1\geq 0$ then for all $s\geq 0$ and $t_0>d/2$ such that $s+m_1\leq t_0+1$, one has (see Prop. 4.2 of [@TaylorM] for instance): $$\label{intro3}
\big\vert [{\mbox{\textnormal{Op}}}(\sigma^1),\sigma^2]u\big\vert_{H^s}
\leq \underline{C}(\sigma^1)\vert \sigma^2\vert_{H^{t_0+1}}
\vert u\vert_{H^{s+m_1-1}}.$$ A drawback of (\[intro1\]), (\[intro2\]) and (\[intro3\]) is that the dependence of the constant $\underline{C}(\sigma^1)$ on $\sigma^1$ is not specified. This may cause these estimates to be inoperative in the study of some nonlinear PDE; indeed, when solving such an equation by an iterative scheme, one is led to study the pseudo-differential operator corresponding to the linearized equations around some reference state. Generally, the symbol of this operator can be written $\sigma(x,\xi)=\Sigma(v(x),\xi)$, where $\Sigma(v,\xi)$ is smooth with respect to $v$ and of order $m$ with respect to $\xi$, while $v(\cdot)$ belongs to some Sobolev space $H^s({{\mathbb R}}^d)$. For instance, in the study of nonlinear water waves, one is led to study the operator associated to the symbol (see [@Lannes]) $$\label{symbDN}
\sigma(x,\xi):=\sqrt{(1+\vert\nabla a\vert^2)\vert\xi\vert^2-
(\nabla a\cdot\xi)^2},$$ which is of the form described above, with $\Sigma(v,\xi)=\sqrt{(1+\vert v\vert^2)\vert\xi\vert^2-(v\cdot\xi)^2}$ and $v(\cdot)=\nabla a$. Such symbols $\sigma(x,\xi)$ are not infinitely smooth with respect to $x$, since their regularity is limited by the regularity of the function $v$. One must therefore be able to handle symbols of limited smoothness to deal with such situations; moreover, one must be able to say which norms of $v(\cdot)$ are involved in the constant $\underline{C}(\sigma^1)$ of (\[intro1\]), (\[intro2\]) and (\[intro3\]).\
But even knowing precisely the way the constants $\underline{C}(\sigma^1)$ depend on $\sigma^1$, the estimates (\[intro1\]) and (\[intro2\]) may not be precise enough in some situations. Indeed, when one has to use, say, a Nash-Moser iterative scheme, *tame* estimates are needed. For instance, in such situations, the product estimate $\vert uv\vert_{H^s}
\lesssim \vert u\vert_{H^s}\vert v\vert_{H^s}$ ($s>d/2$) is inappropriate and must be replaced by Moser’s tame product estimate $\vert uv\vert_{H^s}\lesssim
\vert u\vert_\infty\vert v\vert_{H^s}+\vert u\vert_{H^s}\vert v\vert_\infty$ ($s\geq 0$). Obviously, (\[intro1\]) is not precise enough to contain this latter estimate. Part of this paper is therefore devoted to the derivation of sharper versions of (\[intro1\]).
In the works dealing with pseudo-differential operators with nonregular symbols, the focus is generally on the continuity of such operators on Sobolev or Zygmund spaces (see for instance [@TaylorM0; @TaylorM3]) and not on the derivation of precise (and tame) estimates. In [@Grenier], E. Grenier gave some description of the constants $\underline{C}(\sigma^1)$ in (\[intro1\])-(\[intro2\]) but his results, though sufficient for his purposes, are far from optimal. In this article, we aim at proving more precise versions of (\[intro1\]), (\[intro2\]) and (\[intro3\]), and we also give some extensions of these results. Let us describe roughly some of them:\
[**Action of pseudo-differential operators on Sobolev spaces**]{} (see Corollary \[coroII1\]). Take a symbol $\sigma\in S^m_{1,0}$ of the form $\sigma(x,\xi)=\Sigma(v(x),\xi)$, with $\Sigma$ as described above and $v\in H^\infty$. Then Moser’s tame product estimate can be generalized to pseudo-differential operators of order $m>0$: for all $s>0$, $$\vert {\mbox{\textnormal{Op}}}(\sigma)u\vert_{H^s}
\lesssim C(\vert v\vert_{\infty})
(\vert v\vert_{H^{s+m}}\vert u\vert_\infty+\vert u\vert_{H^{s+m}}).$$ Another estimate which does not assume any restriction on the order $m$ and also holds for negative values of $s$ is the following: for all $t_0>d/2$, one has $$\begin{aligned}
\forall -t_0<s < t_0,&\quad&
\vert {\mbox{\textnormal{Op}}}(\sigma)u\vert_{H^s}\lesssim C(\vert v\vert_{\infty})
\vert v\vert_{H^{t_0}}\vert u\vert_{H^{s+m}},\\
\forall t_0\leq s, &\quad&
\vert {\mbox{\textnormal{Op}}}(\sigma)u\vert_{H^s}
\lesssim C(\vert v\vert_{\infty})
(\vert v\vert_{H^{s}}\vert u\vert_{H^{m+t_0}}
+\vert u\vert_{H^{s+m}}).\end{aligned}$$ [**Commutator estimates.**]{} In this paper, we give a precise description of the constant $\underline{C}(\sigma^1)$ which appears in (\[intro2\]) and (\[intro3\]), and generalize these estimates in three directions:
- We control the symbolic expansion of the commutator in terms of the Poisson brackets. For instance, in the particular case when the symbol $\sigma^1(x,\xi)=\sigma^1(\xi)$ does not depend on $x$, we derive the following estimate (see Th. \[theoIII3\]): if $m_1\in{{\mathbb R}}$ and $n\in{{\mathbb N}}$ are such that $m_1>n$, then for all $s\geq 0$, one has $$\begin{aligned}
\lefteqn{\left\vert [{\mbox{\textnormal{Op}}}(\sigma^1),\sigma^2]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\right\vert_{H^s}}\\
&\lesssim&C(\sigma^1)\left(
\vert\nabla^{n+1}\sigma^2\vert_{{\infty}}
\vert u\vert_{H^{s+m_1-n-1}}+
\vert \sigma^2\vert_{H^{s+m_1}}
\vert u\vert_\infty\right),
\end{aligned}$$ and a precise description of $C(\sigma^1)$ is given; if $\sigma^1(\cdot)$ is regular at the origin, we have a more precise version involving only derivatives of $\sigma^2$, $$\begin{aligned}
\lefteqn{\left\vert [{\mbox{\textnormal{Op}}}(\sigma^1),\sigma^2]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\right\vert_{H^s}}\\
&\lesssim&C'(\sigma^1)\left(
\vert\nabla^{n+1}\sigma^2\vert_{{\infty}}
\vert u\vert_{H^{s+m_1-n-1}}+
\vert \nabla^{n+1}\sigma^2\vert_{H^{s+m_1-n-1}}
\vert u\vert_\infty\right).
\end{aligned}$$ For a similar generalization of (\[intro3\]), see Th. \[theoIII1bis\].
- We allow $\sigma^2$ to be a pseudo-differential operator and not only a function (Ths. \[theoIII1\], \[theoIII2\], \[theoIII1bis\], \[theoIV1\] and \[theoIV1bis\], and Corollaries \[coroIII3\] and \[coroIV1\]);
- We give an alternative to (\[intro2\]) allowing the cases $m_1\leq 0$ and $s<0$ (Ths. \[theoIII1\] and \[theoIV1\] and Corollaries \[coroIII3\] and \[coroIV1\]); similarly, we show that negative values of $s$ and $m_1$ are possible in (\[intro3\]) (see Ths. \[theoIII1bis\] and \[theoIV1bis\]). For instance, if $\sigma^1$ is a Fourier multiplier of order $m_1\in{{\mathbb R}}$ and $\sigma^2$ is of order $m_2\in{{\mathbb R}}$ with $\sigma^2(x,\xi)=\Sigma^2(v(x),\xi)$ and $v\in H^{\infty}({{\mathbb R}}^d)^p$ then for all $s\in{{\mathbb R}}$ such that $\max\{-t_0,-t_0-m_1\}<s$ (with $t_0>d/2$ arbitrary), $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),\sigma^2(x,D)\big]u
-{\mbox{\textnormal{Op}}}({\{\sigma^1,\sigma^2\}_n})
u\big\vert_{H^s}}\\
& \leq &C(\sigma^1,\vert v\vert_{W^{n+1,\infty}})
\big(\vert u\vert_{H^{s+m_1+m_2-n-1}}
+
\vert v\vert_{H^{(s+m_1\wedge n)_+}}
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}\big).
\end{aligned}$$
The above results admit generalization to $L^p$-based Sobolev spaces and Besov spaces, but we deliberately chose to work with classical $L^2$-based Sobolev spaces to ease the readability. We refer the reader interested by this kind of generalizations to [@Yamazaki1; @Yamazaki2], [@Marschall] and [@TaylorM0; @TaylorM3] for instance.\
The methods used to prove the above results rely heavily on Bony’s paradifferential calculus [@Bony] as well as on the works of Coifman and Meyer [@Meyer; @Meyer-Coifman3].\
In Section \[sectsymb\], we introduce the class of symbols adapted to our study; they consist in all the symbols $\sigma(x,\xi)$ such that $\sigma(\cdot,\xi)$ belongs to some Sobolev space for all $\xi$. These symbols are decomposed into four components, one of them being the well-known paradifferential symbol associated to $\sigma$, and some basic properties are given.\
In Section \[sectoper\], we study the action on Sobolev (and Zygmund) spaces of the four components into which each symbol is decomposed, and give precise estimate on the operator norm. These results generalize classical results of paraproduct theory and are in the spirit of [@Marschall] and especially [@Yamazaki1] (but the estimate we give here are different from the ones given in this latter reference). Gathering the estimates obtained on each component, we obtain a tame estimate on the action of the operator associated to the full symbol $\sigma(x,\xi)$.\
Section \[sectcomm\] is devoted to the study of commutator estimates. We first give in Prop. \[lemmIII1\] precise estimates for Meyer’s well-known result on the symbolic calculus for paradifferential operators. In Section \[sectfour\], we address the case of commutators between a Fourier multiplier $\sigma^1(D)$ and a pseudo-differential operator $\sigma^2(x,D)$; we study some particular cases, including the case when $\sigma^2(x,\xi)=\sigma^2(x)$ is a function. The case when $\sigma^1(x,D)$ is a pseudo-differential operator (and not only a Fourier multiplier) is then addressed in Section \[sectfin\].
Throughout this paper, we use the following notations.\
[**Notations.**]{} [**i.**]{} For all $a,b\in{{\mathbb R}}$, we write $a\wedge b:=\max\{a,b\}$;\
[**ii.**]{} For all $a\in {{\mathbb R}}$, we write $a_+:=\max\{a,0\}$, while $[a]$ denotes the biggest integer smaller than $a$;\
[**iii.**]{} If $f\in F$ and $g\in G$, $F$ and $G$ being two Banach spaces, the notation $\vert f\vert_{F}\lesssim \vert g\vert_{G}$ means that $\vert f\vert_{F}\leq C\vert g\vert_{G}$ for some constant $C$ which does not depend on $f$ nor $g$.\
[**iv.**]{} Here, ${{\mathcal S}({{\mathbb R}}^d)}$ denotes the Schwartz space of rapidly decaying functions, and for any distribution $f\in {\mathcal S}'({{\mathbb R}}^d)$, we write respectively $\widehat{f}$ and $\check{f}$ its Fourier and inverse Fourier transform.\
[**v.**]{} We use the classical notation $f(D)$ to denote the Fourier multiplier, namely, $\widehat{f(D)u}(\cdot)=f(\cdot)\widehat{u}(\cdot)$.
Brief reminder of Littlewood-Paley theory
-----------------------------------------
We recall in this section basic facts in Littlewood-Paley theory.
Throughout this article, $\psi\in C_0^\infty({{\mathbb R}}^d)$ denotes a smooth bump function such that $$\label{LP1}
\psi(\xi)=1\quad\mbox{ if } \quad \vert\xi\vert\leq1/2
\quad\mbox{ and } \quad \psi(\xi)=0
\quad\mbox{ if } \quad \vert\xi\vert\geq 1,$$ and we define $\varphi\in C_0^\infty({{\mathbb R}}^d)$ as $$\label{LP2}
\varphi(\xi)=\psi(\xi/2)-\psi(\xi),\qquad\forall\xi\in{{\mathbb R}}^d,$$ so that $\varphi$ is supported in the annulus $1/2\leq \vert\xi\vert\leq 2$, and one has $$\label{LP3}
1=\psi(\xi)+\sum_{p\geq 0}\varphi(2^{-p}\xi),\qquad\forall \xi\in{{\mathbb R}}^d.$$ For all $p\in{{\mathbb Z}}$, we introduce the functions $\varphi_p$, supported in $2^{p-1}\leq\vert\xi\vert\leq 2^{p+1}$, and defined as $$\label{LP4}
\varphi_p=0\quad\mbox{if} \quad p<-1,\qquad
\varphi_{-1}=\psi,\qquad
\varphi_p(\cdot)=\varphi(2^{-p}\cdot)
\quad\mbox{if} \quad p\geq 0.$$ This allows us to give the classical definition of Zygmund spaces:
Let $r\in{{\mathbb R}}$. Then $C^r_*({{\mathbb R}}^d)$ is the set of all $u\in {\mathcal S}'({{\mathbb R}}^d)$ such that $$\vert u\vert_{C^r_*}:=
\sup_{p\geq -1}2^{pr}\vert \varphi_p(D)u\vert_\infty<\infty.$$
We recall the continuous embeddings $H^s({{\mathbb R}}^d)\subset C_*^{s-d/2}({{\mathbb R}}^d)$, for all $s\in{{\mathbb R}}$, and $L^\infty({{\mathbb R}}^d)\subset C^0_*({{\mathbb R}}^d)$.
We now introduce admissible cut-off functions, which play an important role in paradifferential theory ([@Bony; @Meyer] and Appendix B of [@Metivier-Zumbrun]).
\[defiLP0\] A smooth function $\chi(\eta,\xi)$ defined on ${{\mathbb R}}^d\times{{\mathbb R}}^d$ is an admissible cut-off function if and only if:
- There are $\delta_1$ and $\delta_2$ such that $0<\delta_1<\delta_2<1$ and $$\begin{aligned}
\forall \vert\xi\vert\geq 1/2,\qquad
\chi(\eta,\xi)&=&1 \quad\mbox{ for }\quad \vert\eta\vert\leq
\delta_1\vert \xi\vert,\\
\forall \vert\xi\vert\geq 1/2,\qquad
\chi(\eta,\xi)&=&0 \quad\mbox{ for }\quad \vert\eta\vert\geq
\delta_2\vert\xi\vert;
\end{aligned}$$
- For all $\alpha,\beta\in{{\mathbb N}}^d$, there is a constant $C_{\alpha,\beta}$ such that $$\label{estadm}
\forall (\eta,\xi)\in {{\mathbb R}}^{2d},
\quad\vert\xi\vert\geq 1/2,\qquad
\left\vert\partial_\xi^\alpha\partial_\eta^\beta
\chi(\eta,\xi)\right\vert\leq
C_{\alpha,\beta}{\langle\xi\rangle}^{-\vert\alpha\vert-\vert\beta\vert}.$$
A useful admissible cut-off function is given, for all $N\geq 2$, by $$\label{LP7}
\chi(\eta,\xi)=\sum_{p\geq -1}\psi(2^{-p+N}\eta)\varphi_p(\xi);$$ One can check that it satisfies indeed the conditions of Definition \[defiLP0\] with $\delta_1=2^{-N-2}$ and $\delta_2=2^{1-N}$.
An important property satisfied by admissible cut-off functions is that ${\check}\chi(\cdot,\xi)$ and its derivatives with respect to $\xi$ enjoy good estimates in $L^1$-norm. The next lemma is a simple consequence of the estimates imposed in Definition \[defiLP0\]; we refer for instance to Appendix B of [@Metivier-Zumbrun] for a proof.
\[lemmLP0\] Let $\chi(\eta,\xi)$ be an admissible cut-off function. Then for all $\beta\in{{\mathbb N}}^d$, there exists a constant $C_\beta$ such that $$\forall \xi\in {{\mathbb R}}^d,
\qquad
\big\vert \partial_\xi^\beta
{\check}\chi(\cdot,\xi)\big\vert_{L^1({{\mathbb R}}^d)}\leq
C_\beta{\langle\xi\rangle}^{-\vert\beta\vert}.$$
Finally, we end this section with the classical characterization of Sobolev spaces (see for instance Th. 2.2.1 of [@Chemin] or Lemma 9.4 of [@TaylorM3]).
\[lemmLP1\] Let $(u_p)_{p\geq -1}$ be a sequence of ${\mathcal S}'({{\mathbb R}}^d)$ such that for all $p\geq 0$, $\widehat{u_p}$ is supported in $A2^{p-1}\leq\vert\xi\vert\leq B2^{p+1}$, for some $A,B>0$, and such that $\widehat{u_{-1}}$ is compactly supported.\
If, for some $s\in{{\mathbb R}}$, ${\displaystyle}\sum_{p\geq -1}2^{2ps}\vert u_p\vert_2^2<\infty$, then $$\sum_{p\geq -1}u_p=:u\in H^s({{\mathbb R}}^d)\quad\mbox{ and }\quad
\vert u\vert_{H^s}^2\leq {\mbox{\textnormal{Cst }}}\sum_{p\geq -1}2^{2ps}\vert u_p\vert_2^2.$$ Conversely, if $u\in H^s({{\mathbb R}}^d)$ then $$\sum_{p\geq -1}2^{2ps}\vert\varphi_p(D)u\vert_2^2\leq{\mbox{\textnormal{Cst }}}\vert u\vert_{H^s}^2.$$
Symbols {#sectsymb}
=======
As said in the introduction, we are led to consider nonregular symbols $\sigma(x,\xi)$ such that $$\label{A0}
\sigma(x,\xi)={\Sigma}(v(x),\xi),$$ where $v\in C^0({{\mathbb R}}^d)^p$ for some $p\in {{\mathbb N}}$, while ${\Sigma}$ is a smooth function belonging to the class $C^\infty({{\mathbb R}}^p,{\mathcal M}^m)$ defined below.
\[defiA0\] Let $p\in{{\mathbb N}}$, $m\in {{\mathbb R}}$ and let ${\Sigma}$ be a function defined over ${{\mathbb R}}^p_v\times{{\mathbb R}}^d_\xi$. We say that ${\Sigma}\in C^\infty({{\mathbb R}}^p,{\mathcal M}^m)$ if
- $\Sigma_{\vert_{{{\mathbb R}}^p\times\{\vert\xi\vert\leq 1\}}}\in
C^\infty({{\mathbb R}}^p;L^\infty(\{\vert\xi\vert\leq 1\}))$;
- For all $\alpha\in {{\mathbb N}}^p$, $\beta\in {{\mathbb N}}^d$, there exists a nondecreasing function $C_{\alpha,\beta}(\cdot)$ such that $$\sup_{\xi\in{{\mathbb R}}^d,\vert\xi\vert\geq 1/4}
\langle \xi\rangle^{\vert\beta\vert-m}
\left\vert\partial_v^\alpha\partial_\xi^\beta \Sigma(v,\xi)\right\vert
\leq C_{\alpha,\beta}(\vert v \vert).$$
One can write the symbol $\sigma(x,\xi)$ given in (\[symbDN\]) under the form $\sigma(x,\xi)=\Sigma(\nabla a,\xi)$, with $\Sigma(v,\xi)=
\sqrt{(1+\vert v\vert^2)\vert \xi\vert^2-(v\cdot \xi)^2}$. One can check that $\Sigma\in C^\infty({{\mathbb R}}^d,{\mathcal M}^1)$.
[**i.** ]{}We do not assume in this article that the symbols are smooth at the origin with respect to $\xi$ (for instance, (\[symbDN\]) has singular derivatives at the origin). This is the reason why the estimate in Definition \[defiA0\] is taken over frequencies away from the origin, namely $\vert\xi\vert\geq 1/4$.\
[**ii.**]{} When $p=0$, then $C^\infty({{\mathbb R}}^p,{\mathcal M}^m)$ coincides with the class ${\mathcal M}^m$ of symbols of Fourier multipliers of order $m$.
Let us remark now that if $\sigma(x,\xi)$ is as in (\[A0\]), then one can write $$\sigma(x,\xi)=\left[ \sigma(x,\xi)-\Sigma(0,\xi)\right]
+\Sigma(0,\xi);$$ the interest of such a decomposition is that the second term is a simple Fourier multiplier while the first one is in $H^s({{\mathbb R}}^d)$ if $v\in H^s({{\mathbb R}}^d)^p$, $s>d/2$:
\[lemmA0\] Let $p\in{{\mathbb N}}$, $m\in{{\mathbb R}}$, $s_0>d/2$ and take $v\in H^{s_0}({{\mathbb R}}^d)^p$ and $\Sigma\in C^\infty({{\mathbb R}}^p,{\mathcal M}^m)$; set also $\sigma(x,\xi):=\Sigma(v(x),\xi)$.\
Defining $\tau(x,\xi)=\sigma(x,\xi)-\Sigma(0,\xi)$, one has $\tau(\cdot,\xi)\in H^{s_0}({{\mathbb R}}^d)$ for all $\xi\in{{\mathbb R}}^d$; moreover:
- One has $\tau_{\vert_{{{\mathbb R}}^d\times\{\vert\xi\vert\leq 1\}}}
\in L^\infty(\{\vert\xi\vert\leq 1\}; H^{s_0}({{\mathbb R}}^d))$;
- For all $\beta\in {{\mathbb N}}^d$ and $0\leq s\leq s_0$, $$\sup_{\xi\in{{\mathbb R}}^d,\vert\xi\vert\geq 1/4}
\langle \xi\rangle^{\vert\beta\vert-m}
\left\vert\partial_\xi^\beta
\tau(\cdot,\xi)\right\vert_{H^s}
\leq C'_{s,\beta}(\vert v\vert_{L^\infty} \vert)
\vert v\vert_{H^s},$$ where $C'_{s,\beta}(\cdot)$ is some nondecreasing function depending only on the $C_{\alpha,\beta}(\cdot)$, $\vert\alpha\vert\leq [s]+2$, introduced in Definition \[defiA0\].
This is a simple consequence of Moser’s inequality (see e.g. Prop. 3.9 of [@TaylorM3]) and the properties of $\Sigma$ set forth in Definition \[defiA0\].
The previous lemma motivates the introduction of the following class of symbols (see also [@Yamazaki1; @Marschall] for similar symbol classes).
\[defiA1\] Let $m\in{{\mathbb R}}$ and $s_0>d/2$. A symbol $\sigma(x,\xi)$ belongs to the class $\Gamma_{s_0}^m$ if and only if
- One has $\sigma_{\vert_{{{\mathbb R}}^d\times\{\vert\xi\vert\leq 1\}}}
\in L^\infty(\{\vert\xi\vert\leq 1\}; H^{s_0}({{\mathbb R}}^d))$;
- For all $\beta\in {{\mathbb N}}^d$, one has $$\sup_{\xi\in{{\mathbb R}}^d,\vert\xi\vert\geq 1/4}
\langle \xi\rangle^{\vert\beta\vert-m}
\left\vert\partial_\xi^\beta
\sigma(\cdot,\xi)\right\vert_{H^{s_0}}<\infty .$$
We now set some terminology concerning the regularity of the symbols at the origin.
\[defireg\] We say that $\Sigma\in C^\infty({{\mathbb R}}^p,{\mathcal M}^m)$ is *$k$-regular* at the origin if $\Sigma_{\vert_{{{\mathbb R}}^p\times\{\vert\xi\vert\leq 1\}}}\in
C^\infty({{\mathbb R}}^p;W^{k,\infty}(\{\vert\xi\vert\leq 1\}))$.\
Similarly, we say that $\sigma\in \Gamma^m_{s_0}$ is *$k$-regular* at the origin if $\sigma_{\vert_{{{\mathbb R}}^d\times\{\vert\xi\vert\leq 1\}}}
\in W^{k,\infty}(\{\vert\xi\vert\leq 1\}; H^{s_0}({{\mathbb R}}^d))$.
It is quite natural to introduce the seminorms $N_{k,s}^m(\cdot)$ and $M^m_{k,l}(\cdot)$ defined for all $k,l\in{{\mathbb N}}$, $s\in{{\mathbb R}}$ and $m\in {{\mathbb R}}$ as $$\label{A1}
N_{k,s}^m(\sigma):=\sup_{\vert\beta\vert\leq k}
\sup_{\vert\xi\vert\geq 1/4}
\langle\xi\rangle^{\vert\beta\vert-m}
\left\vert \partial_\xi^\beta\sigma(\cdot,\xi)\right\vert_{H^s}$$ and $$\label{A2bis}
M^m_{k,l}(\sigma):=\sup_{\vert \beta \vert\leq k}
\sup_{\vert\xi\vert\geq 1/4} {\langle\xi\rangle}^{\vert\beta\vert-m}
\left\vert\partial_\xi^\beta
\sigma(\cdot,\xi)\right\vert_{W^{l,\infty}}.$$ To get information on the low frequencies, we also define $$\label{A2}
n_{k,s}(\sigma):=
\sup_{\vert\beta\vert\leq k,\vert\xi\vert\leq 1}
\left\vert \partial_\xi^\beta \sigma(\cdot,\xi)\right\vert_{H^s}
\mbox{ and }
m_k(\sigma):=\sup_{\vert\beta\vert\leq k,\vert\xi\vert\leq 1}
\vert\partial_\xi^\beta \sigma(\cdot,\xi)\vert_{\infty}.$$ Note that $M^m_{k,l}(\sigma)$ and $m_k(\sigma)$ still make sense when $\sigma$ is the symbol of a Fourier multiplier (i.e. if $\sigma(x,\xi)=\sigma(\xi)$). When $l=0$, we simply write $M^m_k(\sigma)$ instead of $M^m_{k,0}(\sigma)$.\
Finally, the notation $N_{k,s}^m(\nabla_x^l\sigma)$, $l\in{{\mathbb N}}$, stands for $\sup_{\vert\alpha\vert=l}N^m_{k,s}(\partial_x^\alpha\sigma)$; we use the same convention for the other seminorms defined above.
Associated to the class $\Gamma_s^m$ is the subclass of paradifferential symbols $\Sigma_s^m$ (in the sense of [@Bony; @Meyer], see also [@Meyer-Coifman3] and Appendix B of [@Metivier-Zumbrun]). In the definition below, the notation ${\mbox{\textnormal{Sp }}}$ is used to denote the spectrum of a function, that is, the support of its Fourier transform.
\[defiA2\] Let $m\in{{\mathbb R}}$ and $s_0>d/2$. A symbol $\sigma(x,\xi)$ belongs to the class $\Sigma_{s_0}^m$ if and only if
- One has $\sigma\in\Gamma_{s_0}^m$,
- There exists $\delta\in(0,1)$ such that $$\label{A3}
\forall \xi\in{{\mathbb R}}^d,\qquad
{\mbox{\textnormal{Sp }}}\sigma(\cdot,\xi)\subset
\{\eta\in{{\mathbb R}}^d,\vert\eta\vert\leq\delta\vert\xi\vert\}.$$
If $\psi\in C_0^\infty({{\mathbb R}}^d)$ is as in (\[LP1\]), then the spectral condition (\[A3\]) implies that for all $\sigma\in\Sigma_s^m$, one has $$\label{A4}
\forall \xi\in{{\mathbb R}}^d,\qquad
\sigma(\cdot,\xi)=
\sigma(\cdot,\xi)*\left[(2\delta{\langle\xi\rangle})^d
{\check}{\psi}(2\delta{\langle\xi\rangle}\cdot)\right].$$ It is classical (Bernstein’s lemma) to deduce that for all $\alpha,\beta\in{{\mathbb N}}^d$, one has $$\label{A5}
\forall \xi\in{{\mathbb R}}^d,\vert\xi\vert\geq 1/4,\quad
\left\vert \partial_x^\alpha\partial_\xi^\beta
\sigma(\cdot,\xi)\right\vert_{L^\infty}
\leq {\mbox{\textnormal{Cst }}}M^m_{\vert\beta\vert}(\sigma)
{\langle\xi\rangle}^{m-\vert\beta\vert+\vert\alpha\vert},$$ where $M^m_{\vert\beta\vert}(\cdot)$ is defined in (\[A2bis\]).
It is well known that symbols of $\Gamma_s^m$ can be smoothed into paradifferential symbols of $\Sigma_s^m$. In order to give a precise description of the difference between these two symbols (and of the operator associated to it), we split every $\sigma\in \Gamma_s^m$ into four components: $$\label{A6}
\sigma(x,\xi)={\sigma_{lf}}(x,\xi)+{\sigma_I}(x,\xi)+{\sigma_{II}}(x,\xi)+{\sigma_R}(x,\xi),$$ with, for some $N\in{{\mathbb N}}$, $N\geq 4$, $$\begin{aligned}
\label{A7}
{\sigma_{lf}}(\cdot,\xi)&=&\psi(\xi)\sigma(\cdot,\xi),\\
\label{A8}
{\sigma_I}(\cdot,\xi)&=&\sum_{p\geq -1}\psi(2^{-p+N}D_x)\sigma(\cdot,\xi)
(1-\psi(\xi))\varphi_p(\xi),\\
\label{A9}
{\sigma_{II}}(\cdot,\xi)&=&\sum_{p\geq -1}\varphi_p(D_x)\sigma(\cdot,\xi)
(1-\psi(\xi))\psi(2^{-p+N}\xi),\\
\label{A10}
\sigma_R(\cdot,\xi)&=&\sum_{p\geq-1}\sum_{\vert p-q\vert\leq N}
\varphi_q(D_x)\sigma(\cdot,\xi) (1-\psi(\xi))\varphi_p(\xi),\end{aligned}$$ where $\psi\in C_0^\infty({{\mathbb R}}^d)$ is a bump function satisfying (\[LP1\]).\
We also need sometimes a further decomposition of ${\sigma_R}$ as ${\sigma_R}=\sigma_{R,1}+\sigma_{R,2}$, with $$\label{decompreste}
\sigma_{R,1}(\cdot,\xi):=(1-\psi(D_x)){\sigma_R}(\cdot,\xi)
\quad\mbox{and}\quad
\sigma_{R,2}(\cdot,\xi):=\psi(D_x){\sigma_R}(\cdot,\xi);$$ note that $\sigma_{R,2}$ is given by the *finite* sum $$\sigma_{R,2}(\cdot,\xi)=\sum_{p\leq N+1}\sum_{\vert p-q\vert\leq N}
\psi(D_x)\varphi_q(D_x)\sigma(\cdot,\xi)(1-\psi(\xi))\varphi_p(\xi).$$
\[rempara\] [**i.**]{} The fact that the sum of the four terms given in (\[A7\])-(\[A10\]) equals $\sigma(x,\xi)$ follows directly from (\[LP3\]).\
[**ii.**]{} When $\sigma(x,\xi)=\sigma(x)$ does not depend on $\xi$, one has ${\mbox{\textnormal{Op}}}({\sigma_{lf}})u=\sigma\psi(D)u$, ${\mbox{\textnormal{Op}}}({\sigma_I})u=T_\sigma\widetilde{u}$, ${\mbox{\textnormal{Op}}}({\sigma_{II}})u=T_{\widetilde{u}}\sigma$ and ${\mbox{\textnormal{Op}}}({\sigma_R})u=R(\sigma,\widetilde{u})$, with $\widetilde{u}:=(1-\psi(D))u$ and where $T_f$ denotes the usual paraproduct operator associated to $f$ and $R(f,g)=fg-T_f g-T_gf$ (see [@Bony; @Meyer; @Chemin]).\
[**iii.**]{} In [@Yamazaki1], Yamazaki used a similar decomposition of symbols into three components. We need a fourth one here, namely ${\sigma_{lf}}$, in order to take into account symbols which are not infinitely smooth with respect to $\xi$ at the origin. A fifth component is also introduced in (\[decompreste\]); it is used in the proof of the second parts of Ths. \[theoIII1\]-\[theoIII1bis\].
In the next proposition, we check that ${\sigma_I}$ belongs to the class of paradifferential symbols $\Sigma_s^m$.
\[propA1\] Let $m\in{{\mathbb R}}$, $s_0>d/2$, and let $\sigma\in\Gamma_{s_0}^m$.\
Then, the symbol ${\sigma_I}$ defined in (\[A8\]) belongs to $\Sigma_{s_0}^m$ and, for all $k\in{{\mathbb N}}$ and $s\leq s_0$, $$N^m_{k,s}({\sigma_I})\leq {\mbox{\textnormal{Cst }}}N_{k,s}^m(\sigma)
\quad\mbox{ and }\quad
M_k^m(\sigma_I)\leq {\mbox{\textnormal{Cst }}}M_k^m(\sigma).$$
One can write ${\sigma_I}(\cdot,\xi)
=(1-\psi(\xi)){\check}\chi(\cdot,\xi)*\sigma(\cdot,\xi)$, where $\chi(\eta,\xi)$ denotes the admissible cut-off function constructed in (\[LP7\]). The spectral property (\[A3\]) is thus obviously satisfied by ${\sigma_I}$ and the result follows therefore from simple convolution estimates, together with the bounds on the $L^1$-norm on the derivatives $\partial_\xi^\alpha {\check}\chi(\cdot,\xi)$ given in Lemma \[lemmLP0\].
Together with the decomposition given in (\[A6\]), we shall also need another kind of decomposition, namely, Coifman and Meyer’s decomposition into elementary symbols. The proof of the next proposition is a quite close adaptation of the proof of Prop. II.5 of [@Coifman-Meyer]; it is given in Appendix \[appCM\].
\[propA2\] Let $m\in{{\mathbb R}}$ and $s_0>d/2$, and let $\sigma\in \Gamma_{s_0}^m$. With $\psi\in C^\infty_0({{\mathbb R}}^d)$ as given by (\[LP1\]), one has $$(1-\psi(\xi))\sigma(x,\xi)=
\sum_{k\in{{\mathbb Z}}^d}\frac{1}{(1+\vert k\vert^2)^{[\frac{d}{2}]+1}}
p_k(x,\xi){\langle\xi\rangle}^m,$$ with ${\displaystyle}p_k(x,\xi)=\sum_{q\geq -1}c_{k,q}(x)\lambda_k(2^{-q}\xi)$, and where:\
[**i.**]{} The coefficients $c_{k,q}(\cdot)$ are in $H^{s_0}({{\mathbb R}}^d)$ and for all $s\leq s_0$, one has $$\vert c_{k,q}\vert_{H^s}\leq {\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma);$$ moreover, for all $p\geq -1$, one has ${\displaystyle}\vert \varphi_p(D)c_{k,q}\vert_{H^s}
\leq{\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma_{(p)})$, where the symbol $\sigma_{(p)}$ is defined as $\sigma_{(p)}(\cdot,\xi)=\varphi_p(D_x)\sigma(\cdot,\xi)$.\
[**ii.** ]{} For all $k\in {{\mathbb Z}}^d$, the functions $\lambda_k(\cdot)$ are smooth and supported in $2/5\leq \vert \xi\vert\leq 12/5$. Moreover, ${\displaystyle}\left\vert{\check}\lambda_k\right\vert_{L^1}$ is bounded from above uniformly in $k\in{{\mathbb Z}}^d$.
Operators {#sectoper}
=========
To any symbol $\sigma(x,\xi)\in C^0({{\mathbb R}}^d\times{{\mathbb R}}^d\backslash\{0\})$, one can associate an operator $\sigma(x,D)={\mbox{\textnormal{Op}}}(\sigma)$ acting on functions whose Fourier transform is smooth and compactly supported in ${{\mathbb R}}^d\backslash\{0\}$: $$\forall x\in {{\mathbb R}}^d,\qquad
{\mbox{\textnormal{Op}}}(\sigma)u(x)=(2\pi)^{-d}\int_{{{\mathbb R}}^d}
e^{ix\cdot\xi}\sigma(x,\xi)\widehat{u}(\xi)d\xi.$$ The aim of this section is to study ${\mbox{\textnormal{Op}}}(\sigma)$ when $\sigma\in \Gamma^m_s$. In order to do so,we study successively ${\mbox{\textnormal{Op}}}({\sigma_{lf}})$, ${\mbox{\textnormal{Op}}}({\sigma_I})$, ${\mbox{\textnormal{Op}}}({\sigma_{II}})$ and ${\mbox{\textnormal{Op}}}({\sigma_R})$, where ${\sigma_{lf}}$, ${\sigma_I}$, ${\sigma_{II}}$ and ${\sigma_R}$ are the four components of the decomposition (\[A6\]).\
The operator ${\mbox{\textnormal{Op}}}({\sigma_{lf}})$ is handled as follows:
\[propII1\] Let $m\in {{\mathbb R}}$ and $s_0>d/2$, and let $\sigma\in \Gamma_{s_0}^m$.\
[**i.**]{} The operator ${\mbox{\textnormal{Op}}}({\sigma_{lf}})$ extends as an operator mapping any Sobolev space into $H^s({{\mathbb R}}^d)$, for all $s\leq s_0$. Moreover, $$\forall t\in {{\mathbb R}},\quad\forall s\leq s_0,\quad \forall u\in H^t({{\mathbb R}}^d),
\qquad
\left\vert {\mbox{\textnormal{Op}}}({\sigma_{lf}})u\right\vert_{H^s}\lesssim n_{0,s}(\sigma)
\vert u\vert_{H^t},$$ where $n_{0,s}(\sigma)$ is defined in (\[A2\]).\
[**ii.**]{} If $\sigma$ is ${{2[\frac{d}{2}]+2}}$-regular at the origin, the following estimates also hold, for all $s\leq s_0$, $$\vert {\mbox{\textnormal{Op}}}({\sigma_{lf}})u\vert_{H^s}
\lesssim n_{{{2[\frac{d}{2}]+2}},s}(\sigma)\vert u\vert_{C_*^0}
\lesssim n_{{{2[\frac{d}{2}]+2}},s}(\sigma)\vert u\vert_\infty.$$
By definition, one has $$\begin{aligned}
{\mbox{\textnormal{Op}}}({\sigma_{lf}})u(x)&=&(2\pi)^{-d}\int_{{{\mathbb R}}^d}e^{ix\cdot\xi}
\psi(\xi)\sigma(x,\xi)
\widehat{u}(\xi)d\xi\\
&=&(2\pi)^{-d}\int_{\vert\xi\vert\leq 1}e^{ix\cdot\xi}
\psi(\xi)\sigma(x,\xi)
\widehat{u}(\xi)d\xi.\end{aligned}$$ Since ${\displaystyle}\vert x\mapsto e^{ix\cdot\xi}\sigma(x,\xi)\vert_{H^s}\leq
{\mbox{\textnormal{Cst }}}{\langle\xi\rangle}^{\vert s\vert}\vert \sigma(\cdot,\xi)\vert_{H^s}$, one has $$\vert {\mbox{\textnormal{Op}}}({\sigma_{lf}})u\vert_{H^s}\leq {\mbox{\textnormal{Cst }}}n_{0,s}(\sigma)
\int_{\vert\xi\vert\leq 1}{\langle\xi\rangle}^{\vert s\vert}
\vert \widehat{u}(\xi)\vert d\xi.$$ One can then obtain the first estimate by a simple Cauchy-Schwarz argument.\
In order to prove the second estimate, remark that a simple expansion in Fourier series shows that $$\psi(\xi)\sigma(x,\xi)=
{\mathbf 1}_{\{\vert\xi\vert\leq 1\}}(\xi)\sum_{k\in{{\mathbb Z}}^d}
\frac{1}{(1+\vert k\vert^2)^{1+[\frac{d}{2}]}}
c_k(x)e^{i\xi\cdot k},$$ where $ {\mathbf 1}_{\{\vert\xi\vert\leq 1\}}$ is the characteristic function of the ball $\{\vert\xi\vert\leq 1\}$ and $$c_k(x)=(1+\vert k\vert^2)^{1+[d/2]}(2\pi)^{-d}
\int_{[-\pi,\pi]^d}e^{-i\xi\cdot k}\psi(\xi)\sigma(x,\xi)d\xi.$$ Using methods similar to those used in the proof of Prop. \[propA2\], one obtains that $\vert c_k(\cdot)\vert_{H^s}\leq n_{{{2[\frac{d}{2}]+2}},s}(\sigma)$. Since $${\mbox{\textnormal{Op}}}(c_k(x)e^{i\xi\cdot k}{\mathbf 1}_{\{\vert\xi\vert\leq 1\}}(\xi))u
=c_k(\cdot)({\mathbf 1}_{\{\vert\xi\vert\leq 1\}}(D)u)(\cdot+k),$$ the result follows from the next lemma:
\[lemmII0\] Let $u,v\in{{\mathcal S}({{\mathbb R}}^d)}$ and assume that $\widehat{v}$ is supported in the ball $\{\vert\xi\vert\leq A\}$, for some $A>0$. Then for all $s\in {{\mathbb R}}$, one has $$\vert u v\vert_{H^s}\leq {\mbox{\textnormal{Cst }}}\vert u\vert_{H^s}\vert v\vert_\infty.$$
Write $uv=\sum_{q\geq -1}v\varphi_q(D)u$; except the first ones, each term of this sum has its spectrum included in an annulus of size $\sim 2^q$. Thanks to Lemma \[lemmLP1\], the $H^s$-norm of the product $u v$ can therefore be controlled in terms of $\vert v\varphi_q(D)u \vert_{L^2}$, $q\geq -1$. Since these quantities are easily bounded from above by $\vert v\vert_\infty\vert \varphi_q(D)u\vert_{L^2}$, the lemma follows from another application of Lemma \[lemmLP1\].
We now turn to study ${\mbox{\textnormal{Op}}}({\sigma_I})$. As already said, ${\sigma_I}$ is the paradifferential symbol associated to $\sigma$ so that it is well-known that ${\mbox{\textnormal{Op}}}({\sigma_I})$ maps $H^{s+m}$ into $H^s({{\mathbb R}}^d)$ for all $s\in{{\mathbb R}}$ (see [@Bony; @Meyer] and Prop. B.9 of [@Metivier-Zumbrun]). However, since we need a precise estimate on the operator norm of ${\mbox{\textnormal{Op}}}({\sigma_I})$, we cannot omit the proof.
\[propII2\] Let $m\in{{\mathbb R}}$ and $s_0>d/2$, and let $\sigma\in \Gamma_{s_0}^m$.\
If ${\sigma_I}$ is as defined in (\[A8\]), then ${\mbox{\textnormal{Op}}}({\sigma_I})$ extends as a continuous mapping on $H^{s+m}({{\mathbb R}}^d)$ with values in $H^s({{\mathbb R}}^d)$, for all $s\in{{\mathbb R}}$. Moreover, $$\forall s\in{{\mathbb R}},\quad\forall u\in H^{s+m}({{\mathbb R}}^d),\qquad
\left\vert {\mbox{\textnormal{Op}}}({\sigma_I})u\right\vert_{H^s}\lesssim
M_d^m(\sigma)\vert u\vert_{H^{s+m}},$$ where $M_d^m(\sigma)$ is defined in (\[A2bis\]).
Let us first prove the following lemma, which deals with the action of operators whose symbol satisfies the spectral property (\[A3\]).
\[lemmII1\] Let $m\in{{\mathbb R}}$ and $\sigma(x,\xi)\in
C^\infty({{\mathbb R}}^d\times {{\mathbb R}}^d\backslash\{0\})$ be such that $M^m_d(\sigma)<\infty$, where $M^m_d(\cdot)$ is as defined in (\[A2bis\]).\
If moreover $\sigma(x,\xi)$ vanishes for $\vert\xi\vert\leq 1/2$ and satisfies the spectral condition (\[A3\]) then ${\mbox{\textnormal{Op}}}\left(\sigma\right)$ extends as a continuous mapping on $H^{s+m}({{\mathbb R}}^d)$ with values in $H^s({{\mathbb R}}^d)$ for all $s\in{{\mathbb R}}$ and $$\forall u\in H^{s+m}({{\mathbb R}}^d),\qquad
\left\vert {\mbox{\textnormal{Op}}}\left(\sigma\right)u\right\vert_{H^s}\lesssim
\sup_{\vert\xi\vert\geq 1/2}\sup_{\vert\beta\vert\leq d}
\Big({\langle\xi\rangle}^{\vert\beta\vert-m}
\vert \partial_\xi^\beta\sigma(\cdot,\xi)\vert_\infty\Big)
\vert u\vert_{H^{s+m}}.$$
Using (\[LP3\]), we write ${\displaystyle}\sigma(x,\xi)=
\sum_{p\geq -1}\sigma_p(x,\xi)$, with $\sigma_p(x,\xi)=\varphi_p(\xi)\sigma(x,\xi)$. For all $u\in{{\mathcal S}({{\mathbb R}}^d)}$, (\[LP3\]) and (\[LP4\]) yield $$\label{B1}
{\mbox{\textnormal{Op}}}(\sigma)u=\sum_{p\geq -1}{\mbox{\textnormal{Op}}}(\sigma_p)
\sum_{\vert p-q\vert\leq 1}\varphi_q(D)u.$$ Let us now define $\widetilde{\sigma_p}(x,\xi):=\sigma_p(2^{-p}x,2^p\xi)$ for all $p\in{{\mathbb N}}$. One obviously has (see e.g. Lemma II.1 of [@Coifman-Meyer]) $\Vert {\mbox{\textnormal{Op}}}(\sigma_p)\Vert_{L^2\to L^2}
=\Vert {\mbox{\textnormal{Op}}}(\widetilde{\sigma_p})\Vert_{L^2\to L^2}$. Moreover, Hwang proved in [@Hwang] that $$\Vert {\mbox{\textnormal{Op}}}(\widetilde{\sigma_p})\Vert_{L^2\to L^2}\leq
{\mbox{\textnormal{Cst }}}\sum_{\alpha,\beta\in\{0,1\}^d}
\left\vert\partial_x^\alpha\partial_\xi^\beta
\widetilde{\sigma_p}\right\vert_{L^\infty({{\mathbb R}}^d\times{{\mathbb R}}^d)};$$ it follows therefore from (\[A5\]) that $$\label{B2}
\Vert {\mbox{\textnormal{Op}}}({\sigma_p})\Vert_{L^2\to L^2}
\leq {\mbox{\textnormal{Cst }}}M^m_d(\sigma) 2^{pm}.$$ The result follows therefore from (\[B1\]), (\[B2\]) and Lemma \[lemmLP1\] if we can prove that for all $p\in{{\mathbb N}}$, ${\displaystyle}{\mbox{\textnormal{Op}}}(\sigma_p)\sum_{\vert p-q\vert\leq 1}\varphi_q(D)u$ has its spectrum supported in an annulus $A 2^{p-1}\leq \vert\eta\vert\leq B 2^{p+1}$ for some $A,B>0$. Since this is an easy consequence of the spectral property (\[A3\]), the proof of the lemma is complete.
The proof of the proposition is now very simple. One just has to apply Lemma \[lemmII1\] to ${\sigma_I}$, and to use Prop. \[propA1\].
The next proposition gives details on the action of ${\mbox{\textnormal{Op}}}({\sigma_{II}})$.
\[propII3\] Let $m\in{{\mathbb R}}$ and $s_0>d/2$, and let $\sigma\in \Gamma_{s_0}^m$.\
If ${\sigma_{II}}$ is defined as in (\[A9\]) then ${\mbox{\textnormal{Op}}}({\sigma_{II}})$ extends as an operator on any Sobolev space and one has, for all $s\leq s_0$ and $t>0$, $$\forall u\in C_*^{t+m},\qquad
\left\vert{\mbox{\textnormal{Op}}}({\sigma_{II}})u\right\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},s}^m(\sigma)\vert u\vert_{C_*^{t+m}},$$ and $$\forall u\in C_*^{-t+m},\qquad
\left\vert{\mbox{\textnormal{Op}}}({\sigma_{II}})u\right\vert_{H^{s-t}}\lesssim
N_{{{2[\frac{d}{2}]+2}},s}^m(\sigma)\vert u\vert_{C_*^{-t+m}},$$ where $N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma)$ is defined in (\[A1\]).
\[remagrad\] [**i.**]{} One can replace the quantity $N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma)$ by $N^m_{{{2[\frac{d}{2}]+2}},s-k}(\nabla_x^k\sigma)$, $k\in{{\mathbb N}}$, in the estimates of the proposition. This follows from the fact that $\vert f\vert_{H^s}\leq {\mbox{\textnormal{Cst }}}\vert \nabla^k f\vert_{H^{s-k}}$, $k\in{{\mathbb N}}$, whenever $\widehat{f}$ vanishes in a neighborhood of the origin, and from the observation that one can replace $\sigma$ by $(1-\psi(D_x))\sigma$ in the definition of ${\sigma_{II}}$.\
[**ii.**]{} As said previously, when $\sigma(x,\xi)=\sigma(x)$ does not depend on $\xi$, one has ${\mbox{\textnormal{Op}}}({\sigma_{II}})u=T_{(1-\psi(D))u}\sigma$ and thus $\vert {\mbox{\textnormal{Op}}}({\sigma_{II}})u\vert_{H^s}\lesssim
\vert u\vert_\infty \vert \sigma\vert_{H^s}$, that is, the endpoint case $t=0$ holds in Prop. \[propII3\] if one weakens the $\vert u\vert_{C^0_*}$-control into a $\vert u\vert_\infty$-control. This is no longer true in general when dealing with general symbols.
Prop. \[propA2\] allows us to reduce the study to the case $m=0$ and to the reduced symbols $p_k(x,\xi)$ given in that proposition.\
By definition of ${\sigma_{II}}$, one has, for all $u\in{{\mathcal S}({{\mathbb R}}^d)}$, ${\sigma_{II}}(x,D)u=\sum_{p\geq-1}v_p$, with $v_p=\sigma_{(p)}(x,D)(1-\psi(D))\psi(2^{-p+N}D)u$ and $\sigma_{(p)}(\cdot,\xi)=\varphi_p(D_x)\sigma(\cdot,\xi)$. Since the spectrum of $v_p$ is supported in $(1-2^{1-N})2^{p-1}\leq \vert\xi\vert\leq (1+2^{-1-N})2^{p+1}$, Lemma \[lemmLP1\] reduces the control of $\vert {\sigma_{II}}(x,D)u\vert_{H^s}$ to finding an estimate on each $\vert v_p\vert_2$, and hence on $$\begin{aligned}
I&=&\Big\vert \sum_{q\geq -1}\big[\varphi_p(D_x)c_{k,q}\big]
\lambda_k(2^{-q}D)
\psi(2^{-p+N}D)(1-\psi(D))u\Big\vert_2\\
&\leq&\sum_{q\geq -1}\left\vert \varphi_p(D_x)c_{k,q}\right\vert_2
\left\vert \lambda_k(2^{-q}D)
\psi(2^{-p+N}D)u\right\vert_\infty.\end{aligned}$$ Remarking that $\lambda_k(2^{-q}\xi)\psi(2^{-p+N}\xi)=0$ when $q\geq p-N+2$, and using Prop. \[propA2\], one deduces $$\label{II1}
I\leq {\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(p)})\sum_{q=-1}^{p-N+1}
\left\vert \lambda_k(2^{-q}D)
\psi(2^{-p+N}D)u\right\vert_\infty,$$ where we recall that $\sigma_{(p)}(\cdot,\xi)=\varphi_p(D_x)\sigma(\cdot,\xi)$.\
We now need the following lemma:
\[lemmII2\] Let $A,B>0$ and $\lambda\in C_0^\infty({{\mathbb R}}^d)$ supported in $A\leq\vert\xi\vert\leq B$. Then, for all $t\in{{\mathbb R}}$ and $q\geq -1$, one has, $$\forall u\in C_*^t,\qquad
\left\vert \lambda(2^{-q}D)u\right\vert_\infty
\leq C_t \vert {\check}\lambda\vert_{L^1}2^{-qt}\vert u\vert_{C^t_*}.$$
Since $\lambda$ is supported in $A\leq\vert\xi\vert\leq B$, there exists $n_0\in {{\mathbb N}}$ such that for all $\xi\in{{\mathbb R}}^d$ and $q\geq -1$, one has ${\displaystyle}\lambda(2^{-q}\xi)=\lambda(2^{-q}\xi)\sum_{\vert r-q\vert\leq n_0}
\varphi_r(\xi)$. Therefore, one can write $$\begin{aligned}
\vert\lambda(2^{-q}D)u\vert_\infty&=&
\Big\vert \lambda(2^{-q}D)\sum_{\vert r-q\vert\leq n_0}\varphi_r(D)
u\Big\vert_\infty\\
&\leq&\vert{\check}\lambda\vert_{L^1}\sum_{\vert r-q\vert\leq n_0}
\vert\varphi_r(D)u\vert_\infty,\end{aligned}$$ so that the lemma follows from the very definition of Zygmund spaces.
In order to prove the first part of Prop. \[propII3\], take any $t>0$ and use the lemma to remark that (\[II1\]) yields $$\begin{aligned}
\nonumber
I&\leq& {\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(p)})\sum_{q\geq -1}
2^{-qt}\left\vert
\psi(2^{-p+N}D)u\right\vert_{C^t_*}\\
\label{bis}
&\leq& {\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(p)})
\left\vert u\right\vert_{C_*^{t}},\end{aligned}$$ since $\sum_{q\geq -1}2^{-qt}<\infty$. From Lemma \[lemmLP1\] and the definition of $N^m_{{{2[\frac{d}{2}]+2}},0}(\cdot)$, it is obvious that $$\label{truc}
\sum_{p\geq -1}2^{2ps}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(p)})^2\leq {\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma)^2,$$ so that (\[bis\]) and Lemma \[lemmLP1\] give the result.\
To prove the second part of the proposition, proceed as above to obtain $$\begin{aligned}
I&\leq& {\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(p)})\left(\sum_{q=-1}^{p-N+1}2^{qt}
\right)\left\vert u\right\vert_{C_*^{-t}},\\
&\leq& {\mbox{\textnormal{Cst }}}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(p)})2^{pt}
\left\vert u\right\vert_{C_*^{-t}};\end{aligned}$$ the end of the proof is done as for the first part of the proposition.
We finally turn to study ${\mbox{\textnormal{Op}}}({\sigma_R})$:
\[propII4\] Let $m\in{{\mathbb R}}$ and $s_0>d/2$, and let $\sigma\in\Gamma^m_{s_0}$.\
If ${\sigma_R}$ is as given in (\[A10\]) and if $s+t>0$ and $s\leq s_0$ then ${\mbox{\textnormal{Op}}}({\sigma_R})$ extends as a continuous operator on $H^{m+t}({{\mathbb R}}^d)$ with values in $H^{s+t-\frac{d}{2}}({{\mathbb R}}^d)$. Moreover, $$\forall u\in H^{m+t}({{\mathbb R}}^d),\qquad
\left\vert {\mbox{\textnormal{Op}}}(\sigma_R)u\right\vert_{H^{s+t-d/2}}
\lesssim N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma)\vert u\vert_{H^{m+t}}$$ and $$\forall u\in C_*^{m+t}({{\mathbb R}}^d),\qquad
\left\vert {\mbox{\textnormal{Op}}}(\sigma_R)u\right\vert_{H^{s+t}}
\lesssim N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma)\vert u\vert_{C_*^{m+t}}.$$
\[remagrad1\] For the same reasons as in Remark \[remagrad\], one can replace the quantity $N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma)$ by $N^m_{{{2[\frac{d}{2}]+2}},s-k}(\nabla_x^k\sigma)$, $k\in{{\mathbb N}}$, in the estimates of the proposition, provided that one replaces ${\sigma_R}$ by $\sigma_{R,1}$, where $\sigma_{R,1}$ is defined in (\[decompreste\]).
We only prove the first of the two estimates given in the proposition. The second one is both easier and contained in Th. B of [@Yamazaki1]. The proof we present below is an adaptation of the corresponding result which gives control of the residual term in paraproduct theory (e.g. Th. 2.4.1. of [@Chemin]).\
Using the expression of ${\sigma_R}$ given in (\[A10\]) and a Littlewood-Paley decomposition, one can write $${\mbox{\textnormal{Op}}}({\sigma_R})u=\sum_{r\geq -1}\varphi_r(D){\mbox{\textnormal{Op}}}({\sigma_R})u
=\sum_{r\geq -1}\varphi_r(D)\sum_{p\geq -1}R_p(\sigma)u,$$ where ${\displaystyle}R_p(\sigma)u:=\sum_{\vert p-q\vert\leq N}
\sigma_{(q)}(x,D)(1-\psi(D))\varphi_p(D)u$, and with $\sigma_{(q)}(\cdot,\xi)=\varphi_q(D)\sigma(\cdot,\xi)$.\
Since ${\mbox{\textnormal{Sp }}}R_p(\sigma)u$ is included in $\vert\xi\vert\leq (1+2^N)2^{p+1}$, there exists $n_0\in{{\mathbb N}}$ such that $\varphi_r(D)R_p(\sigma)u=0$ whenever $r>p+n_0$. Thus, one has in fact $${\mbox{\textnormal{Op}}}({\sigma_R})u=\sum_{r\geq -1}\varphi_r(D)\sum_{p\geq r-n_0}R_p(\sigma)u,$$ and the proposition follows from Lemma \[lemmLP1\] and the estimate $$\label{II2}
\Big(\sum_{r\geq -1}2^{2r(s+t-d/2)}
\Big\vert \varphi_r(D)\sum_{p\geq r-n_0}
R_p(\sigma)u\Big\vert_2^2\Big)^{1/2}
\lesssim N^m_{{{2[\frac{d}{2}]+2}},s}(\sigma)\vert u\vert_{m+t}.$$ The end of the proof is thus devoted to establishing (\[II2\]).\
Using Prop. \[propA2\] –and with the same notations– one can see that it suffices to prove (\[II2\]) with $R_p(\sigma)$ replaced by $R_p(p_k{\langle\xi\rangle}^m)$, provided that the estimate is uniform in $k\in{{\mathbb Z}}^d$. Without loss of generality, we can also assume that $m=0$.\
Now, remark that $$\label{II3}
2^{r(s+t-d/2)}\Big\vert \varphi_r(D)\sum_{p\geq r-n_0}
R_p(p_k)u\Big\vert_2
\leq {\mbox{\textnormal{Cst }}}2^{r(s+t)}\Big\vert\sum_{p\geq r-n_0}R_p(p_k)
u\Big\vert_{L^1}$$ and that $$\label{II4}
\left\vert R_p(p_k)u\right\vert_{L^1}=
\Big\vert \sum_{\vert p-q\vert\leq N}\varphi_q(D_x)p_k(x,D)
(1-\psi(D))\varphi_p(D) u\Big\vert_{L^1}.$$ Now, using the expression of $p_k(x,\xi)$ given in Prop. \[propA2\], one can write $$\begin{aligned}
\lefteqn{\varphi_q(D_x)p_k(x,D)(1-\psi(D))
\varphi_p(D)u
=}\\
& &\sum_{l\geq -1}\varphi_q(D_x)c_{k,l}(x)\lambda_k(2^{-l}D)
(1-\psi(D))\varphi_p(D) u,\end{aligned}$$ and since $\lambda_k(2^{-l}\xi)\varphi_p(\xi)=0$ if $\vert p-l\vert>n_1$, for some $n_1\in{{\mathbb N}}$, one deduces that the summation in the r.h.s. of the above inequality is over a finite number of integers $l$; therefore, by Cauchy-Schwartz’s inequality and Prop. \[propA2\], $$\label{II5}
\left\vert\varphi_q(D_x)p_k(x,D)(1-\psi(D))
\varphi_p(D)u\right\vert_{L^1}\lesssim N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(q)})
\left\vert
\varphi_p(D)u\right\vert_{L^2}.$$ From (\[II4\]) and (\[II5\]) one obtains $$\left\vert R_p(p_k)u\right\vert_{L^1}\lesssim
2^{-p(s+t)}\sum_{\vert p-q\vert\leq N}
2^{qs}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(q)})
2^{pt}\left\vert
\varphi_p(D)u\right\vert_2,$$ and the l.h.s. of (\[II3\]) is therefore bounded from above by $$\begin{aligned}
\sum_{p\geq r-n_0} 2^{(r-p)(s+t)}
\sum_{\vert p-q\vert\leq N}
2^{qs}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(q)})
2^{pt}\left\vert
\varphi_p(D)u\right\vert_2. \end{aligned}$$ Since $s+t>0$, Hölder’s inequality yields that the l.h.s. of (\[II2\]) is bounded from above by $$\left\vert\left(
2^{ps}N^m_{{{2[\frac{d}{2}]+2}},0}(\sigma_{(p)})
2^{pt}\Big\vert
\varphi_p(D)u\right\vert_2\right)_{p\geq -1}\Big\vert_{l^1}.$$ By Cauchy-Schwartz’s inequality, Lemma \[lemmLP1\] and an argument similar to the one used in (\[truc\]), one obtains (\[II2\]), which concludes the proof.
A first important consequence of Propositions \[propII3\] and \[propII4\] is that one can control the action of the operator associated to the ’remainder’ symbol $\sigma-{\sigma_{lf}}-{\sigma_I}$, which is more regular than the full operator if $\sigma(x,\xi)$ is smooth enough in the space variables.
\[propII5\] Let $m\in{{\mathbb R}}$, $s_0>d/2$ and $d/2<t_0\leq s_0$. If for some $r\geq 0$, one has $\sigma\in\Gamma^m_{s_0+r}$ then,\
[**i.**]{} For all $-t_0<s\leq s_0$, the following estimate holds: $$\forall u\in H^{m+t_0-r}({{\mathbb R}}^d),\qquad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}})u\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},s+r}^m(\sigma)
\vert u\vert_{H^{m+t_0-r}}.$$ [**ii.**]{} For all $r'\in{{\mathbb R}}$ (such that $t_0+r'\leq s_0+r$) and $-t_0<s\leq t_0+r'$, one has $$\forall u\in H^{s+m-r'}({{\mathbb R}}^d),\qquad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}})u\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},t_0+r'}^m(\sigma)
\vert u\vert_{H^{s+m-r'}}.$$ [**iii.**]{} For symbols of nonnegative order, i.e. when $m>0$, then for all $s>0$ such that $s+m\leq s_0$, one also has $$\forall u\in C_*^{-r}({{\mathbb R}}^d),\qquad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}})u\vert_{H^s}\lesssim
N^m_{{{2[\frac{d}{2}]+2}},s+m+r}(\sigma)\vert u\vert_{C^{-r}_*};$$ this estimate still holds for slightly negative values of $r$, namely, if $-m<r$.
One has $\sigma-{\sigma_I}-{\sigma_{lf}}={\sigma_{II}}+{\sigma_R}$, and we are therefore led to control $\vert {\mbox{\textnormal{Op}}}({\sigma_{II}})u\vert_{H^s}$ and $\vert {\mbox{\textnormal{Op}}}({\sigma_R})u\vert_{H^s}$. We first prove point [**i**]{}.\
The estimate on $\vert{\mbox{\textnormal{Op}}}({\sigma_{II}})u\vert_{H^s}$ is given by the first part of Prop. \[propII3\] when $r=0$. When $r>0$, taking $s=s+r$ and $t=r$ in the second part of this proposition gives the result. The estimate on $\vert{\mbox{\textnormal{Op}}}({\sigma_R})u\vert_{H^s}$ is given by taking $s=s+r$ and $t=t_0-r$ in the first part of Prop. \[propII4\].\
To establish [**ii.**]{}, take $s=t_0+r'$ and $t=t_0-s+r'$ in the second estimate of Prop. \[propII3\] to obtain that $\vert {\mbox{\textnormal{Op}}}({\sigma_{II}})u\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},t_0+r'}^m(\sigma)
\vert u\vert_{H^{s+m-r'}}$ for all $s<t_0+r'$. Taking $t=s-r'$ and $s=t_0+r'$ in the first estimate of Prop. \[propII4\] shows that $\vert {\mbox{\textnormal{Op}}}({\sigma_R})u\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},t_0+r'}^m(\sigma)
\vert u\vert_{H^{s+m-r'}}$ for all $s>-t_0$ and the proof of [**ii.**]{} is complete (the endpoint $s=t_0+r'$ being given by [**i**]{}).\
To prove [**iii.**]{}, take $s=s+m+r$ and $t=m+r>0$ in the second part of Prop. \[propII3\] and $s=s+m+r$ and $t=-m-r$ in the second estimate of Prop. \[propII4\].
The first two points of the following proposition are a close variant of Prop. \[propII5\] which uses the decomposition (\[decompreste\]) of the component ${\sigma_R}$, while the last point addresses the case when $\sigma(x,\xi)=\sigma(x)$ does not depend on $\xi$.
\[propII5bis\] Let $m\in{{\mathbb R}}$, $k\in{{\mathbb N}}$, $s_0>d/2$ and $d/2<t_0\leq s_0$. If for some $r\geq 0$, one has $\sigma\in\Gamma^m_{s_0+r}$ then,\
[**i.**]{} For all $-t_0<s\leq s_0$, the following estimate holds: $$\forall u\in H^{m+t_0-r},\quad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}}-\sigma_{R,2})u\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},s+r-k}^m(\nabla_x^k\sigma)
\vert u\vert_{H^{m+t_0-r}}.$$ [**ii.**]{} For all $r'\in{{\mathbb R}}$ (such that $t_0+r'\leq s_0+r$) and $-t_0<s<t_0+r'$, one has $$\forall u\in H^{s+m-r'},\quad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}}-\sigma_{R,2})u\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},t_0+r'-k}^m(\nabla_x^k\sigma)
\vert u\vert_{H^{s+m-r'}}.$$ [**iii.**]{} For symbols of nonnegative order, i.e. when $m>0$, then for all $s>0$ such that $s+m\leq s_0$, one also has $$\forall u\in C_*^{-r}({{\mathbb R}}^d),\qquad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}}-\sigma_{R,2})u\vert_{H^s}\lesssim
N^m_{{{2[\frac{d}{2}]+2}},s+m+r-k}(\nabla^k_x\sigma)\vert u\vert_{C^{-r}_*}.$$ [**iv.**]{} When $\sigma$ is a function, $\sigma\in H^{s_0}({{\mathbb R}}^d)$, one has $$\forall \quad 0<s\leq s_0, \qquad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}}-\sigma_{R,2})u\vert_{H^s}\lesssim
\vert\nabla^k\sigma\vert_{H^{s-k}}\vert u\vert_{\infty}$$ and, if $\sigma\in H^{s_0+r}({{\mathbb R}}^d)$, with $r>0$, $$\forall \quad 0<s\leq s_0,\qquad
\vert {\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}}-\sigma_{R,2})u\vert_{H^s}\lesssim
\vert\nabla^k\sigma\vert_{H^{s+r-k}}\vert u\vert_{C^{-r}_*};$$ when $s=0$, the above two estimates still hold if one adds $\vert\nabla^{n+1}\sigma\vert_{L^{\infty}}\vert u\vert_{H^{-n-1}}$ to the right-hand-side, for any $n\in{{\mathbb N}}$.
\[remII5bis\] When $k=0$, the estimates of [**iv**]{} still hold if one replaces ${\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}}-\sigma_{R,2})$ by ${\mbox{\textnormal{Op}}}(\sigma-{\sigma_I}-{\sigma_{lf}})$ (and $\vert \nabla^{n+1}\sigma\vert_\infty$ by $\vert \sigma\vert_{W^{n+1,\infty}}$ in the additional term when $s=0$). This is a consequence of the definition of $\sigma_{R,2}$ and of Lemma \[lemmII0\].
One has $\sigma-{\sigma_I}-{\sigma_{lf}}-\sigma_{R,2}={\sigma_{II}}+\sigma_{R,1}$, so that the first three points of the proposition are proved as in Prop. \[propII5\], using Remarks \[remagrad\] and \[remagrad1\].\
We now prove the fourth point of the proposition. Since $\sigma$ is a function, we can write, as in Remark \[rempara\], ${\mbox{\textnormal{Op}}}({\sigma_{II}}+\sigma_{R,1})u=T_{{\widetilde{u}}}\sigma+R(\widetilde{\sigma},{\widetilde{u}})$, with ${\widetilde{u}}:=(1-\psi(D))u$ and $\widetilde{\sigma}:=(1-\psi(D))\sigma$. The estimate for $s>0$ thus follows from the classical properties (e.g. Th. 2.4.1. of [@Chemin], and Prop. 3.5.D of [@TaylorM0] for the last one):
- for all $s\in{{\mathbb R}}$, $\vert T_fg\vert_{H^s}\lesssim \vert f\vert_\infty \vert g\vert_{H^s}$;
- for all $s\in{{\mathbb R}}$ and $r>0$, $\vert T_f g\vert_{H^s}\lesssim \vert f\vert_{C_*^{-r}}\vert g\vert_{H^{s+r}}$;
- for all $s>0$, $r\in{{\mathbb R}}$, $\vert R(f,g)\vert_{H^s}\lesssim\vert f\vert_{H^{s+r}}\vert g\vert_{C_*^{-r}}$;
- for all $n\in{{\mathbb N}}$, $\vert R(f,g)\vert_{L^2}\lesssim
\vert f\vert_{W^{n+1,\infty}}\vert g\vert_{H^{-n-1}}$;
(we also use the fact that $\vert f\vert_{H^s}\leq {\mbox{\textnormal{Cst }}}\vert \nabla^k f\vert_{H^{s-k}}$ and $\vert f\vert_{W^{n,\infty}}\leq {\mbox{\textnormal{Cst }}}\vert \nabla^n f\vert_{L^\infty}$ for all $f$ such that $\widehat{f}$ vanishes in a neighborhood of the origin).
Gathering the results of the previous propositions, one obtains the following theorem, which describes the action of the full operator ${\mbox{\textnormal{Op}}}(\sigma)$, which is of course of order $m$.
\[theoII1\] Let $m\in{{\mathbb R}}$, $d/2<t_0\leq s_0$ and $\sigma\in\Gamma_{s_0}^m$. Then for all $u\in{{\mathcal S}({{\mathbb R}}^d)}$, the following estimates hold: $$\forall -t_0<s<t_0,\qquad
\vert {\mbox{\textnormal{Op}}}(\sigma)u\vert_{H^s}\lesssim \Big(
n_{0,t_0}(\sigma)+N_{{{2[\frac{d}{2}]+2}},t_0}^{m}(\sigma)\Big)
\vert u\vert_{H^{s+m}},$$ and $$\forall t_0\leq s\leq s_0,\quad
\vert {\mbox{\textnormal{Op}}}(\sigma)u\vert_{H^s}\lesssim \big(
n_{0,s}(\sigma)+N_{{{2[\frac{d}{2}]+2}},s}^{m}(\sigma)\big)
\vert u\vert_{H^{m+t_0}}+M_d^m(\sigma)\vert u\vert_{H^{s+m}}.$$
Recall that $\sigma={\sigma_{lf}}+{\sigma_I}+(\sigma-{\sigma_{lf}}-{\sigma_I})$; we use the first two estimates of Prop. \[propII5\] (with $r=0$ and $r'=0$) to control $\sigma-{\sigma_{lf}}-{\sigma_I}$ while $\vert {\mbox{\textnormal{Op}}}({\sigma_{lf}})u\vert_{H^s}$ and $\vert {\mbox{\textnormal{Op}}}({\sigma_I})u\vert_{H^s}$ are easily controlled using Propositions \[propII1\] and \[propII2\] and the observation that by a classical Sobolev embedding, $M_d^m(\sigma)\lesssim N_{{{2[\frac{d}{2}]+2}},t_0}^{m}(\sigma)$.
\[remrem\] [**i.**]{} If $\sigma(x,\xi)=\sigma(x)\in H^{s_0}({{\mathbb R}}^d)$, then the results on the microlocal regularity of products (e.g. [@Hormander], p.240) say that if $u\in H^s({{\mathbb R}}^d)$, then $\sigma u\in H^s({{\mathbb R}}^d)$ if $s+s_0>0$, $s\leq s_0$ and $s_0>d/2$. This result can be deduced from Th. \[theoII1\] (note that the limiting case $s+s_0=0$ is also true, but the proof requires different tools [@Hormander], th. 8.3.1).\
[**ii.**]{} We refer to Prop. 8.1 of [@TaylorMpara] for another kind of estimate on the action of pseudo-differential operators; see also estimate (25) of [@Marschall].
One of the interests of Th. \[theoII1\] is that it gives control of ${\mbox{\textnormal{Op}}}(\sigma)u$ in Sobolev spaces of negative order. The price to pay is that for nonnegative values of the Sobolev index $s$, and when $\sigma(\cdot,\xi)=\sigma(\cdot)\in H^{s_0}({{\mathbb R}}^d)$ does not depend on $\xi$, we do not recover the classical tame estimate $\vert \sigma u\vert_{H^s}\lesssim
\left(\vert u\vert_{H^s}\vert \sigma\vert_\infty
+\vert \sigma\vert_{H^s}\vert u\vert_\infty\right)$ but a weaker one, namely $\vert \sigma u\vert_{H^s}\lesssim \vert u\vert_{H^s}\vert \sigma\vert_{\infty}
+\vert \sigma\vert_{H^s}\vert u\vert_{H^{\frac{d}{2}+\varepsilon}}$, for all $\varepsilon>0$. The difference is slight because the embedding $H^{\frac{d}{2}+\varepsilon}({{\mathbb R}}^d)\subset L^\infty({{\mathbb R}}^d)$ is critical, but can be cumbersome. The next theorem can therefore be a useful alternative to Theorem \[theoII1\]
\[theoII2\] Let $m\in{{\mathbb R}}$, $d/2<t_0\leq s_0$ and $\sigma\in\Gamma_{s_0+m}^m$. Assume that $m>0$ and $\sigma$ is ${{2[\frac{d}{2}]+2}}$-regular at the origin. Then for all $u\in{{\mathcal S}({{\mathbb R}}^d)}$ and $0<s\leq s_0$, one has $$\vert {\mbox{\textnormal{Op}}}(\sigma)u\vert_{H^s} \lesssim
\Big(n_{{{2[\frac{d}{2}]+2}},s}(\sigma)+N^m_{{{2[\frac{d}{2}]+2}},s+m}(\sigma)\Big)
\vert u\vert_{C_*^0}
+M_d^m(\sigma)\vert u\vert_{H^{s+m}}.$$
We just have to control the $H^s$-norm of the four components of ${\mbox{\textnormal{Op}}}(\sigma)u$ by the r.h.s. of the estimate given in the theorem.\
For ${\mbox{\textnormal{Op}}}({\sigma_{lf}})u$ and ${\mbox{\textnormal{Op}}}({\sigma_I})u$, this is a simple consequence of the second part of Prop. \[propII1\] and of Prop. \[propII2\] respectively. The other components are controlled with the help of Prop. \[propII5\].iii.
\[exttheo\] Using the fact that slightly negative values of $r$ are allowed in Prop. \[propII5\].iii, one can check that the estimate given by Th. \[theoII2\] can be extended to $s=0$ provided that the quantity $\vert u\vert_{C_*^0}$ which appears in the r.h.s. of the estimate is replaced by $\vert u\vert_{C^\epsilon_*}$, for any $\epsilon>0$.
The following corollary deals with the case when the symbol $\sigma$ is of the form $\sigma(x,\xi)=\Sigma(v(x),\xi)$.
\[coroII1\] Let $m\in{{\mathbb R}}$, $p\in{{\mathbb N}}$ and $s_0\geq t_0>d/2$. Consider $v\in H^{s_0}({{\mathbb R}}^d)^p$ and assume that $\sigma(x,\xi)=\Sigma(v(x),\xi)$, with $\Sigma\in C^\infty({{\mathbb R}}^p,{\mathcal M}^m)$. Then:\
[**i.**]{} $$\forall -t_0< s <t_0,\qquad
\vert \sigma(x,D)u\vert_{H^s}\lesssim
C_\Sigma(\vert v\vert_\infty)\vert v\vert_{H^{t_0}}
\vert u\vert_{H^{s+m}}$$ and $$\forall t_0\leq s\leq s_0,\qquad
\vert \sigma(x,D)u\vert_{H^s}\lesssim
C_\Sigma(\vert v\vert_\infty)\left(\vert v\vert_{H^{s}}
\vert u\vert_{H^{m+t_0}}
+
\vert u\vert_{H^{s+m}}\right).$$ [**ii.**]{} If moreover $m>0$, $v\in H^{s_0+m}({{\mathbb R}}^d)$, and $\Sigma$ is ${{2[\frac{d}{2}]+2}}$-regular at the origin, then, for $0<s\leq s_0$, $$\vert \sigma(x,D)u\vert_{H^s}\lesssim
C_\Sigma(\vert v\vert_\infty)\big(\vert v\vert_{H^{s+m}}
\vert u\vert_{C_*^0}
+
\vert u\vert_{H^{s+m}}\big).$$ In the above, $C_\Sigma(\cdot)$ denotes a smooth nondecreasing function depending only on a finite number of derivatives of $\Sigma$.
We write $\sigma(x,\xi)=[\sigma(x,\xi)-\Sigma(0,\xi)]+\Sigma(0,\xi)$. Owing to Lemma \[lemmA0\], the first component of this decomposition is in $\Gamma^m_{s_0}$ and we can use Th. \[theoII1\] to study the associated pseudo-differential operator. The estimates of the theorem transform into the estimates stated in the corollary thanks to Lemma \[lemmA0\].\
Since the action of the Fourier multiplier $\Sigma(0,D)$ satisfies obviously these estimates, the first point of the corollary is proved. Using Theorem \[theoII2\], one proves the second estimate in the same way.
Composition and commutator estimates {#sectcomm}
====================================
The composition of two pseudo-differential operators is well-known for classical symbols, and one has ${\mbox{\textnormal{Op}}}(\sigma^1)\circ{\mbox{\textnormal{Op}}}(\sigma^2)\sim{\mbox{\textnormal{Op}}}(\sigma^1\sharp\sigma^2)$, where the symbol $\sigma^1\sharp\sigma^2$ is given by an infinite expansion of $\sigma^1$ and $\sigma^2$. When dealing with symbols of limited regularity, one has to stop this expansion. Therefore, for all $n\in{{\mathbb N}}$, we define $\sigma^1\sharp_n\sigma^2$ as $$\label{III1}
\sigma^1\sharp_n\sigma^2(x,\xi):=\sum_{\vert\alpha\vert\leq n}
\frac{(-i)^{\vert \alpha\vert}}{\alpha !}\partial_\xi^\alpha\sigma^1(x,\xi)
\partial_x^\alpha\sigma^2(x,\xi).$$ Similarly, we introduce the Poisson brackets: $$\label{IIIpoiss}
\{\sigma^1,\sigma^2\}_n(x,\xi):=
\sigma^1\sharp_n\sigma^2(x,\xi)-
\sigma^2\sharp_n\sigma^1(x,\xi).$$
In this section, we describe the composition or commutator of pseudodifferential operators of limited regularity with Fourier multipliers, or with another pseudo-differential operator. A key point in this analysis is the following proposition; the first two points are precise estimates for Meyer’s classical result on the composition of paradifferential operators (e.g. Th. XVI.4 of [@Meyer-Coifman3]).
\[lemmIII1\] Let $m_1,m_2\in{{\mathbb R}}$, $s_0>d/2$, $n\in{{\mathbb N}}$ and $\sigma^2\in \Gamma_{s_0+n+1}^{m_2}$. Then\
[**i.**]{} If $\sigma^1(x,\xi)=\sigma^1(\xi)\in{\mathcal M}^{m_1}$, there exists a symbol $\rho_n(x,\xi)$ such that $$\sigma^1(D)\circ{\mbox{\textnormal{Op}}}({\sigma_I}^2)=
{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n{\sigma_I}^2)+{\mbox{\textnormal{Op}}}(\rho_n);$$ moreover $\rho_n(x,\xi)$ vanishes for $\vert\xi\vert\leq 1/2$ and satisfies the spectral condition (\[A3\]) and the estimate $$\sup_{\vert\xi\vert\geq 1/2}\sup_{\vert\beta\vert\leq d}\Big(
{\langle\xi\rangle}^{\vert\beta\vert+n+1-m_1-m_2}
\vert\partial_\xi^\beta\rho_n(\cdot,\xi)\vert_\infty\Big)\lesssim
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)
M^{m_2}_{d}(\nabla_x^{n+1}\sigma^2).$$ [**ii.**]{} If $\sigma^1\in\Gamma_{s_0}^{m_1}$ then there exists a symbol $\rho_n(x,\xi)$ such that $${\mbox{\textnormal{Op}}}({\sigma_I}^1)\circ{\mbox{\textnormal{Op}}}({\sigma_I}^2)={\mbox{\textnormal{Op}}}({\sigma_I}^1\sharp_n{\sigma_I}^2)+{\mbox{\textnormal{Op}}}(\rho_n),$$ and which satisfies the same properties as in case [**i.**]{}\
[**iii.**]{} If $\sigma^1$ is a function, $\sigma^1\in C_*^r$ for some $r\geq 0$, then the symbol $\rho_n(x,\xi)$ defined in [**ii.**]{} is of order $m_2-n-1-r$ and $$M_d^{m_2-n-1-r}(\rho_n)\lesssim
\vert \sigma^1\vert_{C_*^r}
M^{m_2}_{d}(\nabla_x^{n+1}\sigma^2).$$
\[remkey\] For the sake of simplicity, we stated the above proposition for paradifferential symbols ${\sigma_I}^1$ (and ${\sigma_I}^2$ in ii. and iii.) associated to symbols $\sigma^1$ and $\sigma^2$; the proof below shows that the only specific properties of ${\sigma_I}^1$ and ${\sigma_I}^2$ actually used are the spectral property (\[A3\]) and the cancellation for frequencies $\vert \xi\vert\leq 1/2$. Thus, one can extend the result to all symbols satisfying these conditions.
We omit the proof of the first point of the proposition, which can be be deduced from the proof below with only minor changes. The method we propose here is inspired by the proof of Th. B.2.16 of [@Metivier-Zumbrun] rather than Meyer’s classical one (Th. XVI.4 of [@Meyer-Coifman3]) which would lead to less precise estimates here.\
First remark that since ${\sigma_I}^1$ satisfies the spectral condition (\[A3\]) and vanishes for frequencies $\vert\xi\vert\leq 1/2$, there exists an admissible cut-off function $\chi$ (in the sense of Def. \[defiLP0\]) such that $\widehat{{\sigma_I}^2}(\eta,\xi)\chi(\eta,\xi)=\widehat{{\sigma_I}^2}(\eta,\xi)$; it is then both classical and easy to see that $$\rho_n(x,\xi)=\sum_{\vert\gamma\vert=n+1}
\int_{{{\mathbb R}}^d}G_\gamma(x,x-y,\xi)(\partial_x^\gamma{\sigma_I}^2)(y,\xi)dy,$$ with $$G_\gamma(x,y,\xi):=(-i)^{\vert\gamma\vert}(2\pi)^{-d}
\int_{{{\mathbb R}}^d}e^{iy\cdot\eta}{\sigma_I}^{1,\gamma}(x,\eta,\xi)
\chi(\eta,\xi)d\eta$$ and ${\displaystyle}{\sigma_I}^{1,\gamma}(x,\eta,\xi):=
\int_0^1\frac{(1-t)^n}{n!}\partial_\xi^\gamma{\sigma_I}^1(x,\xi+s\eta)ds$. Therefore, for all $0\leq\vert\beta\vert\leq d$, $$\begin{aligned}
\nonumber
\vert \partial_\xi^\beta\rho_n(x,\xi)\vert&\leq&{\mbox{\textnormal{Cst }}}\sum_{\beta'+\beta''=\beta}
\left\vert \partial_\xi^{\beta'}G_\gamma(x,\cdot,\xi)\right\vert_{L^1}
\left\vert \partial_x^\gamma\partial_\xi^{\beta''}{\sigma_I}^2(\cdot,\xi)
\right\vert_\infty\\
\label{marre1}
&\leq&
{\mbox{\textnormal{Cst }}}\sum_{\beta'+\beta''=\beta}
\left\vert\partial_\xi^{\beta'}G_\gamma(x,\cdot,\xi)\right\vert_{L^1}
M_{d}^{m_2}(\nabla_x^{n+1}\sigma^2){\langle\xi\rangle}^{m_2-\vert\beta''\vert},\end{aligned}$$ where we used Prop. \[propA1\] to obtain the last equality.\
The proposition follows therefore from (\[marre1\]) and the estimate, for all $\vert\beta'\vert\leq d$, $$\label{marre2}
\left\vert\partial_\xi^{\beta'}G_\gamma(x,\cdot,\xi)\right\vert_{L^1}
\leq {\mbox{\textnormal{Cst }}}M_{{n+2+[\frac{d}{2}]+d}}^{m_1}(\sigma^1)
{\langle\xi\rangle}^{m_1-\vert\beta'\vert-n-1}.$$ and, when $\sigma^1$ is a function (case [**iii.**]{} of the lemma), $$\label{marre3}
\left\vert\partial_\xi^{\beta'}G_\gamma(x,\cdot,\xi)\right\vert_{L^1}
\leq{\mbox{\textnormal{Cst }}}\vert \sigma^1\vert_{C_*^r}
{\langle\xi\rangle}^{-\vert\beta'\vert-n-1-r}.$$ Both (\[marre2\]) and (\[marre3\]) follow from the next two lemmas.
For all $\alpha,\beta\in{{\mathbb N}}^d$ such that $\vert\alpha\vert\leq [d/2]+1$ and $\vert\beta\vert\leq d$, one has $$\vert \partial_\eta^\alpha\partial_\xi^\beta
({\sigma_I}^{1,\gamma}(x,\cdot,\cdot)\chi)(\eta,\xi)\vert
\leq {\mbox{\textnormal{Cst }}}M_{{n+2+[\frac{d}{2}]+d}}^{m_1}(\sigma^1)
{\langle\xi\rangle}^{m_1-\vert\alpha\vert-\vert\beta\vert-n-1}.$$ If moreover the symbol is a function, $\sigma^1\in C_*^r$ for some $r\in{{\mathbb R}}$, then $$\vert \partial_\eta^\alpha\partial_\xi^\beta
({\sigma_I}^{1,\gamma}(x,\cdot,\cdot)\chi)(\eta,\xi)\vert
\leq {\mbox{\textnormal{Cst }}}\vert \sigma^1\vert_{C_*^r}
{\langle\xi\rangle}^{-\vert\alpha\vert-\vert\beta\vert-n-1-r}.$$
It suffices to prove the estimate of the lemma for $\partial_\eta^{\alpha'}\partial_\xi^{\beta'}{\sigma_I}^{1,\gamma}
\partial_\eta^{\alpha''}\partial_\xi^{\beta''}\chi$ for all $\alpha'+\alpha''=\alpha$ and $\beta'+\beta''=\beta$. By definition of ${\sigma_I}^{1,\gamma}$, one has $$\partial_\eta^{\alpha'}\partial_\xi^{\beta'}{\sigma_I}^{1,\gamma}
\partial_\eta^{\alpha''}\partial_\xi^{\beta''}\chi(\eta,\xi)
=
\int_0^1\frac{(1-t)^n}{n!}\partial_\xi^{\alpha'+\beta'+\gamma}
{\sigma_I}^1(x,\xi+s\eta)s^{\vert\alpha'\vert}ds \,
\partial_\eta^{\alpha''}\partial_\xi^{\beta''}\chi(\eta,\xi).$$ Since on the support of $\partial_\eta^{\alpha''}\partial_\xi^{\beta''}\chi$ one has $\langle \xi+s\eta\rangle\sim {\langle\xi\rangle}$, the first estimate of the lemma follows from the definition of the seminorms $M_k(\cdot)$ and (\[estadm\]).\
When $\sigma^1$ is a function, and since $\vert\alpha'\vert+\vert\beta'\vert+\vert\gamma\vert\geq 1$, we can use Prop. A.5 of [@Auscher-Taylor] which asserts that $\vert\partial_\xi^{\alpha'+\beta'+\gamma}{\sigma_I}^1(x,\xi)\vert
\leq {\mbox{\textnormal{Cst }}}\vert \sigma^1\vert_{C_*^r}{\langle\xi\rangle}^{-r-\vert\alpha'\vert
-\vert\beta'\vert-\vert\gamma\vert}$, from which one easily obtains the second estimate of the lemma.
Let $F(\cdot,\cdot)$ be a function defined on ${{\mathbb R}}^d_\eta\times{{\mathbb R}}^d_\xi$ and such that
- There exists $0<\delta<1$ such that $F(\eta,\xi)=0$ for all $\vert\eta\vert\geq \delta\vert\xi\vert$;
- For all $\alpha\in{{\mathbb N}}^d$, $\vert\alpha\vert\leq[d/2]+1$, there exists a constant $C_\alpha$ such that $$\forall \eta,\xi\in{{\mathbb R}}^d,\qquad
\left\vert \partial_\eta^\alpha F(\eta,\xi)\right\vert
\leq C_\alpha {\langle\xi\rangle}^{\mu-\vert\alpha\vert}.$$
Then, one has $$\forall\xi\in{{\mathbb R}}^d,\qquad
\left\vert {\check}{F}(\cdot,\xi)\right\vert_{L^1}
\leq {\mbox{\textnormal{Cst }}}\Big(\sup_{\vert\alpha\vert\leq [d/2]+1}C_\alpha\Big)
{\langle\xi\rangle}^\mu.$$
This result can be proved with the techniques used to prove the estimate (2.21) of Appendix B in [@Metivier-Zumbrun]. Briefly, and for the sake of completeness, we sketch the proof. Define $F^\flat(\eta,\xi):=F({\langle\xi\rangle}\eta,\xi)$ and remark that $\vert {\check}{F}(\cdot,\xi)\vert_{L^1}=
\vert {\check}{F^\flat}(\cdot,\xi)\vert_{L^1}\leq {\mbox{\textnormal{Cst }}}\vert F^\flat(\cdot,\xi)\vert_{H^{[d/2]+1}}$. The first assumption made in the statement of the lemma shows that $F^\flat(\cdot,\xi)$ is supported in the ball $\{\vert\eta\vert\leq 1\}$ so that it is easy to conclude using the second assumption.
Commutators with Fourier multipliers {#sectfour}
------------------------------------
We give in this section some commutator estimates between a Fourier multiplier and a pseudo-differential operator of limited regularity. We first set some notations:
[**Notations.**]{} For all $m\in{{\mathbb R}}$, $s_0>d/2$, and all symbols $\sigma\in\Gamma^m_{s_0}$, we define $$\label{defnorm1}
\forall s\leq s_0,\qquad
\Vert \sigma \Vert_{H^s_{(m)}}:=
\frac{1}{2}\left( n_{0,s}(\sigma)+N_{{{2[\frac{d}{2}]+2}},s}^m(\sigma)\right)$$ and, when $\sigma$ is also ${{2[\frac{d}{2}]+2}}$-regular at the origin, $$\label{defnorm2}
\forall s\leq s_0,\qquad
\Vert \sigma \Vert_{H^s_{reg,(m)}}:=
\frac{1}{2}\left( n_{{{2[\frac{d}{2}]+2}},s}(\sigma)+N_{{{2[\frac{d}{2}]+2}},s}^m(\sigma)\right).$$ Finally, if $\sigma$ is $d$-regular at the origin, we set $$\label{defnormnoel}
\Vert \sigma\Vert_{\infty,(m)}:=\frac{1}{2}
\left(
m_d(\sigma)+M^m_d(\sigma)
\right).$$ When no confusion is possible, we omit the subscript $m$ in the above definitions. Remark that when $\sigma$ does not depend on $\xi$ (i.e. when it is a function), then one has $\Vert\sigma\Vert_{H^s}=\Vert\sigma\Vert_{H^s_{reg}}
=\vert \sigma\vert_{H^s}$, and $\Vert\sigma\Vert_\infty=\vert \sigma\vert_\infty$.
The first commutator estimates we state are of Kato-Ponce type:
\[theoIII1\] Let $m_1,m_2\in{{\mathbb R}}$, $n\in{{\mathbb N}}$ and $d/2<t_0\leq s_0$. Let $\sigma^1(\xi)\in{\mathcal M}^{m_1}$ be $n$-regular at the origin and $\sigma^2(x,\xi)\in\Gamma^{m_2}_{s_0+m_1\wedge n+1}$. Then:\
[**i.**]{} For all $s\in{{\mathbb R}}$ such that $\max\{-t_0,-t_0-m_1\}<s\leq s_0+1$, one has $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}}\\
&\lesssim & C(\sigma^1)\big(
M_{d}^{m_2}(\nabla_x^{n+1}\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}
+\Vert \sigma^2\Vert_{H^{s+m_1\wedge n}}
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}\big),
\end{aligned}$$ where $C(\sigma^1):=
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_n(\sigma^1)$.\
[**ii.**]{} If moreover $\sigma^1$ is $({n+2+[\frac{d}{2}]+d})$-regular and $\sigma^2$ is $d$-regular at the origin, then the above estimate can be replaced by $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}}\\
&\lesssim & C'(\sigma^1)\big(
\Vert \nabla_x^{n+1}\sigma^2\Vert_\infty
\vert u\vert_{H^{s+m_1+m_2-n-1}}
+\Vert \nabla_x^{n+1}\sigma^2\Vert_{H^{s+m_1\wedge n-n-1}}
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}\big),
\end{aligned}$$ where $C'(\sigma^1):=
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_{{n+2+[\frac{d}{2}]+d}}(\sigma^1)$.
[**i.**]{} Remark that ${\{\sigma^1,\sigma^2\}_n}
=\sigma^1{\sharp_n}\sigma^2-\sigma^1\sigma^2$ and that ${\mbox{\textnormal{Op}}}(\sigma^2)\circ{\mbox{\textnormal{Op}}}(\sigma^1)={\mbox{\textnormal{Op}}}(\sigma^1\sigma^2)$ since ${\mbox{\textnormal{Op}}}(\sigma^1)$ is a Fourier multiplier. Therefore, one has $$\left[{\mbox{\textnormal{Op}}}(\sigma^1),{\mbox{\textnormal{Op}}}(\sigma^2)\right]-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)
={\mbox{\textnormal{Op}}}(\sigma^1)\circ{\mbox{\textnormal{Op}}}(\sigma^2)-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n\sigma^2).$$ Write now $\sigma^1(D)\circ{\mbox{\textnormal{Op}}}(\sigma^2)-{\mbox{\textnormal{Op}}}(\sigma^1{\sharp_n}\sigma^2)=
\sum_{j=1}^5\tau^j(x,D)$, with $$\begin{aligned}
\tau^1(x,D)&=&\sigma^1(D)\circ{\mbox{\textnormal{Op}}}(\sigma^2-{\sigma_I}^2-{\sigma_{lf}}^2),\\
\tau^2(x,D)&= &\sigma^1(D)\circ {\mbox{\textnormal{Op}}}({\sigma_{lf}}^2),\\
\tau^3(x,D)&=&\sigma^1(D)\circ {\mbox{\textnormal{Op}}}({\sigma_I}^2)
-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n\sigma_I^2),\\
\tau^4(x,D)&=&{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n{\sigma_I}^2
-(1-\psi(\xi))\sigma^1{\sharp_n}\sigma^2),\\
\tau^5(x,D)&=&-{\mbox{\textnormal{Op}}}(\psi(\xi)\sigma^1\sharp_n\sigma^2).\end{aligned}$$ We now turn to control the operator norms of $\tau^j(x,D)$, $j=1,\dots,5$.\
$\bullet$ [**Control of $\mathbf{\tau^1(x,D)}$.**]{} Since $\sigma^1(D)$ is a Fourier multiplier, one obtains easily that for all $s\in{{\mathbb R}}$ and $u\in{{\mathcal S}({{\mathbb R}}^d)}$, $$\vert \tau^1(x,D)u\vert_{H^s}\leq (m_0(\sigma^1)+M_0^{m_1}(\sigma^1))
\left\vert {\mbox{\textnormal{Op}}}(\sigma^2-{\sigma_I}^2-{\sigma_{lf}}^2)u\right\vert_{H^{s+m_1}}.$$ Using Prop. \[propII5\].i (with $r=m_1\wedge n-m_1$) gives therefore, for all $-t_0-m_1<s\leq s_0+1$, $$\label{bor1}
\vert \tau^1(x,D)u\vert_{H^s}\lesssim
(m_0(\sigma^1)+M_0^{m_1}(\sigma^1))
N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n}(\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ $\bullet$ [**Control of $\mathbf{\tau^2(x,D)}$.**]{} One has $$\vert \tau^2(x,D)u\vert_{H^s}\leq
(m_0(\sigma^1)+M^{m_1}_0(\sigma^1))
\vert {\mbox{\textnormal{Op}}}({\sigma_{lf}}^2)u\vert_{H^{s+m_1}},$$ so that it is a direct consequence of Prop. \[propII1\] that one has, for all $s\leq s_0+1$, $$\label{bor2}
\vert {\mbox{\textnormal{Op}}}(\tau^2)u\vert_{H^s}\lesssim
(m_0(\sigma^1)+M_0^{m_1}(\sigma^1))
n_{0,s+m_1}(\sigma^2) \vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ $\bullet$ [**Control of $\mathbf{\tau^3(x,D)}$.**]{} We have ${\mbox{\textnormal{Op}}}(\tau^3)={\mbox{\textnormal{Op}}}(\rho_n)$ with $\rho_n$ as given in the first part of Prop. \[lemmIII1\]. This lemma asserts that the symbol $\rho_n(x,\xi)$ satisfies the conditions of application of Lemma \[lemmII1\], which states that $\vert {\mbox{\textnormal{Op}}}(\rho_n)u\vert_{H^s}\lesssim
M^{m_1+m_2-n-1}_d(\rho_n)\vert u\vert_{H^{s+m_1+m_2-n-1}}$. Using the estimate of $M^{m_1+m_2-n-1}_d(\rho_n)$ given in Prop. \[lemmIII1\] shows therefore that for all $s\in{{\mathbb R}}$, $$\label{bor3}
\vert {\mbox{\textnormal{Op}}}(\tau^3)u\vert_{H^s}\lesssim
M^{m_1}_{{n+2+[\frac{d}{2}]+d}}(\sigma^1)M^{m_2}_{d}(\nabla_x^{n+1}\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}.$$ $\bullet$ [**Control of $\mathbf{\tau^4(x,D)}$.**]{} By definition of the product law $\sharp_n$, one has $$\begin{aligned}
\tau^4(x,\xi)&=&
-\sum_{\vert\alpha\vert\leq n}\frac{(-i)^\alpha}{\alpha!}
\partial_\xi^\alpha \sigma^1(\xi)
\partial_x^\alpha (\sigma^2-{\sigma_I}^2-{\sigma_{lf}}^2)(x,\xi)\\
&=&-\sum_{\vert\alpha\vert\leq n}\frac{(-i)^\alpha}{\alpha!}
{\mathbf 1}_{[1/2,\infty)}(\xi)
\partial_\xi^\alpha \sigma^1(\xi)
\partial_x^\alpha (\sigma^2-{\sigma_I}^2-{\sigma_{lf}}^2)(x,\xi),\end{aligned}$$ where ${\mathbf 1}_{[1/2,\infty)}(\cdot)$ denotes the characteristic function of the interval $[1/2,\infty)$. It follows that $$\big\vert \tau^4(x,D)u\big\vert_{H^s}\lesssim
\sum_{\vert\alpha\vert\leq n}
\big\vert{\mbox{\textnormal{Op}}}(\partial_x^\alpha\sigma^2-(\partial_x^\alpha\sigma^2)_I
-(\partial_x^\alpha\sigma^2)_{lf})v
\big\vert_{H^s}$$ with $v={\mbox{\textnormal{Op}}}({\mathbf 1}_{[1/2,\infty)}\partial_\xi^\alpha\sigma^1)u$; we now use the first estimate of Prop. \[propII5\] (with $r=m_1\wedge n-\vert\alpha\vert$) to obtain that the terms of the above sum are bounded from above by $N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n-\vert\alpha\vert}
(\partial_x^\alpha\sigma^2)\vert v
\vert_{H^{m_2+t_0-m_1\wedge n+\vert\alpha\vert}}$, for all $-t_0<s\leq s_0+1$. It is now straightforward to conclude that for all $-t_0<s\leq s_0+1$, $$\label{bor5}
\big\vert \tau^4(x,D)u\big\vert_{H^s}\lesssim
M^{m_1}_{n}(\sigma^1) N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n}(\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ $\bullet$ [**Control of $\mathbf{\tau^5(x,D)}$.**]{} One has ${\mbox{\textnormal{Op}}}(\psi(\xi)\sigma^1\sharp_n\sigma^2)u
={\mbox{\textnormal{Op}}}((\sigma^1\sharp_n\sigma^2)_{lf})u$, so that Prop. \[propII1\] can be used to obtain for all $s\leq s_0+1$, $$\label{bor6}
\left\vert{\mbox{\textnormal{Op}}}(\psi(\xi)\sigma^1\sharp_n\sigma^2)u\right\vert_{H^s}
\lesssim m_n(\sigma^1)n_{0,s+n}(\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$
Recalling that $\Vert \sigma^2\Vert_{H^s}$ is defined in (\[defnorm1\]), the estimate given in [**i.**]{} of the theorem now follows directly from (\[bor1\]), (\[bor2\]), (\[bor3\]), (\[bor5\]) and (\[bor6\]).\
[**ii.**]{} We use here another decomposition, namely $\sigma^1(D)\circ{\mbox{\textnormal{Op}}}(\sigma^2)-{\mbox{\textnormal{Op}}}(\sigma^1{\sharp_n}\sigma^2)=
\sum_{j=1}^6{\underline{\tau}}^j(x,D)$, with $$\begin{aligned}
{\underline{\tau}}^1(x,D)&=&\sigma^1(D)\circ
{\mbox{\textnormal{Op}}}(\sigma^2-{\sigma_I}^2-{\sigma_{lf}}^2-\sigma^2_{R,2}),\\
{\underline{\tau}}^2(x,D)&= &\sigma^1(D)\circ {\mbox{\textnormal{Op}}}(\Psi(D_x){\sigma_{lf}}^2+\sigma^2_{R,2})
-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n(\psi(D_x)\sigma^2_{lf}+\sigma^2_{R,2})),\\
{\underline{\tau}}^3(x,D)&=&\sigma^1(D)\circ {\mbox{\textnormal{Op}}}({\sigma_I}^2)
-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n\sigma_I^2),\\
{\underline{\tau}}^4(x,D)&=&-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n(\sigma^2-
{\sigma_I}^2-{\sigma_{lf}}^2
-\sigma^2_{R,2})),\\
{\underline{\tau}}^5(x,D)&=&\sigma^1(D)\circ{\mbox{\textnormal{Op}}}((1-\psi(D_x)\sigma^2_{lf})),\\
{\underline{\tau}}^6(x,D)&=&-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n(1-\psi(D_x)){\sigma_{lf}}^2).\end{aligned}$$ We now turn to control the operator norms of ${\underline{\tau}}^j(x,D)$, $j=1,\dots,6$.\
[**Control of $\mathbf{{\underline{\tau}}^1(x,D)}$.**]{} Proceeding as for the control of $\tau^1(x,D)$ in [**i.**]{} above, but using Prop. \[propII5bis\] instead of Prop. \[propII5\], one can replace $N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n}(\sigma^2)$ by $N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n-n-1}(\nabla_x^{n+1}\sigma^2)$ in (\[bor1\]).\
[**Control of $\mathbf{{\underline{\tau}}^2(x,D)}$.**]{} We need here two lemmas:
Let $m_1\in{{\mathbb R}}$, $n\in{{\mathbb N}}$ and $\sigma^1(\xi)\in{\mathcal M}^{m_1}$ be $({n+2+[\frac{d}{2}]+d})$-regular at the origin. Let $\sigma^2(x,\xi)$ be a symbol $d$-regular at the origin and such that $\widehat{\sigma^2}(\eta,\xi)$ is supported in the ball $\vert \eta\vert+\vert\xi\vert\leq A$, for some $A>0$.\
Then, $\sigma^1(D)\circ {\mbox{\textnormal{Op}}}(\sigma^2)={\mbox{\textnormal{Op}}}(\sigma^1\sharp_n\sigma^2)
+{\mbox{\textnormal{Op}}}(\rho_n)$, where the symbol $\rho_n(x,\xi)$ is such that $\widehat{\rho_n}(\eta,\xi)$ vanishes outside the ball $\vert\eta\vert+\vert\xi\vert\leq A$ and satisfies the estimate $$\begin{aligned}
\sup_{\vert\xi\vert\leq A}\sup_{\vert\beta\vert\leq d}
\vert \partial_\xi^\beta\rho_n(\cdot,\xi)\vert_\infty
&\leq& {\mbox{\textnormal{Cst }}}\sup_{\vert\xi\vert\leq 2A}
\sup_{\vert\alpha\vert\leq n+2+[\frac{d}{2}]+d}
\vert\partial_\xi^\alpha\sigma^1(\xi)\vert\\
& &\times\sup_{\vert\xi\vert\leq A}\sup_{\vert\alpha\vert\leq d}
\vert\nabla^{n+1}_x\partial_\xi^\alpha
\sigma^2(\cdot,\xi)\vert_\infty.
\end{aligned}$$
The proof is a close adaptation of the proof of Prop. \[lemmIII1\]. First replace the admissible cut-off function $\chi(\eta,\xi)$ used there by a smooth function $\widetilde{\chi}(\eta,\xi)$ supported in the ball $\vert\eta\vert+\vert\xi\vert\leq A$. Inequality (\[marre1\]) must then be replaced by $$\vert\partial_\xi^\beta\rho_n(x,\xi)\vert\leq{\mbox{\textnormal{Cst }}}\Big(\sum_{\vert\beta'\vert\leq d}\big\vert\partial_\xi^{\beta'}G_\gamma(\cdot,\xi)\big\vert_{L^1}\Big)
\sup_{\vert\xi\vert\leq A}\sup_{\vert \alpha\vert\leq d}\vert\nabla^{n+1}_x\partial_\xi^\alpha\sigma^2(\cdot,\xi)\vert_\infty,$$ for all $\vert\beta\vert\leq d$ and $\vert\xi\vert\leq A$.\
Finally, one concludes the proof as in Prop. \[lemmIII1\] after remarking that (\[marre2\]) can be replaced here by $$\big\vert\partial_\xi^{\beta'}G_\gamma(\cdot,\xi)\big\vert_{L^1}
\leq{\mbox{\textnormal{Cst }}}\sup_{\vert\xi\vert\leq 2A}
\sup_{\vert\alpha\vert\leq n+2+[\frac{d}{2}]+d}
\vert\partial_\xi^\alpha\sigma^1(\xi)\vert.$$
The proof of the following lemma is a straightforward adaptation of the proof of Lemma \[lemmII1\].
Let $\rho(x,\xi)$ be a symbol such that $\widehat{\rho}(\eta,\xi)$ is supported in the ball $\vert\eta\vert+\vert\xi\vert\leq A$, for some $A>0$. Then ${\mbox{\textnormal{Op}}}(\rho)$ extends as a continuous operator on every Sobolev space, with values in $H^\infty({{\mathbb R}}^d)$. Moreover, $$\forall s,t\in{{\mathbb R}},\forall u\in H^t({{\mathbb R}}^d),\qquad
\vert {\mbox{\textnormal{Op}}}(\rho)u\vert_{H^s}\leq {\mbox{\textnormal{Cst }}}\sup_{\vert\xi\vert\leq A}\sup_{\vert\alpha\vert\leq d}
\vert \partial_\xi^\alpha\rho(\cdot,\xi)\vert_\infty
\vert u\vert_{H^t},$$ where the constant depends only on $A$, $s$ and $t$.
From these two lemmas, one easily gets $$\vert {\underline{\tau}}^2(x,D)u\vert_{H^s}\lesssim C'(\sigma^1)
\Vert \nabla_x^{n+1}\sigma^2\Vert_\infty
\vert u\vert_{H^{s+m_1+m_2-n-1}}.$$ where $C'(\sigma^1)$ is as given in the statement of the theorem.\
[**Control of $\mathbf{{\underline{\tau}}^3(x,D)}$.**]{} One has ${\underline{\tau}}^3=\tau^3$, so that ${\underline{\tau}}^3(x,D)$ is controlled via (\[bor3\]).\
[**Control of $\mathbf{{\underline{\tau}}^4(x,D)}$.**]{} To control this term, proceed exactly as for the control of $\tau^4(x,D)$ above, but use Prop. \[propII5bis\] instead of Prop. \[propII5\]. This yields $$\big\vert {\underline{\tau}}^4(x,D)u\big\vert_{H^s}\lesssim
M^{m_1}_{n}(\sigma^1)
N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n-n-1}(\nabla_x^{n+1}\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ [**Control of $\mathbf{{\underline{\tau}}^5(x,D)}$ and $\mathbf{{\underline{\tau}}^6(x,D)}$.**]{} The difference between ${\underline{\tau}}^5(x,D)$ and $\tau^2(x,D)$ is that the operator $(1-\psi(D_x))$ is applied to ${\sigma_{lf}}^2(\cdot,\xi)$ in the former. This allows us to replace $n_{0,s+m_1}(\sigma^2)$ by $n_{0,s+m_1-n-1}(\nabla_x^{n+1}\sigma^2)$ in (\[bor2\]).\
A similar adaptation of (\[bor6\]) gives $$\big\vert{\mbox{\textnormal{Op}}}({\underline{\tau}}^6)u\big\vert_{H^s}
\lesssim m_n(\sigma^1)n_{0,s-1}(\nabla_x^{n+1}\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$
Point [**ii**]{} of the theorem thus follows from the estimates on ${\underline{\tau}}^j(x,D)$, $j=1,\dots,6$, proved above.
When $m_2\geq 0$, an easy adaptation of the above proof shows that the quantity $\Vert \sigma^2\Vert_{H^{s+m_1\wedge n}}
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}$ which appears in the r.h.s. of the first estimate of the theorem can be replaced by $\Vert \sigma^2\Vert_{H^{s+m_1\wedge n+m_2}}
\vert u\vert_{H^{t_0+m_1-m_1\wedge n}}$.
In the spirit of Theorem \[theoII2\], the following two theorems can be a useful alternative to Theorem \[theoIII1\]. Theorem \[theoIII2\] deals with the case of pseudo-differential operators $\sigma^2(x,D)$ of nonnegative order, while Th. \[theoIII3\] addresses the case $\sigma^2(x,\xi)=\sigma^2(x)$.
\[theoIII2\] Let $m_1\in{{\mathbb R}}$, $m_2>0$, $n\in{{\mathbb N}}$ and $s_0>d/2$. Let $\sigma^1(\xi)\in{\mathcal M}^{m_1}$ be $n$-regular at the origin and let $\sigma^2(x,\xi)\in \Gamma^{m_2}_{s_0+m_1\wedge n+1}$ be ${{2[\frac{d}{2}]+2}}$-regular at the origin. Then:\
[**i.**]{} For all $s$ such that $0<s+m_1$, $0<s$ and $s+m_2\leq s_0+1$, one has $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}}\\
&\lesssim& C(\sigma^1)\big(M_{d}^{m_2}(\nabla_x^{n+1}\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}
+
\Vert \sigma^2\Vert_{H^{s+m_1\wedge n+m_2}_{reg}}
\vert u\vert_{\infty}\big),
\end{aligned}$$ where $C(\sigma^1)=M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)
+m_n(\sigma^1)$.\
[**ii.**]{} If moreover $\sigma^1$ is $({n+2+[\frac{d}{2}]+d})$-regular at the origin, then the above estimate can be replaced by $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}}\\
&\lesssim& C'(\sigma^1)\big(\Vert \nabla_x^{n+1}\sigma^2\Vert_\infty
\vert u\vert_{H^{s+m_1+m_2-n-1}}
+
\Vert \nabla_x^{n+1}\sigma^2\Vert_{H^{s+m_1\wedge n+m_2-n-1}_{reg}}
\vert u\vert_{\infty}\big),
\end{aligned}$$ where $C'(\sigma^1):=
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_{{n+2+[\frac{d}{2}]+d}}(\sigma^1)$.
We only give a sketch of the proof of [**i.**]{} which follows the same lines as the proof of Theorem \[theoIII1\].i; [**ii.**]{} is obtained similarly. The modifications to be made are:\
- Inequality (\[bor1\]) must be replaced by $$\label{modif1}
\vert \tau^1(x,D)u\vert_{H^s}\lesssim
(m_0(\sigma^1)+M_0^{m_1}(\sigma^1))
N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_1+m_2}(\sigma^2)\vert u\vert_{\infty},$$ which holds for all $s\in{{\mathbb R}}$ such that $s+m_1>0$ and $s+m_2\leq s_0+1$. This is a consequence of Prop. \[propII5\].iii, which can be used since we assumed $m_2>0$.\
- Similarly, the second estimate of Prop. \[propII1\] allows us to replace (\[bor2\]) by $$\vert {\mbox{\textnormal{Op}}}(\tau^2)u\vert_{H^s}\lesssim
(m_0(\sigma^1)+M_0^{m_1}(\sigma^1))
n_{{{2[\frac{d}{2}]+2}},s+m_1}(\sigma^2) \vert u\vert_{\infty}.$$ - Inequality (\[bor3\]) is left unchanged.\
- Estimate (\[bor5\]) must be replaced by $$\label{modif2}
\big\vert \tau^4(x,D)u\big\vert_{H^s}\lesssim
M^{m_1}_n(\sigma^1) N^{m_2}_{{{2[\frac{d}{2}]+2}},s+m_2+m_1\wedge n}(\sigma^2)
\vert u\vert_{\infty},$$ which holds for all $0<s$ and $s+m_2\leq s_0+1$. One proves (\[modif2\]) in the same way as (\[bor5\]), using the third point of Prop. \[propII5\] rather than the first one (this is possible because $m_2>0$ here).\
- Finally, inequality (\[bor6\]) is replaced, using the second part of Prop. \[propII1\], by $$\big\vert{\mbox{\textnormal{Op}}}(\psi(\xi)\sigma^1\sharp_n\sigma^2)u\big\vert_{H^s}
\lesssim m_n(\sigma^1)n_{{{2[\frac{d}{2}]+2}},s+n}(\sigma^2)
\vert u\vert_{\infty}.$$
An interesting particular case is obtained when the symbol $\sigma^2(x,\xi)$ does not depend on $\xi$ (i.e., it is a function).
\[theoIII3\] Let $m_1\in{{\mathbb R}}$, $n\in{{\mathbb N}}$ and $s_0>d/2$. Let $\sigma^1(\xi)\in{\mathcal M}^{m_1}$ be $n$-regular at the origin and let $\sigma^2\in H^{s_0+m_1\wedge n+1}({{\mathbb R}}^d)$.\
[**i.**]{} If $m_1>n$ then for all $s$ such that $0\leq s\leq s_0+1$, one has $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),\sigma^2\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}}\\
&\leq &C(\sigma^1)\big(
\vert\sigma^2\vert_{W^{n+1,\infty}}
\vert u\vert_{H^{s+m_1-n-1}}
+\vert \sigma^2\vert_{H^{s+m_1}}
\vert u\vert_{\infty}\big),
\end{aligned}$$ where $C(\sigma^1)=
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_n(\sigma^1)$.\
[**ii.**]{} If moreover $\sigma^1$ is $({n+2+[\frac{d}{2}]+d})$-regular at the origin, then the above estimate can be replaced by $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),\sigma^2\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}}\\
&\leq &C'(\sigma^1)\big(
\vert\nabla^{n+1}\sigma^2\vert_{\infty}
\vert u\vert_{H^{s+m_1-n-1}}
+\vert \nabla^{n+1}\sigma^2\vert_{H^{s+m_1-n-1}}
\vert u\vert_{\infty}\big),
\end{aligned}$$ where $C'(\sigma^1):=
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_{{n+2+[\frac{d}{2}]+d}}(\sigma^1)$.
Here again, we only prove the first point of the theorem since the proof of the second one can be deduced similarly from the proof of Th. \[theoIII1\].ii.\
Remark that for all $k\in{{\mathbb N}}$, $s\in{{\mathbb R}}$ and $v\in{{\mathcal S}({{\mathbb R}}^d)}$, one has $N^0_{k,s}(v)=\vert v\vert_{H^s}$. Therefore, we just have to adapt the points of the proof of Th. \[theoIII2\] which use the assumption $m_2>0$, namely the obtention of (\[modif1\]) and (\[modif2\]).\
One can check that (\[modif1\]) remains true here. This is a consequence of Prop. \[propII5bis\].iv and Remark \[remII5bis\].\
We now prove that (\[modif2\]) can be replaced by $$\vert \tau^4(x,D)u\vert_{H^s}\lesssim M_n^{m_1}(\sigma^1)
\vert \sigma^2\vert_{H^{s+m_1}}\vert u\vert_{\infty},$$ which holds for all $0<s\leq s_0+1$ and provided that $m_1>n$, and which remains true when $s=0$ provided one adds $\vert \sigma^2\vert_{W^{n+1,\infty}}\vert u\vert_{H^{m_1-n-1}}$ to the right-hand-side. To obtain this, we need to control in $H^s$-norm, and for all $0\leq \vert\alpha\vert\leq n$, the terms ${\mbox{\textnormal{Op}}}(\partial_x^\alpha\sigma^2-(\partial_x^\alpha \sigma^2)_{I}
-(\partial_x^\alpha\sigma^2)_{lf})(\partial_\xi^\alpha \sigma^1(D))u$, which is done using Prop. \[propII5bis\].iv (with $r=m_1-\vert\alpha\vert>0$) and Remark \[remII5bis\].
We finally give commutator estimates of Calderon-Coifman-Meyer type:
\[theoIII1bis\] Let $m_1,m_2\in{{\mathbb R}}$, $n\in{{\mathbb N}}$ and $d/2<t_0\leq s_0$. Let $\sigma^1(\xi)\in{\mathcal M}^{m_1}$ be $n$-regular at the origin and $\sigma^2(x,\xi)\in\Gamma^{m_2}_{s_0+m_1\wedge n+1}$. Then:\
[**i.**]{} For all $s\in{{\mathbb R}}$ such that $-t_0<s\leq t_0+1$ and $-t_0<s+m_1\leq t_0+n+1$, $$\big\vert \big[\sigma^1(D),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}
\lesssim C(\sigma^1)
\Vert \sigma^2\Vert_{H^{t_0+n+1}}
\vert u\vert_{H^{s+m_1+m_2-n-1}},$$ where $C(\sigma^1):=
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_n(\sigma^1)$.\
[**ii.**]{} If moreover $\sigma^1$ is $({n+2+[\frac{d}{2}]+d})$-regular and $\sigma^2$ is $d$-regular at the origin, then the above estimate can be replaced by $$\big\vert \big[\sigma^1(D),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}
\lesssim C'(\sigma^1)
\Vert \nabla_x^{n+1}\sigma^2\Vert_{H^{t_0}}
\vert u\vert_{H^{s+m_1+m_2-n-1}},$$ where $C'(\sigma^1):=
M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_{{n+2+[\frac{d}{2}]+d}}(\sigma^1)$.
The proof follows closely the proof of Th. \[theoIII1\], so that we just mention the adaptations that have to be made.\
[**i.**]{} Below is the list of changes one must perform in the control of the operators $\tau^j(x,D)$.\
$\bullet$ [**Control of $\mathbf{\tau^1(x,D)}$.**]{} For all $-t_0<s+m_1\leq t_0+n+1$ and using Prop. \[propII5\].ii with $r'=n+1$ instead of Prop. \[propII5\].i, one obtains instead of (\[bor1\]) a control in terms of $N^{m_2}_{{{2[\frac{d}{2}]+2}},t_0+n+1}(\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}$.\
$\bullet$ [**Control of $\mathbf{\tau^2(x,D)}$.**]{} For $s+m_1\leq
t_0+n+1$, one just has to remark that Prop. \[propII1\] gives a control in terms of $n_{0,t_0+n+1}(\sigma^2) \vert u\vert_{H^{s+m_1+m_2-n-1}}$.\
$\bullet$ [**Control of $\mathbf{\tau^4(x,D)}$.**]{} If $-t_0<s\leq t_0+1$, one can replace (\[bor5\]) by a control in terms of $N^{m_2}_{{{2[\frac{d}{2}]+2}},t_0+n+1}(\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}$ provided that one uses Prop. \[propII5\].ii with $r'=n+1-\vert\alpha\vert$ instead of Prop. \[propII5\].i.\
$\bullet$ [**Control of $\mathbf{\tau^5(x,D)}$.**]{} When $s\leq t_0+1$ one gets easily a control in terms of $\Vert \sigma^2\Vert_{H^{t_0+1}}\vert u\vert_{H^{s+m_1+m_2-n-1}}$.\
[**ii.**]{} One deduces the second point of the theorem from Th. \[theoIII1\].ii exactly as we adapted the proof of the first point from the proof of Th. \[theoIII1\].i.
[**i.**]{} Extending results of Moser [@Moser] and Kato-Ponce [@Kato-Ponce], Taylor proved in [@TaylorM0] the following generalized Kato-Ponce estimates (which also hold in $L^p$-based Sobolev spaces): for all Fourier multiplier $\sigma^1(D)$ of order $m_1>0$, and all $\sigma^2\in H^\infty({{\mathbb R}}^d)$, one has for all $s\geq0$, $$\label{estTay}
\big\vert [{\mbox{\textnormal{Op}}}(\sigma^1),\sigma^2]u\big\vert_{H^s}
\leq \underline{C}(\sigma^1)\big(\vert\nabla\sigma^2\vert_{\infty}
\vert u\vert_{H^{s+m_1-1}}
+\vert \sigma^2\vert_{H^{s+m_1}}\vert u\vert_\infty\big),$$ where $\underline{C}(\sigma^1)$ is some constant depending on $\sigma^1$ (in [@TaylorM0], Taylor also deals with classical pseudo-differential operators $\sigma^1(x,D)$ –and not only Fourier multipliers–; we address this problem in Section \[sectfin\]).\
The estimate of Th. \[theoIII3\] coincides with (\[estTay\]) when $n=0$ (it is in fact more precise since it allows one to replace the term $\vert \sigma^2\vert_{H^{s+m_1}}$ by $\vert \nabla\sigma^2\vert_{H^{s+m_1-1}}$ [^1]); the general case $n\in {{\mathbb N}}$ gives an extension of this result involving the Poisson bracket of $\sigma^1$ and $\sigma^2$. Th. \[theoIII2\] extends (\[estTay\]) in another way, allowing $\sigma^2(x,D)$ to be a pseudo-differential operator of nonnegative order instead of a function. Finally, the most general extension of (\[estTay\]) is given by Th. \[theoIII1\], since it contains the two improvements just mentioned, allows estimates in Sobolev spaces of negative order, and does not assume cumbersome restrictions on the order of $\sigma^1(\xi)$ and $\sigma^2(x,\xi)$. For instance, (\[estTay\]) does not hold when $s<0$ or $m_1\leq 0$ but can be replaced by: for all Fourier multiplier $\sigma^1(D)$ of order $m_1\in{{\mathbb R}}$ regular at the origin, and all $\sigma^2\in H^\infty({{\mathbb R}}^d)$, one has for all $t_0>d/2$ and $s>max\{-t_0,-t_0-m_1\}$, $$\big\vert [{\mbox{\textnormal{Op}}}(\sigma^1),\sigma^2]u\big\vert_{H^s}
\leq C(\sigma^1)\big(\vert \nabla\sigma^2\vert_\infty
\vert u\vert_{H^{s+m_1-1}}
+\vert \nabla\sigma^2\vert_{H^{s+(m_1)_+-1}}
\vert u\vert_{H^{t_0+(m_1)_-}}\big),$$ with $(m_1)_+=\max\{m_1,0\}$ and $(m_1)_-=\min\{m_1,0\}$.\
[**ii.**]{} The Calderon-Coifman-Meyer commutator estimate of Th. \[theoIII1bis\] coincides with (\[intro3\]) when $n=0$ and $\sigma^1$ is a Fourier multiplier (the general case is addressed in Section \[sectfin\]) but its range of validity is wider since it allows negative values of $s$ and $m_1$. We also have the same kind of generalization as for the Kato-Ponce estimates.
In the particular case when the symbol $\sigma^2$ is of the form $\sigma^2(x,\xi)=\Sigma(v(x),\xi)$, one can now obtain easily, proceeding as in the proof of Corollary \[coroII1\]:
\[coroIII3\] Let $m_1,m_2\in{{\mathbb R}}$, $p\in{{\mathbb N}}$ and $d/2<t_0\leq s_0$. Let $\sigma^1(\xi)\in{\mathcal M}^{m_1}$ be $n$-regular at the origin and $\sigma^2(x,\xi)=\Sigma^2(v(x),\xi)$ with $\Sigma^2\in C^\infty({{\mathbb R}}^p,{\mathcal M}^{m_2})$ and $v\in H^{s_0+m_1\wedge n+1}({{\mathbb R}}^d)^p$. Then,\
[**i.**]{} For all $s\in{{\mathbb R}}$ such that $\max\{-t_0,-t_0-m_1\}<s\leq s_0+1$ $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),\sigma^2(x,D)\big]u
-{\mbox{\textnormal{Op}}}({\{\sigma^1,\sigma^2\}_n})
u\big\vert_{H^s}}\\
&\leq& C(\sigma^1)C_{\Sigma^2}(\vert v\vert_{W^{n+1,\infty}})
\big(\vert u\vert_{H^{s+m_1+m_2-n-1}}
+
\vert v\vert_{H^{(s+m_1\wedge n)_+}}
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}\big).
\end{aligned}$$ [**ii.**]{} For all $s\in{{\mathbb R}}$ such that $-t_0<s<t_0+1$ and $-t_0<s+m_1\leq t_0+n+1$, one also has $$\big\vert \big[\sigma^1(D),\sigma^2(x,D)\big]u
-{\mbox{\textnormal{Op}}}({\{\sigma^1,\sigma^2\}_n})
u\big\vert_{H^s}
\leq C(\sigma^1)C_{\Sigma^2}(\vert v\vert_{\infty})
\vert v\vert_{H^{t_0+n+1}}\vert u\vert_{H^{s+m_1+m_2-n-1}}.$$ [**iii.**]{} If moreover $m_2>0$ and $\Sigma^2$ is ${{2[\frac{d}{2}]+2}}$-regular at the origin, one has, for all $s\in{{\mathbb R}}$ such that $0<s+m_1$, $0<s$, and $s+m_2\leq s_0+1$, $$\begin{aligned}
\lefteqn{\big\vert \big[\sigma^1(D),\sigma^2(x,D)\big]u
-{\mbox{\textnormal{Op}}}({\{\sigma^1,\sigma^2\}_n})
u\big\vert_{H^s}}\\
& \leq& C(\sigma^1)C_{\Sigma^2}(\vert v\vert_{W^{n+1,\infty}})
\big(\vert u\vert_{H^{s+m_1+m_2-n-1}}
+
\vert v\vert_{H^{s+m_2+m_1\wedge n}}\vert u\vert_{\infty}\big).
\end{aligned}$$ In the above, $C_{\Sigma^2}(\cdot)$ denotes a smooth nondecreasing function depending only on a finite number of derivatives of $\Sigma^2$ and $C(\sigma^1)=M^{m_1}_{n+2+[\frac{d}{2}]+d}(\sigma^1)+m_n(\sigma^1)$.
Let $\sigma(x,\xi)$ be the symbol given by (\[symbDN\]), with $a\in H^\infty({{\mathbb R}}^d)$, and let $m\geq 0$. Then for all $t_0>d/2$ and $s>-t_0$, one has $$\left\vert \left[\langle D\rangle^m,\sigma(x,D)\right]u
\right\vert_{H^s}
\leq C(\vert \nabla a\vert_{W^{1,\infty}})
\left(
\vert u\vert_{H^{s+m}}+\vert \nabla a\vert_{H^{(s+m)_+}}
\vert u\vert_{H^{1+t_0}}\right),$$ and, writing $\widetilde{m}:=\max\{m,1\}$ $$\left\vert \left[\langle D\rangle^m,{\mbox{\textnormal{Op}}}(\sigma)\right]u
-{\mbox{\textnormal{Op}}}(\tau)u\right\vert_{H^s}
\leq C(\vert \nabla a\vert_{W^{2,\infty}})
\left(
\vert u\vert_{H^{s+m-1}}+\vert \nabla a\vert_{H^{(s+\widetilde{m})_+}}
\vert u\vert_{H^{1+t_0}}\right),$$ where the symbol $\tau(x,\xi)$ is given by $$\tau(x,\xi):=m\langle \xi\rangle^{m-2}
\frac{\vert\xi\vert^2 d^2a(\nabla a,\xi)-
(\nabla a\cdot \xi)d^2a(\xi,\xi)}{\sigma(x,\xi)}.$$ For $s>0$, we also have $$\left\vert \left[\langle D\rangle^m,\sigma(x,D)\right]u
\right\vert_{H^s}
\leq C(\vert \nabla a\vert_{W^{1,\infty}})
\left(
\vert u\vert_{H^{s+m}}+\vert \nabla a\vert_{H^{1+s+m}}
\vert u\vert_{{\infty}}\right).$$ Finally, if $s$ and $t_0$ are such that $-t_0< s$ and $s+m\leq t_0+1$, then $$\left\vert \left[\langle D\rangle^m,\sigma(x,D)\right]u
\right\vert_{H^s}
\leq C(\vert \nabla a\vert_{W^{1,\infty}})
\vert \nabla a\vert_{H^{t_0+1}}\vert u\vert_{H^{s+m}}.$$
Composition and commutators between two pseudo-differential operators of limited regularity {#sectfin}
-------------------------------------------------------------------------------------------
This section is devoted to the proof of composition and commutator estimates involving two pseudo-differential operators of limited regularity. We first introduce the following notations:
[**Notations.**]{} For all $s_0>d/2$, $n\in{{\mathbb N}}$, and all symbols $\sigma\in\Gamma^m_{s_0}$ $n$-regular at the origin, we define $$\label{defnorm3}
\forall s\leq s_0,\qquad
\Vert \sigma \Vert_{H^s_{n,(m)}}:=
\frac{1}{2}\left( n_{n,s}(\sigma)+N_{{n+2+[\frac{d}{2}]+d},s}^m(\sigma)\right)$$ and $$\label{defnorm4}
\forall k\in{{\mathbb N}}, \quad 0\leq k\leq n,\qquad
\Vert \sigma \Vert_{W^{k,\infty}_{n,(m)}}:=
\frac{1}{2}\left( m_{n}(\sigma)+M_{{n+2+[\frac{d}{2}]+d},k}^m(\sigma)\right).$$ When no confusion is possible, we omit the subscript $m$. Remark that when $\sigma$ does not depend on $\xi$ (i.e. when it is a function), then one has $\Vert\sigma\Vert_{H^s_n}=\vert \sigma\vert_{H^s}$ and $\Vert \sigma\Vert_{W^{k,\infty}_n}=\vert \sigma\vert_{W^{k,\infty}}$.
We first give commutator estimates of Kato-Ponce type:
\[theoIV1\] Let $m_1,m_2\in{{\mathbb R}}$, $n\in{{\mathbb N}}$ and $d/2<t_0\leq s_0$. Define $m:=m_1\wedge m_2$ and let $\sigma^j\in \Gamma^{m_j}_{s_0+m\wedge n+1}$ ($j=1,2$) be two symbols $n$-regular at the origin.\
[**i.**]{} For all $s$ such that $\max\{-t_0,-t_0-m_1\}<s\leq t_0+1$, one has $$\begin{aligned}
\lefteqn{\big\vert {\mbox{\textnormal{Op}}}(\sigma^1)\circ {\mbox{\textnormal{Op}}}(\sigma^2)u
-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n\sigma^2)u
\big\vert_{H^s}}\\
&\lesssim&\Vert\sigma^1\Vert_{W^{n+1,\infty}_n}
\Vert\sigma^2\Vert_{W^{n+1,\infty}}
\vert u\vert_{H^{s+m_1+m_2-n-1}}\\
&+&\big(\Vert \sigma^1\Vert_{H^{t_0+1}_n}
\Vert\sigma^2\Vert_{H^{s_++m_1\wedge n}}
+\Vert \sigma^1\Vert_{H^{s_++m_1\wedge n}_n}
\Vert \sigma^2\Vert_{H^{t_0}}\big)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}},
\end{aligned}$$ where $s_+=\max\{s,0\}$ and with the notations (\[defnorm1\])-(\[defnorm2\]) and (\[defnorm3\])-(\[defnorm4\]).\
[**i.bis.**]{} Under the same assumptions, one also has $$\begin{aligned}
\lefteqn{\big\vert {\mbox{\textnormal{Op}}}(\sigma^1)\circ {\mbox{\textnormal{Op}}}(\sigma^2)u
-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n\sigma^2)u
\big\vert_{H^s}}\\
&\lesssim&\Vert \sigma^1\Vert_{H^{t_0+m_1\wedge n+1}_n}\big(
\Vert\sigma^2\Vert_{H^{s+m_1\wedge n}}
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}
+\Vert\sigma^2\Vert_{W^{n+1,\infty}}
\vert u\vert_{H^{s+m_1+m_2-n-1}}\big).
\end{aligned}$$ [**i.ter.**]{} For $s\in{{\mathbb R}}$ such that $max\{t_0+1,-t_0-m_1\}\leq s \leq s_0+1$, the estimate of [**i.**]{} still holds if one adds $\Vert \sigma^1\Vert_{H^s}\Vert \sigma^2\Vert_{H^{m_1\wedge n+t_0+1}} \vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}$ to the right-hand–side.\
[**ii.**]{} For all $s\in{{\mathbb R}}$ such that $\max\{-t_0,-t_0-m_1,-t_0-m_2\}<s\leq t_0+1$, one has $$\big\vert \big[{\mbox{\textnormal{Op}}}(\sigma^1),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}\lesssim I(\sigma^1,\sigma^2,m_1,m_2)+
I(\sigma^2,\sigma^1,m_2,m_1),$$ where $I(\sigma^1,\sigma^2,m_1,m_2)$ denotes the r.h.s. or the estimate of [**i**]{}, [**i.bis**]{} or [**i.ter.**]{} In this latter case, the range of application of the estimate is bounded from above by $s_0+1$ instead of $t_0+1$.\
[**iii.**]{} If $m_1>0$, $\sigma^1$ is ${{2[\frac{d}{2}]+2}}$ regular at the origin, and $\sigma^2(x,\xi)=\sigma^2(x)$ does not depend on $\xi$ then, for all $0\leq s\leq t_0+1$, $$\big\vert \big[{\mbox{\textnormal{Op}}}(\sigma^1),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
\big\vert_{H^s}
\lesssim\Vert \sigma^1\Vert_{H^{t_0+1}_{reg}}\big(
\vert\sigma^2\vert_{H^{s+m_1}}
\vert u\vert_{\infty}
+\vert\sigma^2\vert_{W^{1,\infty}}
\vert u\vert_{H^{s+m_1-1}}\big).$$
One has $$\begin{aligned}
[{\mbox{\textnormal{Op}}}(\sigma^1),{\mbox{\textnormal{Op}}}(\sigma^2)]-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\})&=&
{\mbox{\textnormal{Op}}}(\sigma^1)\circ{\mbox{\textnormal{Op}}}(\sigma^2)-{\mbox{\textnormal{Op}}}(\sigma^1\sharp_n\sigma^2)\\
& &-\left({\mbox{\textnormal{Op}}}(\sigma^2)\circ{\mbox{\textnormal{Op}}}(\sigma^1)
-{\mbox{\textnormal{Op}}}(\sigma^2\sharp_n\sigma^1)\right)\\
&:=&\tau_1(x,D)-\tau_2(x,D).\end{aligned}$$ Controlling the operator norm of $\tau_1(x,D)$ gives the composition estimates of the theorem. To obtain the commutators estimates, we need also a control on $\tau_2(x,D)$; since this latter is obtained by a simple permutation of $\sigma^1$ and $\sigma^2$, we just have to treat $\tau_1(x,D)$. We decompose this operator into $ \tau_1(x,D)=\sum_{j=1}^7\tau_1^j(x,D)$ with $$\begin{aligned}
\tau_1^1(x,D)&:=&{\mbox{\textnormal{Op}}}(\sigma^1)\circ{\mbox{\textnormal{Op}}}(\sigma^2-{\sigma_{lf}}^2-{\sigma_I}^2)\\
\tau_1^2(x,D)&:=&{\mbox{\textnormal{Op}}}(\sigma^1)\circ{\mbox{\textnormal{Op}}}({\sigma_{lf}}^2)\\
\tau_1^3(x,D)&:= &
\left[{\mbox{\textnormal{Op}}}({\sigma_I}^1)\circ{\mbox{\textnormal{Op}}}({\sigma_I}^2)-{\mbox{\textnormal{Op}}}({\sigma_I}^1\sharp_n{\sigma_I}^2)\right]\\
\tau_1^4(x,D)&:= &\left[{\mbox{\textnormal{Op}}}({\sigma_I}^1\sharp_n{\sigma_I}^2)
-{\mbox{\textnormal{Op}}}((1-\psi(\xi))\sigma^1\sharp_n\sigma^2)\right]\\
\tau_1^5(x,D)&:= &{\mbox{\textnormal{Op}}}((\sigma^1\sharp_n\sigma^2)_{lf})\\
\tau_1^6(x,D)&:=&{\mbox{\textnormal{Op}}}({\sigma_{lf}}^1)\circ{\mbox{\textnormal{Op}}}({\sigma_I}^2)\\
\tau_1^7(x,D)&:=&{\mbox{\textnormal{Op}}}(\sigma^1-{\sigma_{lf}}^1-{\sigma_I}^1)\circ {\mbox{\textnormal{Op}}}({\sigma_I}^2).\end{aligned}$$ The proof reduces therefore to the control of $\vert \tau_1^j(x,D)u\vert_{H^s}$ for all $j=1,\dots,7$.\
$\bullet$ [**Control of $\mathbf{\tau_1^1(x,D)}$ and $\mathbf{\tau_1^2(x,D)}$.**]{} Using the first estimate of Th. \[theoII1\], one gets that for all $-t_0<s\leq t_0+1$, $$\label{blob}
\vert\tau_1^j(x,D)u\vert_{H^s}
\lesssim \Vert \sigma^1\Vert_{H^{t_0+1}}
\vert v_j\vert_{H^{s+m_1}},
\qquad (j=1,2)$$ with $v_1:={\mbox{\textnormal{Op}}}(\sigma^2-{\sigma_{lf}}^2-{\sigma_I}^2)u$ and $v_2:={\mbox{\textnormal{Op}}}({\sigma_{lf}}^2)u$.\
Proceeding as for the estimates of $\tau^1$ and $\tau^2$ in the proof of Th. \[theoIII1\], one obtains, for all $\max\{-t_0,-t_0-m_1\}<s\leq t_0+1$, $$\label{IV1}
\vert \tau_1^j(x,D)u\vert_{H^s}\lesssim
\Vert \sigma^1\Vert_{H^{t_0+1}}\Vert \sigma^2\Vert_{H^{s+m_1\wedge n}}
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ $\bullet$ [**Control of $\mathbf{\tau_1^3(x,D)}$.**]{} A direct use of Prop. \[lemmIII1\] and Lemma \[lemmII1\] yields, for all $s\in{{\mathbb R}}$, $$\label{IV3}
\vert\tau_1^3(x,D)u\vert_{H^s}\lesssim
M^{m_1}_{{n+2+[\frac{d}{2}]+d}}(\sigma^1)M_{d}^{m_2}(\nabla_x^{n+1}\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}.$$ $\bullet$ [**Control of $\mathbf{\tau_1^4(x,D)}$.**]{} Obviously, it suffices to control the operator norm of $A_\alpha(x,D):={\mbox{\textnormal{Op}}}(\partial_\xi^\alpha\sigma^1_I\partial_x^\alpha\sigma_I^2)
-{\mbox{\textnormal{Op}}}((1-\psi(\xi))\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2)$ for all $0\leq\vert\alpha\vert\leq n$. Let us introduce here ${\widetilde{N}}:=N+3$; we can assume that the paradifferential decomposition (\[A6\]) used in this proof is done using the integer ${\widetilde{N}}$ instead of $N$. To enhance this fact, we write momentaneously ${\sigma_{{\widetilde{I}}}}$ the paradifferential symbol associated to any symbol $\sigma$ using (\[A7\]) with $N$ replaced by ${\widetilde{N}}$. We denote ${\sigma_I}$ when $N$ is used. We can write $A_\alpha(x,D)=A_\alpha^1(x,D)+A_\alpha^2(x,D)$ with $$\begin{aligned}
\label{compo1}
A_\alpha^1(x,D)&:=&
{\mbox{\textnormal{Op}}}\left(\partial_\xi^\alpha{\sigma_{{\widetilde{I}}}}^1\partial_x^\alpha{\sigma_{{\widetilde{I}}}}^2
-(\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2)_I\right),\\
\label{compo2}
A_\alpha^2(x,D)&:=&-
{\mbox{\textnormal{Op}}}\left(\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2
-(\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2)_{lf}
-(\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2)_I\right).\end{aligned}$$ The operator norm of $A_\alpha^1(x,D)$ is controlled using the next lemma, whose proof is postponed to Appendix \[applemm\] to ease the readability of the present proof. Note that this result is in the spirit of the main result of [@Yamazaki2], but that the estimate given in this latter reference is not useful here.
\[lemmfond\] Let $\sigma^1(x,\xi)$ and $\sigma^2(x,\xi)$ be as in the statement of the theorem, and let $\alpha\in{{\mathbb N}}^d$ be such that $0\leq\vert\alpha\vert\leq n$. Then, $$\forall u\in {{\mathcal S}({{\mathbb R}}^d)},\qquad \left\vert A^1_\alpha(x,D)u
\right\vert_{H^s}
\lesssim
M^{m_1}_{n+d,n+1}(\sigma^1)M^{m_2}_{d,n+1}(\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}},$$ where $A_\alpha^1(x,D)$ is defined in (\[compo1\]).
To control $A_\alpha^2(x,D)$, we use Prop. \[propII5\].i. with $r=n\wedge m_1-\vert \alpha\vert$ to obtain, for all $-t_0\leq s\leq s_0+1$, $$\left\vert A^2_\alpha(x,D)u
\right\vert_{H^s}
\lesssim
N^{m_1+m_2-\vert\alpha\vert}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n-\vert\alpha\vert}
(\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ Using classical tame product estimates, one gets easily $$\label{adapt}
N^{m_1+m_2-\vert\alpha\vert}_{{{2[\frac{d}{2}]+2}},s+m_1\wedge n-\vert\alpha\vert}
(\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2)
\lesssim
\Vert \sigma^1\Vert_{W^{0,\infty}_n}
\Vert \sigma^2\Vert_{H^{s_+ +m_1\wedge n}}
+
\Vert \sigma^1\Vert_{H^{s_+ +m_1\wedge n}_n}
\Vert \sigma^2\Vert_{W^{0,\infty}_0}.$$ We have thus proved the following estimate on ${\mbox{\textnormal{Op}}}(\tau^4_1)$, for all $-t_0<s\leq s_0+1$: $$\begin{aligned}
\label{IV4}
\vert {\mbox{\textnormal{Op}}}(\tau^4_1)u\vert_{H^s}\lesssim
M^{m_1}_{n+d,n+1}(\sigma^1)M^{m_2}_{d,n+1}(\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}\\
\nonumber
+\big(\Vert \sigma^1\Vert_{W^{0,\infty}_n}
\Vert \sigma^2\Vert_{H^{s_+ +m_1\wedge n}}
+
\Vert \sigma^1\Vert_{H^{s_+ +m_1\wedge n}_n}
\Vert \sigma^2\Vert_{W^{0,\infty}_0}\big)\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}} \end{aligned}$$ $\bullet$ [**Control of $\mathbf{\tau_1^5(x,D)}$.**]{} Using again classical tame product estimates, one gets $$n_{0,s}((\sigma^1\sharp_n\sigma^2)_{lf})
\lesssim
m_n(\sigma^1)n_{0,s_++n}(\sigma^2)
+n_{n,s_++n}(\sigma^1)m_{0,0}(\sigma^2)$$ so that by Prop. \[propII1\], one obtains, for all $s\leq s_0+1$, $$\label{IV5}
\vert {\mbox{\textnormal{Op}}}(\tau^5_1)u\vert_{H^s}\lesssim
\big(m_n(\sigma^1)n_{0,s_++n}(\sigma^2)
+n_{n,s_++n}(\sigma^1)m_{0,0}(\sigma^2)\big)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ $\bullet$ [**Control of $\mathbf{\tau_1^6(x,D)}$.**]{} Using successively Props. \[propII1\] and \[propII2\] one obtains, for all $s\leq s_0+n+1$, $$\label{IV6}
\vert {\mbox{\textnormal{Op}}}(\tau^6_1)u\vert_{H^s}\lesssim
n_{0,s}(\sigma^1)M_d^{m_2}(\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ $\bullet$ [**Control of $\mathbf{\tau_1^7(x,D)}$.**]{} Using successively Props. \[propII5\].i (with $r=m_1\wedge n$) and \[propII2\] one obtains, for all $-t_0<s\leq s_0+n+1$, $$\label{IV7}
\vert {\mbox{\textnormal{Op}}}(\tau^7_1)u\vert_{H^s}\lesssim
N_{0,s+m_1\wedge n}^{m_1}(\sigma^1)M_d^{m_2}(\sigma^2)
\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}.$$ $\bullet$ [**Proof of i.**]{} Gathering the estimates (\[IV1\])-(\[IV3\]) and (\[IV4\])-(\[IV7\]), and using standard Sobolev embeddings yields the result.\
$\bullet$ [**Proof of i.bis**]{} When $-t_0<s\leq t_0+1$ one can replace the r.h.s. of (\[adapt\]) by $\Vert \sigma^1\Vert_{H^{t_0+m_1\wedge n+1}_n}
\Vert \sigma^2\Vert_{H^{s+m_1\wedge n}}$ and modify subsequently (\[IV4\]). Similarly, one can modify (\[IV5\]) remarking that $n_{0,s}((\sigma^1\sharp_n\sigma^2)_n)\lesssim
\Vert \sigma^1\Vert_{H^{t_0+1}_n}\Vert \sigma^2\Vert_{H^{s+n}}$. Remarking also that in (\[IV6\]) on can replace $\vert u\vert_{H^{m_1+m_2+t_0-m_1\wedge n}}$ by $\vert u\vert_{H^{s+m_1+m_2-n-1}}$ and that (\[IV7\]) can be replaced by $$\vert {\mbox{\textnormal{Op}}}(\tau^7_1)u\vert_{H^s}\lesssim
N_{0,t_0+n+1}^{m_1}(\sigma^1)M_d^{m_2}(\sigma^2)
\vert u\vert_{H^{s+m_1+m_2-n-1}}$$ if one uses Prop. \[propII5\].ii (with $r'=n+1$) rather than Prop. \[propII5\].i, one gets [**i.bis**]{}.\
[**Proof of i.ter**]{} In the proof of [**i.**]{}, the only estimate which is not valid when $t_0+1< s\leq s_0+1$ is (\[IV1\]). To give control of $\tau_1^1$ and $\tau_1^2$ one now has to use the second estimate of Th. \[theoII1\] instead of the first one, whence the additional term in the final estimate.\
[**Proof of ii.**]{} As said above, the control of $\tau_2$ is deduced from the control of $\tau_1$ by a simple permutation of $\sigma^1$ and $\sigma^2$, whence the result.\
[**Proof of iii.**]{} When $n=0$, one has $\tau_1^4=\tau^4_2$, so that the control of both term is not needed to estimate the commutator. We keep the same control of $\tau_1^3$ as in the proof of [**i**]{} and explain the modifications that must be performed to control $\tau_1^j$, $j=1,2,5,6,7$ (of course, the components of $\tau_2$ are treated the same way).\
[**Control of $\mathbf{\tau^1_1(x,D)}$ and $\mathbf{\tau_1^2(x,D)}$.**]{} Since $\sigma^2$ is a function, one can invoke Prop. \[propII5bis\].iv (see also Remark \[remII5bis\]) and Lemma \[lemmII0\] to deduce from (\[blob\]) that for all $0\leq s\leq t_0+1$, $$\label{bay1}
\vert \tau_1^j(x,D)u\vert_{H^s}\lesssim
\Vert \sigma^1\Vert_{H^{t_0+1}}\vert \sigma^2\vert_{H^{s+m_1}}
\vert u\vert_{\infty}\qquad (j=1,2).$$ [**Control of $\mathbf{\tau^5_1(x,D)}$** ]{} Since $\sigma_1\sharp_0\sigma^2=\sigma^1\sigma^2$, and because $\sigma^1$ is ${{2[\frac{d}{2}]+2}}$-regular at the origin, we can use Prop. \[propII1\] to get $\vert \tau_1^5(x,D)u\vert_{H^s}\lesssim
N_{{{2[\frac{d}{2}]+2}},s}(\sigma^1\sigma^2)\vert u\vert_\infty$. It follows easily that for all $0\leq s\leq t_0+1$, $$\label{bay2}
\vert \tau_1^5(x,D)u\vert_{H^s}\lesssim
\Vert \sigma^1\Vert_{H^{t_0+1}_{reg}}
\vert \sigma^2\vert_{H^{s}}\vert u\vert_\infty.$$ $\bullet$ [**Control of $\mathbf{\tau_1^6(x,D)}$.**]{} Using successively Props. \[propII1\] and \[propII2\] as in [**i**]{} one can also obtain, for all $s\leq t_0+1$, $$\label{bay3}
\vert {\mbox{\textnormal{Op}}}(\tau^6_1)u\vert_{H^s}\lesssim
n_{0,s}(\sigma^1)\vert \sigma^2\vert_\infty
\vert u\vert_{H^{s+m_1-1}}.$$ $\bullet$ [**Control of $\mathbf{\tau_1^7(x,D)}$.**]{} Using successively Props. \[propII5\].ii (with $r'=1$) and \[propII2\] one obtains, for all $0\leq s\leq t_0+1$, $$\label{bay4}
\vert {\mbox{\textnormal{Op}}}(\tau^7_1)u\vert_{H^s}\lesssim
N_{0,t_0+1}^{m_1}(\sigma^1)\vert \sigma^2\vert_\infty
\vert u\vert_{H^{s+m_1-1}}.$$
Point [**iii**]{} of the theorem follows from (\[IV3\]) and (\[bay1\])-(\[bay4\]).
Taylor proved in [@TaylorM0] that for all classical pseudo-differential operator $\sigma^1(x,D)$ of order $m_1>0$ and $\sigma^2\in H^\infty({{\mathbb R}}^d)$, one has, for all $s\geq 0$, $$\left\vert [{\mbox{\textnormal{Op}}}(\sigma^1),\sigma^2]u\right\vert_{H^s}
\leq \underline{C}(\sigma^1)\left(\vert\sigma^2\vert_{W^{1,\infty}}
\vert u\vert_{H^{s+m_1-1}}
+\vert \sigma^2\vert_{H^{s+m_1}}\vert u\vert_\infty\right).$$ This is exactly the estimate of Th. \[theoIV1\].iii, which also gives a description of the constant $\underline{C}(\sigma^1)$. The commutator estimate corresponding to [**i.bis**]{} generalizes this result taking into account the smoothing effect of the Poisson bracket. It turns out that this estimate is not tame with respect to $m_1$, while the commutator estimate corresponding to [**i**]{}, which is not *stricto sensu* of Kato-Ponce type, is tame.
We also give commutator estimates of Calderon-Coifman-Meyer type:
\[theoIV1bis\] Let $m_1,m_2\in{{\mathbb R}}$, $n\in{{\mathbb N}}$ and $d/2<t_0\leq s_0$. Let $\sigma^j(x,\xi)\in\Gamma^{m_2}_{s_0+m_j\wedge n+1}$ ($j=1,2$) be $n$-regular at the origin.\
For all $s\in{{\mathbb R}}$ such that $-t_0<s+m_j\leq t_0+n+1$ ($j=1,2$) and $-t_0<s\leq t_0+1$, one has $$\big\vert \big[{\mbox{\textnormal{Op}}}(\sigma^1),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}
\lesssim \Vert \sigma^1\Vert_{H^{t_0+n+1}_n}
\Vert \sigma^2\Vert_{H^{t_0+n+1}}
\vert u\vert_{H^{s+m_1+m_2-n-1}}.$$
The proof follows closely the proof of Th. \[theoIV1\], so that we just mention the adaptations that have to be made.\
$\bullet$ [**Control of $\mathbf{\tau^j_1(x,D)}$, $j=1,2,4,5$.**]{} One just has to do as in the proof of Th. \[theoIII1bis\].\
$\bullet$ [**Control of $\mathbf{\tau^6_1(x,D)}$.**]{} When gets easily from Props. \[propII1\] and \[propII2\] that for all $s\leq t_0+1$, one has $\vert \tau_1^6(x,D)u\vert_{H^s}\lesssim
\Vert \sigma^1\Vert_{H^{t_0+1}}
\Vert \sigma^2\Vert_\infty \vert u\vert_{H^{s+m_1+m_2-n-1}}$.\
$\bullet$ [**Control of $\mathbf{\tau^7_1(x,D)}$.**]{} By Prop. \[propII5\].ii (with $r'=n+1$) and Prop. \[propII2\], one obtains $\vert \tau_1^7(x,D)u\vert_{H^s}\lesssim
\Vert \sigma^1\Vert_{H^{t_0+n+1}}
\Vert \sigma^2\Vert_\infty \vert u\vert_{H^{s+m_1+m_2-n-1}}$, for all $-t_0<s\leq t_0+1$.
When $n=0$ and $\sigma^2$ is a function, the estimate of Th. \[theoIV1bis\] is exactly the Calderon-Coifman-Meyer estimate (\[intro3\]), with extended range of validity. Th. \[theoIV1bis\] is also more general in the sense that it allows $n>0$ and $\sigma^2$ to be a pseudo-differential operator.
When the symbols $\sigma^1$ and $\sigma^2$ are of the form (\[A0\]), one gets the following corollary:
\[coroIV1\] Let $m_1,m_2\in{{\mathbb R}}$, $m:=m_1\wedge m_2$, and $d/2<t_0\leq s_0$. Let also $\sigma^j(x,\xi)=\Sigma^j(v^j(x),\xi)$ with $p_j\in{{\mathbb N}}$, $\Sigma^j\in C^\infty({{\mathbb R}}^{p_j},{\mathcal M}^{m_j})$ and $v_j\in H^{s_0+m\wedge n+1}({{\mathbb R}}^d)^{p_j}$ ($j=1,2$). Assume moreover that $\Sigma^1$ and $\Sigma^2$ are $n$-regular at the origin.\
[**i.**]{} For all $s\in{{\mathbb R}}$ such that $\min\{-t_0,-t_0-m_1,-t_0-m_2\}\leq s\leq s_0+1$ the following estimate holds (writing $v:=(v^1,v^2)$) $$\begin{aligned}
\left\vert [{\mbox{\textnormal{Op}}}(\sigma^1),{\mbox{\textnormal{Op}}}(\sigma^2)]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)
\right\vert_{H^s}\lesssim C(\vert v\vert_{W^{n+1,\infty}})
\vert u\vert_{H^{s+m_1+m_2-n-1}}\\
+
C(\vert v\vert_{W^{n+1,\infty}})\big(\vert v^1\vert_{H^{t_0+1}}
\vert v^2\vert_{H^{s_++m\wedge n}}+\vert v^2\vert_{H^{t_0+1}}
\vert v^1\vert_{H^{s_++m\wedge n}}\big)
\vert u\vert_{H^{m+t_0}};
\end{aligned}$$ [**ii.**]{} For all $s\in{{\mathbb R}}$ such that $-t_0<s+m_j\leq t_0+n+1$ ($j=1,2$) and $-t_0<s\leq t_0+1$, one has $$\big\vert \big[{\mbox{\textnormal{Op}}}(\sigma^1),{\mbox{\textnormal{Op}}}(\sigma^2)\big]u
-{\mbox{\textnormal{Op}}}(\{\sigma^1,\sigma^2\}_n)u
\big\vert_{H^s}
\lesssim C(\vert v\vert_\infty)\vert \sigma^1\vert_{H^{t_0+n+1}}
\vert \sigma^2\vert_{H^{t_0+n+1}}
\vert u\vert_{H^{s+m_1+m_2-n-1}}.$$
Writing $\sigma^j(x,\xi)=\left[\sigma^j(x,\xi)-\Sigma^j(0,\xi)\right]+\Sigma^j(0,\xi)$, the result follows from Lemma \[lemmA0\], Ths. \[theoIII1\] and \[theoIV1\] (for [**i**]{}) and Ths. \[theoIII1bis\] and \[theoIV1bis\] (for [**ii**]{}).
Proof of Prop. \[propA2\] {#appCM}
=========================
Owing to (\[LP3\])-(\[LP4\]), we can write ${\displaystyle}(1-\psi(\xi))\sigma(x,\xi)=\sum_{q\geq -1}\sigma_q(x,\xi){\langle\xi\rangle}^m$, with $\sigma_q(x,\xi):=(1-\psi(\xi))\sigma(x,\xi)\varphi_q(\xi){\langle\xi\rangle}^{-m}$. Obviously, for all $q\geq -1$, $\sigma_q(x,\xi)=0$ if $\vert\xi\vert\geq 2^{q+1}$ or $\vert\xi\vert\leq 2^{q-1}$; it follows that the function ${\displaystyle}A_q(x,\xi):=\sum_{k\in{{\mathbb Z}}^d}\sigma_q(x,2^{q+1}(\xi-2k\pi))$ is $2\pi$-periodic with respect to $\xi$ and coincides with $\sigma_q(x,2^{q+1}\xi)$ in the box ${\mathcal C}:=\{\xi\in{{\mathbb R}}^d, -\pi\leq \xi_j\leq\pi, j=1,\dots,d\}$. Therefore, we can write $\sigma_q(x,2^{q+1}\xi)=A_q(x,\xi)\lambda(\xi)$, where $\lambda\in C_0^\infty({{\mathbb R}}^d)$ is supported in $1/5\leq\vert\xi\vert\leq 6/5$ and $\lambda(\xi)=1$ in $1/4\leq\vert\xi\vert\leq 1$.\
Expending $A_q(x,\xi)$ into Fourier series, one obtains $$A_q(x,\xi)
=\sum_{k\in{{\mathbb Z}}^d}\frac{1}{(1+\vert k\vert^2)^{1+[d/2]}}c_{k,q}(x)
e^{i\xi\cdot k},$$ with $$c_{k,q}(x)=(1+\vert k\vert^2)^{1+[d/2]}(2\pi)^{-d}\int_{{\mathcal C}}
e^{-i\xi\cdot k}\sigma_q(x,2^{q+1}\xi)d\xi,$$ so that $$\begin{aligned}
\sigma_q(x,\xi)&=&A_q(x,2^{-q-1}\xi)\lambda(2^{-q-1}\xi)\\
&=&\sum_{k\in{{\mathbb Z}}^d}\frac{1}{(1+\vert k\vert^2)^{1+[d/2]}}
c_{k,q}(x)\lambda_k(2^{-q}\xi),\end{aligned}$$ where $\lambda_k(\xi):=e^{i\xi\cdot k/2}\lambda(\xi/2)$ and satisfies therefore the properties announced in the statement of the proposition.\
The last step is therefore to obtain the desired estimates on the Fourier coefficients $c_{k,q}$. By integration by parts, one obtains first $$c_{k,q}(x)=
(2\pi)^{-d}\int_{{\mathcal C}}
e^{-i\xi\cdot k}\left[(1-2^{2(q+1)}\Delta_\xi)^{1+[d/2]}\sigma_q\right]
(x,2^{q+1}\xi)d\xi,$$ which we can rewrite as $$\label{enfin1}
c_{k,q}(x)=(2\pi)^{-d}
\int_{{\mathcal C}} e^{-i\xi\cdot k}
\sum_{\vert\alpha\vert\leq 2+2[d/2]}*_\alpha 2^{(q+1)\vert\alpha\vert}
(\partial_\xi^\alpha\sigma_q)(x,2^{q+1}\xi)d\xi,$$ where, here and in the following, $*_{\alpha}$ denotes some numerical coefficient depending on $\alpha$ and whose precise value is not important.\
Recalling that for $q\geq 0$ (we omit the case $q=-1$ which does not raise any difficulty), $\sigma_q(x,\xi)=\widetilde{\sigma}(x,\xi)\varphi(2^{-q}\xi)$ with $\widetilde{\sigma}(x,\xi):=(1-\psi(\xi))\sigma(x,\xi){\langle\xi\rangle}^{-m}$, one obtains, for all $\alpha\in{{\mathbb N}}^d$, $$\partial_\xi^\alpha\sigma_q(x,\xi)
=\sum_{\alpha'+\alpha''=\alpha}*_{\alpha',\alpha''}
\partial_\xi^{\alpha'}\widetilde{\sigma}(x,\xi)
2^{-q\vert \alpha''\vert}(\partial_\xi^{\alpha''}\varphi)(2^{-q}\xi);$$ it follows that $$\begin{aligned}
\label{enfin2}
\lefteqn{2^{(q+1)\vert\alpha\vert}
\partial_\xi^\alpha\sigma_q(x,2^{q+1}\xi)
=}\\
\nonumber
& &\sum_{\alpha'+\alpha''=\alpha}*_{\alpha',\alpha''}
\frac{2^{(q+1)\vert\alpha'\vert}}
{\langle2^{q+1}\xi\rangle^{\vert\alpha'\vert}}
\langle 2^{q+1}\xi\rangle^{\vert\alpha'\vert}
\partial_\xi^{\alpha'}\widetilde{\sigma}(x,2^{q+1}\xi)
(\partial_\xi^{\alpha''}\varphi)(2\xi).\end{aligned}$$ Since ${\displaystyle}\frac{2^{(q+1)\vert\alpha'\vert}}
{\langle2^{q+1}\xi\rangle^{\vert\alpha'\vert}}\leq {\mbox{\textnormal{Cst }}}$ on the support of $\partial_\xi^{\alpha''}\varphi(2\cdot)$, it follows from (\[enfin1\]) and (\[enfin2\]) that $$\vert c_{k,q}\vert_{H^s}
\leq (2\pi)^{-d}\sum_{\vert\alpha\vert\leq2+2[d/2]}
\sum_{\alpha'+\alpha''=\alpha}*_{\alpha',\alpha''}
\int_{{\mathcal C}}
\langle 2^{q+1}\xi\rangle^{\vert\alpha'\vert}
\left\vert \partial_\xi^{\alpha'}
\widetilde{\sigma}(\cdot,2^{q+1}\xi)\right\vert_{H^s}
d\xi,$$ from which the estimate on $c_{k,q}$ follows. The estimate on $\varphi_p(D)c_{k,q}$ is proved in a similar way.
Proof of Lemma \[lemmfond\] {#applemm}
===========================
Throughout this proof, we write, $\psi_p(\cdot):=\psi(2^{-p}\cdot)$, for all $p\in{{\mathbb Z}}$. By definition of ${\sigma_{{\widetilde{I}}}}^1$ and ${\sigma_{{\widetilde{I}}}}^2$, one has $$\begin{aligned}
\lefteqn{\partial_\xi^\alpha{\sigma_{{\widetilde{I}}}}^1(\cdot,\xi)
\partial_x^\alpha {\sigma_{{\widetilde{I}}}}^2(\cdot,\xi)=
\sum_{q\geq -1}\partial_\xi^\alpha\left(
\psi_{q-{\widetilde{N}}}(D_x)\sigma^1(\cdot,\xi)\varphi_q(\xi)
(1-\psi(\xi))\right)}\\
& &\times \sum_{p\geq -1}\psi_{p-{\widetilde{N}}}(D_x)
\partial_x^\alpha\sigma^2(\cdot,\xi)\varphi_p(\xi)
(1-\psi(\xi))\\
&= &\sum_{p\geq -1} \sum_{h=0,\pm 1}
\partial_\xi^\alpha\left(
\psi_{p+h-{\widetilde{N}}}(D_x)\sigma^1(\cdot,\xi)\varphi_{p+h}(\xi)
(1-\psi(\xi))\right)\\
& &\times \psi_{p-{\widetilde{N}}}(D_x)
\partial_x^\alpha\sigma^2(\cdot,\xi)\varphi_p(\xi)
(1-\psi(\xi)),\end{aligned}$$ the last equality being a consequence of the fact that for all $p\geq -1$, one has $\varphi_p\varphi_q=0$ for all $q\neq p,p\pm 1$.\
Remarking that for all $p\geq -1$, the $p$-th term of the above summation has a spectrum included in the ball $\{\vert\eta\vert\leq 2^{p+2-{\widetilde{N}}}\}$, and recalling that ${\widetilde{N}}=N+3$, one deduces that $$\label{lemmfond1}
\partial_\xi^\alpha{\sigma_{{\widetilde{I}}}}^1(\cdot,\xi)
\partial_x^\alpha {\sigma_{{\widetilde{I}}}}^2(\cdot,\xi)=\sum_{p\geq -1}
\psi_{p-N}(D_x)\theta_p(\cdot,\xi)
\varphi_p(\xi)(1-\psi(\xi))$$ with $$\begin{aligned}
\theta_p(\cdot,\xi)&:=& \psi_{p-N-3}(D_x)
\partial_x^\alpha\sigma^2(\cdot,\xi)\\
& &\times \sum_{h=0,\pm 1}
\partial_\xi^\alpha\left(
\psi_{p+h-N-3}(D_x)\sigma^1(\cdot,\xi)\varphi_{p+h}(\xi)
(1-\psi(\xi))\right).\end{aligned}$$ We now turn to study the term $(\partial_\xi^\alpha \sigma^1\partial_x^\alpha\sigma^2)_I$. By definition, one has $$\begin{aligned}
(\partial_\xi^\alpha \sigma^1\partial_x^\alpha\sigma^2)_I
(\cdot,\xi)
&=&
\sum_{p\geq -1} \psi_{p-N}(D_x)
\left(\partial_\xi^\alpha\sigma^1\partial_x^\alpha\sigma^2\right)
(\cdot,\xi)
\varphi_p(\xi)(1-\psi(\xi))\\
&=&\sum_{p\geq -1} \psi_{p-N}(D_x)
\Theta_p(\cdot,\xi)
\varphi_p(\xi)(1-\psi(\xi)),\end{aligned}$$ with $$\Theta_p(\cdot,\xi):=
\sum_{h=0,\pm 1}\partial_\xi^\alpha\left(\sigma^1(\cdot,\xi)
\varphi_{p+h}(\xi)\right)
\partial_x^\alpha\sigma^2(\cdot,\xi),$$ where we used (\[LP3\]) and the fact that $\varphi_p\varphi_q=0$ if $\vert p-q\vert\geq 2$.\
Decomposing $\sigma^1$ into $\sigma^1=\psi_{p+h-N-3}(D_x)\sigma^1+(1-\psi_{p+h-N-3}(D_x))\sigma^1$, one obtains $$\begin{aligned}
\Theta_p(\cdot,\xi)&=&
\sum_{h=0,\pm 1}\partial_\xi^\alpha
\left(\psi_{p+h-N-3}(D_x)\sigma^1(\cdot,\xi)\varphi_{p+h}(\xi)\right)
\partial_x^\alpha\sigma^2(\cdot,\xi)\\
&+&
\sum_{h=0,\pm 1}\partial_\xi^\alpha\left((1-\psi_{p+h-N-3}(D_x))
\sigma^1(\cdot,\xi)\varphi_{p+h}(\xi)\right)
\partial_x^\alpha\sigma^2(\cdot,\xi).\end{aligned}$$ Remark that in the first term of the r.h.s. of the above identity, one can replace $\partial_x^\alpha\sigma^2$ by $\psi_{p-N+1}(D_x)\partial_x^\alpha\sigma^2$, so that one finally gets $$\begin{aligned}
\nonumber
\Theta_p(\cdot,\xi)&=&\theta_p(\cdot,\xi)\\
\nonumber
&+&\sum_{h=0,\pm 1}\partial_\xi^\alpha
\left(\psi_{p+h-N-3}(D_x)\sigma^1(\cdot,\xi)\varphi_{p+h}(\xi)\right)\\
\nonumber
& &\times
(\psi_{p-N+1}(D_x)-\psi_{p-N-3}(D_x))
\partial_x^\alpha\sigma^2(\cdot,\xi)\\
\nonumber
&+&
\sum_{h=0,\pm 1}\partial_\xi^\alpha((1-\psi_{p+h-N-3}(D_x))
\sigma^1(\cdot,\xi)\varphi_{p+h}(\xi))
\partial_x^\alpha\sigma^2(\cdot,\xi)\\
\label{lemmfond2}
&:=&\theta_p(\cdot,\xi)+\Theta^1_p(\cdot,\xi)+\Theta^2_p(\cdot,\xi).\end{aligned}$$ It follows therefore from (\[lemmfond1\]) and (\[lemmfond2\]) that $${\mbox{\textnormal{Op}}}\big(
(\partial_\xi^\alpha\sigma^1\partial_x^\alpha \sigma^2)_I-
\partial_\xi^\alpha{\sigma_{{\widetilde{I}}}}^1\partial_x^\alpha {\sigma_{{\widetilde{I}}}}^2\big)
={\mbox{\textnormal{Op}}}\big(\Theta^1(x,\xi))+{\mbox{\textnormal{Op}}}(\Theta^2(x,\xi)\big),$$ where ${\displaystyle}\Theta^j(\cdot,\xi)=\sum_{p\geq -1}\psi_{p-N}(D_x)
\Theta^j_p(\cdot,\xi)\varphi_p(\xi)(1-\psi(\xi))$, $j=1,2$.\
Quite obviously, the symbols $\Theta^1(x,\xi)$ and $\Theta^2(x,\xi)$ satisfy the assumptions of Lemma \[lemmII1\]. The result follows therefore from this lemma and the estimates $$\label{lemmfond3}
M_d^{m_1+m_2-n-1}(\Theta^1)
\leq {\mbox{\textnormal{Cst }}}M^{m_1}_{n+d}(\sigma^1)
M^{m_2}_{d,n+1}(\sigma^2)$$ and $$M_d^{m_1+m_2-n-1}(\Theta^2)
\leq {\mbox{\textnormal{Cst }}}M^{m_1}_{n+d,n+1}(\sigma^1)
M^{m_2}_{d,n}(\sigma^2).$$ We only prove the first of these two estimates, the second one being obtained in a similar way. One easily obtains that $$\begin{aligned}
\vert \Theta^1(\cdot ,\xi)\vert_\infty&\leq& {\mbox{\textnormal{Cst }}}\sup_{p\geq -1}
\vert \Theta^1_p(\cdot,\xi)\varphi_p(\xi)\vert_\infty\\
&\leq&{\mbox{\textnormal{Cst }}}\sup_{p\geq -1}
\vert \partial_\xi^\alpha(\sigma^1(\cdot,\xi)\varphi_{p}(\xi))
\vert_\infty\\
& &\times
\sup_{p\geq -1} \vert (\psi_{p-N-1}(D_x)-\psi_{p-N-3}(D_x))
\partial_x^\alpha\sigma^2(\cdot,\xi)\varphi_p(\xi)\vert_\infty.\end{aligned}$$ Since for all $r\in{{\mathbb N}}$, $p\geq -1$, and $f\in{{\mathcal S}({{\mathbb R}}^d)}$, one has $\vert (\psi_{p+2}(D)-\psi_p(D))f\vert_\infty\leq
{\mbox{\textnormal{Cst }}}2^{-pr}\vert f\vert_{W^{r,\infty}}$, it follows that (taking $r=n+1-\vert\alpha\vert$), $$\vert \Theta^1(\cdot ,\xi)\vert_\infty
\leq {\mbox{\textnormal{Cst }}}{\langle\xi\rangle}^{m_1+m_2-n-1} M^{m_1}_n(\sigma^1)
M^{m_2}_{0,n}(\nabla_x\sigma^2).$$ The derivatives of $\Theta^1$ with respect to $\xi$ can be handled in the same way, thus proving (\[lemmfond3\]).
[**Acknowledgments.**]{} The author warmly thanks T. Alazard and G. Métivier for fruitful discussions about this work, and B. Texier for his remarks on a previous version of this work. This work was partially supported by the ACI Jeunes Chercheuses et Jeunes Chercheuses “Dispersion et nonlinéarités”.
[00]{} <span style="font-variant:small-caps;">P. Auscher, M. Taylor</span>, [*Paradifferential operators and commutator estimates*]{}, Comm. Partial Differential Equations [**20**]{} (1995), no. 9-10, 1743–1775.
<span style="font-variant:small-caps;">S. Benzoni-Gavage, R. Danchin, S. Descombes</span>, [*Multi-dimensional Korteweg model*]{}, Preprint.
<span style="font-variant:small-caps;">J.-M. Bony</span>, [*Calcul symbolique et propagation des singularités pour les équations aux dérivés partielles non linéaires*]{}, Ann. scient. École. Norm. Sup. (4) [**14**]{} (1981), no. 2, 209–246.
<span style="font-variant:small-caps;">G. Bourdaud</span>, [*Une algèbre maximale d’opérateurs pseudo-différentiels*]{}, Comm. PDE [**13**]{} (1988), 1059-1083.
<span style="font-variant:small-caps;">J.-Y. Chemin</span>, [*Fluides parfaits incompressibles*]{}, Astérisque No. 230, (1995), 177 pp.
<span style="font-variant:small-caps;">R. R. Coifman, Y. Meyer</span>, [*Au-delà des opérateurs pseudo-différentiels*]{}, Astérisque, 57. Société Mathématique de France, Paris, 1978. i+185 pp.
<span style="font-variant:small-caps;">E. Grenier</span>, [*Pseudo-differential energy estimates of singular perturbations*]{}, Comm. Pure Appl. Math. [**50**]{} (1997), no. 9, 821–865.
<span style="font-variant:small-caps;">L. Hörmander</span>, [*Lectures on Nonlinear Hyperbolic Differential Equations*]{}, Mathématiques & Applications, [**26**]{}. Springer-Verlag, Berlin, 1997.
<span style="font-variant:small-caps;">I. L. Hwang</span>, [*The $L\sp 2$-boundedness of pseudodifferential operators*]{}, Trans. Amer. Math. Soc. [**302**]{} (1987), no. 1, 55–76.
<span style="font-variant:small-caps;">T. Kato, G. Ponce</span>, [*Commutator estimates and the Euler and Navier-Stokes equations*]{}, Comm. Pure Appl. Math. [**41**]{} (1988), no. 7, 891–907.
<span style="font-variant:small-caps;">D. Lannes</span>, [*Well-posedness of the water-waves equations*]{}, J. Amer. Math. Soc. [**18**]{} (2005), 605-654.
<span style="font-variant:small-caps;">J. Marschall</span>, [*Pseudodifferential operators with coefficients in Sobolev spaces*]{}, Trans. Amer. Math. Soc. [**307**]{} (1988), no. 1, 335–361.
<span style="font-variant:small-caps;">G. Métivier, K. Zumbrun</span>, [*Large viscous boundary layers for noncharacteristic nonlinear hyperbolic problems*]{}, to appear in Memoirs of the Amer. Math. Soc.
<span style="font-variant:small-caps;">Y. Meyer</span>, [*Remarques sur un théorème de J. M. Bony*]{}, Supplemento al Rendiconti der Circolo Matematico di Palermo, Serie II, No1, 1981.
<span style="font-variant:small-caps;">Y. Meyer, R.R. Coifman</span>, [*Ondelettes et opérateurs. III, Opérateurs multilinéaires*]{}, Actualités Mathématiques. Hermann, Paris, 1991. pp. i–xii and 383–538.
<span style="font-variant:small-caps;">J. Moser</span>, [*A rapidly convergent iteration method and nonlinear partial differential equations, I*]{}, Ann. Scuola Norm. Sup. Pisa [**20**]{} (1966), 265-315.
<span style="font-variant:small-caps;">M. Taylor</span>, [*Pseudodifferential Operators and Nonlinear PDE*]{}, Progress in Mathematics, [**100**]{}. Birkhäuser, Boston-Basel-Berlin, 1991.
<span style="font-variant:small-caps;">M. Taylor</span>, [*Partial differential equations. III. Nonlinear equations*]{}, Applied Mathematical Sciences, [**117**]{}. Springer-Verlag, New York, 1997.
<span style="font-variant:small-caps;">M. Taylor</span>, [*Tools for PDE. Pseudodifferential operators, paradifferential operators, and layer potentials*]{}, Mathematical Surveys and Monographs, [**81**]{}. American Mathematical Society, Providence, RI, 2000. x+257 pp.
<span style="font-variant:small-caps;">M. Taylor</span>, [*Commutator estimates*]{}, Proc. Amer. Math. Soc. [**131**]{} (2003), no. 5, 1501–1507.
<span style="font-variant:small-caps;">M. Yamazaki</span>, [*A quasi-homogeneous version of paradifferential operators, I. Boundedness on spaces of Besov type*]{}, J. Fac. Sci. Univ. Tokyo, Sect. IA, Math. [**33**]{} (1986), 131-174.
<span style="font-variant:small-caps;">M. Yamazaki</span>, [*A quasi-homogeneous version of paradifferential operators, II. A symbol calculus*]{}, J. Fac. Sci. Univ. Tokyo, Sect. IA, Math. [**33**]{} (1986), 311-345.
[^1]: Such an improvement has been proved recently in [@Benzoni-Danchin-Descombes] for $n=0$ or $n=1$, in the case where $\sigma^1(\xi)={\langle\xi\rangle}^{m_1}$ and $\sigma^2$ does not depend on $\xi$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we study the regularity of non-linear parabolic PDEs and stochastic PDEs on metric measure spaces admitting heat kernel estimates. In particular we consider mild function solutions to abstract Cauchy problems and show that the unique solution is Hölder continuous in time with values in a suitable fractional Sobolev space. As this analysis is done via a-priori estimates, we can apply this result to stochastic PDEs on metric measure spaces and solve the equation in a pathwise sense for almost all paths. The main example of noise term is of fractional Brownian type and the metric measure spaces can be classical as well as given by various fractal structures. The whole approach is low dimensional and works for spectral dimensions less than 4.'
address:
- 'University of Leeds, School of Mathematics, Leeds, LS2 9JT, UK'
- 'Friedrich Schiller University, Institute of Mathematics, 07743 Jena, Germany'
author:
- Elena Issoglio
- Martina Zähle
title: Regularity of the solutions to SPDEs in metric measure spaces
---
Introduction
============
In this paper the following non-linear Cauchy problem $$\label{eq: introduction: Cauchy pb}
\begin{cases}
\frac{\partial u}{\partial t} = -A u + F(u) + G(u)\cdot \dot{z}, & t\in(0,t_0]\\
u(0)=f
\end{cases}$$ is considered on $\sigma $-finite metric measure spaces $(X,\mu ,d)$. Here $t_0>0$ is arbitrary, $-A$ is the generator of a [*Markovian strongly continuous symmetric semigroup*]{} $\{T(t), t\geq0\}$ on $L_2(\mu)$, $F$ and $G$ are sufficiently regular functions. The term $\dot{z}$ denotes a fractional space-time perturbation which will be made more precise later on. In the case of linear spaces it can be interpreted as a formal time derivative of a spatial distribution. Solutions to are considered in the mild form, formally given by $$\begin{gathered}
\label{eq: introduction: mild sol}
u(t)= T(t)f + \int_0^t T(t-s)F(u(s)) {\mathrm{d}}s + \int_0^t T(t-s)G(u(s)) {\mathrm{d}}z( s).\end{gathered}$$ This formulation is, in a first place, only formal. We will give it a proper mathematical meaning as it is done in [@HZ12], in particular the last term involving the noise $z$ is defined by means of fractional derivatives and it is shown to be indeed well defined using the notion of pointwise product of functions and “distributions”. The spaces which we will use to describe the space-regularity of the solution are fractional Sobolev spaces defined on metric measure spaces by means of the associated semigroups.
The main aim of this paper is to show that the mild solution $u$ of the Cauchy problem given by is $\gamma$-Hölder continuous in time with respect to the $H^\delta(\mu)$-norm in space (Theorem \[thm: gamma holder regularity\]). This result is achieved under the same assumptions as in Theorem 1.2 in [@HZ12], where it was shown that a unique solution exists and belongs to the space $W^\gamma([0,t_0], H^\delta_\infty(\mu))$. Moreover, under slightly stricter conditions, we can show that the solution in fact belongs to any space $W^\gamma([0,t_0], H^\delta_\infty(\mu))\cap C^\gamma([0,t_0], H^\delta(\mu))$ for all $\gamma$ and $\delta$ smaller than certain parameters determined by the regularity properties of the distributional noise $z$ and the initial function $f$ (Corollary \[cor: extra holder regularity\]). In the case of noises of fractional Brownian type these parameters are determined my means of the Hurst exponents in space and in time. Note that the fractional time regularity is always greater than $1/2$.\
In Remark \[rem: low spectral dim\] we outline an extension of the results to an appropriate parameter condition for the case of spectral dimensions $\le 1$, which has not been considered in [@HZ12]. In particular, white noise in space can be treated in this low dimensional case.
Deterministic elliptic equations and some parabolic equations without noises on classes of fractals have been studied, e.g. in [@Ba98; @Fa99; @FH01; @GHL03; @St06; @FHS12].\
Abstract problems with Brownian and fractional Brownian noises have been considered in many papers with various approaches, in particular, in [@Wa86; @dPZ92; @GA99; @MN03; @TTV03; @GLT06; @PR07; @FKN11]. None of these covers the results of the present paper.\
Some relationships have been discussed in [@HZ12], see also [@HIZ14]. To these two references, we only add a brief comparison between the present paper and the rough path approach developed in recent years. For example in a series of papers [@HV11; @HW13; @HMW14] the rough path approach has been applied to study a stochastic Burger-type equation with multiplicative white noise. Even though the equation considered there is of a different kind than the one studied in this paper (the setting is Euclidean and not metric measure space, the noise is white and not coloured and there is a non-linear product term of the form $g(u)\partial_x u$ which we do not have), it is interesting to notice that one of the main difficulties is to give a meaning to the solution, in particular to the non-linear product term $g(u)\partial_x u$, and this is done using the notion of *paraproduct*.
Preliminaries
=============
Semigroups and potential spaces {#semigroups_potential spaces}
-------------------------------
Throughout the paper we use the letter $c$ for a general finite constant which might change value from line to line.\
We first recall some basic notions and relationships which are known from the literature. In the case of metric measure spaces the analogues of the classical fractional Sobolev (or Bessel potential) spaces in the literature are introduced by means of the given semigroup $\{T(t), t\geq0\}$, i.e., of its generator $-A$:
The *generalized Bessel potential operator* on $L_2(\mu)$ is defined for $\sigma\geq 0$ as $$J^\sigma (\mu):= (A+{\operatorname{Id}})^{-\sigma/2}.$$ To each operator there corresponds a *potential space* defined as $$H^\sigma(\mu):= J^\sigma(L_2(\mu))$$ and equipped with the norm $\|u\|_{H^\sigma(\mu)}:= \| u\|_{L_2(\mu)}+\|A^{\sigma/2} u\|_{L_2(\mu)}$, which is equivalent to $\|(A+{\operatorname{Id}})^{\sigma/2}u\|_{L_2(\mu)}$. In fact these spaces correspond to the domains of fractional powers of $A$, i.e., $D((A+{\operatorname{Id}})^{\sigma/2}) =D(A^{\sigma/2})= H^{\sigma}(\mu)$. In particular, for any $\alpha\geq0$ the operator $J^\alpha$ acts as an isomorphisms between $H^\sigma(\mu)$ and $H^{\alpha + \sigma}(\mu)$. Analogously one can define the potential spaces corresponding to the generators $-A_p, 1<p<\infty$, of Markovian semigroups on $L_p(\mu)$. They are denoted by $H_p^\sigma(\mu)$ and clearly $H_2^\sigma(\mu)=H^\sigma(\mu)$. We will also consider the spaces $$H^\sigma_\infty(\mu) := H^\sigma(\mu)\cap L_\infty(\mu)$$ normed by $\|\cdot\|_{H^\sigma(\mu)}+\|\cdot\|_\infty$, with slight abuse of notation. Here the norm $\|\cdot\|_\infty$ in $L_\infty(\mu)$ is given by the essential supremum.\
The dual spaces of $H^\sigma_p(\mu)$ will be used in the sequel: for $1<p<\infty, \sigma\geq0$ they are denoted by $$H^{-\sigma}_{p'}(\mu):= \left(H^\sigma_p(\mu)\right)^*,$$ where $\frac1p + \frac{1}{p'}=1$. In case $p=2$ we do not write $p$ explicitly. Note that $H^{-\sigma}(\mu)\subseteq \left(H^\sigma_\infty(\mu)\right)^*$ often being a strict inclusion.
For the regularity in time of the solution we consider the following spaces frequently used in the literature: Let $0<\eta<1$ and $(X, \|\cdot\|_X)$ be a normed space. Then $W^\eta([0, t_0], X)$ denotes the space of functions $v:[0, t_0]\to X$ such that $\|v\|_{\eta, X}<\infty$, where $$\|v\|_{\eta, X}:= \sup_{0\leq t\leq t_0} \left(\|v(t)\|_X + \int_0^t \frac{\|v(t)-v(s)\|_X }{(t-s)^{\eta+1}}{\mathrm{d}}s \right)$$ is the norm in $W^\eta([0, t_0], X)$.\
We will use the [*short notations*]{} for the following norms: $$\|\cdot\|_{\alpha, \infty}:= \|\cdot\|_{ H^\alpha_\infty (\mu)}~~\mbox{and}~~ \|\cdot\|_{\alpha}:= \|\cdot\|_{ H^{\alpha} (\mu)}~~\mbox{for each}~~ \alpha\in
\mathbb{R}\, .$$ Then we recall that for $\nu\ge 0$ and $t>0$ the operators $T(t)$ and $A^\nu$ commute on $D(A^\nu)$ and satisfy the following well-known estimates (see e.g. [@Pa83]) for $u,v\in D(A^\nu)$: $$\label{eq: T(t)- extra regularity} \|T(t)v||_\nu\le c t^{-\frac{\nu}{2}}\|v\|_0\, ,$$ and for $0<\nu< 1$, $$\label{eq: T(t)- alpha continuity in L2} \|T(t)u-u\|_0\le c t^\nu\|u\|_{2\nu}\, ,$$ where $0<t\leq t_0$.\
The symmetry of the semigroup $\{T(t), t\geq 0\}$ has been used in order to extend it to elements from the dual spaces. If $w\in H^{-\beta}(\mu)$ then $T(t)w$ is the element of $L_2(\mu)$ determined by the duality relation $$(v,T(t)w):=(T(t)v,w)\, ,~~v\in L_2(\mu)\, .$$ Then we get $$|(v,T(t)w)|=|(T(t)v,w)|\le \|T(t)v\|_\beta \|w\|_{-\beta}$$ and hence, $$\|T(t)w\|_0\le c t^{-\frac{\beta}{2}}\|w\|_{-\beta}$$ in view of . Applying the latter again and using $T(t)=T(\frac{t}{2})\circ T(\frac{t}{2})$ we infer $$\label{eq: T(t) delta-beta bound}
\|T(t)w\|_\delta\le c t^{-\frac{\delta}{2}-\frac{\beta}{2}}\|w\|_{-\beta}$$ for any $\delta,\beta>0$.\
Similarly one obtains from $$\label{eq: T(t)-alpha continuity}
\|T(t)w-w\|_{-\beta-2\nu}\le c t^\nu\|w\|_{-\beta}\, ,$$ for any $\beta>0$, $w\in H^{-\beta}(\mu)$ and $0<\nu<1$. Note that the constants in the estimates depend on the related parameters. Throughout the paper we make the following standing assumptions which are the same as in [@HZ12]:
**Assumption (MMS)**
: $(X,d)$ is a locally compact separable metric space. We consider the Borel $\sigma$-field on $X$ and a Radon measure $\mu$ on $(X,d)$.
**Assumption (HKE($\beta$))**
: The transition kernel $P_t(x, dy)$ associated with the semigroup $T(t), t\geq0$ admits a transition density $P_t(x, dy) = p(t,x,y) \mu(dy)$ which satisfies for almost all $x, y \in X$ the following [*heat kernel estimate*]{} $$t^{-\frac{d_H}{w}} \Phi_1 (t^{-\frac{1}{w}}d(x,y) ) \leq p(t,x,y) \leq t^{-\frac{d_H}{w}} \Phi_2 (t^{-\frac{1}{w}}d (x,y) )$$
if $0<t<R_0$ for some constants $R_0>0$, $w\ge 2$ and nonnegative bounded decreasing functions $\Phi_i$ on $[0,\infty)$, where $d_H$ is the [*Hausdorff dimension*]{} of $(X,d)$. For $t\ge R_0$, $$p(t,x,y)\le p_t$$ and $p_t$ decreases in $t$. (In this case the semigroup is [*ultracontractive*]{}, i.e.,$$\|T(t)\|_{L_\infty(\mu)}\le p_t\|f\|_{L_2(\mu)}\, ,$$ where $p_t:=c\, t^{-d_S/4}$, if $t<R_0$, and the value $d_S=\frac{2d_H}{w}$ agrees with its [*spectral dimension*]{}. $w$ is called [*walk dimension*]{} of the semigroup.) For a given $\beta>0 $ we further assume the [*integrability condition*]{} $$\int_0^\infty s^{d_H+\beta w/2-1}\Phi_2(s) \mathrm d s<\infty\, .\footnote{In \cite{HZ12} the $w$ is missing in the exponent.}$$
Heat kernels of this type have been studied in Grigor’yan and Kumagai [@GK08] and related references therein. Further relationships are presented in the recent survey [@GHL14] of Grigor’yan, Hu and Lau. In order to make the integral in precise we need [*pointwise products*]{} of functions and dual elements from the potential spaces. In [@HZ12] the following is proved which also extends related results for the Euclidean case.
\[prop: hinz zaehle\_paraproducts\] [@HZ12 Corollary 4.1]\
Suppose (MMS) and (HKE($\beta$)) for $0<\beta <\delta<\min(\frac{d_S}{2}, 1)$. Then for $q=\frac{d_S}{\delta}$ the product $gh$ of $g\in H^\delta(\mu)$ and $h\in H^{-\beta}_q(\mu)$ is well defined in $H^{-\beta}(\mu)$ by the duality relation $(f,gh):= (fg,h)\, ,~ f\in H^\delta(\mu)$, and the following estimate holds true: $$\|gh\|_{-\beta}\leq c \|g\|_\delta \|h\|_{H^{-\beta}_q(\mu)}\, .$$
The integral equation and mild solution
---------------------------------------
A rigorous definition for the integral and a contraction principle for the solution to equation are given in [@HZ12] by means of fractional calculus in Banach spaces, in particular, under the following additional conditions.
**Assumption (FG)**
: The nonlinear functions $F$ and $G$ are such that $F\in C^1(\mathbb R^n)$, $F(0)=0$ and $F$ has bounded Lipschitz derivative $F'$ and $G\in C^2(\mathbb R^n), G(0)=0$ and $G$ has bounded Lipschitz second derivative $G^{''}$.
For the parameters we here consider the case II from [@HZ12].
**Assumption (P)**
: $0<\alpha<\gamma$, $0<\beta<\delta<\min(\frac{d_S}{2},1)$, $\gamma<1-\alpha-\frac{\beta}{2}-\frac{d_S}{4}$, where $\beta$ and $d_S$ are from (HKE($\beta$)), and $q=\frac{d_S}{\delta}$.
We now will briefly summarize the construction. If $u\in W^\gamma([0,t_0], H^{\delta}_\infty(\mu) )$ the operator $U(t;s): H^{-\beta}_q(\mu)\to H^{\delta}_\infty(\mu) $ is defined as $$\label{eq: U(t;s)}
U(t;s)w:=T(t-s) \left( G(u(s))w \right)$$ for $w\in H^{-\beta}_q(\mu)$. Then under the above assumptions on the function $G$ and the parameters (P) for any $0<\eta<\gamma$ the [*left-sided Weyl-Marchaud fractional derivative*]{} of order $\eta$ is determined by $$D^\eta_{0+} U(t;s) := \frac{{{\mathbbm{1}}}_{(0,t)}(s)}{\Gamma(1-\eta)} \left( \frac{U(t;s)}{s^\eta} + \eta \int_0^s \frac{U(t;s) -U(t;\tau)}{(s-\tau)^{\eta+1}} \mathrm d \tau \right)$$ as an element of $L_1([0,t],L(H^{-\beta}_q(\mu),H^\delta(\mu))$ (in the sense of Bochner integration). This is shown in [@HZ12 Lemma 5.2, (ii)][^1].\
Let us now consider the regulated version of $z\in C^{1-\alpha}([0,t_0],H_q^{-\beta})$ on $[0,t]$ given by $z_t(s) := {\mathbbm{1}}_{(0,t)}(s) (z(s)-z(t))$. If additionally $1-\eta<1-\alpha$, which is always possible in view of (P), one can define the [*right-sided Weyl-Marchaud fractional derivative*]{} of $z_t$ of order $1-\eta$ by $$D^{1-\eta}_{t-} z_t(s) := \frac{(-1)^{1-\eta}{\mathbbm{1}}_{(0,t)}(s)}{\Gamma(\eta)} \left( \frac{z(s)-z(t)}{(t-s)^{1-\eta}} + (1-\eta) \int_s^t \frac{z(s) -z(\tau)}{(\tau-s)^{(1-\eta)+1}} \mathrm d \tau \right)$$ as an element of $L_\infty([0,t],H^{-\beta}_q(\mu))$.\
For more details on these fractional derivative we refer the reader to [@SKM93], [@Za98], [@Za01], and [@HZ09_I] for the Banach space version. They are used for one of the results in [@HZ12]:
\[prop: hinz zaehle\] Suppose (MMS), (HKE($\beta$)), (FG), the parameter conditions (P) and $z\in C^{1-\alpha}([0,t_0], H^{-\beta}_q(\mu))$. Then we have the following.
- [@HZ12 Lemma 5.1] For the operator $U(t;x)=T(t-x)(G(u(x)\cdot)$ as in with $u\in W^\gamma([0, t_0],H_\infty^\delta(\mu))$ the integral $\int_s^t U(t;x)\mathrm d z(x) $ is well defined by $$\label{eq: def of convolution intergal}
\int_s^t U(t;x)\mathrm d z(x) := (-1)^\eta\int_s^t D^\eta_{s+} U(t;x) D^{1-\eta}_{t-} z_t(x) \mathrm d x,$$ independently of the choice of $\eta$ with $\eta<\gamma$ and $1-\eta<1-\alpha$. (In particular, the integrand on the right side is a Lebesgue integrable real function.)
- [@HZ12 Theorem 1.2] For any initial function $f\in H^{2\gamma+\delta+\varepsilon}(\mu)$ with some $\varepsilon>0$ there exists a unique solution $u$ to equation for the definition of the integral such that $u\in W^\gamma([0, t_0],H^\delta_\infty(\mu))$.
The main result
===============
Regularity of the solution {#subsc: regularity}
--------------------------
The main results are stated in Theorem \[thm: gamma holder regularity\] and Corollary \[cor: extra holder regularity\]. In this section we use [*short notations*]{} for the following norms: $$\|\cdot\|_L:= \|\cdot\|_{L(H^{-\beta}_q(\mu),H^\delta(\mu))}~~,~~\|\cdot\|_{W^\gamma}:=\|\cdot\|_{W^\gamma([0, t_0], H^\delta_\infty(\mu))}$$ with the specified parameters as in Assumption (P).\
First recall that under Assumption (FG) the nonlinear operators $F$ and $G$ are bounded from $H^\delta_\infty(\mu)$ into itself (see [@HZ12 Proposition 3.1]).
\[lm: holder regularity of F term\] Suppose (MMS), (HKE($\beta$)), (FG), the parameter conditions (P) and let $z\in C^{1-\alpha}([0,t_0], H^{-\beta}_q(\mu))$. Then there is a positive constant $c$ such that $$\label{eq: holder regularity of F term}
\left\| \int_s^t T(t-x) F(u(x))\mathrm d x \right\|_{ \delta} \leq c||u||_{W^\gamma}(t-s).$$
Since the semigroup $T(t)$ is a contraction on $H^\delta(\mu)$ we have $$\begin{aligned}
\left\| \int_s^t T(t-x) F(u(x))\mathrm d x \right\|_{ \delta}
\leq& \int_s^t \| T(t-x) F(u(x)) \|_{ \delta} \mathrm d x \\
\leq& \int_s^t \|F(u(x)) \|_{\delta} \mathrm d x\\
\leq& c \int_s^t \|u(x) \|_{\delta} \mathrm d x \leq c \|u\|_{W^\gamma} (t-s),\end{aligned}$$ where the latter bound follows from the definition of the $W^\gamma$-norm.
\[lm: epsilon argument\] Under the same conditions as in Lemma \[lm: holder regularity of F term\] we have $$\begin{aligned}
\| T(t-x)& \left( G(u(x)) \right) - T(t-y) \left( G(u(x)) \right) \|_L \\
&\leq c ||u||_{W^\gamma} (t-x)^{-\frac\delta2-\frac\beta2-\nu} (x-y)^\nu
\end{aligned}$$ for any $0<\nu<1$.
Let $h$ be an arbitrary element of $H^{-\beta}_q(\mu)$ with $\|h\|_{H^{-\beta}_q(\mu)}\le 1$. Then Proposition \[prop: hinz zaehle\_paraproducts\] implies $$\|G(u(x))h\|_{-\beta} \le c \|u\|_{W^\gamma}$$ uniformly in the time argument $x\in[0,t_0]$ by the mapping property of $G$ and the definition of the $W^\gamma$-norm. Using , and the last estimate we infer $$\begin{aligned}
\| T(t-x) & (\mathrm{Id} - T(x-y) ) G(u(x))h \|_{\delta}\\
&\le c (t-x)^{-\delta/2-\beta/2-\nu} \|(\mathrm{Id} - T(x-y) ) G(u(x))h \|_{-\beta-2\nu}\\
&\le c (t-x)^{-\delta/2-\beta/2-\nu} (x-y)^\nu\|G(u(x))h\|_{-\beta}\\
&\le c (t-x)^{-\delta/2-\beta/2-\nu} (x-y)^\nu\|u\|_{W^\gamma}\, .\end{aligned}$$
This bound together with the definition of the $L$-norm $$\begin{aligned}
\| T(t-x) & (\mathrm{Id} - T(x-y) ) G(u(x)) \|_{L}\\
&= \sup_{\|h\|_{H^{-\beta}_q}\leq 1} \| T(t-x) (\mathrm{Id} - T(x-y) ) \left( G(u(x)) h\right) \|_{\delta}
\end{aligned}$$ completes the proof.
\[lm: holder regularity of G term\] Suppose (MMS), (HKE($\beta$)), (FG), the parameter conditions (P) and let $z\in C^{1-\alpha}([0,t_0], H^{-\beta}_q(\mu))$. Then there is a positive constant $c$ such that $$\label{eq: holder regularity of G term}
\left\| \int_s^t T(t-x) G(u(x)) \mathrm d z(x) \right\|_{ \delta} \leq c (t-s)^\gamma.$$
Since by assumption $z\in C^{1-\alpha}([0,t_0], H^{-\beta}_q(\mu))$ then for any $\eta$ such that $1-\eta < 1-\alpha$ we have $$\sup_{t\in[0,t_0]}\sup_{x\in[0,t]} \left\| D^{1-\eta}_{t-} z_t(x) \right\|_{H^{-\beta}_q(\mu)} \leq c<\infty\, .$$ Let us fix $\eta$ throughout the proof as some number slightly bigger than $\alpha$ such that $\alpha<\eta<\gamma$ and at the same time $\gamma<1-\eta-\frac{\delta}{2} - \frac{\beta}{2}$ which is always possible in view of (P). We then get $$\begin{aligned}
&\left\| \int_s^t U(t;x)\mathrm d z(x) \right\|_{\delta} \\
=&\left\| \int_s^t D^\eta_{s+} U(t;x) D^{1-\eta}_{t-} z_t(x) \mathrm d x \right\|_{\delta} \\
\leq& \sup_{t\in[0,t_0]}\sup_{x\in[0,t]} \left\| D^{1-\eta}_{t-} z_t(x) \right\|_{H^{-\beta}_q(\mu)} \int_s^t \left\| D^\eta_{s+} U(t;x) \right\|_{L} \mathrm d x \\
\leq& c \int_s^t \left\| D^\eta_{s+} T(t-x) G(u(x)) \right\|_{L} \mathrm d x \\
\le& c \int_s^t \frac{\| T(t-x) G(u(x)) \|_L}{(x-s)^{\eta} }\mathrm d x \\
&+ c \int_s^t \int_s^x \frac{\| T(t-x) G(u(x)) - T(t-y) G(u(y)) \|_L}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x \\
=&: S_1 + S_2.\end{aligned}$$ Consider $S_1$ first. Using and Proposition \[prop: hinz zaehle\_paraproducts\] we obtain $$\begin{aligned}
S_1 \leq &c\int_s^t \sup_{\|w\|_{H^{-\beta}_q(\mu)}\leq 1}\frac{\| T(t-x)\left( G(u(x))w\right) \|_{\delta} }{(x-s)^{\eta}}\mathrm d x\\
\leq & c \int_s^t (t-x)^{-\frac\delta2 - \frac\beta2} \sup_{\|w\|_{H^{-\beta}_q(\mu)} \leq 1}\frac{\| G(u(x)) w \|_{-\beta}}{(x-s)^{\eta}}\mathrm d x\\
\leq & c \int_s^t (t-x)^{-\frac\delta2 - \frac\beta2} (x-s)^{-\eta} \| G(u(x)) \|_{\delta} \sup_{\|w\|_{H^{-\beta}_q(\mu)}\leq 1} \|w\|_{H^{-\beta}_q(\mu)} \mathrm d x\\
\leq & c \int_s^t (t-x)^{-\frac\delta2 - \frac\beta2} (x-s)^{-\eta} \|u\|_{W^\gamma} \mathrm d x\\
\leq & c (t-s)^{1-\frac\delta2 - \frac\beta2-\eta} \leq c(t-s)^\gamma,\end{aligned}$$ the latter following from $1-\frac\delta2 - \frac\beta2-\eta>\gamma$ by construction. Moreover the integral is finite since $\gamma>0$.\
Consider $S_2$. The numerator inside the integral can be bounded as follows $$\begin{aligned}
\label{eq: difference of T}
\| T(t-x) & G(u(x)) - T(t-y) G(u(y)) \|_L \\ \nonumber
\leq& \| T(t-x) G(u(x)) - T(t-y) G(u(x)) \|_L \\ \nonumber
+&\| T(t-y) G(u(x)) - T(t-y) G(u(y)) \|_L,\end{aligned}$$ so that we have $$\begin{aligned}
S_2=& c \int_s^t \int_s^x \frac{\| T(t-x) G(u(x)) - T(t-y) G(u(y)) \|_L}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x \\
\leq & c \int_s^t \int_s^x \frac{\| T(t-x) G(u(x)) - T(t-y) G(u(x)) \|_L}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x \\
+& c \int_s^t \int_s^x \frac{\| T(t-y) G(u(x)) - T(t-y) G(u(y)) \|_L}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x\\
=: & S_3 + S_4.\end{aligned}$$ Let us consider the term $S_4$ first. We have (with similar computations as for $S_1$) $$\begin{aligned}
S_4=& c \int_s^t \int_s^x \frac{\| T(t-y)G(u(x)) - T(t-y) G(u(y)) \|_L}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x\\
\leq & c \int_s^t \int_s^x \sup_{\|w\|_{H^{-\beta}_q}\leq1} \frac{\| T(t-y) \left( \left[ G(u(x)) -G(u(y))\right]w\right) \|_{\delta}}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x\\
\leq & c \int_s^t(t-x)^{-\frac\delta2 - \frac\beta2} \int_s^x \frac{\| G(u(x)) - G(u(y)) \|_{\delta}}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x\\
\intertext{recall that $\eta<\gamma$ by definition of $\eta$, thus}
\leq & c \int_s^t(t-x)^{-\frac\delta2 - \frac\beta2} \| u \|_{W^\gamma} \mathrm d x\\
\leq & c (t-s)^{1-\frac\delta2 - \frac\beta2} \leq c (t-s)^\gamma,\end{aligned}$$ the latter being true as $\gamma<1-\frac\delta2 - \frac\beta2$ by assumption. Regarding the term $S_3$, we apply Lemma \[lm: epsilon argument\] with $\nu>\eta$ to the numerator inside the integral of $S_3$, so that $S_3$ can be bounded by $$\begin{aligned}
S_3 & \leq c \int_s^t \int_s^x \frac{\| T(t-x) G(u(x)) - T(t-y) G(u(x)) \|_L}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x \\
& \leq c \int_s^t \int_s^x \frac{(t-x)^{-\frac\delta2 - \frac\beta2-\nu} (x-y)^\nu}{(x-y)^{1+\eta} } \mathrm d y \mathrm d x \\
& \leq c \int_s^t (t-x)^{-\frac\delta2 - \frac\beta2-\nu} \int_s^x (x-y)^{\nu -1-\eta} \mathrm d y \mathrm d x \\
& \leq c \int_s^t (t-x)^{-\frac\delta2 - \frac\beta2-\nu} (x-s)^{\nu-\eta} \mathrm d x \\
& \leq c (t-s)^{1-\frac\delta2 - \frac\beta2-\nu+\nu - \eta} \leq c (t-s)^\gamma,\end{aligned}$$ the latter bound being true as $\gamma<1-\eta-\frac\delta2 - \frac\beta2$. The proof is complete.
We are now about to state and prove the main regularity property of the solution $u$ under the Assumptions (MMS), (HKE($\beta$)), (FG) and (P).
\[thm: gamma holder regularity\] Suppose (MMS), (HKE($\beta$)) and (FG). Let $0<\alpha<\gamma$ , $0<\beta < \delta<\min(\frac{d_S}{2},1)$ and $\gamma<1-\alpha-\frac{\beta}{2} - \frac{d_S}{4}$ and set $q=\frac{d_S}{\delta}$. If $z\in C^{1-\alpha}([0,t_0], H^{-\beta}_q(\mu))$ and the initial condition $f$ is an element of $H^{\delta+2\gamma+\varepsilon}(\mu)$ for some $\varepsilon>0$, then the unique solution $u \in W^\gamma([0, t_0], H^\delta_\infty(\mu))$ for is also an element of $C^{\gamma}([0, t_0], H^{\delta}(\mu))$.
For $u \in W^\gamma([0, t_0], H^\delta_\infty(\mu))$ the second integral in equation has to be interpreted as above. For $u$ considered as element of $C^{\gamma}([0, t_0], H^{\delta}(\mu))$ this integral agrees with the corresponding Riemann-Stieltjes integral with values in the Banach space $H^{\delta}(\mu)$. The latter has been used in Gubinelly, Lejay and Tindel [@GLT06] in an abstract setting in the sense of Young.
Let $0\leq s<t\leq t_0$. We consider the solution at time $t$ as the evolution of $u$ according to with initial condition at time $s$ being $u(s)$, that is $$u(t)= T(t-s)u(s) + \int_s^t T(t-r)F(u(r)) {\mathrm{d}}r + \int_s^t T(t-r)G(u(r)) {\mathrm{d}}z( r),$$ so that $$\begin{aligned}
\label{eq: difference ut-us}
u(t)-u(s) =&( T(t-s)- {\operatorname{Id}}) u(s) \nonumber\\
+& \int_s^t T(t-r)F(u(r)) {\mathrm{d}}r + \int_s^t T(t-r)G(u(r)) {\mathrm{d}}z( r)\end{aligned}$$ The $H^\delta(\mu)$-norms of the two integrals are bounded by $ c (t-s)$ and $ c (t-s)^\gamma$ according to Lemma \[lm: holder regularity of F term\] and Lemma \[lm: holder regularity of G term\], respectively. If we show that $\|u(t)\|_{\delta+2\gamma}<c$ uniformly in $t$ for $t\in[0, t_0]$, then the $H^\delta(\mu)$-norm of the term involving the initial condition can be easily bounded. In fact note that in this case $(A+{\operatorname{Id}})^{\frac\delta2}u\in D(A^{\gamma}) $ thus we can apply to get $$\begin{aligned}
\|(T(t-s)- {\operatorname{Id}})u(s)\|_{\delta} \leq & c \|(A+{\operatorname{Id}})^{\delta/2}(T(t-s)- {\operatorname{Id}})u(s)\|_{0}\\
\leq & c \|(T(t-s)- {\operatorname{Id}})(A+{\operatorname{Id}})^{\delta/2 }u(s)\|_{0}\\
\leq & c (t-s)^\gamma \|(A+{\operatorname{Id}})^{\delta/2} u(s)\|_{2\gamma} \\
\leq & c (t-s)^\gamma \|u(s)\|_{\delta+2\gamma} \leq c (t-s)^\gamma\end{aligned}$$ as wanted. It remains to prove that $\|u(t)\|_{\delta+2\gamma}<c$ uniformly in $t$ for $t\in[0, t_0]$. Recall that
$$u(t) = T(t)f + \int_0^t T(t-r)F(u(r)) {\mathrm{d}}r + \int_0^t T(t-r)G(u(r)) {\mathrm{d}}z( r) ,$$ thus $$\begin{aligned}
\|u(t)\|_{\delta+2\gamma} &\leq \|T(t)f\|_{\delta+2\gamma} \\
&+ \int_0^t\| T(t-r)F(u(r))\|_{\delta+2\gamma} {\mathrm{d}}r \\
&+ \int_0^t\| D^\eta_{0+} U(t;r) D^{1-\eta}_{t-} z_t(r) \|_{\delta+2\gamma} {\mathrm{d}}r\\
=: S_1 + S_2 + S_3.
\end{aligned}$$ The term $S_1$ is easily bounded by $c\|f\|_{\delta+2\gamma}$. The term $S_2$ is bounded recalling that $\|T(t-r) F(u(r))\|_{\delta+2\gamma} \leq (t-r)^{-\gamma} \|F(u(r))\|_\delta$ because of the smoothing action of the semigroup. Thus $S_2\leq c t^{1-\gamma}$. The last term can be treated in a similar way as the proof of Lemma \[lm: holder regularity of G term\] with the difference that the space $H^{\delta}(\mu)$ is replaced by $H^{\delta+2\gamma}(\mu)$. All computations for the $H^{\delta+2\gamma}(\mu)$-norm term carry out in the same way, except that the exponent $-\frac\delta2-\frac\beta2$ is replaced by $-\frac\delta2-\frac\beta2 -\gamma$ so that $S_3\leq c t^{1-\eta-\frac\delta2-\frac\beta2-\gamma}$ with $1-\eta-\frac\delta2-\frac\beta2-\gamma>0$ by construction of $\eta$. Clipping the result together we have $$\|u(t)\|_{\delta+2\gamma} \leq c+ ct^{1-\gamma} + c t^{1-\eta-\frac\delta2-\frac\beta2-\gamma},$$ and finally taking the supremum over $t\in[0,t_0]$ we get the uniform bound.
With slightly more restrictive assumptions on the noise we can show that the unique solution $u$ belongs to the spaces $W^\gamma([0,t_0], H^\delta_\infty(\mu))$, and thus to $C^\gamma([0,t_0], H^\delta(\mu))$, for all $(\gamma,\delta)$ such that $0<\gamma<1-\alpha-\frac{\beta}{2}-\frac{d_S}{4}$ and $\beta<\delta<\min(\frac{d_S}{2},1)$.
\[cor: extra holder regularity\] Suppose (MMS), (HKE($\beta$)) and (FG).
- Let $0<\alpha<\frac{1}{2}$ and $0<\beta<\min(\frac{d_S}{2},1-2\alpha, 2(1-\alpha)-\frac{d_S}{2})$ be given. Suppose that $z\in C^{1-\alpha}([0,t_0],H_q^{-\beta}(\mu))$ for any $1<q<\frac{d_S}{\beta}$ and $f\in H^{2(1-\alpha)-\beta}(\mu)$. Then for any $\beta<\delta<\min(\frac{d_S}{2},1)$ and $0<\gamma<1-\alpha-\frac{\beta}{2}-\frac{d_S}{4}$ Equation has a unique solution in the space $W^\gamma([0,t_0], H^\delta_\infty(\mu))$ and hence, it has a unique solution belonging to all these spaces.
- Moreover, this solution is an element of $C^\gamma([0,t_0], H^\delta(\mu))$ for any $\gamma$ and $\delta$ as before.
Part (a). Take $\delta$ and $\gamma$ as in the assumption. Then $2(1-\alpha) -\beta > 2\gamma + \frac{d_S}{2}>2\gamma + \delta + \varepsilon$ for some $\varepsilon>0$ implies that $f\in H^{2\gamma + \delta+\varepsilon}(\mu)$. Moreover $H^{-\beta}_{q}(\mu) \subset H^{-\beta}_{\frac{d_S}{\delta}}(\mu)$ for $1<q<\frac{d_S}{\beta}$ since $\beta<\delta<d_S$, thus $z\in C^{1-\alpha}([0,t_0], H^{-\beta}_{\frac{d_S}{\delta}}(\mu))$. Then we are under the assumptions of Theorem \[thm: gamma holder regularity\] and thus there exists a unique solution to which belongs to $ W^\gamma([0,t_0], H^\delta_\infty(\mu))$. Because of the embedding of the spaces involved, clearly $u\in W^{\gamma'}([0,t_0], H^{\delta'}_\infty(\mu))$ for any $0<\delta'\le\delta$ and $0<\gamma'\le\gamma$, too. We also know that for $\delta'$ and $\gamma'$ satisfying the assumptions there exists a unique solution $u'$ to which is in $W^{\gamma'}([0,t_0], H^{\delta'}_\infty(
\mu))$. As the initial condition $f$ and the noise term $z$ are the same, then by uniqueness we must have $u=u'$ in the larger space $ W^{\gamma'}([0,t_0], H^{\delta'}_\infty(\mu))$.
Part (b). It follows directly form part (a) and Theorem \[thm: gamma holder regularity\]. Obviously, due to the embedding of the fractional Sobolev spaces $H^{\delta}(\mu)\subset H^{\delta'}(\mu)$ for $\delta'<\delta$ and of the Hölder spaces $C^{\gamma}\subset C^{\gamma'}$ for $\gamma'<\gamma$, we have that $u\in C^{\gamma'}([0, t_0], H^{\delta'}(\mu))$ for all $0<\delta'<\delta$ and $0<\gamma'<\gamma$.
\[rem: low spectral dim\] The parameter condition (P), in particular $\delta<d_S/2$, and the integrability condition on the function $\Phi_2$ in the heat kernel estimate (HKE) have been used only for the product estimate in Proposition \[prop: hinz zaehle\_paraproducts\]. An analysis of the proofs, in particular those of [@HZ12], shows that the assertions of Theorem \[thm: gamma holder regularity\] and Corollary \[cor: extra holder regularity\] remain valid under the parameter condition $$\label{eq: parameter low spectral dim}
0<\frac{d_S}{2}\le \beta\le\delta<1\, ,~~0<\alpha<\gamma<1-\alpha-\frac{\beta}{2}-\frac{\delta}{2}$$ provided $z\in C^{1-\alpha}([0,t_0], H^{-\beta}_q(\mu))$ for some $q>2$ such that the multiplication property $$\label{eq: product assumption}
\|vz\|_{-\beta}\le c\|v\|_\delta\, \|z\|_{H_q^{-\beta}(\mu)}$$ holds true for any $v\in H^\delta(\mu)$.\
Then one obtains a complement to the former assertions for the case $d_S\le 1$: Here we can choose $\frac{1}{2}\le \beta<1-2\alpha$, which implies $1-\alpha>\frac{3}{4}$, in order to get a Hölder continuous solution to equation . In particular, in the Gaussian setting (see below) such a $\dot{z}$ may be interpreted as a noise white in space and coloured in time.
Applications and extensions
---------------------------
With similar techniques as in the previous section it is possible to treat equation with $A$ replaced by a fractional power of $A$, that is for instance $A^\theta$ for $0<\theta\le1$. This is done in [@HZ12]. In this case the semigroup associated with the fractional power $-A^\theta$ will be the subordinated semigroup $T^{(\theta)}(t)$ and the power $\theta$ must be taken into account in Assumption (P) and all following theorems accordingly. Then similar regularity properties as shown for the case $\theta=1$ follow. These results go in the direction of [@RZZ14] where the authors study a stochastic equation with fractional dissipation (that is with a term $A^\theta, 0<\theta\leq 1$) but with the difference that in the present paper the noise is coloured, whereas in [@RZZ14] the noise is white in time and solutions are strong and local.\
Clearly, the special case of linear $F$ and $G$ can be considered. In this case the $L_\infty$-boundedness of the solution is no longer needed (see [@HZ12 Theorem 1.3]) and the conditions on the parameters are weaker, in particular in Assumption (P) it is sufficient that $\gamma<1-\alpha -\frac\beta2-\frac\delta2$. Moreover the spectral dimension restriction $d_S<4$ can be lifted. Using the aforementioned corresponding results from [@HZ12] and the present methods one can get similar results as in the previous section, the proofs being completely analogous.\
Moreover, one can easily consider linear combinations of noise terms such as $$\sum_{i=1}^N G_i(u)\cdot \dot z_i$$ for any finite integer $N$ and $G_i$ and $z_i$ as $G$ and $z$ in this paper.
Here we list a few examples. The results of Section \[subsc: regularity\] can be applied to the same kind of stochastic equations considered in the paper by Hinz and Zähle [@HZ12 Section 7]. In particular, we can consider stochastic partial differential equations driven by fractional Brownian noises on metric measure spaces. The equations are studied in the pathwise sense and the results are valid $\mathbb P$-a.s.
A classical example is the nonlinear heat equation on a smooth bounded domain $D\subset\mathbb R^n$, $n\le 3$, provided with the Lebesgue measure $\mu$, driven by fractional Brownian field, see e.g. [@HZ09_II Section 6]. We consider a real valued fractional Brownian sheet $\{B^{H, K}(t,x), [0,t_0]\times \mathbb R^n\}$ with Hurst indices $0<H<1$ and $0<K<1$ for time and space respectively. This is a centered Gaussian field on $[0,t_0]\times \mathbb R^n$ with stationary rectangular increments satisfying $$E\big( B^{H,K}(s,x)-B^{H,K}(s,y)-B^{H,K}(t,x)+B^{H,K}(t,y)\big)^2= c |s-t|^{2H} |x-y|^{2K}.$$ It can be shown [@HZ09_II Section 6] that there exits a version such that for almost all trajectories $\omega\in\Omega$ one has $B^{H,K}(\omega)\in C^{1-\alpha}([0,t_0], H^\sigma_q(\mu))$ for $0<1-\alpha<H, \, 0<\sigma<K$ and $1<q<\infty$. Thus the distributional spatial partial derivatives $\frac{\partial}{\partial x_i}B^{H,K}$ for $i=1, \ldots,n$ belong to $C^{1-\alpha}([0,t_0], H^{-\beta}_q(\mu))$, where $-\beta=\sigma-1$. Let $F:\mathbb R^n \to \mathbb R$ and $G:\mathbb R^n \to \mathbb R^n$ such that $F$ and each component $G_i$ satisfy Assumption (FG). Then Theorem \[thm: gamma holder regularity\] and Corollary \[cor: extra holder regularity\] can be applied in the pathwise sense to $$\label{eq: example in Rn}
\left\{
\begin{array}{l}
\frac{\partial u}{\partial t} = \Delta_D u +F(u) +\left\langle G(u), \frac\partial{\partial t}\nabla{B}^{H,K}\right\rangle\\
u(0,x)=0, \textrm{ for } t\in(0,t_0) \\
u(t,x)=0, \textrm{ for } x\in \partial D
\end{array}
\right.$$ for almost all paths. Here $\Delta_D$ is the classical Dirichlet Laplacian, i.e., $-\Delta_D$ generates a heat semigroup on $L_2(D)$ with Gaussian estimates. $\nabla B^{H,K}$ denotes the distributional gradient of the fractional Brownian sheet. The term $\left\langle G(u), \frac\partial{\partial t}\nabla{B}^{H,K}\right\rangle$ in is given by $$\sum_{i=1}^n G_i(u) \frac{\partial^2}{ \partial t \partial x_i} B^{H,K}\, ,$$ where $\frac{\partial^2}{ \partial t \partial x_i} B^{H,K}$ is interpreted as $\dot{z}_i$ in the above sense.\
Since the spectral dimension of $\mathbb R^n$ is $d_S=n$, for almost all sample paths the unique solution is an element of $$W^\gamma([0,t_0], H^\delta_\infty(\mu))\cap C^\gamma([0,t_0], H^\delta(\mu))$$ for all $\gamma$ and $\delta$ such that $1-H<\gamma<H-\frac{1-K}{2}-\frac{n}{4}$ and $1-K<\delta<\min (\frac{n}{2}, 1)$. This can be satisfied only if $n\le3$.\
According to Remark \[rem: low spectral dim\] for $d_S=n=1$ the result remains valid under the parameter condition , since the multiplication property is fulfilled in this Euclidean case. The latter can be seen as follows. Theorem 4.5.2 in Runst and Sickel [@RS96] and restriction to functions vanishing on $\partial D$ lead to the estimate $$\|uv\|_{H^\beta_p(\mu)}\le c\|u\|_\delta\|v\|_\beta\ ,~~~~
1<p<2\, ,~~~~ \frac{1}{2}\le\beta<\delta<1\ .$$ Then for these parameters follows by duality arguments. Hence, for $n=1$ and $K\le\frac{1}{2}$ Hölder continuous solutions to equation can be obtained if $H>\frac{2-K}{2}$. Note that $K=\frac{1}{2}$ means white noise in space.
A more sophisticated case mentioned in [@HZ12 Example 2]) is the following. Let $(X,d,\mu)$ be a compact metric measure space satisfying Assumption (MMS) and admitting a semigroup $\{T(t), t\geq 0\}$ generated by a (fractal) Neumann Laplacian $\Delta$ associated to a local regular Dirichlet form $(\mathcal E, D(\mathcal E))$ on $X$, i.e., $-A=\Delta$. For various classes of fractals the corresponding heat kernels exist and satisfy Assumption (HKE($\beta$)) for any $\beta>0$ (see, e.g., Barlow and Bass [@BB92] and [@BB99], Barlow, Bass, Kumagai and Teplyaev [@BBKT10], Fitzsimmons, Hambly and Kumagai [@FHK94], Hambly and Kumagai [@HK99], Kigami [@Ki12], Barlow, Grigor’yan and Kumagai [@BGK12] and the references therein).\
A standard example for the noise process $z$, modified for our situation, is the following: Let $e_0,e_1,e_2,\dots$ be a complete orthonormal system of eigenfunctions of $A$ in $L_2(\mu)$ and $\lambda_i$ be the corresponding eigenvalues, $B_1(t)^H,B_2(t)^H,\dots$ are i.i.d. fractional Brownian motions in $\mathbb{R}$ with Hurst exponent $\frac{1}{2}<H<1$, and consider the formal series $$b^H(t)=\sum_{i=1}^\infty B_i^H(t)\, q_i\, e_i$$
for real coefficients $q_i$. Then we get a modification $b^H$ such that a.s. $$z:=b^H\in C^{1-\alpha}\left([0,t_0], H_q^{-\beta}(\mu)\right)$$ (with convergence of the series in these spaces) for any $0<1-\alpha<H$, $\beta_*<\beta<1$ and $q>2$ under the following conditions on the measure $\mu$, the parameter $\beta_*$, the eigenfunctions $e_i$ and the coefficients $q_i$ for $i\geq 1$:\
(a) $||e_i||_\infty\le c_1 \lambda_i^{a_1/2}$ and\
(b) $|e_i(x)-e_i(y)|\le c_2 \lambda_i^{a_2/2}d(x,y)^b$ (up to an exceptional set)\
for some positive constants $a_1, a_2, b, c_0, c_1, c_2$, and for $a:=\max(a_1,a_2)$, $$\sum_{i=1}^\infty q_i^2\, \lambda_i^{-\beta_*+a-b2/w} <\infty \, .$$ (Note that in the case $q=2$, which is not relevant for our purposes, Conditions (a) and (b) are not needed and the convergence $\sum_{i=1}^\infty q_i^2\, \lambda_i^{-\beta_*} <\infty$ would be sufficient for the above property of $z$.) ([*Idea of proof:*]{} By the mapping properties of the resolvent operators $J^\sigma(\mu)$, which may be replaced here by $I^\sigma(\mu):=(-\Delta)^{\sigma/2}$ since the included eigenvalues are strictly positive, it is equivalent to get a modification, which satisfies a.s. $$I^{\beta_*+b2/w}(\mu)b^H\in C^{1-\alpha}([0,t_0], H_q^{-\beta+\beta_*+b2/w}(\mu))$$ for any $q$ and $\beta>\beta_*$. For $0<\delta'<\delta$ the embedding of the Hölder space $C^\delta(X)$ into $H_q^{\delta'2/w}(\mu)$ can be seen, e.g., from the arguments in the proof of [@HZ09 Proposition 5.6] (using there only the upper estimates taking into regard that the lower heat kernel estimates imply $\mu(B(x,r))\le c_0 r^{d_H}$ for any ball with centre $x$ and radius $r$). Therefore a sufficient condition for the above one is that $$I^{\beta_*+b2/w}(\mu)b^H\in C^{1-\alpha}([0,t_0], C^\delta(X))$$ for any $\delta<b$, and the latter can be proved by means of the Kolmogorov principle for the random function $$Y(t,x):=I^{\beta_*+b2/w}b^H(t,x)=\sum_{i=1}^\infty B_i^H(t)\, q_i\,\lambda_i^{-(\beta_*+b2/w)/2}\, e_i(x)\, .$$ We get $$\begin{aligned}
& &\mathbb{E}\big(Y(t_1,x_1)-Y(t_2,x_1)-(Y(t_1,x_2)-Y(t_2,x_2)\big)^2\\
&=& \sum_{i=1}^\infty q_i^2\lambda_i^{-(\beta_*+b2/w)}\mathbb{E}(B^H_i(t_1)-B^H_i(t_2))^2(e_i(x_1)-e_i(x_2))^2\\
&\le& \sum_{i=1}^\infty q_i^2\lambda_i^{-(\beta_*+b2/w)}|t_1-t_2|^{2H}c_2\lambda_i^a\,d(x_1,x_2)^{2b}\\
&\le& c\,|t_1-t_2|^{2H}d(x_1,x_2)^{2b}\, .\end{aligned}$$ Moreover, $$\begin{aligned}
& &\mathbb{E}(Y(t_1,x)-Y(t_2,x))^2\\
&=& \sum_{i=1}^\infty q_i^2\lambda_i^{-(\beta_*+b2/w)}\mathbb{E}(B^H_i(t_1)-B^H_i(t_2))^2|e_i(x)|^2\\
&\le& \sum_{i=1}^\infty q_i^2\lambda_i^{-(\beta_*+b2/w)}|t_1-t_2|^{2H}c_1^2\lambda_i^a\\
&\le& c\,|t_1-t_2|^{2H}\, .\end{aligned}$$ Using that the higher moments of centered Gaussian random variables are powers of the second moments this ensures the usual construction of a modification of $Y$ with the desired Hölder regularity by means of an extension of the values on a countable dense subset of $[0,t_0]\times X$.) Note that because of the above ultracontractivity of the semigroup Condition (a) is always fulfilled for $a_1:=\frac{d_S}{2}$. Furthermore, if we work with the resistance metric $R(x,y)$ w.r.t. the Dirichlet form $\mathcal E$ then Condition (b) is satisfied for $a_2=b=1$.\
For p.c.f. fractals with regular harmonic structures we have $d_S=\frac{2d_H}{d_H+1}<2$, see Kigami [@Ki01]. Moreover, under some mild additional assumptions on such fractals in Euclidean spaces the resistance metric $R$ satisfies $R(x,y)\le|x-y|^b$ for some $b> 0$, see Hu and Wang [@HW06]. Hence, in this case Condition (b) is also fulfilled for the Euclidean metric.\
Examples with spectral dimension greater than 2 are provided by generalized Sierpinski carpets, see Barlow and Bass [@BB99], or by certain products of fractals, see Strichartz [@St05]. According to Theorem \[thm: gamma holder regularity\] function solutions to Equation which are Hölder regular in time can be found for $d_S= \frac{2d_H }{w}<4$ (recall that $d_H$ denotes the Hausdorff dimension of $X$ and $w$ the walk dimension of the semigroup) and Hurst exponent $H>\frac{1}{2}+\frac{d_S}{8}$. Recall that in this case we have $\beta<\frac{d_S}{2}$.\
If $\frac{d_S}{2}\le 1$ then Remark \[rem: low spectral dim\] provides the alternative parameter condition $\frac{1}{2}\le\beta< 2H-1$ for existence of Hölder continuous solutions. In particular, for $\beta=\frac{1}{2}$ the noise is “white” in space.\
For the classical case of the Dirichlet Laplace operator on the unit interval this example has been treated in Gubinelly, Lejay and Tindel [@GLT06] with different methods.\
[20]{} Barlow M.T. *Diffusions on fractals*, LNM 1690, Springer, New York (1998) Barlow M.T., Bass R.F. *Transition densities for Brownian motion on the Sierpinski carpet*, Probab. Theor. Rel. Fields 91, 307–330 (1992) Barlow M.T., Bass R.F. *Brownian motion and harmonic analysis on Sierpinski carpets*, Canad. J. Math. 51, 673–744 (1999) Barlow M.T., Bass R.F., Kumagai T., Teplyaev A. *Uniqueness of Brownian motion on Sierpinski carpets*, J. Eur. Math. Soc. 12, 655–701 (2010) Barlow M.T., Grigor’yan A., Kumagai T. *On the equivalence of parabolic Harnack inequalities and heat kernel estimates*, J. Math. Soc. Japan 64, 1091–1146 (2012) Da Prato G., Zabzcyk J. *Stochastic equations in infinite dimensions*, Cambridge Univ. Press, Cambridge (1992) Falconer K.J. *Semilinear PDEs on self-similar fractals*, Commun. Math. Phys. 206, 235–245 (1999) Falconer K., Hu J. *Nonlinear diffusion equations on unbounded fractal domains*, J. Math. Anal. Appl. 256, no. 2, 606-–624 (2001) Falconer K., Hu J., Sun Y. *Inhomogeneous parabolic equations on unbounded metric measure spaces*, Proc. Roy. Soc. Edinburgh Sect. A 142, no. 5, 1003-–1025 (2012) Fitzsimmons P.J., Hambly B.M., Kumagai T. *Transition density esimates for Brownian motion on affine nested fractals*, Commun. Math. Phys. 165, 595–620 (1994) Foondun M., Khoshnevisan D., Nualart E. *A local-time correspondence for stochastic partial differential equations*, Trans. Amer. Math. Soc. 363, 2481–2515 (2011) Grecksch W., Anh V.V. *Parabolic stochastic differential equation with fractional Brownian motion input*, Stat. Probab. 41, 337-–346 (1999) Grigor’yan A., Hu J., Lau K.S. *Heat kernels on metric-measure spaces and an application to semi-linear elliptic equations*, Trans. Amer. Math. Soc. 355, 2065–2095 (2003) Grigor’yan A., Hu J., Lau K.-S., *Heat kernels on metric measure spaces*, in “Geometry and Analysis on Fractals”, Springer Proceedings in Mathematics and Statistics, 88, 147-208 (2014) Grigor’yan A., Kumagai T., *On the dichotomy in the heat kernel two sided estimates*, Proceedings of Symposia in Pure Mathematics 77, 199-210 (2008) Gubinelli M., Lejay A., Tindel S. *Young integrals and [SPDE]{}s*, Potential Anal. 25 (4), 307–326 (2006) Hairer M., Maas J., Weber, H., *Approximating rough stochastic PDEs* Comm. Pure Appl. Math. 67, no. 5, 776–870 (2014) Hairer M., Voss, J., *Approximations to the stochastic Burgers equation* J. Nonlinear Sci. 21, no. 6, 897-920 (2011) Hairer M., Weber, H., *Rough Burgers-like equations with multiplicative noise* Probab. Theory Related Fields 155, no. 1-2, 71-126 (2013) Hambly B.M., Kumagai T. *Transition density estimates for diffusion processes on post critically finite self-similar fractals*, Proc. London Math. Soc. (3)78, 431–-458, (1999) Hinz M., Issoglio E., Zähle M. *Elementary pathwise methods for nonlinear parabolic and transport type SPDE with fractal noise* Modern Stochastics and Applications, Springer Optimization and Its Applications 90, 123–141 (2014) Hinz M., Zähle M. *Gradient type noises I - Partial and hybrid integrals*, Compl. Var. Ell. Equations 54, 561–583 (2009) Hinz M., Zähle M. *Gradient type noises II - Systems of stochastic partial differential equations*, J. Funct. Anal. 256, 3192–3235 (2009) Hinz M., Zähle M. *Semigroups, potential spaces and applications to (S)PDE* Potential Anal. 36, 483–515 (2012) Hu J., Wang X. *Domains of Dirichlet forms and effective resistance estimates on p.c.f. fractals*, Studia Math. 177, 153–172 (2006) Hu J., Zähle M. *Generalized Bessel and Riesz potentials on metric measure spaces*, Potential. Anal. 30, 315–340 (2009) Kigami J. *Analysis on fractals*, Cambridge Tracts in Mathematics 143, Cambridge University Press (2001) Kigami J. *Resistance forms, quasisymmetric maps and heat kernel estimates*, Memoirs Amer. Math. Soc. 216, No. 1015 (2012) Maslowski B., Nualart D. *Evolution equations driven by fractional Brownian motion*, J. Funct. Anal. 202, 277–305 (2003) Pazy A. *Semigroups of Linear Operators and Applications to Partial Differential Equations*, Springer, New York (1983) Prévôt C., Röckner M. *A Consise Course on Stochastic Partial Differential Equations*, LNM 1905, Springer, New York (2007) Röckner M., Zhu R., Zhu X. *Local existence and non-explosion of solutions for stochastic fractional partial differential equations driven by multiplicative noise* Stochastic Process. Appl. 124, no. 5, 1974–2002 (2014) Runst T., Sickel W. *Sobolev Spaces of Fractional Order, Nemytskij Operators, and Nonlinear Partial Differential Equations*, deGruyter, Berlin, New York (1996) Samko S.G., Kilbas A.A., Marichev O.I. *Fractional integrals and derivatives*, Gordon and Breach Science Publishers, Yverdon (1993) Strichartz R.S., *Analysis on products of fractals*, Trans. Amer. Math. Soc. 357, 571-615 (2005) Strichartz R.S., *Differential equations on fractals*, Princeton Univ. Press (2006) Tindel S., Tudor C.A., Viens F. *Stochastic evolution equations with fractional Brownian motion*, Probab. Theor. Relat. Fields 127, 186–204 (2003) Walsh J.B. *An introduction to stochastic partial differential equations*, École d’été de probabilités de Saint-Flour, XIV-1984, LNM 1180, Springer (1986) Zähle M. *Integration with respect to Fractal Functions and Stochastic Calculus I*, Probab. Theor. Relat. Fields 111, 333–374 (1998) Zähle M. *Integration with respect to Fractal Functions and Stochastic Calculus II*, Math. Nachr. 225, 145–183 (2001)
[^1]: We remark that there is a typo in [@HZ12 Lemma 5.2], namely in (ii) and (iii) the right hand side of the main condition on the parameters should read $2-2\eta -\beta$ instead of $2-2\eta -(\beta\vee\frac{d_S}{2})$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Traffic flow data collected by an induction loop detector on the highway close to Köln-Nord are investigated with respect to their dynamics including the stochastic content. In particular we present a new method, with which the flow dynamics can be extracted directly from the measured data. As a result a Langevin equation for the traffic flow is obtained. From the deterministic part of the flow dynamics, stable fixed points are extracted and set into relation with common features of the fundamental diagram.'
address: |
$^1$ Fachbereich 8 Physik, Universität Oldenburg, 26111 Oldenburg, Germany\
$^2$Institut für Theoretische Physik und Synergetik, Universität Stuttgart, 70550 Stuttgart, Germany\
$^3$ Deutsches Zentrum für Luft- und Raumfahrt (DLR) e.V., Institut für Verkehrsforschung, Rutherfordstrasse 2, 12489 Berlin, Germany
author:
- 'S. Kriso$^1$, R. Friedrich$^2$, J. Peinke$^1$ and P. Wagner$^3$'
title: Reconstruction of dynamical equations for traffic flow
---
\#1[[**\[\[\#1\]\]**]{}]{}
Introduction
============
By increasing the number of licensed vehicles on our roads it becomes more and more necessary to reduce appearances of traffic congestion. Getting higher capacities of highways means looking for optimized flow rates of cars. In order to do so, it is necessary to investigate the complex system of traffic flow and to understand its regularities. In a subsequent step, intelligent traffic control systems may use these laws to influence the traffic flow and thus to increase the highway’s capacity utilization.
For our investigation a large amount of traffic flow data were collected at the highway near Köln-Nord (Germany) over more than one week. For each car crossing an induction loop detector the following data were recorded: (i) the time, when the car crossed the detector, with an resolution of 1 sec, (ii) the type of car (passenger car, truck), (iii) the lane number, (iv) the velocity, truncated to an 8-bit integer (0...255 km/h), accuracy of approximately 3 %, (v) the length of the car in meters (8-bit integer, not calibrated) and (vi) the distance to the car driving ahead (an integer in the range 0…999 m, not calibrated). The lanes are labeled from A to C, where lane A is the right driving lane where mostly trucks are found (because of a German law), and lanes B and C are the fast drivers’ lanes.
In the following we present a new method to derive from the measured data dynamical equations for the traffic situation. We start with the common presentation of the fundamental diagram. Next, we evaluate the deterministic and the stochastic content of of the traffic dynamics by means of a Langevin equation. Finally, we give an interpretation of the determined Langevin equation.
Fundamental diagrams
====================
Here and in the following we calculate the car density $k$ \[km$^{-1}$\] and the flux state $q$ \[h$^{-1}$\] of each car using the measured data: the velocity $v$ \[km/h\], the length of the car $l$ \[m\] and the distance $d$ \[m\] to the car ahead: $$\begin{aligned}
k & = & \frac{1000}{l+d} \;\; ,\\ q & = & k \cdot v\;\; .
\label{Def}\end{aligned}$$
To avoid an overload of the presentation we restrict the diagrams and the following calculations to two cases. Firstly, only the traffic of a single lane C, secondly the cumulative traffic of all three lanes A, B and C are considered. Furthermore we calculate for each density state $k$ the mean flux state
$$\langle q \rangle_{k} = \frac{1}{n(k)} \sum q(k) \;\; .
\label{eins}$$
The obtained fundamental diagrams (q$_{i}$;k$_{i}$) and ($<$q$>$;k) are shown in Fig. \[FundDia\]. We note that for this presentation no significant difference in the traffic dynamics of one lane and the cumulative dynamics of three lanes can be detected. In both diagrams we find a maximum flux in free traffic flow of $q_{\rm max} = 8000$ cars/h, according to the results of [@Kerner]. For the flux out of traffic jams we find in both cases $q_{\rm out} = 5000 $ cars/h. So we have a ratio of $\eta := q^{\rm free}_{\rm max}/q_{\rm out} =
1.6$, which meets the value $\eta \approx 1.5$ found in [@Kerner] quite well.
Langevin equation for the traffic flow
======================================
In order to grasp the underlying dynamics of the traffic flow (one lane and all three lanes), we utilize a new method to analyze the traffic data more extensively. In particular, the iterative dynamics of the traffic state ${\mathbf
x}_{N}$, given by the velocity ${\mathbf v}_N$ and the flux ${\mathbf
q}_N$ of the $N$-th car, as a function of traffic state ${\mathbf
x}_{N-1}$ of the $N-1$-th car is investigated. Note, that other state variables could have been choosen as well. For the iterative dynamics of the traffic state ${\bf x}_{N}$ (which may be taken in a generalized case as a $r$-dimensional variable) we propose a description by a stationary Langevin equation, taking into account a combination of deterministic and random (noisy) forces cf. [@Risken]: $$x_{i,N+1} = x_{i,N} + h_{i}({\bf x}_{N}) + \sum_{j=1}^{r}
g_{ij}({\bf x}_{N}) \cdot \Gamma_{j,N} \;\; ,
\label{Langevin}$$ where the indices $i$ and $j$ denote components of the multidimensional variables, $N$ the time, and $\Gamma_{j,N}$ are $r
\times N$ independent Gaussian noise variables with zero mean and with variance 2, i.e., $$\langle \Gamma_{j,N} \rangle = 0 \; , \;\;
\langle \Gamma_{j,M} \cdot \Gamma_{k,N} \rangle =
2 \delta_{j,k}\delta_{M,N} \;\; .$$
The central part of our following work is that it is possible to determine the functions $h_{i}$ and $g_{ij}$ directly from empirical data. Taking (\[Langevin\]) as the Ito presentation of the stochastic process the following relation to the Kramers-Moyal coefficients can be given, $$\begin{aligned}
D_{i}^{(1)}({\bf x}) & = & h_{i}({\bf x}) \\
D_{ij}^{(2)}({\bf x}) & = &\sum_{k=1}^{r}g_{ik}({\bf
x})g_{jk}( {\bf x}) \;\; ,
\label{Dxh}\end{aligned}$$ where $D^{(1)}$ and $D^{(2)}$ are called drift and diffusion coefficient. These coefficients can be evaluated by the conditional moments $$\begin{aligned}
\label{D1} D_{i}^{(1)}({\bf x}) = \langle x_{i,N+1}-x_{i}
\rangle \Big|_{{\bf
x}_{N}={\bf x}} \;\; , \\
D_{ij}^{(2)}({\bf x}) =
\langle (x_{i,N+1}-x_{i})(x_{j,N+1}-x_{j}) \rangle \Big|_{{\bf
x}_{N}={\bf x}} \;\; .
\label{DefD}\end{aligned}$$
Recently it has been shown that with the analogous definition of these Kramers-Moyal coefficients it is possible to reconstruct from time continuous dynamics the underling stochastic differential equation [@us].
Before presenting our results on the dynamics we want to comment on the validity of this ansatz to describe the traffic flow by the Langevin equation (\[Langevin\]). This ansatz implies that the dynamics is in the class of Markovian processes, i.e. the system does not have a memory. This can be tested by conditional probabilities $$p(\bf{x}_{N}|\bf{x}_{N-1}, \ldots, \bf{x}_{N-m}) =
p(\bf{x}_{N}|\bf{x}_{N-1})$$ or by the necessary condition of the Chapman-Kolmogorov equation $$p(\bf{x}_{N}|\bf{x}_{N-r}) = \sum_{\bf{x}_
{N-s}} p(\bf{x}_{N}|\bf{x}_{N-s})
p(\bf{x}_{N-s}|\bf{x}_{N-r}) \;\; ,$$ where $r>s$. From our data, conditional probabilities have been evaluated and the validity of the Chapman-Kolmogorov equation was found for the iterative dynamics of both quantities, the velocity and the flux. If this Markovian property holds, the inherent noise of the dynamics ($\Gamma$) can be taken as $\delta$-correlated. It should be noted, that even in the case where the noise is not $\delta$-correlated, the deterministic part of the dynamics can be reconstructed from given data (\[D1\]), as we found by analysing numerically generated test data, [@Malte].
As expressed by (\[DefD\]), the knowledge of the conditional probabilities $p(\bf{x}_N | \bf{x}_{N-1})$ provides the basis to estimate the Kramers-Moyal coefficients from the traffic data. First we consider the simplified case of the onedimensional dynamics of the velocity only. The results for and $D^{(2)}$(v) for the traffic of one lane and for the cumulative traffic of all three lanes are shown in Fig. \[Drift\]. The one dimensional deterministic dynamics can also be expressed by the potential $\Phi_{D}$(v), defined as - $\frac{\delta \Phi_{D}}{\delta v}$ = D$^{(1)}$(v). The corresponding potentials are shown in Fig. \[Pot\]. From these results three noticable velocities $v_1 \approx 37$ km/h, $v_{2} \approx $ 75 km/h and $v_{3} \approx $ 107 km/h appear, which allows to identify the so called congested flow for $v\leq v_1$, correlated flow for $v_{1} < v \leq v_{2}$ and the free flow for $v > v_{2}$ [@flowDef; @flowDef1], respectively. Note these velocities can be defined as fixed points ($D^{(1)}(v)=0$) of the deterministic part of the cumulative traffic dynamics (see Fig. 2b).
For the traffic dynamics of lane C we find in the congested and in the correlated regime metastable traffic states, the deterministic drift term $D^{(1)}$ gets zero over finite intervals. This corresponds to the plateau structure in the potential, see Fig. 3a. A clearly different behaviour is found for the cumulative traffic dynamics of all three lanes, see Figs. \[Drift\]b and \[Pot\]b. The different flow regimes are seperated by two fixed points at $v_{1}$ and $v_{2}$. The slope of these fixed points defines the stability, thus the congested and the correlated flow regimes are separated by a stable fixed point, whereas the correlated and the free flow regime are separated by an instable fixed point. For the free flow, in both cases of one lane or three lane traffic a stable fixed point is found at $v_{3}$, correspondingly the drift potential has its local minimum. Because of a speed limit of 100 km/h given on the inspected highway we see a great increase of the potential for $v > v_{3}$: the faster a car is driving, the stronger the attraction is to the potential’s minimum [@comment] .
To get an understanding of the real traffic dynamics grasped by these drift coefficients or drift potentials, the additional noise has to be taken into account. In Fig. \[Drift\]c and d the corresponding magnitude of the noise are expressed by the evalutaed diffusion coefficients $D^{(2)}$. The noise will now cause transitions between different flow states. For the traffics dynamics of one lane, the noise will effect larger fluctuations as it is the case for the cumulative traffic dynamics, which has two clear minima in the potential. A further interesting detail is that the magnitude of $D^{(2)}$ has a minimum around the stable fixed point at $v_{3}$. This indicates a pronounced stability of this traffic situation.
Next we present the results of a higher dimensional analysis by taking ${\bf x}$ with the components $x_{1} = v$ and $x_{2} = q$. Now also the drift coefficient $D^{(1)}$ becomes a vector depending on $v$ and $q$, as shown in Fig. \[2DDrift\]. These results were obtained by binning the velocity flux data into a $25 \times 25$ matrix, corresponding to a binning of the velocity into intervals of $\Delta
\, v = 5$ km/h.
The solid lines in Fig. \[2DDrift\] show the states where we have no drift of the flux component: $D_{q}^{(1)} = 0$. On these lines we find only a velocity drift with a constant flux. In accordance to Fig. \[Drift\]a we find in Fig. \[2DDrift\]a for slow velocities mainly no velocity drift, in Fig. \[2DDrift\]b (like in Fig. \[Drift\]b) there seem to exist stable velocity drift states at the same velocities ($v_i \approx 37$km/h,$107$ km/h). The topology of the instable fixed point gets now a saddle point which is attractive for larger and smaller flux values but instable in the direction of larger and smaller velocity values. Again clear differences of the dynamics of on lane and three lanes is found.
Discussion and Conclusion
=========================
By the investigation of traffic flow data as an iterative stochastic process we were able to calculate from the given data the 1-dimensional drift and diffusion coefficients and thus to find the deterministic and stochastic part of the corresponding Langevin equation. We are able to find stable, metastable and unstable states (fixed points) in the deterministic part of free, correlated and congested traffic flow. For a fully description of the whole dynamics also the diffusion coefficient has to be taken into account, which provides transition probabilities between the different (meta-)stable states. Without this noisy part of the traffic flow dynamics, the stable states would never be left, i.e. a congestion would stay forever if once prepared. To see the dependency of velocity and flow from each other, the investigations were expanded to a higher dimensional analysis. In this case we find the deterministic and stochastic part of the 2-dimensional Langevin equation. Now we are able to identify stable velocity and flux states of the deterministic part.
Interestingly, the results found in this study are in agreement with empirical investigations that have identified three phases of traffic [@phastrans], together with transitions connecting theses phases. Especially the transition from free flow to correlated flow (or synchronized flow in the terminology of [@phastrans]) has similarities.
Finally we want to point out, that we presented here a new method to analyse traffic data with respect to a derivation of dynamical equations from pure data analysis. Furthermore we could show that our method provides more insight into the traffic dynamics than the presentations of diagrams like the fundamental diagram. A clear difference in the dynamics of one lane and the dynamics of cumulative three lanes was found. Our analysis provides evidence of the presence of fixed points, which are of practical importance if a control of a traffic should be achieved. At last one may conclude that this method will be helpful to perform a more thorough comparison between traffic flow models and empirical data.
Acknowledgement: Helpful discussions with Ch. Renner, St. Lück and M. Siefert are acknowledged. We also would like to thank the Landschaftsverband Rheinland and the Northrhine-Westfalia Ministry for Economy and Transport for providing the data used in this study.
B. S. Kerner, Traffic Flow, Experiment and Theory, in Proceedings of the Workshop on Traffic and Granular Flow ’97, (Springer Verlag, Berlin, 1998) pp. 239-267
H. Risken, [*The Fokker–Planck Equation*]{} (Springer, Berlin, 1989).
S. Siegert, R. Friedrich, and J. Peinke, Phys. Lett. A [**243**]{}, 275 (1998); [**271**]{}, 217 (2000).
M. Siefert, Diplomarbeit Oldenburg, 2000.
B. S. Kerner, Phys. Rev. Lett. [**81**]{}, 3797 (1998). It is commonly known, that in Germany people like to drive about 10 percent faster than given speed limits. With this speed there is still no punishment by the police.
L. Neubert, L. Santen, A. Schadschneider, and M. Schreckenberg, Phys. Rev. E [**60**]{} 6480 (1999).
B. S. Kerner, J. Phys. A [**33**]{}, L221-L228 (2000).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Fitting autoregressive moving average (ARMA) time series models requires model identification before parameter estimation. Model identification involves determining the order for the autoregressive and moving average components which is generally performed by visual inspection of the autocorrelation and partial autocorrelation functions, or by other offline methods. In many of today’s big data regime applications of time series models, however, there is a need to model one or multiple streams of data in an iterative fashion. Hence, the offline model identification step is significantly prohibitive. In this work, we regularize the objective of the optimization behind the ARMA parameter estimation problem with a nonsmooth hierarchical sparsity inducing penalty based on two path graphs that allows incorporating the identification into the estimation step. A proximal block coordinate descent algorithm is then proposed to solve the underlying optimization problem. The resulting model satisfies the required stationarity and invertibility conditions for ARMA models. Numerical results supporting the proposed method are presented.'
author:
- |
[Yin Liu[^1^]{} and Sam Davanloo Tajbakhsh[^1^]{}[^1] ]{}\
[[^1^]{}Department of Integrated Systems Engineering]{}\
[`{liu.6630,davanloo.1}@osu.edu `]{}\
[The Ohio State University]{}
bibliography:
- 'refs.bib'
title: ' **Fitting ARMA Time Series Models without Identification: A Proximal Approach**'
---
[ ***Keywords—*** ARMA time series models, Proximal methods, hierarchical sparsity. ]{}
Introduction {#sec:intro}
============
ARIMA time series models have a multitude of applications, e.g., in epidemiological surveillance [@zhang2014applications], water resource management [@wang2015improving], transportation systems [@billings2006application], drought forecasting [@han2010drought], stock price forecasting [@adebiyi2014comparison], business planning [@calheiros2014workload], and power systems [@chen2009arima], to name a few. Even emergence of deep neural networks and their customized architectures for time series modeling, e.g., Recurrent Neural Nets (RNN) and Long Short-Term Memory (LSTM) has not decreased the popularity of ARIMA models [@makridakis2018statistical].
Fitting ARMA(p,q) time series models requires a two-step process: 1. Model identification, 2. Parameter estimation. The model identification step determines the order of the autoregressive (AR) component (p) and moving average (MA) component (q). Next, given the underlying ARMA model, the parameters are estimated by solving an optimization problem for the maximum likelihood or least square estimates [@box2015time; @del2002statistical]. We should note that ARMA models are to model stationary processes; however, there exists a more general class of ARIMA models for *homogenous nonstationary* processes (which are stationary in the mean). Such processes become stationary after $d$ times differencing; hence, the corresponding ARIMA(p,d,q) model includes differencing of order $d$. The results of this paper are mainly for stationary processes with potential extension to the homogenous nonstatioanry processes.
Model identification is based primarily on visual inspection of the sample autocorrelation function (ACF) and partial autocorrelation (PACF) plots. For the AR(p) process, the sample ACF follows an exponential decay and the sample PACF cuts off after lag $p$, while for the MA(q) process, the sample ACF cuts off after lag $q$ and the sample PACF decays exponentially [@del2002statistical]. When the process involves both AR and MA components, it is more difficult to identify the correct orders. Next, model parameters are estimated by minimizing a loss function (e.g., negative loglikelihood or least square). @box2015time stepped even further and recommended an *iterative* approach between model identification and parameter estimation which involves inspection of the the residuals from the fitted model to make sure that they are indeed white noise.
In many of today’s applications, ARMA models should be fitted to many times series of data $\{y_t^j\}_{j=1}^J$ where $J$ is very large. For instance, the data could be the demand for more than thousands of product over time which are not necessarily correlated; hence, fitting Vector ARMA models are unnecessary, and separate modeling is more parsimonious. In such scenarios, model identifications become a significant bottleneck in the fitting process. This work is about a novel approach for fitting ARMA time series models that allows automating the fitting procedure by eliminating an explicit identification step. Indeed, with the aid of a single tuning parameter, the proposed algorithms allows data to identify the appropriate model.
Contributions
-------------
The contributions of this work are as follows:
- We develop a novel approach to fit ARMA time series models that identifies the model only by tuning a single continuous parameter ($\lambda$). The main idea behind this approach is to merge the model identification into the parameter estimation step by introducing a hierarchical sparsity inducing penalty into the optimization problem. The sparsity inducing penalty preserves the hierarchical structure of the nonzero parameter, e.g., it will not set the first AR parameter to zero while the second AR parameter is nonzero.
- We propose an efficient proximal block coordinate descent (BCD) algorithm to solve the underlying nonsmooth and nonconvex optimization problem to a stationary point – see Algorithm \[alg:master\_alg\]. The proximal map of the nonsmooth hierarchical sparsity inducing penalty is shown to be separable on the AR and MA components.
- The proposed approach will automate the ARMA time series modeling without a need for offline inspection for model identification, and allows fitting ARMA time series models to large number of time series data.
Notations
---------
Lowercase boldface letters denote vectors, and uppercase greek letters denote sets, except for $\cB$ which denotes the back-shift operator. The set of all real and complex numbers are denoted by $\mR$ and $\mC$ , respectively. Given a set $g\subseteq\cG$, $|g|$ denotes its cardinality and $g^c$ denotes its complement. Given $\b\in\mR^d$ and $g\subseteq\{1,\cdots,d\}$, $\b_g\in\mR^{|g|}$ is a vector with its elements selected from $\b$ over the index set $g$.
Problem definition {#sec:problem_def}
==================
We consider a stationary ARMA(p,q) time series process with a zero mean as $$\label{eq:arma_1}
y_t=\phi_1 y_{t-1}+\phi_2 y_{t-2}+\cdots+\phi_p y_{t-p} - \theta_1\epsilon_{t-1}-\theta_2\epsilon_{t-2}-\cdots-\theta_q\epsilon_{t-q}+\epsilon_t,$$ where $\phi_\ell, \ell=1,...,p$ are the parameters of the AR component, and $\theta_\ell,\ \ell=1,...,q$ are the parameters of the MA component, and $\epsilon_t$ is a white noise with zero mean and variance $\sigma^2$. The process can also be written as $$\label{eq:arma_2}
P_\ph^p(\cB) y_t = P_\th^q(\cB)\epsilon_t,$$ where $\cB$ is the *back-shift operator*, i.e., $\cB y_t=y_{t-1}$, and $$P_{\boldsymbol{\alpha}}^d(z)\triangleq1-\alpha_1z-\alpha_2z^2-\cdots-\alpha_d z^d,$$ is a polynomial of degree $d$ with the parameter $\boldsymbol{\alpha}$. The process is *stationary* if the AR component is stationary which is the case if all roots of the $P_\ph^p(z)$ polynomial are outside the unit circle; furthermore, the process is *invertible* if the the MA component is invertible which is the case if all roots of the $P_\th^d(z)$ polynomial are outside the unit circle [@del2002statistical]. Requiring the two polynomials to have roots outside of the unit circle in the $\cB$ space translates to some constraints on $\ph=[\phi_1,\cdots,\phi_p]^\top$ and $\th=[\theta_1,\cdots,\theta_q]^\top$, i.e. $\ph\in\cX_{\ph}^p\subseteq\mR^p$ and $\th\in\cX_\th^q\subseteq\mR^q$, where $\cX_{\boldsymbol{\alpha}}^d$ is defined as $$\label{eq:opt}
\cX_{\boldsymbol{\alpha}}^d \triangleq \{\boldsymbol{\alpha}\in\mR^d:\ \forall z\in\mC, \ P_{\boldsymbol{\alpha}}^d(z)=0 \Rightarrow |z|>1 \}.$$ We should note that there is also another (maybe more common) representation for $\cX_{\boldsymbol{\alpha}}^d$ based on the monic polynomial $$\bar{P}_{\boldsymbol{\alpha}}^d (z)\triangleq z^d+\alpha_1z^{d-1}+\cdots+\alpha_{d-1}z+\alpha_d,$$ of degree $d$, where it can be shown that $$\label{eq:opt_2}
\cX_{\boldsymbol{\alpha}}^d = \{\boldsymbol{\alpha}\in\mR^d:\ \forall z\in\mC, \ \bar{P}_{\boldsymbol{\alpha}}^d(z)=0 \Rightarrow |z|<1 \}.$$ Note that the new representation requires roots of the polynomial to be *inside* the unit circle. For an arbitrary $d$, geometrical complexity of $\cX_{\boldsymbol{\alpha}}^d$ makes projection onto this set very difficult [@combettes1992best]. Indeed, [@combettes1992best] discussed that $\cX_{\boldsymbol{\alpha}}^d$ is open, bounded, and *not necessarily convex* – see also [@moses1991determining; @blondel2012explicit]. To deal with the openness of $\cX_{\boldsymbol{\alpha}}^d$, it is common to approximate it with a closed set from inside – see . However, projection onto this set or its approximation *may not be unique* due to their potential nonconvexities. A method for projection onto the $\cX_{\boldsymbol{\alpha}}^d$ set was developed in [@moses1991determining]. While their scheme is easy to implement, the convergence of this iterative method is slower than steepest descent – see also [@combettes1992best]. To conclude, imposing stationarity and invertibility of the model is performed by projecting $\ph$ and $\th$ onto (inner approximate of) $\cX_\ph^p$ and $\cX_\th^q$, respectively, which my not be unique.
The above discussion is for an ARMA model that is already identifies, i.e., $p$ and $q$ are known. For a model that is not identified, we also need $$\begin{aligned}
\label{eq:parent_child}
\begin{split}
\text{if } \phi_\ell=0 \text{ then } \phi_{\ell'}=0, \ \forall \ell<\ell', \iff \text{if } \phi_{\ell'}\neq0 \text{ then } \phi_{\ell}\neq0, \ \forall \ell<\ell', \\
\text{if } \theta_\ell=0 \text{ then } \theta_{\ell'}=0, \ \forall \ell<\ell', \iff \text{if } \theta_{\ell'}\neq0 \text{ then } \theta_{\ell}\neq0, \ \forall \ell<\ell',
\end{split}\end{aligned}$$ i.e., the sparsity of $\ph$ and $\th$ follow hierarchical structures.
Before discussing how these sparsity structures are enforced, we will briefly discuss the loss function for fitting ARMA models. Provided an identified model, i.e., $p$ and $q$ are known, fitting ARMA models are generally performed by finding the conditional maximum likelihood or conditional least-square estimates, which are close to each other assuming that $\epsilon_t$ in follow a Normal distribution and the size of the time series $T$ is reasonably large. The conditional least-square estimate (for an identified model) requires solving $$\begin{aligned}
\label{eq:cond_LS}
\begin{split}
\min_{\ph,\th} \quad \cL(\ph,\th) & = (1/2)\hspace{-0.4cm}\sum_{t=\max\{p,q\}}^T \hspace{-0.4cm}\epsilon_t^2 = (1/2)\hspace{-0.4cm}\sum_{t=\max\{p,q\}}^T\hspace{-0.3cm}\left(y_t-\hat{y}_{t|t-1}(\ph,\th)\right)^2 \\ \text{s.t.} & \quad \ph\in\cX_\ph^p, \quad \th\in\cX_\th^q,
\end{split}\end{aligned}$$ where $\hat{y}_{t|t-1}(\ph,\th)$ is the model prediction for $y_t$ using the data $\{y_t\}_{t=1}^{t-1}$, and is called *conditional* since it depends on the $p$ initial values for $y_t$ and $q$ initial values for $\epsilon_t$. Note that in the absence of MA terms (i.e., $q=0$), the objective function of is convex in the parameters of the AR model $\ph$. However, if $q>0$ then the objective function of is nonconvex, and optimization routines are not guaranteed to converge to the global optimum [@hamilton1994time; @box2015time; @benidir1990nonconvexity; @georgiou2008convex]. To sum up, in its most general case, problem involves nonconvex minimization over a nonconvex set and, hence, it is difficult to solve.
This paper is concerned with an optimization problem solution of which preserves the hierarchical sparsity structure discussed above. In the next section, we propose a method that allows learning $p$ and $q$ within the parameter estimation step.
Proposed method {#sec:prop_method}
===============
Before discussing the proposed method, we should briefly discuss the notion of *hierarchical sparsity*. Let $D=(\cS,\cE)$ be a Directed Acyclic Graph (DAG) where $\cS=\{s_1,\cdots,s_n\}$ is the set of graph nodes and $\cE$ be the set of ordered pair of nodes denoting edges where each pair denotes an edge from the node in the first element to the node in the second element. Each $s_i$ is an index set of the parameters of the model such that $s_i\cap s_j=\varnothing,\ \forall (i,j)$ and $\cup_{i=1}^n s_i=\{1,\cdots,d\}$ where $d$ is the number of parameters. DAG shows the sparsity structures of interest in the parent/child relationship. Assuming one variable per node, the variable in a child node can only be nonzero if the variable in the parent node is nonzero. For instance, given a parameter $\b\in\mR^3$, the left plot in Figure \[fig:simple\_path\_graphs\] requires $\beta_1\neq 0$ if $\beta_2\neq 0$ and $\beta_2\neq 0$ if $\beta_3\neq 0$.
![Path graphs showing hierarchical sparsities: **(Left)** A graph with a variable per node for $\b\in\mR^3$. **(Right)** A graph with multiple variables per node for $\b\in\mR^5$.[]{data-label="fig:simple_path_graphs"}](path_graph_1.pdf "fig:") ![Path graphs showing hierarchical sparsities: **(Left)** A graph with a variable per node for $\b\in\mR^3$. **(Right)** A graph with multiple variables per node for $\b\in\mR^5$.[]{data-label="fig:simple_path_graphs"}](path_graph_2.pdf "fig:")
For a DAG that contains more than one variable per node (e.g. the right plot in Figure \[fig:simple\_path\_graphs\]), two different hierarchies can be considered: 1. [*Strong hierarchy*]{}: the parameters in the child node can only be nonzero if *all* of the parameters in its parent node(s) are nonzero. 2. [*Weak hierarchy*]{}: the parameters in the child node can be nonzero if *at least one* of the parameters in its parent node(s) is nonzero [@bien2013lasso]. For more information about hierarchical sparsity structures refer to [@zhao2009composite; @jenatton2011proximal; @jenatton2011structured; @bach2012structured; @yan2017hierarchical].
Hierarchical sparsity for ARMA models {#sec:hierarchical_arma}
-------------------------------------
In this work, we want to include the model identification of ARMA models into the parameter estimation step. We assume the knowledge about some upper bounds on the true $p^*$ and $q^*$, i.e., $\bar{p}\geq p^*$ and $\bar{q}\geq q^*$, respectively. Considering $\text{ARMA}(\bar{p},\bar{q})$, the estimated parameters should satisfy the condition . To do so, we define two path graphs as shown in Figure \[fig:arma\_DAG\]. Since this DAG consists of two path graphs and there is only one variable in each node, weak and strong hierarchies are equivalent.
![DAG for the $\text{ARMA}(\bar{p},\bar{q})$ process. The red dotted rectangles illustrate the ascending grouping scheme for the LOG penalty. []{data-label="fig:arma_DAG"}](arma_hierarchy.pdf)
Enforcing the sparsity structure shown in Figure \[fig:arma\_DAG\] *exactly* requires introducing binary variables into the optimization problem and solving a Mixed Integer Program (MIP). For instance, to model the parent/child hierarchy between $\phi_1$ and $\phi_2$, one need to introduce a binary variable $z\in\{0,1\}$ and two constraints as $z\epsilon\leq|\phi_1|$ and $|\phi_2|\leq z\mu$ for some reasonably small and large parameters $\epsilon$ and $\mu$, respectively. Provided that the the underlying optimization problem is already very difficult to solve, introducing $\bar{p}+\bar{q}-2$ binary variable makes the problem even more challenging. Hence, despite the significant recent advances in MIP algorithms (see e.g., [@manzour2019integer; @bertsimas2016best; @mazumder2017thediscrete; @bertsimas2017sparse]), we use a nonsmooth but convex regularizer that induces hierarchical sparsity structures of interest.
Latent Overlapping Group (LOG) Lasso
------------------------------------
The hierarchical sparsity structure shown in Figures \[fig:arma\_DAG\] is induced by regularizing the objective function in by the LOG penalty – see [@jacob2009group]. Let $\b\triangleq[\ph,\th]\in\mR^{(\bar{p}+\bar{q})}$ denote all of the parameters of the ARMA model. The LOG penalty function is defined as $$\label{eq:LOG}
\Omega_{\text{LOG}}(\b)=\inf_{\n^{(g)},\ g\in\cG} \left\{\sum_{g\in\cG}w_g\norm{\n^{(g)}}_2 \ \ \text{s.t.} \ \sum_{g\in\cG}\n^{(g)}=\b,\ \n^{(g)}_{g^c}=0 \right\}$$ where $\cG\subseteq\{1,\cdots,(\bar{p}+\bar{q})\}$ is the set of groups (of the nodes of the DAG which is discussed next), $g\in\cG$ is itself a set, $\n^{(g)}\in\mR^{(\bar{p}+\bar{q})}$ is a latent vector indexed by $g$, and $w_g$ is the weight for group $g$. The groups should follow an *ascending* structure, i.e., for each node there is a group that contains that node and all of its ascendants. For the DAG in Figure \[fig:arma\_DAG\], the groups are $$\cG=\Big\{ \{1\}, \{1,2\},\cdots,\{1,\cdots,\bar{p}\},\{\bar{p}+1\},\{\bar{p}+1,\bar{p}+2\},\cdots,\{\bar{p}+1,\cdots,\bar{p}+\bar{d}\} \Big\},$$ where given the definition of $\b$, we assumed the indices of the AR parameter $\ph$ first, and those of the MA parameter $\th$ next. These groups are shown with the red dotted rectangles in Figure \[fig:arma\_DAG\]. Given that $\ell_2$-norm induces block sparsity, the LOG penalty tries to find block sparse combinations of the latent variables the sum of which is equal to $\b$ [@jacob2009group; @yan2017hierarchical]. For instance, for an ARMA model with $\bar{p}=2$ and $\bar{q}=2$, $\cG=\{\{1\},\{1,2\},\{3\},\{3,4\}\}$, the objective of the infimum is $|\n^{\{1\}}_1|+\norm{[\n^{\{1,2\}}_1,\n^{\{1,2\}}_2]}+|\n^{\{3\}}_3|+\norm{[\n^{\{3,4\}}_3,\n^{\{3,4\}}_4]}$ (where for simplicity $w_g=1,\ \forall g\in\cG$) and the constraints are $$\begin{bmatrix}
\n^{\{1\}}_1 \\
0 \\
0 \\
0
\end{bmatrix}
+
\begin{bmatrix}
\n^{\{1,2\}}_1 \\
\n^{\{1,2\}}_2 \\
0 \\
0
\end{bmatrix}
+
\begin{bmatrix}
0 \\
0 \\
\n^{\{3\}}_3 \\
0
\end{bmatrix}
+
\begin{bmatrix}
0 \\
0 \\
\n^{\{3,4\}}_3 \\
\n^{\{3,4\}}_4
\end{bmatrix}
=
\begin{bmatrix}
\phi_1 \\
\phi_2 \\
\theta_1 \\
\theta_2
\end{bmatrix}.$$
The proposed Hierarchically Sparse learning problem
---------------------------------------------------
The proposed Hierarchically Sparse (HS) learning problem is $$\tag{HS-ARMA}
\label{eq:prop_opt}
\begin{split}
\min_{\ph,\th} \quad &\cL(\ph,\th) +\lambda\Omega_{\text{LOG}}(\ph,\th) \\
\text{s.t.} & \quad \ph\in\cX_\ph^p, \quad \th\in\cX_\th^q,
\end{split}$$ where $\lambda>0$ is a tuning parameter, $\cX_\ph^p$ and $\cX_\th^q$ are defined based on , and $\Omega_{\text{LOG}}(\cdot)$ is defined in . $\lambda$ controls the tradeoff between the loss and penalty functions and, hence, allows model identification and parameter estimation simultaneously. Obviously increasing $\lambda$ result in sparser models where these nested models satisfy the hierarchical sparsity structure shown in Figure \[fig:arma\_DAG\]. As discussed in Section \[sec:hierarchical\_arma\], $\bar{p}$ and $\bar{q}$ are some upper bounds on the true $p^*$ and $q^*$ known a priori.
Given the convex nonsmooth function $\Omega_{\text{LOG}}(\cdot)$, we propose to solve using a proximal method [@nesterov2013gradient; @beck2009fast; @parikh2014proximal]. Similar to gradient methods which requires iterative evaluation of the gradient, proximal methods require iterative evaluation of the proximal operator. Proximal operator of the $\Omega_{\text{LOG}}(\bb)$ at $\bb\in\mR^{(\bar{p}+\bar{q})}$ is defined as $$\label{eq:prox_def}
\prox{\lambda\Omega_{\text{LOG}}}(\bb)\triangleq \operatorname*{argmin}_{\b\in\mR^{(\bar{p}+\bar{q})}}\left\{\lambda\Omega_{\text{LOG}}(\b)+\frac{1}{2}\norm{\b-\bb}_2^2\right\}.$$ In [@zhang2020firstorder], authors developed a two-block alternating direction method of multiplier (ADMM) with a sharing scheme [@boyd2011distributed] to solve – see Algorithm \[alg:sharing\_Prox\_LOG\]. The proposed algorithm can be parallelized over all groups in $\cG$ in the update of the first-block; furthermore, it converges *linearly* – see [@zhang2020firstorder] for more details.
$\bb, \lambda, \alpha, w_g \ \forall g\in\cG$ $k=0,\ U^0_{.g}=\bo,\ X^{2,0}_{.g}=\bo~\forall g\in\cG$ $k\gets k+1$\
$X_{gg}^{1,k+1} \gets \prox{\lambda w_g\norm{\cdot}_2}(X^{2,k}_{gg}-U^k_{gg}), \ \ \forall g\in\cG$\
$X_{g^cg}^{1,k+1} \gets \bo, \ \ \forall g\in\cG$ $\bar{\bx}^{2,k+1}\gets\frac{1}{|\cG|+\rho}\Big(\bb+\frac{\rho}{|\cG|}\sum_{g\in\cG}(X^{1,k+1}_{.g}+U^k_{.g})\Big)$\
$X_{.g}^{2,k+1}\gets\bar{\bx}^{2,k+1}+X^{1,k+1}_{.g}+U^k_{.g}-(1/|\cG|)\sum_{g\in\cG}(X^{1,k+1}_{.g}+U^k_{.g}),\ \forall g\in\cG$\
$U_{.g}^{k+1}=U_{.g}^{k}+(\alpha/\rho)\big(\frac{1}{|\cG|}\sum_{g\in\cG}(X^{1,k+1}_{.g}+U^k_{.g})-\bar{\bx}^2\big), \ \forall g\in\cG.$\
\
$\b=\sum_{g\in\cG} X^{1,k+1}_{.g}$\
**Output:** $\b$
Let $\Omega_{\text{LOG}}^{\text{AR}}$ and $\Omega_{\text{LOG}}^{\text{MA}}$ be the LOG penalties for the pure AR, i.e, ARMA$(\bar{p},0)$, and pure MA, i.e, ARMA$(0,\bar{q})$, models, respectively. In Lemma \[lem:separable\_prox\] below, we show that the proximal operator of $\Omega_{\text{LOG}}$ is separable over $\ph$ and $\th$.
\[lem:separable\_prox\] The proximal operator of the LOG penalty defined over the ARMA DAG is separable, i.e., $\prox{\Omega_{\text{LOG}}}(\bb_1,\bb_2)=( \prox{\Omega_{\text{LOG}}^{\text{AR}}}(\bb_1),\prox{\Omega_{\text{LOG}}^{\text{MA}}}(\bb_2) )$.
With a slight abuse of notation, let $\cG^{\text{AR}}$ be the set of groups for $\Omega_{\text{LOG}}^{\text{AR}}$ such that $\sum_{g\in\cG^{\text{AR}}}\n^{(g)} = \ph$ (the top path graph in Figure \[fig:arma\_DAG\]). Similarly, let $\cG^{\text{MA}}$ be the set of groups for $\Omega_{\text{LOG}}^{\text{MA}}$ such that $\sum_{g\in\cG^{\text{MA}}}\boldsymbol{\omega}^{(g)} = \th$. Given that the objective of the infimum in the definition of $\Omega_{\text{LOG}}$ for the ARMA DAG is separable in $\cG^{\text{AR}}$ and $\cG^{\text{MA}}$, we have $\Omega_{\text{LOG}}(\ph,\th)=\Omega_{\text{LOG}}^{\text{AR}}(\ph)+\Omega_{\text{LOG}}^{\text{MA}}(\th)$. Hence, the result follows from the separable sum property of the proximal operator.
Indeed, in Algorithm \[alg:master\_alg\], the proximal operator of LOG is not evaluated in one step while the algorithm evaluates $\prox{\lambda\Omega_{\text{LOG}}^{\text{AR}}}$ and $\prox{\lambda\Omega_{\text{LOG}}^{\text{MA}}}$ sequentially in a Gauss-Seidel manner.
The algorithm to solve problem is a two-block proximal block coordinate descent (BCD) with projection, shown in Algorithm \[alg:master\_alg\].
$\lambda, \bar{p}, \bar{q}, \ph_0\in\cX_{\ph}^{\bar{p}}, \th_0\in\cX_{\th}^{\bar{q}}$ $k=1$\
$\ph^{k+1/2} \gets \prox{\lambda\Omega_{\text{LOG}}^{\text{AR}}}(\ph^k-\gamma_k\grad_{\ph}\cL(\ph^k,\th^k))$\
$p \gets \text{card}(\{i: \ph^{k+1/2}_i\neq 0\})$\
$\ph^{k+1} \gets \text{Proj}_{\tilde{\cX}_\ph^p}(\ph^{k+1/2})$\
$\th^{k+1/2} \gets \prox{\lambda\Omega_{\text{LOG}}^{\text{MA}}}(\th^k-\gamma_k\grad_{\th}\cL(\ph^{k+1},\th^k))$\
$q \gets \text{card}(\{i: \th^{k+1/2}_i\neq 0\})$\
$\th^{k+1} \gets \text{Proj}_{\tilde{\cX}_\th^q}(\th^{k+1/2})$\
$k\gets k+1$\
\
**Output:** $(\ph_k,\th_k)$
From , since $\epsilon_t=y_t-\ph^\top\by_{t-p}^{t-1}-\th^\top\e_{t-q}^{t-1}$ where $\by_{t-p}^{t-1}=[y_{t-1},\cdots,y_{t-p}]$ and $\e_{t-q}^{t-1}=[\epsilon_{t-1},\cdots,\epsilon_{t-q}]$, we have
$$\begin{aligned}
\grad_{\ph} \cL(\ph,\th) &= -\hspace{-0.4cm}\sum_{t=\max\{\bar{p},\bar{q}\}}^T (y_t-\ph^\top\by_{t-p}^{t-1}-\th^\top\e_{t-q}^{t-1})\by_{t-p}^{t-1}, \label{eq:grad_wrt_phi} \\
\grad_{\th} \cL(\ph,\th) &= -\hspace{-0.4cm}\sum_{t=\max\{\bar{p},\bar{q}\}}^T (y_t-\ph^\top\by_{t-p}^{t-1}-\th^\top\e_{t-q}^{t-1})\e_{t-q}^{t-1}. \label{eq:grad_wrt_theta}\end{aligned}$$
The gradient updates are passed to proximal operators as arguments which is indeed proximal gradient steps [@beck2009fast; @parikh2014proximal]. Note that the solution of the proximal operators are sparse vectors that conform to the hierarchical sparsity of Figure \[fig:arma\_DAG\].
The solutions of the proximal gradient steps for the AR and MA components, i.e., $\ph^{k+1/2}$ and $\th^{k+1/2}$ are not necessarily stationary or invertible, respectively. The stationarity and invertibility of AR and MA are regained by projection on $\cX_{\ph}^p$ and $\cX_{\th}^q$ where $p$ and $q$ are the order of AR and MA components from the proximal steps. For projection, we use the second definition of $\cX_{\boldsymbol{\alpha}}^d$ in . Since $\cX_{\boldsymbol{\alpha}}^d$ is an open set, following [@combettes1992best], we find its approximation with a closed set from inside as $$\label{eq:apprx_cX}
\tilde{\cX}_{\boldsymbol{\alpha}}^d(\delta) \triangleq \{\boldsymbol{\alpha}\in\mR^d:\ \forall z\in\mC, \ \bar{P}_{\boldsymbol{\alpha}}^d(z)=0 \Rightarrow -1+\delta\leq z \leq1-\delta \},$$ where $\delta>0$ determines the approximation gap. Euclidean projection on $\tilde{\cX}_{\ph}^p(\delta)$ and $\tilde{\cX}_{\th}^q(\delta)$ sets guarantee stationarity and invertibility of $\ph^{t+1}$ and $\th^{t+1}$, respectively. Note that these projections do *not* change sparsity of the parameters.
Finally, since $\epsilon_t$ in the objective of is calculated based on ARMA$(\bar{p},\bar{q})$ while the iterates $\ph^{t+1}$ and $\th^{t+1}$ are feasible with respect to $\cX_{\ph}^p$ and $\cX_{\th}^q$ respectively, we need to show $(\ph^{k+1},\th^{k+1})\in\cX_\ph^{\bar{p}}\times\cX_\th^{\bar{q}}$. This is established in Lemma \[lem:nested\] below.
\[lem:nested\] For any $d\in\{1,2,...\}$, we have $\cX_{\boldsymbol{\alpha}}^d\subseteq\cX_{\boldsymbol{\alpha}}^{d+1}$.
Proof follows from the definition of $\cX_{\boldsymbol{\alpha}}^d$ in , and that if $\boldsymbol{\alpha}\in\cX_{\boldsymbol{\alpha}}^d$ then $[\boldsymbol{\alpha},0]\in\cX_{\boldsymbol{\alpha}}^{d+1}$.
Therefore, $\{\cX_{\boldsymbol{\alpha}}^d\}_{d=1}^{\bar{d}}$ is a sequence of nested sets as $\cX_{\boldsymbol{\alpha}}^1\subseteq\cdots\subseteq\cX_{\boldsymbol{\alpha}}^{\bar{d}}$. However, the reverse is not true, i.e., $\boldsymbol{\alpha}\in\cX^d_{\boldsymbol{\alpha}}$ is *not sufficient* for $[\alpha_1,\cdots,\alpha_{d-1}]\in\cX^{d-1}_{\boldsymbol{\alpha}}$, which can be shown by counter examples.
A note on the optimization problem {#sec:note}
-----------------------------------
Problem requires nonconvex and nonsmooth optimization over a nonconvex set. To be specific, if $q=0$ the loss function is convex in $\ph$; otherwise, $\cL(\ph,\th)$ is nonconvex in *both* $\ph$ and $\th$. Indeed, when $q>0$ the objective function is a *polynomial* function of degree $T-\max\{p,q\}$. The $\Omega_{\text{LOG}}(\ph,\th)$ penalty is a nonsmooth but jointly convex function in its arguments. Finally, $\cX_{\ph}^p$ and $\cX_{\th}^q$ are open nonconvex sets and their approximations $\tilde{\cX}_{\ph}^p$ and $\tilde{\cX}_{\th}^q$ (defined in ) are closed but still nonconvex.
To Deal with nonconvexities of $\tilde{\cX}_{\ph}^p$ and $\tilde{\cX}_{\th}^q$, one may try to approximate them with some inscribed convex sets which requires generalizations of the *potato peeling problem* [@goodman1981largest] and the algorithm in [@chang1986polynomial] to non-polygon geometries – see also [@cabello2017peeling]. Note that optimization over the convex hulls of these sets may result in nonstationary or noninvertible solutions.
Under some convex approximations of the sets $\tilde{\cX}_{\ph}^p$ and $\tilde{\cX}_{\th}^q$, the problem under investigation is a nonconvex nonsmooth optimization over a convex set. For such a setting, algorithms are settled with finding solutions that satisfy some necessary optimality conditions, e.g., stationary solutions which are those that lack a feasible descent direction. To the best of our knowledge, the only study that provides a method that converges to stationary points in this setting is [@razaviyayn2013unified], which involves iterative minimization of a consistent majorizer of the objective function over the feasible set.
Numerical Studies {#sec:numerical}
=================
Data generation process {#sec:data_gen}
-----------------------
To generate a statioanry and an invertible ARMA($p^*,q^*$) model, we first generate $p^*+q^*$ numbers uniformly at random on $[-1,-0.1]\cup [0.1,1]$ for all parameters. The samples are then rejected if the stationary and invertibility conditions, based on , are not satisfied. Given an accepted sampled parameter $(\ph^{*,i},\th^{*,i})$, a realization of the time series with length $T=4000$ is simulated with a zero mean and variance equal to one.
Model identification and parameter estimation accuracy
------------------------------------------------------
To evaluate the estimation error of the proposed method, we simulate $n=20$ realizations of ARMA models with orders $(p^*,q^*)$ such that $p^*\leq \bar{p}=10$ and $q^*\leq\bar{q}=10$ following our discussion in Section \[sec:data\_gen\]. The tuning parameter of the $\Omega_{\text{LOG}}$ penalty is set as $\lambda = \lambda_0 \sqrt{T}$ with $\lambda_0 \in \{0.5,1,2,3,5,10\}$ and $w_g$ in its definition is set to $|g|^{1/2}$. The estimation error is calculated as $ \epsilon_{\lambda_0} = \|(\hat{\ph}_{\lambda_0},\hat{\th}_{\lambda_0})-(\ph^{*},\th^{*}) \|_2$, where $(\ph^*,\th^*)$ are the true and $(\hat{\ph},\hat{\th})$ are the estimated parameters based on Algorithm \[alg:master\_alg\]. Table \[tab:average\_error\] reports the mean and standard deviation of the estimation errors for different $\lambda_0$ values.
To provide a better understanding of the quality of parameter estimates and how they conform to the induced sparsity structure in Figure \[fig:arma\_DAG\], we conduct another study. First, we sampled one realization from 10 different ARMA(3,2) models. Then, with $\bar{p}=\bar{q}=5$ and $\lambda_0 \in \{0.5,1,2,3,5,10\}$, the HS-ARMA parameter estimates $(\hat{\ph}^i_{\lambda_0}, \hat{\th}^i_{\lambda_0})$ are calculated using Algorithm \[alg:master\_alg\] and reported along with the true parameters $(\ph^{*,i},\th^{*,i})$ in Table \[tab:sample\_detail\], where $i$ is the simulation index. Simple tuning of $\lambda_0$ allows the method to correctly identify the true orders $(p^*,q^*)$ and the estimated parameters conform to the underlying sparsity structure. Furthermore, the estimation errors are reasonably small. We also compared the estimation errors with pre-identified models where their parameters are estimated using a package – see Figure \[fig:distance\]. The mean of the HS-ARMA estimation error lies between those of the correctly and incorrectly identified (by one order in the AR component) models. For some samples with $\lambda_0$ around 2 or 3, the error of HS-ARMA is very close to the correctly identified ARMA model.
![The estimation error of HS-ARMA and two pre-identified models. The three thicker lines are the mean estimation errors and thinner lines represent estimation errors for each sample.[]{data-label="fig:distance"}](error_sample.pdf){width="60.00000%"}
Prediction performance
----------------------
We also compare the prediction performance of the HS-ARMA with those of correctly and incorrectly identified models using 10 realizations of one ARMA(3,2) model. For each realization, the estimated parameters with $\lambda_0\in \{0.5,1,2,3,5\}$ are used to forecast the process for the next 20 time points. Note that $\lambda_0=10$ is omitted because the fitted parameters were too sparse. Figure \[fig:prediction\] illustrates Root Mean Square Error (RMSE) for these methods.
For some $\lambda_0$, the RMSE of HS-ARMA is smaller than that of the correctly identified ARMA model. Furthermore, all HS-ARMA predictions for different $\lambda_0$ values have significantly lower RMSE compared to the incorrectly identified model.
![Prediction RMSEs for the HS-ARMA method vs. the correctly and an incorrectly identified models. Each grey thin line is the RMSE of HS-ARMA with one $\lambda_0$ from $\{0.5,1,2,3,5\}$ from ten realizations and the black thick line is the average of the grey lines. The green and red lines are the RMSEs from the ten realizations for correctly and incorrectly identified models.[]{data-label="fig:prediction"}](prediction_error.pdf){width="60.00000%"}
Concluding remarks {#sec:conclusion}
==================
This work presents a new learning framework that allows model identification and parameter estimation for ARMA time series models simultaneously. To do so, we introduced a hierarchical sparsity inducing penalty, namely the Latent Overlapping Group (LOG) lasso, in the objective of the parameter estimation problem. While the addition of a nonsmooth (but convex) function to the objective of an already difficult nonconvex optimization seems restrictive, we propose a proximal block coordinate descent algorithm that can solve the problem to a potential stationary point efficiently. Numerical simulation studies confirm capabilties of the proposed learning framework to identify the true model and estimate its parameters with reasonably high accuracy.
We believe that this study sheds some light on the hard optimization problem behind the parameter estimation of ARMA time series models (see our brief discussion in Section \[sec:note\]). Furthermore, we hope it motivates future studies to look into convergence analysis of the proposed proximal BCD or other algorithms for such problem structures. Finally, the proposed framework can be extended to fit Vector ARMA (VARMA) models where the underlying path graphs would contain multiple variables per nodes (see e.g. the right plot in Figure \[fig:simple\_path\_graphs\]), which we also leave for future studies.
[^1]: The corresponding author
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Differences in the confinement of electrons and holes in quantum dots are shown to profoundly impact the magnitude of scattering with acoustic phonons in materials where crystal deformation shifts the conduction and valence band in the same direction. Using an extensive model that includes the non-Markovian nature of the phonon reservoir, we show how the effect may be addressed by photoluminescence excitation spectroscopy of a single quantum dot. We also investigate the implications for cavity QED, i.e. a coupled quantum dot-cavity system, and demonstrate that the phonon scattering may be strongly quenched. The quenching is explained by a balancing between the deformation potential interaction strengths and the carrier confinement and depends on the quantum dot shape. Numerical examples suggest a route towards engineering the phonon scattering.'
author:
- 'A. Nysteen, P. Kaer, and J. Mork'
title: 'Quenching of phonon-induced processes in quantum dots due to electron-hole asymmetries'
---
It is well-known that cavity QED systems undergo dephasing due to interactions with the environment, and that this leads to loss of quantum mechanical coherence between the different states of the system. In particular, solid-state based cavity QED systems such as semiconductor micropillars [@Reithmaier2004; @Gerard1996; @Lermer2012] and photonic crystal cavities [@Yoshie2004; @Vuckovic2003] are strongly affected by dephasing induced by phonon scattering [@Hohenester2010; @Ates2009; @Hughes2011; @Majumdar2011; @WilsonRae2002; @Muljarov2004; @Kaer2010; @Gauger2008]. Phonon-induced decoherence thus makes the observation of such effects as vacuum Rabi oscillations [@Norris1994] much more difficult than in atomic cavity QED systems and impairs the realization of a scalable solid-state platform for quantum information technology. For instance, the requirement of indistinguishability of subsequent emission events from a single-photon source places strict limitations on the amount of dephasing that can be accepted [@KaerLang; @Kiraz2004; @Bylander2003].
In this Letter we investigate the role of carrier confinement on phonon scattering in a semiconductor quantum dot (QD). We employ a comprehensive theoretical model that takes into account the non-Markovian nature of the phonon reservoir [@Kaer2010], thus avoiding the standard approach of describing phonon-induced decoherence by a pure dephasing rate [@Krummheuer2002; @Auffeves2010; @Naesby2008]. Surprisingly, we find that acoustic deformation potential scattering may be completely quenched under certain conditions, depending on the degree of confinement of the involved electronic states. The effect is investigated in detail for two cases.
Firstly, we consider the case of photoluminescence spectroscopy of a single QD and find that the luminescence is suppressed at certain detunings due to a quenching of phonon scattering. For spherical confinement potentials we derive approximate analytical results for the scattering rate. These results show that the effect originates from the difference in the spatial confinement of electrons and holes, usually neglected in theoretical treatments.
Secondly, we consider a coupled QD-cavity system. We find that the quenching of the phonon scattering strongly affects pure dephasing with consequences e.g. for the indistinguishability of single photon sources. We further expand the description to realistic QD structures, which are analyzed numerically, and the conditions for reducing phonon scattering are established.
Other approaches towards controlling the degree of phonon scattering include the use of phononic bandgap structures for suppressing vibrational modes [@Kushwaha1993]. Structures with simultaneous photonic and phononic bandgaps have been discussed theoretically [@Maldovan2006] and experimentally realized [@Safavi2010] with phononic bandgaps in the GHz-regime. In cavity QED, the most relevant phonons belong to the acoustic branch and have energies in the THz-regime, hence the present phononic band gap structures are primarily of interest for improving the optomechanical coupling [@Gavartin2011]. Another approach towards manipulating the electron-phonon interaction has been demonstrated by placing the QD near a surface, thereby changing the phononic dispersion [@Krummheuer2005].
At first we consider a two-level QD with transition energy $\hbar\omega_\textrm{QD}$ illuminated by a CW laser with frequency $\omega_\textrm{L}$ in a standard photoluminescence excitation experiment. We limit ourselves to detunings below $10\,\textrm{meV}$, where the the deformation potential coupling between electrons and longitudinal acoustic phonons constitutes the dominating coupling mechanism [@Krummheuer2005_2]. Initially we consider spherical-parabolic confinement potentials, giving ground state wavefunctions $$\phi_\nu(\mathbf{r})=\pi^{-3/4}w_\nu^{-3/2}\exp\left[-r^2/(2w_\nu^2)\right], \label{eq:wavef}$$ with $\nu\in \{e,h\}$ representing the electron and the hole respectively. Numerical results for more realistic QD structures will be presented later. The differences in carrier confinement lead to wavefunctions with different effective widths for electrons ($w_e$) and holes ($w_h$). The equations of motion for the system are derived using a quantum master equation approach [@Breuer1999] exploiting recent results [@Kaer2010] for the reduced density matrix, $\rho(t)$, where the phononic degrees of freedom have been traced out. The system evolves according to $$\begin{aligned}
\partial_t\rho(t)=-\textrm{i}\hbar^{-1}[H_\textrm{s},\rho(t)]+S_\textrm{LA}(t)+L(t), \label{eq:reduceddensity}\end{aligned}$$ where $H_\textrm{s}=\hbar(\omega_\textrm{QD}-\omega_\textrm{L})\sigma_{ee}+\hbar\Omega(\sigma_{eg}+\sigma_{ge})$ is the system Hamiltonian for the QD and the optical excitation [^1], corresponding to a Rabi-frequency $\hbar\Omega=10\,\mu\textrm{eV}$. $S_\textrm{LA}(t)$ is a time-local phonon-induced scattering term, and $L(t)$ represents losses included through the Lindblad formalism [@Lindblad1976]. The latter accounts, through the rate $\Gamma$, for the decay of the excited QD state due to spontaneous radiation into other modes than the cavity mode as well as non-radiative processes.
. (**a**) Stationary population of a bare QD, when excited by a CW field, corresponding to a Rabi-frequency $\hbar\Omega=10\,\mu\textrm{eV}$. The black curve describes the case without phonon-assisted coupling, and the QD decay rate is $\Gamma=1\,\textrm{ns}^{-1}$. (**b**) The effective phonon spectrum for spherical wavefunctions for varying $w_h$ and $w_e$. For both plots, $w_e^3+w_h^3$ is kept constant, and $T=4\,\textrm{K}$.](Fig1.eps){width="48.00000%"}
Solving Eq. (\[eq:reduceddensity\]) numerically [^2], we calculate the stationary population of an InAs QD embedded in GaAs in depending on the detuning between the laser and the QD ground state resonance, see Fig. \[PLE\]a. Comparing the black and blue curve in Fig. \[PLE\]a shows that the QD indeed, as is well-known [@Hughes2011_PrX; @HughesNye], may be populated by a phonon-assisted process, even when the laser and the QD are off-resonant. However, when the widths of the electron and hole wavefunction differ, as for the red curve, the QD population and thus the photoluminescence intensity, is suppressed at a specific detuning. As we shall show, this implies that the phonon scattering is quenched, which introduces a new and interesting approach for manipulating the electron-phonon interaction.
The strength of the carrier-phonon scattering is quantified by the effective phonon spectrum, $\mathcal{D}(\omega)$, describing how the phonon modes interact with the QD at a given temperature. Positive phonon frequencies refer to a phonon of energy $\hbar\omega$ being emitted into the surroundings, whereas negative phonon frequencies indicates the absorption of a phonon of energy $\hbar|\omega|$ from the surrounding environment. $\mathcal{D}(\omega)$ is calculated as the real part of the Fourier transform of the phonon reservoir correlation function $\mathcal{D}(t)$ [@Kaer2010]. We have $$\mathcal{D}(\omega)=\pi\sum_\mathbf{k}|M^\mathbf{k}|^2\bigg[n_\mathbf{k}\delta(\omega+\omega_\mathbf{k})+(n_\mathbf{k}+1)\delta(\omega-\omega_\mathbf{k})\bigg],$$ where $\mathbf{k}$ is the phonon wavevector, $n_\mathbf{k}=1/[\exp(\hbar\omega_\mathbf{k}/k_BT)-1]$ is the average thermal occupation number of the phonon mode $\mathbf{k}$ at temperature $T$. $M^\mathbf{k}=M^\mathbf{k}_{ee}-M^\mathbf{k}_{hh}$ is the effective electron/hole-phonon coupling matrix element with $$M^\mathbf{k}_{\nu\nu}=D_\nu\sqrt{\frac{\hbar k}{2dc_lV}}\mathcal{F}_\nu(\mathbf{k}). \label{eq:intmat}$$ The electronic form factor, $$\mathcal{F}_\nu(\mathbf{k})=\int\,\textrm{d}\mathbf{r}|\phi_\nu(\mathbf{r})|^2 \textrm{e}^{\textrm{i}\mathbf{k}\cdot\mathbf{r}}, \label{eq:formfac}$$ describes how a phonon with wavevector $\mathbf{k}$ interacts with the confined carriers, governed by the overlap between the carrier and phonon wavefunctions. We assume bulk phonons with a linear dispersion relation, $\omega_\mathbf{k}=c_l|\mathbf{k}|$. $D_e$ ($D_h$) is the deformation potential constant of a conduction band electron (valence band hole), $c_l$ is the velocity of longitudinal acoustic waves, $d$ is the bulk density, and $V$ is the phonon quantization volume.
For spherical wavefunctions, cf. Eq. (\[eq:wavef\]), the effective phonon spectrum reduces to $$\begin{aligned}
\mathcal{D}(\omega)=\frac{\hbar}{4\pi dc_l^5}\frac{\omega^3}{1-\text{e}^{-\beta \hbar\omega}} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \nonumber\\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \times \left[D_e\text{e}^{-\omega^2w_e^2/(4c_l^2)}-D_h\text{e}^{-\omega^2w_h^2/(4c_l^2)}\right]^2, \label{eq:dspher}\end{aligned}$$ and is shown in Fig. \[PLE\]b. The magnitude of $\mathcal{D}(\omega)$ is generally smaller for $\omega<0$ compared to $\omega>0$ due to the low probability of thermally excited phonons. From Eq. (\[eq:formfac\]) we note that a spatially narrow wavefuntion is wide in $\mathbf{k}$-space and thus interacts with many phonon modes, explaining why $\mathcal{D}(\omega)$ broadens and increases in magnitude as $w_h$ decreases. For $T>0$, $\mathcal{D}(\omega)$ has three zeros at $$\omega=0, \;\;\;\omega^2=\frac{4c_l^2}{w_e^2-w_h^2}\ln\left(\frac{D_e}{D_h}\right). \label{eq:dspectr}$$ In materials where $D_e/D_h>0$, like GaAs [@Hohenester2010; @Krummheuer2002], dips thus appear in the effective phonon spectrum for non-zero phonon energies for $w_e\neq w_h$. The zeros in the effective phonon spectrum appear exactly when $|M^\mathbf{k}|\propto |D_e\mathcal{F}_e(k=\omega/c_l)-D_h\mathcal{F}_h(k=\omega/c_l)|=0$ $\Rightarrow D_e\mathcal{F}_e(k=\omega/c_l)=D_h\mathcal{F}_h(k=\omega/c_l)$.
Based on this we can explain the quenching of phonon-induced processes. The deformation potential interaction occurs in general due to the different values of the deformation potential constant in the conduction and valence band. However, for a specific detuning, this can be compensated by a difference in the confinement of the electron and the hole through the form factor in Eq. (\[eq:formfac\]). In the case of identical electron and hole wavefunctions, $w_e=w_h$, the zeros appear at infinity, and the effect is not apparent.
The quenching of phonon processes has interesting consequences for cavity QED. Instead of illuminating the QD with a laser, the QD is now placed inside a single-mode optical cavity, detuned by $\Delta=\omega_\textrm{QD}-\omega_\textrm{cav}$ from the cavity resonance, see Fig. \[figSketch\]. In this case, $H_\textrm{s}$ in Eq. (\[eq:reduceddensity\]) is replaced by the combined QD-cavity system Hamiltonian (corresponding to replacing $\Omega$ with the optical interaction strength $g$ with the cavity mode), see e.g. [@KaerLang]. Furthermore, the escape of photons through the cavity is included via a rate $\kappa$ in the Lindblad term $L(t)$ in Eq. (\[eq:reduceddensity\]).
. Coupled cavity–QD system interacting with a phonon reservoir. The optical coupling strength is $g$, the QD–cavity detuning is $\Delta$ (positive for the case shown), and $M_{ee/hh}$ are electron/hole–phonon coupling matrix elements. $\Gamma$ is the QD population decay rate, and $\kappa$ is the leakage rate from the optical cavity.](Fig2.eps){width="45.00000%"}
By solving Eq. (\[eq:reduceddensity\]), we obtain the QD decay curves shown in Fig. \[fig1\]a. The structure of the phonon bath results in an asymmetry, expressed in the possibility of the QD to couple to a red-tuned (lower energy, $\Delta>0$) cavity by the emission of an acoustic phonon, but a lack of coupling to a blue-tuned (higher energy, $\Delta<0$) cavity by phonon absorption at low temperatures, where the population of thermally excited phonons is low [@Hohenester2010; @Kaer2010]. This is in accordance with $\mathcal{D}(\omega)$ in Fig. \[PLE\]b. In the limit of $|\Delta|\rightarrow\infty$, the QD and the cavity decouple, and the QD decays with the rate $\Gamma$. By single-exponential fits to the QD decay curves, the lifetime of the excited QD state is calculated, see Fig. \[fig1\]b. From this we determine the lifetime ratio $\tau_{\Delta<0}/\tau_{\Delta>0}$, see Fig. \[fig1\]c. Surprisingly, we observe that when $w_e\neq w_h$, a non-zero QD–cavity detuning exists, where no lifetime asymmetry is present.
. (**a**) Time evolution of the population of an initially excited QD for different detunings. The solid (dashed) curve is for negative (positive) detuning, and the electron and the hole are equally confined, $w_e=w_h=5\,\textrm{nm}$. The temperature is $T = 4 \,\textrm{K}$, and we use $\hbar g=150\,\mu\textrm{eV}$, $\Gamma=1\,\textrm{ns}^{-1}$, and $\hbar\kappa=100\,\mu\textrm{eV}$. (**b**) Lifetimes, $\tau$, of the excited QD state plotted versus detuning for different values of $w_e$ and $w_h$, keeping the volume parameter $w_e^3+w_h^3$ constant. The solid (dashed) curves indicate negative (positive) detunings. (**c**) The degree of asymmetry versus detuning.](Fig3.eps){width="48.00000%"}
To examine the physical origin of the lifetime asymmetry, we consider the total decay rate of the excited QD state in the large detuning limit $\Delta\gg g$ [@KaerLang], $$\Gamma_\text{tot}\approx\Gamma+2g^2\frac{\gamma_\text{tot}}{\gamma_\text{tot}^2+\Delta^2}\left[1+\frac{1}{\hbar^2\gamma_\text{tot}}\mathcal{D}(\omega=\Delta)\right], \label{eq:PersExp}$$ with $\gamma_\text{tot}=(\Gamma+\kappa)/2$. The total QD decay rate, $\Gamma_\textrm{tot}$, has two contributions besides the decay rate $\Gamma$. The first term in the square brackets represents the usual Purcell enhancement rate, while the second term represents the decay of the QD through the cavity by the simultaneous emission/absorption of a phonon. Thus the behaviour of $\mathcal{D}(\omega)$ translates directly into the behaviour of the lifetime ratio plot in Fig. \[fig1\]c, such that non-zero QD-cavity detunings exist, where phonons do not affect the lifetime of the excited QD state.
. (**a**) Time evolution of phonon-induced dephasing rate for different QD-cavity detunings. Inset show the long-time value of the dephasing rate versus detunings, $w_h=3\,\textrm{nm}$, $w_e=6\,\textrm{nm}$, and $T=4\,\textrm{K}$. (**b**) Time evolution of phonon-induced dephasing rate for different wavefunction widths, keeping $w_e^3+w_h^3$ constant, and $\Delta=1\,\textrm{meV}$. The inset quantifies the short-time scattering by showing the integrated value of $\gamma_\textrm{ph}(t)$ from $t=0$ to 3 ps for different wavefunction widths.](Fig4.eps){width="49.00000%"}
The suppression of the effective phonon density is expected to affect not only the QD lifetime, but all phonon-induced effects such as pure dephasing. The degree of single-photon indistinguishability [@Kiraz2004] is typically quantified as the ratio between the coherence time and lifetime of the emitter [@Bylander2003]. For our system, this ratio is $\Gamma_\textrm{tot}/(\Gamma_\textrm{tot}+2\gamma_\textrm{ph})$, with $\gamma_\textrm{ph}$ being the pure dephasing contribution from the phonons. This hints at the importance of having a high decay rate compared to the dephasing rate.
The phonon-induced pure dephasing rate [^3] is calculated from the phonon bath correlation function, $$\begin{aligned}
\gamma_\textrm{ph}(t)=\textrm{Re}\bigg\{\hbar^{-2}(1-K)\int_0^t\mathrm{d}t'\,\mathcal{D}(t')\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\nonumber\\
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;+\hbar^{-2}K\int_0^t\mathrm{d}t'\,\cos\left(t'\sqrt{\Delta^2+4g^2}\right)\mathcal{D}(t')\bigg\}, \label{eq:inds}\end{aligned}$$ where $K$ is a parameter depending solely on the ratio $\Delta/g$. For bulk phonons, $\mathcal{D}(t)$ decays within 5 ps [@KaerLang]. As seen from Fig. \[fig444\_ny\]a, the initial behaviour of $\gamma_\textrm{ph}(t)$ is governed by the bare QD-phonon coupling, whereas the long-time value depends on the cavity detuning. On the short time scale, the phonons participate in non-energy conserving virtual processes. Thus the electrons interact with the full phonon spectrum $\mathcal{D}(\omega)$. The long-time ($t\rightarrow\infty$) value of $\gamma_\textrm{ph}(t)$, on the other hand, implies a Fourier transform of $\mathcal{D}(t)$ and depends only on $\mathcal{D}(\omega=0)$ and $\mathcal{D}(\omega=\pm\sqrt{\Delta^2+4g^2})$, where the former is zero. The phonon-induced dephasing of the QD–cavity system may be reduced by minimizing both the short-time and long-time scattering. The former may be achieved by reducing the overall amplitude of $\mathcal{D}(\omega)$, as illustrated in Fig. \[fig444\_ny\]b for $w_h\sim 3.4\,\textrm{nm}$, compare with Fig. \[PLE\]b. The reduction of long-time scattering is obtained by using a QD–cavity detuning such that $\sqrt{\Delta^2+4g^2}$ coincides with a dip in $\mathcal{D}(\omega)$, as seen from the inset in Fig. \[fig444\_ny\]a. It turns out that for the indistinguishability of single photons, the reduction of the overall amplitude is the most important [@KaerIndist].
QD structures realized by epitaxial growth are not spherically symmetric [@Stobbe2009], and to investigate the quenching for more realistic structures, we model truncated conical QD structures by solving the one-band effective mass Schrödinger equation using the Finite Element Method [@Melnik2004].
The effective phonon spectrum for the FEM wavefunctions is shown in Fig. \[fig6\]a for an InAs QD and wetting layer embedded in a GaAs barrier [^4]. The overall amplitude of the spectrum is the largest, when the carriers are strongly confined in two directions, i.e. for the tallest dot, compared to the shallow dot where the carriers are only strongly confined in one direction. These results are qualitatively explained by the form factor in Eq. (\[eq:formfac\]).
. (**a**) The effective phonon spectrum calculated from the FEM wavefunctions of the three QDs shown in the insets with matching colors, $T=4\,\textrm{K}$. All QDs have constant volume and side wall slope. (**b**) The effective phonon spectrum for the shallow dot in (a) for different material compositions, In$_y$Ga$_{1-y}$As, of the QD.](Fig5.eps){width="48.50000%"}
To examine the role of material composition, we consider an In$_y$Ga$_{1-y}$As QD, where the amount of gallium in the QD and wetting layer is varied. By increasing the gallium concentration the band offsets shrink [@Vurgaftman2001] and the effective mass changes, which results in the phonon spectrum in Fig. \[fig6\]b. Thus, $w_e$ and $w_h$ both increase, resulting in a lower overall amplitude of the effective phonon spectrum. Due to the heavier hole mass, the asymmetry between the electron and hole wavefunctions also increases, moving the dip in the spectrum towards lower frequencies, as predicted by the spherical model in Eq. (\[eq:dspectr\]) when $w_e^2-w_h^2$ increases. The position of the dips may in this way be changed depending on the growth parameters of the QD. For small gallium concentrations, the confinement energies may become comparable to the Coulomb energy, in which case the exitonic nature of the electron–hole pair will have to be taken into account.
In conclusion, through the analysis of a comprehensive model, we have learned that electron-phonon scattering may be suppressed due to differences in the spatial confinement of electrons and holes. We suggest that the effect may be measured by photoluminescence excitation spectroscopy, but also that it should strongly affect the decoherence properties of cavity QED systems. We expect this phonon-quenching to be important for the engineering of the phononic properties of devices for quantum information technology, like single-photon sources and switches.
The authors acknowledge helpful discussions with Jørn M. Hvam and financial support from the Villum foundation via the NATEC Centre of Excellence.
[34]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, , , , , ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , , , , ****, ().
, , , ****, ().
, , , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , , , , ().
, , , , ().
, , , , , ****, ().
, ****, ().
, ****, ().
[^1]: The system Hamiltonian is similar to the coupled QD-cavity system Hamiltonian in [@KaerLang] with $g\rightarrow \Omega$
[^2]: We consider parameters for GaAs, with deformation potentials $D_e=-14.6\,\textrm{eV}$, $D_h=-4.8\,\textrm{eV}$ [@Hohenester2010; @Krummheuer2002], $c_l=5110\,\textrm{m/s}$, and $d=5370\,\textrm{kg/m}^3$.
[^3]: $\gamma_\textrm{ph}$ corresponds to $\textrm{Re}\{\gamma_{12}\}$ in [@Kaer2010].
[^4]: We use $E_{g,\text{GaAs}}=1.424\,\text{eV}$, $E_{g,\text{InAs}}=0.359\,\text{eV}$, $m^*_{c,\text{GaAs}}=0.0665m_0$, $m^*_{c,\text{InAs}}=0.027m_0$, $m^*_{v,\text{GaAs}}=0.38m_0$, and $m^*_{v,\text{InAs}}=0.34m_0$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The motion of domain walls in ferromagnetic, cylindrical nanowires is investigated numerically by solving the Landau-Lifshitz-Gilbert equation for a classical spin model in which energy contributions from exchange, crystalline anisotropy, dipole-dipole interaction, and a driving magnetic field are considered. Depending on the diameter, either transverse domain walls or vortex walls are found. The transverse domain wall is observed for diameters smaller than the exchange length of the given material. Here, the system behaves effectively one-dimensional and the domain wall mobility agrees with a result derived for a one-dimensional wall by Slonczewski. For low damping the domain wall mobility decreases with decreasing damping constant. With increasing diameter, a crossover to a vortex wall sets in which enhances the domain wall mobility drastically. For a vortex wall the domain wall mobility is described by the Walker-formula, with a domain wall width depending on the diameter of the wire. The main difference is the dependence on damping: for a vortex wall the domain wall mobility can be drastically increased for small values of the damping constant up to a factor of $1/\alpha^2$.'
author:
- 'R. Wieser, U. Nowak and K. D. Usadel'
title: 'Domain wall mobility in nanowires: transverse versus vortex walls'
---
Arrays of magnetic nanowires are possible candidates for patterned magnetic storage media [@rossPRB00; @nielschJMMM02]. For these nanowires and also for other future magneto-electronic devices the understanding of domain wall motion and mobility is important for the controlled switching of the nanostructure. In a recent experiment the velocity of a domain wall in a NiFe/Cu/NiFe trilayer was investigated using the GMR effect [@onoSCIENCE99]. The measured velocities were compared with the Landau-Lifshitz formula for domain wall motion [@landauPZS35]. This comparison was used to determine the damping constant of the trilayer, a quantity which is usually not known a priori. However, several formulas for the velocity of a domain wall can be found in the literature [@landauPZS35; @malozemoffBOOK79; @schryerJAP74; @dillonBOOK63; @garaninPA91b] which are derived in different limits and all in (quasi) one-dimensional models neglecting the possible influence of non-uniform spin structures within the domain wall. Thus the question arises in how far these formulas are applicable to real three dimensional domain structures. To shed some light onto this problem we numerically investigate the domain wall mobility in nanowires starting from a three dimensional local spin model.
In the following we consider a classical spin model with energy contributions from exchange, crystalline anisotropy, dipole-dipole interaction, and a driving magnetic field. Such a spin model for the description of magnetic nanostructures [@nowakARCP01] can be justified following different lines: on the one hand it is the classical limit of a quantum mechanical, localized spin model, on the other hand it might be interpreted as the discretized version of a micromagnetic continuum model, where the charge distribution for a single cell of the discretized lattice is approximated by a point dipole. For certain magnetic systems their description in terms of a lattice of magnetic moments may even be based on the mesoscopic structure of the material, especially when a particulate medium is described.
However, our intention is not to describe a particular material but to investigate a general model Hamiltonian which is $$\begin{aligned}
\label{H}
{\cal H} = &-& J \sum\limits_{\langle ij \rangle} {\mathbf S}_i
\cdot {\mathbf S}_j
- \mu_s {\mathbf B} \cdot \sum\limits_i {\mathbf S}_i
- D_e \sum\limits_i (S_i^z)^2 \nonumber \\
& - & \omega \sum\limits_{i<j} \frac{3({\mathbf S}_i \cdot {\mathbf
e}_{ij})({\mathbf e}_{ij} \cdot {\mathbf S}_j) - {\mathbf S}_i
\cdot {\mathbf S}_j}{r^3_{ij}},\end{aligned}$$ where the ${\mathbf S}_i = {\boldsymbol \mu}_i/\mu_s$ are three dimensional magnetic moments of unit length on a cubic lattice.
The first sum is the ferromagnetic exchange between nearest neighbors with coupling constant $J$. The second sum is the coupling of the spins to an external magnetic field $B$, the third sum represents a uniaxial anisotropy, here, with $D_e > 0$, favoring the $z$ axis as easy axis of the system, and the last sum is the dipolar interaction where $w = \mu_0 \mu_s^2 /(4 \pi a^3)$ describes the strength of the dipole-dipole interaction. The ${\mathbf e}_{ij}$ are unit vectors pointing from lattice site $i$ to $j$ and $r_{ij}$ is the distance between these lattice sites in units of the lattice constant $a$.
The underlying equation of motion for magnetic moments which we consider in the following is the Landau-Lifshitz-Gilbert (LLG) equation, $$\frac{\partial {\mathbf S}_i}{\partial t} =
- \frac{\gamma}{(1+\alpha^2)\mu_s}
{\mathbf S}_i \times \Big[{\mathbf H}_i(t) +
\alpha \big ({\mathbf S}_i \times {\mathbf H}_i(t) \big) \Big],
\label{e:llg}$$ with the gyromagnetic ratio $\gamma = 1.76 \times 10^{11} (\mathrm{
Ts})^{-1}$, the dimensionless damping constant $\alpha$ (after Gilbert), and the internal field $ {\mathbf H}_i(t) = - \partial {\cal
H} /\partial {\mathbf S}_i$.
We simulate cylindrical systems being parallel to the $z$-axis with a typical length of 256 lattice sites and different diameters $d$. Due to shape as well as crystalline anisotropy the equilibrium magnetization is aligned with the long axis of the system. However, we start the simulation with an abrupt, head-to-head domain wall as initial configuration, letting the wall relax until a stable state is reached. The distance of the initial wall position from the end is approximately 1/3 of the system length. Then we switch on the driving magnetic field $B$ along the easy axis and wait until a stationary state is reached for some time interval in which the velocity $v$ of the wall is constant while the wall is moving through the central part of the wire. We calculate the domain wall velocity from the magnetization versus time data, averaged over a period of time where no influence of the finite system size on the domain wall can be observed, i. e. until the wall approaches the other end of the wire.
Inspection of the stationary state of the moving domain wall shows that, depending on the ratio $\omega/J$, either transverse domain walls or vortex walls are found. Representative spin configurations are shown in Fig. \[f:snaps\]. The transverse domain wall (left hand side) is observed for diameters smaller than the exchange length $d_{ex}/a = \pi \sqrt{J/(6 \omega \zeta(3))}$ of the system [@hinzkeJMMM00] where $\zeta(3) \approx 1.2$ is Riemann’s Zeta-function (see also [@hubertBOOK98] for the exchange length in continuum theory where $3 \zeta(3)$ is replaced by $\pi$). Here all spins within cross-sectional planes perpendicular to the wire axis are parallel so that the system is effectively one dimensional. Note, that the spin precession leads to a rotation of the spin direction within the wall while it is moving.
![Snapshots of a transverse (left, $\omega/J = 0.003$) and a vortex (right, $\omega/J = 0.2$) wall. The diameter $d=8$ is kept constant while the exchange length of the system is varied. Shown is only a part of the system below the current wall position. $D_e/J = 0.05$.[]{data-label="f:snaps"}](transverse.ps "fig:"){height="8cm"} ![Snapshots of a transverse (left, $\omega/J = 0.003$) and a vortex (right, $\omega/J = 0.2$) wall. The diameter $d=8$ is kept constant while the exchange length of the system is varied. Shown is only a part of the system below the current wall position. $D_e/J = 0.05$.[]{data-label="f:snaps"}](vortex.ps "fig:"){height="8cm"}
With increasing dipolar interaction, a crossover to a vortex wall sets in (right hand side of Fig. \[f:snaps\]) which is now energetically favorable since the vortex structure leads to a flux closure. These findings are in agreement with corresponding spin model simulations of thermally activated reversal [@hinzkeJMMM00] and micromagnetic results [@forsterJAP02; @hertelJMMM02] obtained from simulations of the LLG equation using a micromagnetic continuum model.
In the following we turn to the investigation of the influence of the domain wall width and structure on its velocity. Fig. \[f:vpropd\] compares the dependence of the domain wall width $\Delta$ and domain wall velocity $v$ on the strength of the dipolar coupling. Here, the domain wall width was determined numerically by fitting an $\tanh$-profile to the easy axis magnetization of the moving wall in the stationary state where the magnetization is averaged over cross-sectional planes. However, it should be mentioned that for large dipolar interaction in a vortex wall the wall profile cannot accurately be described by a simple tanh-profile. Note, that even in the limit $\omega \to 0$ the wall is stabilized by the additional crystalline anisotropy $D_e$.
For a spin chain ($d = 1$) the domain wall is necessarily always planar while for the system with larger diameter a crossover to a vortex wall occurs. The crossover can be identified as a jump of the domain wall velocity for the $d=8$ data. Fig. \[f:vpropd\] demonstrates that for a transverse wall the domain wall velocity is proportional to the wall width. For a vortex wall this is at least qualitatively the case. Width and velocity of transverse walls decrease with increasing dipolar interaction while for vortex walls the opposite is true. The crossover itself leads to a jump of the wall velocity not the wall width.
![Domain wall velocity and domain wall width versus dipolar coupling for a spin chain and a cylindrical system, respectively. $D_e/J = 0.05$, $\mu_s B/J = 0.05$, $\alpha = 1$. The solid lines correspond to Eqs. \[e:wall\] and \[e:slon\], error bars are smaller than the symbol size.[]{data-label="f:vpropd"}](DeltaUndVuW.eps){width="8cm"}
For a transverse wall the domain wall velocity is well described by an equation derived by Slonczewski as the lower limit for one-dimensional domain wall motion [@malozemoffBOOK79], $$ v = \frac{\gamma}{\alpha + 1/\alpha} \Delta_{\mathrm B} B.
\label{e:slon}$$ Here, $\Delta_B$ is the well-known Bloch wall width $$\Delta_{\mathrm B} = a \sqrt{\frac{J}{2 \left(D_e + 3 \omega
\zeta(3)\right)}},
\label{e:wall}$$ where, for our case, the denominator $D_e + 3 \omega \zeta(3)$ estimates the effective anisotropy coming from shape as well as crystalline contributions (as before in a continuum theory $3
\zeta(3)$ is replaced by $\pi$). Both equations above are drawn in Fig. \[f:vpropd\] as solid lines and they agree very well with the numerical data for transverse walls.
In the following we focus on the mobility of vortex walls. The crossover from transverse to vortex wall can also be observed while varying the diameter of the system keeping the exchange length constant. Since for sufficiently small driving fields the domain wall velocity is proportional to the field in Fig. \[f:mobility\] we directly show the domain wall mobility ${\mathrm d} v / {\mathrm d} B$ versus diameter of the system. Obviously there are two distinct regions with distinct wall mobility behavior. For low diameters where the transverse wall is found the system behaves effectively one-dimensional and in the limit $d \to 1$ the domain wall mobility follows Eq. \[e:slon\]. With increasing diameter the observed width of the transverse domain wall increases little due to the dipolar interaction leading to small deviations from the analytic Slonczewski result assuming a Bloch wall width. Nevertheless, we confirmed numerically, that Eq. \[e:slon\] is still valid when the Bloch wall width $\Delta_{\mathrm B}$ is replaced by the actual (numerically determined) width of the transverse wall.
![Domain wall mobility versus diameter of cylindrical systems. The model parameters are $\omega/J = 0.003$, $D_e/J =
0.1$, $\alpha = 1$. The constant line corresponds to Eq. \[e:slon\], the dashed lines are guides to the eye, error bars are smaller than the symbol size.[]{data-label="f:mobility"}](BewUdE.eps){width="8cm"}
Increasing the diameter of the system a crossover from transverse to vortex wall is observed with a drastic increase of the domain wall mobility. As seen in Fig. \[f:vpropd\], the reason for this effect is not a comparably drastic change of the wall width. Instead, as we will discuss in the following for a vortex wall the domain wall mobility follows a law with a different dependence on the damping constant, namely the Landau-Lifshitz formula [@landauPZS35], $$ \frac{{\mathrm d} v}{{\mathrm d} B} = \frac{\gamma}{\alpha} \Delta,
\label{e:ll}$$ where in our case $\Delta$ is the actual domain wall width of the vortex wall. The equation above is a limit of the more general Walker equation [@schryerJAP74; @dillonBOOK63; @garaninPA91b], $$ v = \frac{\gamma B a}{\alpha}\sqrt{\frac{J}{2(D_e + D_h
\sin^2\phi)}},
\label{e:walker}$$ which was derived for sufficiently small driving fields for a system with an additional hard-axis anisotropy $D_h$. This anisotropy forces the equilibrium magnetization into an easy plane. Walkers formula is valid as long as the spin motion takes place in one plane which is defined by a constant angle $\phi$ to the easy plane of the system. $\phi$ is given as $$\sin \phi \cos \phi = \frac{\mu_s B}{\alpha 2 D_h},
\label{e:phi}$$ where this equation also defines a condition for the validity of the walker formula. For a given $\alpha$ there exists a maximum field value (or vice versa for a given field a minimum $\alpha$ value) beyond which the spin motion is no longer restricted to one plane and instead an irregular precessional motion starts [@schryerJAP74]. Note, that the Landau-Lifshitz formula is the $\phi = 0$ limit of the Walker equation, i. e. the limit of a strong hard axis anisotropy which forces the spin motion into the easy plane.
![Two different paths for the reversal of a spin. While (2) is dominated by precession, as in a transverse wall, the path (1) is restricted to one plane as in a vortex wall.[]{data-label="f:paths"}](Ummag_grau.ps){width="5cm"}
The equations above were derived for one dimensional systems and the question arises why these formula should be valid for the motion of a vortex wall with a non-uniform spin structure in cross-sectional planes. For a qualitative understanding we note that the motion of the spins within each spin chain which is parallel to the wire axis is indeed restricted to a certain plane passing through the spins positions. For a spin chain at the surface of the cylinder and in the limit of small driving fields these are tangential planes of the cylinder surface. The responsible force which keeps the spin motion of each chain in this plane is for a vortex wall not a hard axis anisotropy — as in the original calculation — but the energetical principle which forms the vortex, i. e., the combination of exchange and dipolar interaction. Since this is the condition under which Walkers formula was derived it seems to be plausible that Eq. \[e:walker\] describes the wall mobility in the case of an extended spin system as long as the spin motion during the reversal takes place in one plane. For a transverse wall, on the other hand, the situation is different: the precession of the wall leads to the fact that the motion of each single spin consists of precession and relaxation with no restriction to one single plane. These two different paths for the reversal of a spin are sketched in Fig. \[f:paths\].
The main difference between Eqs. \[e:slon\] and \[e:ll\] is the dependence on damping. This is demonstrated in Fig. \[f:damping\] which shows the ratio of the calculated wall mobility and the numerically determined domain wall width for two different strengths of dipolar interaction leading to the two different wall shapes. In the high damping limit both formulas agree. For a transverse wall the mobility shows a maximum at $\alpha = 1$ and for lower damping the domain wall mobility decreases with decreasing damping constant. In the limit $\alpha \to 0$ only a precession of the domain wall remains without an effective wall motion along the wire.
![Reduced domain wall mobility versus damping constant. The model parameters are $\omega/J = 0.01$ (transverse wall) and $\omega/J =
0.7$ (vortex wall) respectively, $D_e/J = 0.05$, $\mu_s B/J =
0.05$, $d = 4$. Error bars are smaller than the symbol size.[]{data-label="f:damping"}](vua2.eps){width="8cm"}
For a vortex wall the domain wall mobility increases with decreasing damping constant following a $1/\alpha$ law as long as one is above a critical value $\alpha_c$. As was discussed in connection with Eq. \[e:phi\] this value $\alpha_c$ sets the limit of pure relaxational spin motion. As was discussed before, the role of the hard axis anisotropy $D_h$ in Eq. \[e:phi\] in our case is played by the combination of exchange and dipolar interaction which forms the vortex and forces the spin motion into one plane. We would like to stress that for experimental systems the low damping limit is more relevant. Here, the difference between the two domain wall mobilities (reduced to the domain wall width) can be extremely large, up to a factor of $1/\alpha^2$.
For smaller values of $\alpha$ below the critical one the mobility decreases again and finally converges to a finite value since even for $\alpha = 0$ the wall can move. In this limit the LLG equation conserves the energy of the system, and lowering the Zeeman energy leads to an increase of exchange energy, leaving an excited spin system behind the wall. These observations are also in agreement with the calculations of Walker [@schryerJAP74].
\
This effect is demonstrated in Fig. \[f:winding\]. Here, profiles of the moving walls are shown as well as the so-called winding number $$n = \frac{1}{2 \pi R} \int ({\rm rot}\underline S)_zdxdy$$ which is calculated numerically over all perpendicular planes of the wire. The winding number is a measure for the existence of vortices. $n=1$ means that all spins along the boundary are aligned building a ring with perfect flux closure. Fig. \[f:winding\] a) shows the high damping limit with a perfect vortex in the center of the wall. In the very low-damping limit, Fig. \[f:winding\] b) the situation is much more complicated. Here, behind the moving wall (smaller $z$) an excited spin system is left with vortex-type spin waves which are ejected from the moving wall.
To conclude, in agreement with prior work [@hinzkeJMMM00; @forsterJAP02; @hertelJMMM02] we have found different wall structures for driven domain walls in cylindrical systems, transverse and vortex walls, depending on the diameter of the system as compared to the exchange length of the given material. While for vortex walls the domain wall velocity is described by the formula from Walker, transverse walls follow a formula from Slonczewski. In both cases the domain wall velocity is proportional to the domain wall width. The main difference is the dependence on the damping constant. For small values of the damping constant this difference can lead to drastic differences where the velocity of the vortex wall is up to a factor of $1/\alpha^2$ larger. The reason for this difference is probably the fact that each spins motion in the case of the vortex wall is completely within one single plane as it is the case for the model where the Walker formula was derived for, while this is not the case for a transverse wall where the precession of the wall leads to a three dimensional spin motion.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank D. Garanin and S. Lübeck for helpful discussions. This work has been supported by the Deutsche Forschungsgemeinschaft (SFB 491 and NO290).
[10]{}
C. A. Ross, R. W. Chantrell, M. Hwang, M. Farhoud, T. A. Savas, Y. Hao, H. I. Smith, F. M. Ross, M. Redjdal, and F. B. Humphrey, Phys. Rev. B [**62**]{}, 14252 (2000).
K. Nielsch, R. B. Wehrspohn, J. Barthel, J. Kirschner, S. F. Fischer, H. Kronmüller, T. Schweinböck, D. Weiss, and U. Gösele, J. Magn. Magn. Mat. [**249**]{}, 234 (2002).
T. Ono, H. Miyajima, K. Shigeto, K. Mibu, N. Hosoito, and T. Shinjo, Science [**284**]{}, 468 (1999).
D. L. Landau and E. Lifshitz, Phys. Z. Sowjetunion [**8**]{}, 153 (1935).
A. P. Malozemoff and J. C. Slonczewski, [*Magnetic Domain Walls in Bubble Materials*]{} (Academic Press, New York, 1979).
and [L.R. Walker]{}, J. Appl. Phys. [**45**]{}, 5406 (1974).
J. F. Dillon, in [*Magnetism*]{}, edited by G. T. Rado and H. Suhl (Academic Press, New York, 1963), Vol. 1, p. 149.
, Physica A [**178**]{}, 467 (1991).
U. Nowak, in [*Annual Reviews of Computational Physics IX*]{}, edited by D. Stauffer (World Scientific, Singapore, 2001), p. 105.
D. Hinzke and U. Nowak, J. Magn. Magn. Mat. [**221**]{}, 365 (2000).
A. Hubert and R. Schäfer, [*Magnetic Domains*]{} (Springer-Verlag, Berlin, 1998).
H. Forster, T. Schrefl, D. Suess, W. Scholz, V. Tsiantos, R. Dittrich, and J. Fidler, J. Appl. Phys. [**91**]{}, 6914 (2002).
R. Hertel, J. Magn. Magn. Mat. [**249**]{}, 251 (2002).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove that, at every positive temperature, the infinite-volume free energy of the one dimensional log-gas, or beta-ensemble, has a unique minimiser, which is the Sine-beta process arising from random matrix theory. We rely on a quantitative displacement convexity argument at the level of point processes, and on the screening procedure introduced by Sandier-Serfaty.'
address:
- 'Institut für angewandte Mathematik, Universität Bonn, Endenicher Allee 60, 53115 Bonn, Germany'
- 'Universität Wien, Oskar-Morgenstern-Platz 1, 1090 Wien, Austria'
- 'Courant Institute of Mathematical Sciences, 251 Mercer Street, New York University, New York, NY 10012-1110, USA'
author:
- 'Matthias Erbar, Martin Huesmann, Thomas Leblé'
bibliography:
- 'uniqueness.bib'
title: 'The one-dimensional log-gas free energy has a unique minimiser'
---
[^1]
Introduction and main result
============================
The one-dimensional log-gas
---------------------------
The one-dimensional log-gas in finite volume can be defined as a system of particles interacting through a repulsive pairwise potential proportional to the logarithm of the distance, and confined by some external field.
For a fixed value of $\beta > 0$, called the *inverse temperature* parameter, and for $N \geq 1$, we consider the probability measure $\PNbeta$ on $\XN = (x_1, \dots, x_N) \in \R^N$ defined by the density $$\label{def:PNbeta}
d\PNbeta(\XN) := \frac{1}{\ZNbeta} \exp\left( - \beta \left( \sum_{i < j} - \log |x_i - x_j| + \sum_{i=1}^N N \frac{x_i^2}{2} \right) \right),$$ with respect to the Lebesgue measure on $\R^N$. The quantity $\ZNbeta$ is a normalization constant, the *partition function*. We call $\PNbeta$ the *canonical Gibbs measure* of the log-gas.
Part of the motivation for studying log-gases comes from Random Matrix Theory (RMT), for which $\PNbeta$ describes the joint law of $N$ eigenvalues in certain classical models: the Gaussian orthogonal, unitary, symplectic ensemble respectively for $\beta = 1,2,4$, and the “tridiagonal model” discovered in [@dumitriu2002matrix] for arbitrary $\beta$. We refer to the book [@forrester2010log] for a comprehensive presentation of the connection between log-gases and random matrices. Log-gases are also interesting from a statistical physics point of view, as a toy model with singular, long-range interaction.
Questions about such systems usually deal with the large $N$ limit (also called thermodynamic, or infinite-volume limit) of the system, as encoded by certain observables. For example, in order to understand the “global” behavior, one may look at the empirical measure $\frac{1}{N} \sum_{i=1}^N \delta_{x_i}$, and asks about the typical behavior of this *random* probability measure on $\R$ as $N$ tends to infinity. By now, this is fairly well understood, we refer e.g. to the recent lecture notes [@serfaty2017microscopic] and the references therein. In the present paper, we are rather interested in the asymptotic behavior at *microscopic* scale.
The Sine-beta process
---------------------
Let ${\mathcal{C}}_{N,0}$ be the point configuration $${\mathcal{C}}_{N, 0} := \sum_{i=1}^N \delta_{N x_i},$$ where $\XN = (x_1, \dots, x_N)$ is distributed according to $\PNbeta$. The limit in law of ${\mathcal{C}}_{N,0}$ as $N \to \infty$ was constructed in [@valko2009continuum] and named the ${\mathrm{Sine}_{\beta}}$ process. We refer to [@killip2009eigenvalue] for a different construction of a process that turns out to be same, and to [@valko2017sine; @valko2017operator] for recent developments concerning ${\mathrm{Sine}_{\beta}}$. This process is the universal behavior of log-gases (in the bulk), in the sense that replacing the $\frac{x_i^2}{2}$ term in by a general potential $V(x_i)$ yields the same microscopic limit, up to a scaling on the average density of points (our convention is that ${\mathrm{Sine}_{\beta}}$ has intensity $1$) and mild assumptions on $V$, see [@Bourgade1d1; @Bourgade1d2].
In [@dereudre2018dlr], a different description of ${\mathrm{Sine}_{\beta}}$ is given using the Dobrushin-Landford-Ruelle (DLR) formalism, but the question of whether ${\mathrm{Sine}_{\beta}}$ is the unique solution to DLR equations is left open. The main result of the present paper answers positively to a slightly different uniqueness question, phrased in terms of the log-gas free energy.
The log-gas free energy
-----------------------
In [@leble2017large], the infinite-volume free energy of the log-gas (and of other related systems) was introduced as the weighted sum ${\mathcal{F}_{\beta}}:= \beta {\mathcal{W}}+ {\mathcal{E}}$, where the functionals ${\mathcal{W}}, {\mathcal{E}}$ and the free energy ${\mathcal{F}_{\beta}}$ are defined on the space of stationary random point processes. The functional ${\mathcal{W}}$ corresponds to the “renormalized energy” introduced in [@sandier2012ginzburg], and adapted to this context in [@sandier20151d; @petrache2017next], and ${\mathcal{E}}$ is the usual *specific relative entropy*. Both terms are defined below, see Section \[sec:defoffree\].
The free energy ${\mathcal{F}_{\beta}}$ appears in [@leble2017large] as the rate function for a large deviation principle concerning the behavior of log-gases at the microscopic level. If $\XN = (x_1, \dots, x_N)$ is an $N$-tuple of particles distributed according to the Gibbs measure of a log-gas, they are known to typically arrange themselves on an interval approximately given by $[-2, 2]$. For $\xx$ in this interval, we let ${\mathcal{C}}_{N, \xx}$ be the point configuration $(x_1, \dots, x_N)$ “seen from $\xx$”, namely $
{\mathcal{C}}_{N, \xx} := \sum_{i=1}^N \delta_{N(x_i - \xx)}$. We may then consider the *empirical field* of the system in the state $\XN$, defined as $$\Emp_N(\XN) := \frac{1}{4}
\int_{-2}^{2} \delta_{{\mathcal{C}}_{N, \xx}} d \xx.$$ The empirical field $\Emp_N(\XN)$ is a probability measure on (finite) point configurations in $\R$, and it was proven in [@leble2017large] that its law satisfies a large deviation principle, at speed $N$, with a rate function built using ${\mathcal{F}_{\beta}}$. We refer to the paper cited above for a precise statement, here it suffices to say that *understanding the minimisers* of ${\mathcal{F}_{\beta}}$ gives an *understanding of the typical microscopic behavior* of a finite $N$ log-gas at temperature $\beta$, when $N$ is large.
For any $\beta$ in $(0, +\infty)$, the functional ${\mathcal{F}_{\beta}}$ is known to be lower semi-continuous, with compact sub-level sets. In particular, it admits a compact subset of minimisers. However, the question of uniqueness of minimisers for ${\mathcal{F}_{\beta}}$ remained open, and we address it in this paper.
Main result
-----------
\[theo:main\] For any $\beta$ in $(0, + \infty)$, the free energy ${\mathcal{F}_{\beta}}$ has a unique minimiser.
Since it was proven in [@leble2017large Corollary 1.2] that ${\mathrm{Sine}_{\beta}}$ minimises ${\mathcal{F}_{\beta}}$, we deduce that:
\[coro:sinebeta\] For any $\beta$ in $(0, + \infty)$, the ${\mathrm{Sine}_{\beta}}$ process is the unique minimiser of ${\mathcal{F}_{\beta}}$.
This provides a variational characterization of ${\mathrm{Sine}_{\beta}}$. A weaker variational property of ${\mathrm{Sine}_{\beta}}$ had been used e.g. in [@leble2016logarithmic] to prove the convergence of ${\mathrm{Sine}_{\beta}}$ to a Poisson point process as $\beta \to 0$, retrieving a result of [@allez2014sine].
Elements of proof and plan of the paper
---------------------------------------
The proof of Theorem \[theo:main\] goes by contradiction. We assume that ${\mathcal{F}_{\beta}}$ admits two distinct minimisers, and we construct a stationary point process whose free energy is stricly less than $\min {\mathcal{F}_{\beta}}$, which is of course absurd. Since the free energy ${\mathcal{F}_{\beta}}$ is affine, it is not strictly convex for the usual linear interpolation of probability measures. We use instead the notion of *displacement convexity*, which was introduced in [@MCCANN1997153] to remedy situations where energy functionals are not convex in the usual sense. This idea was suggested to us by Alice Guionnet, and we warmly thank her for her insight.
### Strategy of the proof {#strategy-of-the-proof .unnumbered}
The proof goes by contradiction. We start with two stationary point processes $\Pz, \Pu$ such that $\Pz \neq \Pu$, and assume that both are minimisers of ${\mathcal{F}_{\beta}}$. We cannot argue via displacement convexity directly on the level of $\Pz, \Pu$ since they are probability measures on *infinite* point configurations. Optimal transport theory for random stationary measures as initiated in [@HuSt13; @Hu16; @ErHu15] is not yet developed well enough to be directly applicable. Instead, we use transport theory between *finite* measures together with a careful approximation argument relying on screening of electric fields. More precisely, we write $${\mathcal{F}_{\beta}}({\mathsf{P}}) = \lim_{R \to \infty} \frac{1}{|{\Lambda}_R|} \left(\beta {\mathcal{W}}_R({\mathsf{P}}) + {\mathcal{E}}_R({\mathsf{P}}) \right),$$ where ${\mathcal{W}}_R, {\mathcal{E}}_R$ are quantities (the energy, and the relative entropy) depending on the restriction of ${\mathsf{P}}$ to the line segment ${\Lambda}_R := [-R, R]$.
1. Let $\Pz_R, \Pu_R$ be the restriction of $\Pz, \Pu$ to ${\Lambda}_R$. Assume that there are almost surely $2R$ points in $[-R,R]$. We may thus see $\Pz_R, \Pu_R$ as probability measures on $[-R, R]^{2R}$, apply classical results about optimal transportation, and find an optimal transport map $\Trans_R$ which pushes $\Pz_R$ onto $\Pu_R$.
In fact, it is not true that $\Pz, \Pu$ have almost surely $2R$ points in ${\Lambda}_R$, so we first perform a version of the “screening procedure” of Sandier-Serfaty. It has for effect to modify the configurations near the extremities of $[-R, R]$ in order to enforce the correct number of points, while not changing too much the average energy, nor the entropy. The screening procedure requires some conditions in order to be applied, we will need to guarantee that they are verified with high probability under $\Pz, \Pu$.
2. We let $\Phal_R$ be the half-interpolate of $\Pz_R, \Pu_R$ along the displacement $\Trans_R$, i.e. the push-forward of $\Pz_R$ by ${\frac{1}{2}}(\Id + \Trans_R)$. A result of [@MCCANN1997153] ensures that the relative entropy is displacement convex, so $${\mathcal{E}}_R[\Phal_R] \leq {\frac{1}{2}}\left( {\mathcal{E}}_R(\Pz) + {\mathcal{E}}_R(\Pu)\right).$$ Moreover, the interaction potential $-\log |x-y|$ is strictly convex, so again by a result of [@MCCANN1997153], the energy ${\mathcal{W}}_R$ is also displacement convex. More precisely, we have $${\mathcal{W}}_R[\Phal_R] \leq {\frac{1}{2}}\left( {\mathcal{W}}_R(\Pz) + {\mathcal{W}}_R(\Pu)\right) - \mathrm{Gain}_R,$$ where $\mathrm{Gain}_R>0$ is some quantitative positive gain due to the *strict* convexity of the interaction. With some work, using the fact that $\Pz, \Pu$ are stationary, we are able to show that the gain is at least proportional to $R$.
3. We turn $\Phal_R$ into a process on the full line by pasting independent copies of itself on disjoint intervals of length $2R$. The relative entropy is additive, and we can show that the interaction of two independent copies is almost zero. Thanks to the quantitative convexity estimate, we obtain a global candidate $\Phal$ for which $${\mathcal{F}_{\beta}}(\Phal) < {\frac{1}{2}}\left( {\mathcal{F}_{\beta}}(\Pz) + {\mathcal{F}_{\beta}}(\Pu)\right),$$ which is absurd, hence the minimiser of ${\mathcal{F}_{\beta}}$ is unique.
### Plan of the paper {#plan-of-the-paper .unnumbered}
- In Section \[sec:defoffree\], we define formally the logarithmic energy and the specific relative entropy, which are the two ingredients of the log-gas free energy.
- In Section \[sec:descriscri\], we present the screening procedure.
- In Section \[sec:discregaps\], we state a key discrepancy estimate, and collect auxiliary results controlling the typical size of the gaps, and positions of points, for processes of finite logarithmic energy.
- In Section \[sec:largeboxscreening\], we implement the first step of the proof, by performing the screening procedure on a large box for any given point process with finite energy.
- In Section \[sec:displacement\], we recall results from optimal transportation and the theory of displacement convexity.
- In Section \[sec:interpolateprocess\], we apply displacement interpolation to build a local candidate: the interpolate process in a large box, whose local energy and entropy are better than the two processes it was constructed from.
- In Section \[sec:conclusionproof\], we use the interpolate process to build a global candidate, that is a stationary point process whose free energy is strictly smaller than the minimum of ${\mathcal{F}_{\beta}}$, which is the desired contradiction.
- Finally, in Appendix \[sec:annex\], we gather the proofs of some technical lemmas, and in Appendix \[sec:proofscreening\] we sketch a proof of the screening procedure adapted to the present setting.
Preliminary definitions and notation
------------------------------------
### Notation {#notation .unnumbered}
- For $R > 0$, we denote by ${\Lambda}_R$ the interval $[-R,R]$.
- If $I$ is an interval of $\R$, we let $|I|$ be its length.
- If $A$ is a subset of $\R$ or $\R^2$, we denote the boundary of $A$ by $\partial A$.
- We let $\diamond$ be the diagonal in $\R \times \R$.
- We will use the symbol $\preceq$ to denote an inequality that holds up to some universal (positive) multiplicative constant.
- Points in $\R \times \R$ are denoted by upper case letters, e.g. $X = (x,y)$.
- If $\mu$ is a signed measure on some interval $I \subset \R$, we will sometimes treat it as a signed measure on $I \times \R$, where it is understood that, $\varphi : I \times \R \to \R$ being a test function, we define $$\label{conventionmeasure}
\int_{I \times \R} \varphi(x,y) d\mu(x,y) := \int_{I} \varphi(x,0) d\mu(x).$$
- If $\mu$ is a probability measure on $X$ and $F$ is a map from $X$ to $Y$, we let $F_* \mu$ be the push-forward of $\mu$ by $F$, which is a probability measure on $Y$.
### Point configurations and processes {#point-configurations-and-processes .unnumbered}
- A point configuration ${\mathcal{C}}$ on a set $S\subset \R$ is defined as a purely atomic Radon measure on $S$, giving integer mass to singletons and finite mass to any compact set. Any configuration can be written as ${\mathcal{C}}= \sum_{i\in I} \delta_{p_i}$, where $(p_i)_i$ is a collection of points in $S$ and $I$ is finite or countable. Here, multiple points are allowed. A configuration is simple if no multiple points occur, i.e. ${\mathcal{C}}(\{x\})\in\{0,1\}$ for all $x$. In this case we also consider ${\mathcal{C}}$ as a locally finite collection of points, i.e. the set $\{p_i, i\in
I\}$. Abusing notation we write ${\mathcal{C}}=\sum_{p\in{\mathcal{C}}}\delta_p$. For all random configurations we consider, multiple points do not occur almost surely.
- For any interval $I \subset \R$, we let ${\mathrm{Conf}}(I)$ be the space of point configurations supported on $I$. It is endowed with the initial topology for the family of maps ${\mathcal{C}}\mapsto \int \varphi d{\mathcal{C}}$, for $\varphi$ continuous and compactly supported on $\R$ (topology for which it is a Polish space), and with the associated Borel sigma-algebra.
- Given a subset $J\subset I$ and ${\mathcal{C}}\in{\mathrm{Conf}}(I)$, we denote by $|_J{\mathcal{C}}$ the restriction of ${\mathcal{C}}$ to $J$ (viewed as a measure).
- If $u \in \R$ and ${\mathcal{C}}\in {\mathrm{Conf}}(\R)$, we let ${\mathcal{C}}- u$ be the translation of ${\mathcal{C}}$ by $u$, which is the point configuration corresponding to shifting all the points of ${\mathcal{C}}$ by $-u$, or equivalently, for any test function $\varphi$ $$\int \varphi d\left({\mathcal{C}}-u\right) := \int \varphi( \cdot + u ) d{\mathcal{C}}.$$
- We will use two ways of enumerating points of a configuration:
- In a fixed interval $[-R,R]$, we will write $-R \leq \Xx_0 \leq \Xx_1 \leq \dots \leq \Xx_k \leq \dots \leq R$, enumerating points from the leftmost one.
- On $\R$, we will write $\dots x_{-k} \leq \dots \leq x_{-1} < 0 \leq x_0 \leq x_1 \leq \dots \leq x_k \leq \dots$, enumerating points starting from the origin.
We will need to pass from one enumeration to the other, for example in .
- A point process on $I$ is a probability measure on ${\mathrm{Conf}}(I)$. For $J\subset I$ and a point process ${\mathsf{P}}$ on $I$ we denote by ${\mathsf{P}}_J$ its restriction to $J$, i.e. ${\mathsf{P}}_J=(|_J)_*{\mathsf{P}}$ the point process on $J$ obtained as the image under the restriction map $|_J:{\mathrm{Conf}}(I)\to{\mathrm{Conf}}(J)$.
- We let ${\mathsf{\Pi}}$ be the Poisson point process on $\R$ of intensity $1$.
### Discrepancy {#discrepancy .unnumbered}
If ${\mathcal{C}}$ is a finite point configuration, we let $|{\mathcal{C}}|$ be the number of points of ${\mathcal{C}}$. If ${\mathcal{C}}$ is a point configuration and $I$ an interval, we let ${\mathcal{C}}_I$ be the restriction of ${\mathcal{C}}$ to $I$. In particular, $|{\mathcal{C}}_I|$ denotes the number of points of ${\mathcal{C}}$ in $I$. We then define the discrepancy of ${\mathcal{C}}$ in $I$ as the difference: $$\label{def:Discr}
\Discr_I({\mathcal{C}}) := |{\mathcal{C}}_I| - |I|.$$
Definition of the free energy {#sec:defoffree}
=============================
Electric fields and electric energy
-----------------------------------
Let us recall that, in the sense of distributions, the following identity holds $$- \div \left( \nabla \left(- \log | \cdot | \right) \right) = 2\pi \delta_0 \text{ in } \R \times \R,$$ which corresponds to the fact that $X \mapsto \frac{1}{2\pi} \log|X|$ is the fundamental solution of Laplace’s equation $\Delta f = \delta_0$ in dimension two.
### Electric fields {#electric-fields .unnumbered}
\[def:electricfields\] Let $I$ be an interval of $\R$.
- We call *electric fields on $I$* the set of all vector fields $\El$ in $\Lploc(I \times \R, \R \times \R)$, for some $1 < p < 2$ fixed.
- Let ${\mathcal{C}}$ be a finite point configuration in $I$ and $\El$ an electric field on $I$. We say that *$\El$ is compatible with ${\mathcal{C}}$ in $I$* provided $$\label{def:compatible}
-\div(\El) = \c \left({\mathcal{C}}- dx \right) \quad \text{in } I \times \R,$$ in the sense of distributions, and using the convention of . See for an example.
- If $\El$ is compatible with ${\mathcal{C}}$ in $I$, for $\eta \in (0,1)$ we define the $\eta$-truncation of the electric field $\El$ as $$\label{def:Eeta}
\El_{\eta}(X) := \El(X) - \sum_{p \in {\mathcal{C}}\cap I} \nabla \f_{\eta} (X - (p,0)),$$ where $\f_{\eta}$ is the function $$\f_{\eta}(x) = \max \left(- \log \left( \frac{|x|}{\eta} \right), 0\right).$$
- In particular, if ${\mathcal{C}}= \sum_{x \in {\mathcal{C}}} \delta_x$ is a point configuration in $I$ and $E$ is compatible with ${\mathcal{C}}$, the truncation $\El_{\eta}$ satisfies the equation (compare with ) $$\label{compa2}
-\div(\El_{\eta}) = \c \left(\sum_{x \in {\mathcal{C}}} \sigma_{x,\eta} - dx \right) \quad \text{in } I \times \R,$$ where $\sigma_{x,\eta}$ is the “smeared out charge” as in [@rougerie2016higher], the uniform measure on the circle of radius $\eta$ around $x$ (the measure $\sigma_{x,\eta}$ is truly supported on $\R \times \R$).
Let $R > 0$, let ${\mathcal{C}}$ be a point configuration in ${\Lambda}_R$.
- We define the *local electric field generated by ${\mathcal{C}}$* as $$\label{def:Eloc}
X \in \R \times \R \mapsto \Eloc(X ; {\mathcal{C}}; {\Lambda}_R) := \int_{{\Lambda}_R} - \nabla \log |X-(u,0)| (d{\mathcal{C}}(u) - du).$$
- For any $\eta > 0$, we define the local energy of ${\mathcal{C}}$ (truncated at $\eta$) as $$\label{def:Wel}
\Wel_{\eta}({\mathcal{C}}; {\Lambda}_R) := \frac{1}{2\pi} \int_{\R \times \R} |\Eloc_{\eta}|^2 + |{\mathcal{C}}| \log \eta.$$
$\Eloc$ is a model case for the definition of electric fields. Indeed, $\Eloc$ is an electric field compatible with ${\mathcal{C}}$ in ${\Lambda}_R$, in the sense of , and it is in $\Lploc$ for all $p \in (1,2)$ but fails to be in $L^2$ near each point of ${\mathcal{C}}$.
### The intrinsic energy, monotonicity estimates {#the-intrinsic-energy-monotonicity-estimates .unnumbered}
Let ${\mathcal{C}}$ be a point configuration in ${\Lambda}_R$. We define the *intrinsic energy of ${\mathcal{C}}$* as the quantity $$\label{def:Wint}
{\mathcal{W}^{\mathrm{int}}}({\mathcal{C}};{\Lambda}_R) := \iint_{\mdiag} - \log|x-y| (d{\mathcal{C}}(x) - dx) (d{\mathcal{C}}(y) - dy)\;,$$ where $\diamond$ denotes the diagonal in $\Lambda_R\times\Lambda_R$.
\[lem:monoto\] Let ${\mathcal{C}}$ be a finite point configuration in ${\Lambda}_R$, with exactly $|{\Lambda}_R|$ points.
1. The limit $\lim_{\eta \to 0} \Wel_{\eta}({\mathcal{C}};{\Lambda}_R)$ exists and satisfies $$\label{WelWint}
\lim_{\eta \to 0} \Wel_{\eta}({\mathcal{C}};{\Lambda}_R) = {\mathcal{W}^{\mathrm{int}}}({\mathcal{C}};{\Lambda}_R).$$
2. We have, for $\eta > 0$ $$\label{WelleqWint}
\Wel_{\eta}({\mathcal{C}};{\Lambda}_R) \leq {\mathcal{W}^{\mathrm{int}}}({\mathcal{C}};{\Lambda}_R) + |{\Lambda}_R| \ErMon(\eta),$$ where $\ErMon$ is some function independent of ${\mathcal{C}}$, satisfying $$\label{ErMon0}
\lim_{\eta \to 0} \ErMon(\eta) = 0.$$
3. Conversely, we have, for $\eta > 0$ $$\label{WintleqWel}
{\mathcal{W}^{\mathrm{int}}}({\mathcal{C}};{\Lambda}_R) \leq \Wel_{\eta}({\mathcal{C}};{\Lambda}_R) + |{\Lambda}_R| \ErMon(\eta) + \ErTrun({\mathcal{C}}, \eta; {\Lambda}_R),$$ where $\ErMon$ is as above, and $\ErTrun$ satisfies $$\label{ErTrun}
\ErTrun({\mathcal{C}},\eta;{\Lambda}_R) \leq \iint_{(x,y) \in {\Lambda}_R \times {\Lambda}_R, 0 < |x-y| < 2\eta} - \log \left|x-y \right| d{\mathcal{C}}(x) d{\mathcal{C}}(y).$$
The existence of the limit and is proven e.g. in [@sandier20151d]. The monotonicity estimates , follow from [@petrache2017next Lemma 2.3](by sending the parameter $\alpha$ appearing there to $0$).
### Electric energy of a point process {#electric-energy-of-a-point-process .unnumbered}
We introduce the first component of the free energy functional: the electric energy of a point process.
- Let ${\mathcal{C}}$ be a point configuration on $\R$. We define the global electric energy of ${\mathcal{C}}$ as $$\label{def:tWel}
\tWel({\mathcal{C}}) := \inf_{E} \left( \lim_{\eta \to 0} \left( \limsup_{R \to \infty} \frac{1}{|{\Lambda}_R|} \frac{1}{\c} \int_{{\Lambda}_R \times \R} |\El_{\eta}|^2 + \log \eta \right) \right),$$ where the $\inf$ is taken over electric fields $E$ that are compatible with ${\mathcal{C}}$ in $\R$, in the sense of Definition \[def:electricfields\]. As usual, if there is no such electric fields, we let $\tWel({\mathcal{C}}) := + \infty$.
- Let ${\mathsf{P}}$ be a point process. We define the electric energy of ${\mathsf{P}}$ as the quantity $$\label{def:Welec}
{\mathcal{W}}({\mathsf{P}}) := \E_{{\mathsf{P}}} \left[ \tWel({\mathcal{C}}) \right].$$
We refer to [@leble2017large Section 2.7] for more details. The following remark is a consequence of [@leble2017large Lemmas 2.3 and 3.8].
If ${\mathcal{C}}$ is a point configuration such that $\tWel({\mathcal{C}})$ is finite, then the infimum in is attained for exactly one electric field $\El$, and in fact all other compatible fields have infinite energy. If ${\mathsf{P}}$ is a point process with finite energy, there is a measurable choice ${\mathcal{C}}\mapsto \El$ of compatible electric fields such that $${\mathcal{W}}({\mathsf{P}}) = \E_{{\mathsf{P}}} \left[ \lim_{\eta \to 0} \left( \limsup_{R \to \infty} \frac{1}{|{\Lambda}_R|} \frac{1}{\c} \int_{{\Lambda}_R \times \R} |\El_{\eta}|^2 + \log \eta \right) \right].$$ Moreover, if ${\mathsf{P}}$ is stationary, we can write for any $R > 1$, $$\label{EnergyStatCase}
{\mathcal{W}}({\mathsf{P}}) = \lim_{\eta \to 0} \E_{{\mathsf{P}}}\left[ \frac{1}{|{\Lambda}_R|} \frac{1}{\c} \int_{{\Lambda}_R \times \R} |\El_{\eta}|^2 + \log \eta \right].$$
An alternative definition for the logarithmic energy of a point process ${\mathsf{P}}$ was introduced in [@leble2016logarithmic], inspired by [@MR3046995]. We define the intrinsic energy of ${\mathsf{P}}$ as $$\label{def:Wintglob}
\Wintg({\mathsf{P}}) := \liminf_{L \to \infty} \frac{1}{|{\Lambda}_L|} \E_{{\mathsf{P}}} \left[ {\mathcal{W}^{\mathrm{int}}}({\mathcal{C}}; {\Lambda}_L)\right].$$ The energies ${\mathcal{W}}$ and $\Wintg$ are related to each other, in particular it is proven in [@leble2016logarithmic]\[Prop. 5.1.\] that if ${\mathsf{P}}$ is a stationary process with small discrepancies, for example such that $\sup_L \E_{{\mathsf{P}}} \left[ \Discr_{{\Lambda}_L}^2 \right] < \infty$, we have $$\label{welecwintg}
{\mathcal{W}}({\mathsf{P}}) \leq \Wintg({\mathsf{P}}),$$ and in fact one can obtain some weak form of the converse inequality, via a recovery sequence, we refer to [@leble2016logarithmic] for more details.
The specific relative entropy
-----------------------------
Let ${\mathsf{P}}$ be a stationary point process. The following limit exists in $[0, + \infty]$ and defines the *specific relative entropy* of ${\mathsf{P}}$ with respect to the Poisson point process. \[def:SRE\] $$\label{SRE}
{\mathcal{E}}[{\mathsf{P}}] := \lim_{R \to \infty} \frac{1}{|{\Lambda}_R|} {\mathrm{Ent}}\left[ {\mathsf{P}}_{| {\Lambda}_R} \big| {\mathsf{\Pi}}_{|{\Lambda}_R} \right].$$
It is a lower semi-continuous, affine function of ${\mathsf{P}}$. Its unique minimiser – among stationary processes – is the Poisson process itself, for which ${\mathcal{E}}[{\mathsf{\Pi}}] = 0$. We refer e.g. to [@friedli2017statistical Sec. 6.9.2] for elementary properties of the specific entropy.
The free energy
---------------
Let $\beta > 0$ be fixed. For any stationary point process ${\mathsf{P}}$, we define the free energy ${\mathcal{F}_{\beta}}({\mathsf{P}})$ as the weighted sum $$\label{def:fbeta}
{\mathcal{F}_{\beta}}({\mathsf{P}}) := \beta {\mathcal{W}}({\mathsf{P}}) + {\mathcal{E}}({\mathsf{P}}).$$
It is lower semi-continuous, affine, and has compact sub-level sets. In particular, it has a minimum, and admits a compact set of minimisers (our point is to prove that there is only one).
The screening procedure {#sec:descriscri}
=======================
The screening technique has been introduced in [@ss2d], followed by several adaptations in e.g. [@rougerie2016higher; @petrache2017next]. In this section, we state a version of the procedure suitable for us. In particular, the stationary nature of our problem allows us to bypass the usual step of “finding a good boundary”, and we work with a fixed good boundary. In Claim \[claim:discrenearboundary\], we add a technical remark concerning the “new” points produced by the screening construction, that is needed here.
Screened fields {#sec:notationscreen}
---------------
Let ${\mathcal{C}}$ be a point configuration in an interval $I$, and let $\El$ be an electric field, compatible with ${\mathcal{C}}$ in $I$, i.e. holds. We say that $\El$ is screened, and write $\El \in \Screen({\mathcal{C}}, I)$ when $$\label{lescreening}
\El \cdot \vec{\nu} = 0 \quad \text{on } \partial (I \times \R),$$ where $\vec{\nu}$ is the outer unit normal vector.
The original motivation for working with *screened* electric fields is that one can paste such fields together: since their normal component agree on the boundary, they form a globally compatible electric field. In the present paper, we use this screening property rather incidentally, only to compare the “screened” energy with the “local” energy in the proof of Proposition \[prop:largeboxapproximation\]. Here we mostly use the screening procedure as a way to transform a configuration with possibly more or less than $2R$ points in $[-R,R]$ into a configuration with exactly $2R$ points, in such a way that we keep the “old” configuration on a large segment, and that we do not augment much the energy.
The screening procedure {#the-screening-procedure}
-----------------------
\[prop:screening\] Let $s \in \lp 0, \frac{1}{4} \rp$ and let $M > 1$. There exists a universal constant $\etac > 0$ and $R_0=R_0(s,M)>0$ such that for any $\eta\in(0, \etac)$ and any integer $R\geq R_0$, the following holds:\
Put $R' := R(1-s)$, $\Old := {\Lambda}_{R'}$, and $\New := {\Lambda}_R \setminus \Old$. Let ${\mathcal{C}}$ be a point configuration in ${\Lambda}_R$, let $\El$ be compatible with ${\mathcal{C}}$ in $\Old$, and let $\El_{\eta}$ be the truncation as in .
Assume that $$\begin{aligned}
\label{definiM} \Mec := \int_{\{-R', R'\} \times [-R, R]} |\El_{\eta}|^2 & \leq M \\
\label{decrvert} \Hec := \frac{1}{s^4 R} \int_{{\Lambda}_R \times \left( \R \backslash (-{\frac{1}{2}}s^2 R, {\frac{1}{2}}s^2 R)\right)} |\El|^2 & \leq 1.\end{aligned}$$ and that furthermore $$\label{nochargeneargoodboundary}
\left| {\mathcal{C}}_{[-R'-2\eta, -R' + 2 \eta]} \right| = 0, \quad \left| {\mathcal{C}}_{[R'-2\eta, R' + 2 \eta]} \right| = 0.$$ Then, there exists a probability measure $\Phiscr_{{\mathcal{C}}, s, \eta, R}$ on point configurations in ${\Lambda}_R$ such that for $\Phiscr_{{\mathcal{C}}, s, \eta, R}$-a.e. configuration $\Cscr$, the following holds:
- The number of points is given by $|\Cscr_{{\Lambda}_R}| = |{\Lambda}_R|$.
- The configurations ${\mathcal{C}}$ and $\Cscr$ coincide on $\Old$.
- There is no point at distance less than $\frac{1}{10}$ of $\{-R, R\}$.
- The pairs of points of $\Cscr$ which are at distance less than $\eta$ form a subset of the pairs of points of ${\mathcal{C}}$ with the same property, in particular $$\label{screeninggoodtrunc}
\iint_{0 < |x-y| < \eta} - \log |x-y| d\Cscr(x) d\Cscr(y) \leq \iint_{0 < |x-y| < \eta} - \log |x-y| d{\mathcal{C}}(x) d{\mathcal{C}}(y).$$
- There exists an electric field $\Escr$ in $\Screen(\Cscr, {\Lambda}_R)$ such that $$\label{erreurEcrantage}
\int_{{\Lambda}_R \times \R} |\Escr_{\eta}|^2 \leq \int_{{\Lambda}_R \times [-R, R]} |\El_{\eta}|^2 + \ErrScr,$$ with an error term $\ErrScr$ bounded by $$\label{ErrScr}
\ErrScr \preceq |\log \eta| M s R.$$
Moreover, the restriction of $\Phiscr_{{\mathcal{C}}, s, \eta, R}$ to $\New$ can be written as the point process given by $$\label{CscrNew}
\Cscr_{\New} := \sum_{i=1}^{n} \delta_{p_i + r_i \eta},$$ where $n :=|\Cscr_{\New}|$ is the number of points of $\Cscr$ in $\New$, the $p_i$’s are points in $\New$, and the $r_i$’s are i.i.d uniform random variables in $\left[-\frac{1}{4},\frac{1}{4}\right]$. The number $n$ and the points $p_i$ are deterministic for $\Phiscr_{{\mathcal{C}}, s, \eta, R}$, namely they depend only on ${\mathcal{C}}$.
\[fig:screenstatement\]
In the screening procedure of [@petrache2017next], designed for being applied to a finite gas, the condition is replaced by a bound of the form $$\frac{1}{R} \int_{{\Lambda}_R \times [-R, R]} |\El_{\eta}|^2 \leq M.$$ Then a mean value argument is used to find a “good boundary” on which an estimate of the type holds. Taking advantage of the stationary character of the infinite gas, we may skip this step and impose a fixed good boundary. It shortens the procedure a bit, and turns out to be very convenient for the proof of entropy estimates below.
Assuming is not mandatory, and not used in [@petrache2017next], but it simplifies the technical details of the proof.
A technical remark on the screening construction
------------------------------------------------
The set $\New$ consists of two intervals $[-R, -R']$ and $[R',R]$, with $R - R' = sR$. We state here the results for the “left side” $[-R, -R']$, they extend readily to the other side.
\[claim:position\] Let $\kkmax$ be the number of points of $\Cscr$ in $[-R, -R']$ and for $1 \leq k \leq \kkmax$, let $\Xx_k$ be the $k$-th point of $\Cscr$ in $[-R, R']$, starting from the leftmost one. We also write $$\bXx_k = -R + k - 1/2,$$ which corresponds to the “ideal” position of $\Xx_k$ if the points were regularly spaced.
\[claim:discrenearboundary\]
- The number $\kkmax$ satisfies $$\label{bornekmax}
|\kkmax - sR| \preceq M^{1/2} sR^{1/2}.$$
- For $k$ such that $s^2R \preceq |sR - k|$ (for points that are close to $-R$ and far from the boundary point $-R'$ between $\New$ and $\Old$), we have $$\label{xkmoinsk} \left|\Xx_k - \bXx_k \right| \preceq \frac{k}{R^{1/2}},$$ so in particular we have $\left|\Xx_k - \bXx_k \right| \preceq \frac{k}{2}$ and $\left|\Xx_k - \bXx_k \right| \preceq sR^{1/2}$. It also implies $$\label{discrNewA}
\left|\Discr_{[-R, -R + k]}\right| \preceq \frac{k}{R^{1/2}},$$
- For $k$ such that $|sR - k| \preceq s^2R$ (for points that are close to the boundary point $-R'$ between $\New$ and $\Old$ and far from $-R$), we have: $$\label{xkmoinskB}
\left| \Xx_k - \bXx_k \right| \preceq M^{1/2} sR^{1/2}.$$ It also implies $$\label{discrNewB}
\left| \Discr_{[-R, -R + k]} \right| \preceq M^{1/2} sR^{1/2}.$$
Proposition \[prop:screening\] is based on the existing screening procedure. A sketch of the proof with, in particular, a justification of Claim \[claim:discrenearboundary\], is given in Section \[sec:proofscreening\].
Discrepancy estimate, gaps {#sec:discregaps}
==========================
A discrepancy estimate
----------------------
Let us recall that for an interval $I$, the discrepancy $\Discr_{I}({\mathcal{C}})$ is defined as $|{\mathcal{C}}_I| - |I|$, hence $\E_{{\mathsf{P}}} \left[ \Discr_{I}^2 \right]$ can be thought of as the variance, under ${\mathsf{P}}$, of the number of points in $I$. \[sec:discrepancy\] It is observed in [@leble2017large Lemma 3.2], that if ${\mathcal{W}}({\mathsf{P}})$ is finite, we have $$\label{discrbasic}
\limsup_{R \to \infty} \frac{1}{|{\Lambda}_R|} \E_{{\mathsf{P}}} \left[\Discr_{{\Lambda}_R}^2\right] < \infty.$$ Moreover, in [@leble2017large Remark 3.3], it was proven that if ${\mathsf{P}}$ is a stationary point process on $\R$ with finite energy, then $$\label{discrPfiniA}
\liminf_{R \to \infty} \frac{1}{|{\Lambda}_R|} \E_{{\mathsf{P}}} \left[\Discr_{{\Lambda}_R}^2\right] = 0.$$ In fact, an examination of the proof allows for a stronger statement.
\[lem:variancesublinear\] Assume that ${\mathcal{W}}({\mathsf{P}})$ is finite. Then $$\label{discrPfiniB}
\lim_{R \to \infty} \frac{1}{|{\Lambda}_R|} \E_{{\mathsf{P}}} \left[\Discr_{{\Lambda}_R}^2\right] = 0.$$
We give the proof of Lemma \[lem:variancesublinear\] in Section \[sec:proofvariancesublinear\]. The gain from to may seem minor, but it turns out to be crucial here (it is also used in a central way in [@Leble:2018aa]).
Points, gaps and positions
--------------------------
### Definition of the gaps {#definition-of-the-gaps .unnumbered}
\[def:gaps\] For any point configuration ${\mathcal{C}}$ on $\R$, let us enumerate the points of ${\mathcal{C}}$, counted from the origin, as $$\label{enumerationB}
\dots < x_{-k} < \dots < x_{-2} < x_{-1} < 0 \leq x_0 < x_1 < \dots < x_k < \dots.$$ For ${\mathcal{C}}$ fixed and $k \in \Z$, we will write $x_k({\mathcal{C}})$ to denote the $k$-th point of ${\mathcal{C}}$ in the sense of this enumeration. The enumeration is well-defined if all the points are simple, which is an event of full measure for all the processes considered here.
We define the sequence of gaps “counted from 0” for ${\mathcal{C}}$ as the sequence $\{\Gap_k\}_{k \in \Z}$, where $$\label{GapsFrom0}
\Gap_k = x_{k+1} - x_k.$$ If ${\mathcal{C}}$ is a finite configuration, we let $\Gap_k = + \infty$ after the last point is reached in either direction.
For a given point configuration ${\mathcal{C}}$, and a real number $u$, we let $\Pos({\mathcal{C}};u)$ be the integer such that $$\label{def:Pos}
\Pos({\mathcal{C}};u) = m \iff x_{0}({\mathcal{C}}- u) = x_{m}({\mathcal{C}}) \iff \Gap_0({\mathcal{C}}- u) = \Gap_m({\mathcal{C}}).$$ In plain words, $\Pos({\mathcal{C}};u)$ is the position in ${\mathcal{C}}$ of the first point (counted from $0$) of the translated configuration ${\mathcal{C}}-u$.
### Average size of the gaps {#average-size-of-the-gaps .unnumbered}
\[lem:gapL2\] Let $0 < \eta < 1/2$, $R > 0$, let ${\mathcal{C}}$ be a configuration in ${\Lambda}_R$ that has at least $R/2$ points[^2] in $[0,R]$ and in $[-R, 0]$. We have: $$\label{gapsL2}
\sum_{i=-\frac{R}{2}}^{\frac{R}{2}} \left(\Gap_i\right)^2 \preceq R + \int_{[-R, R] \times \R} |\El_{\eta}|^2.$$
This result says that gaps are “typically of order 1”, because in view of we expect the second term in the right-hand side of to be of order $R$. We give the proof of Lemma \[lem:gapL2\] in Section \[sec:proofGapL2\].
### Deviations estimates for the positions {#deviations-estimates-for-the-positions .unnumbered}
Intuitively, since the processes have intensity $1$, the first point of a configuration, counted from the origin, is close to $0$, and the $k$-th point is at a distance $\approx k$. It would imply that the first point of ${\mathcal{C}}-u$ is $\approx u$, and more generally that the $v$-th point of ${\mathcal{C}}-u$ is $\approx u + v$. The following lemma shows that these heuristics are correct.
\[lem:discrUVW\] Let ${\mathsf{P}}$ be a stationary point process and assume that ${\mathcal{W}}({\mathsf{P}})$ is finite. It holds, for any $u, v$ in $\Z$ $$\label{PposfarA}
\lim_{w \to + \infty} {\mathsf{P}}\left( x_{\Pos({\mathcal{C}}, u) + v}({\mathcal{C}}) \geq (u + v) + w\right) = 0.$$ Similarly, we have $$\label{PposfarB}
\lim_{w \to + \infty} {\mathsf{P}}\left( x_{\Pos({\mathcal{C}}, u) + v}({\mathcal{C}}) \leq (u + v) - w \right) = 0.$$
The fact that $ x_{\Pos({\mathcal{C}}, u) + v}({\mathcal{C}}) \geq (u+v) + w$ implies that there are less than $v$ points between $u$ and $u+v +w$, and hence $$\llbr x_{\Pos({\mathcal{C}}, u)+v}({\mathcal{C}}) \geq w + u + v \rrbr \subset \llbr \Discr_{[u,u+v+w]} \geq w. \rrbr$$ In view of , and since ${\mathsf{P}}$ is stationary, the probability of the latter event tends to $0$ as $w \to \infty$, for $v$ fixed, independently of $u$. The proof of is identical.
In particular:
- Taking $u, v = 0$, so that $\Pos({\mathcal{C}},u) + v = 0$ and sending $w \to \infty$, we get $$\label{firstoftennotfar}
\lim_{w \to \infty} {\mathsf{P}}\left( x_0({\mathcal{C}}) \geq w \right) = 0.$$ This shows that indeed the first point of a configuration is typically at a bounded distance from $0$.
- Taking $v$ fixed, $w = u - v$, and sending $u \to \infty$, we obtain $$\label{xposugeq2u}
\lim_{u \to \infty} {\mathsf{P}}\left( x_{\Pos({\mathcal{C}}, u)+ v}({\mathcal{C}}) \geq 2u \right) = 0.$$ This shows that for $v$ fixed and $u$ large, the $v$-th point of ${\mathcal{C}}-u$ cannot be much further than $2u$. This estimate is far from being sharp, but it is enough for our purposes.
Distinct stationary processes have distinct gap distributions
-------------------------------------------------------------
\[prop:gapsdiff\] Let $\Pz, \Pu$ be two stationary point processes such that ${\mathcal{W}}(\Pz)$ and ${\mathcal{W}}(\Pu)$ are finite, and assume that $\Pz \neq \Pu$. Then there exists $c > 0$, an integer $r \geq 1$, and a function $H : \R_+^{2r+1} \to \R$, such that
- $H$ is compactly supported, bounded by $1$, and $1$-Lipschitz with respect to the $\|\cdot\|_{1}$ norm on $\R_+^{2r+1}$, i.e. $$\label{HLip}
\| H \|_{\infty} \leq 1, \quad \left| H(a_{-r}, \dots, a_{r}) - H(b_{-r}, \dots, b_{r})\right| \leq \sum_{i=-r}^{r} |a_{i} - b_i|.$$
- $H$ detects the difference of distribution of the gaps of $\Pz$ and $\Pu$, namely $$\label{gapsdiff}
\E_{\Pz}\left[ H(\Gap_{-r}, \dots, \Gap_r) \right] - \E_{\Pu}\left[ H(\Gap_{-r}, \dots, \Gap_r) \right] > c.$$
We postpone the proof of Lemma \[prop:gapsdiff\] to Section \[sec:proofgapsdiff\].
\[prop:gaps2\] Let $\Pz, \Pu$ be two stationary point processes such that ${\mathcal{W}}(\Pz)$ and ${\mathcal{W}}(\Pu)$ are finite, and assume that $\Pz \neq \Pu$. There exists $\g > 0$ such that for $R$ large enough, any coupling $Q$ of the restrictions of $\Pz,\Pu$ to ${\Lambda}_R$, and any random variable $S$ bounded by $R^{1/2}$, then $$\label{EPigaps}
\E_{\Qo} \left[ \sum_{i=-R/2}^{R/2} \frac{\left|\Gap_i(\Cz) - \Gap_{i+S}(\Cu) \right|^2}{|\Gap_i(\Cz)|^2 + |\Gap_{i+S}(\Cu)|^2} \right] \geq \g R.$$
We give the proof of Proposition \[prop:gaps2\] in Section \[sec:proofgaps2\]. Intuitively, the reason that holds true is the following: we know that the gaps are typically of order $1$ (see Lemma \[lem:gapL2\]), so the denominator is of order $1$, and $\Pz, \Pu$ are stationary so the difference between their gaps distribution on a given interval should be proportional to its length.
Since the quantity $\frac{\left|\Gap_i(\Cz) - \Gap_{i+S}(\Cu) \right|^2}{|\Gap_i(\Cz)|^2 + |\Gap_{i+S}(\Cu)|^2}$ is bounded, the following extension of Proposition \[prop:gaps2\] is straightforward.
\[coro:gaps\] In the situation of Proposition \[prop:gaps2\], if $A, B$ are events such that $$\Pz(A) \geq 1 - \frac{\g}{100}, \quad \Pu(B) \geq 1 - \frac{\g}{100},$$ and if $\tQo$ is a coupling of the restrictions to ${\Lambda}_R$ of $\Pz$ conditioned to $A$ and $\Pu$ conditioned to $B$, then $$\E_{\tQo} \left[ \sum_{i=-R/2}^{R/2} \frac{\left|\Gap_i(\Cz) - \Gap_{i+S}(\Cu) \right|^2}{|\Gap_i(\Cz)|^2 + |\Gap_{i+S}(\Cu)|^2} \right] \geq {\frac{1}{2}}\g R.$$
Indeed, we can lift the coupling $\tQo$ of the conditioned processes (restricted to ${\Lambda}_R$) to a coupling of $\Pz, \Pu$ (restricted to ${\Lambda}_R$) by defining $$\Qo := \Pz(A)\Pu(B)\cdot \tQo +\1_{(A\times B)^c}\cdot\Pz\big|_{{\Lambda}_R}\otimes\Pu\big|_{{\Lambda}_R},$$ and we apply Proposition \[prop:gaps2\] to $\Qo$. Letting $$G : = \sum_{i=-R/2}^{R/2} \frac{\left|\Gap_i(\Cz) - \Gap_{i+S}(\Cu) \right|^2}{|\Gap_i(\Cz)|^2 + |\Gap_{i+S}(\Cu)|^2},$$ which is a quantity bounded by $R$, we have $\E_{\Qo} [ G ] \geq \g R$, but also $$\E_{\tQo} [ G ]= \E_{ \Qo}[\1_{A\times B} G ]\frac{1}{\Pz(A)\Pu(B)}\geq \E_{\Qo}[\1_{A\times B} G] \geq \E_{\Qo}[G] - \Qo( (A \times B)^c)R,$$ and using the assumptions on $\Pz(A), \Pu(B)$ we get the result.
Large box approximation
=======================
In this section, we introduce an approximation scheme: given a point process ${\mathsf{P}}$, and parameters $\epsilon, s > 0$, for $R$ large enough we are able, after conditioning on a likely event $\EveR$, to restrict ${\mathsf{P}}$ in ${\Lambda}_R$, to modify the configurations near the edges of ${\Lambda}_R$, and to produce a “local candidate” $\tPR$ in ${\Lambda}_R$ which
1. “Looks like” ${\mathsf{P}}$ (in a sense that is controlled by the parameter $s$).
2. Has an energy, entropy comparable to those of ${\mathsf{P}}$, up to small errors (controlled by $\epsilon$).
3. Has various good discrepancies bounds (controlled by $s$).
4. Has almost surely $2R$ points in ${\Lambda}_R$.
Points 1 and 4 are obtained through the screening procedure described in Section \[sec:descriscri\]. Points 2 and 3 are guaranteed by conditioning on a good event, and using properties that hold for processes with finite energy, as mentioned in Section \[sec:discregaps\]. The first step is to check that certain conditions, related to the applicability of the screening procedure, or to the good controls on the discrepancy that we want to enforce, are often satisfied under ${\mathsf{P}}$, if ${\mathsf{P}}$ has finite energy.
\[sec:largeboxscreening\]
Conditions that are often satisfied
-----------------------------------
Let ${\mathsf{P}}$ be a stationary point process such that ${\mathcal{W}}({\mathsf{P}})$ is finite. We recall that to ${\mathsf{P}}$-almost every configuration is associated an electric field $\El$ such that holds.
### Good energy {#good-energy .unnumbered}
This control is related to the condition in the statement of the screening procedure.
\[claim:goodenergy\] For all $\eta$ in $(0,1)$, for all $\delta > 0$, for all $M$ large enough (depending on ${\mathsf{P}}, \eta, \delta$), for all $s \in \lp 0, \frac{1}{4} \rp$ for all $R > 1$, letting $R' = R(1-s)$ (as in Proposition \[prop:screening\]), the event $$\EventEnergy(R, s, M, \eta) := \llbr \int_{\{-R', R'\} \times [-R, R]} |\El_{\eta}|^2 \leq M \rrbr$$ satisfies $${\mathsf{P}}\left( \EventEnergy(R, s, M, \eta) \right) \geq 1 - \delta.$$
Since ${\mathsf{P}}$ is stationary and has finite energy we have, in view of , for any fixed $\eta$, for any $x > 0$ $$\E_{{\mathsf{P}}} \left[ \int_{\{-x, x\} \times \R} |\El_{\eta}|^2 \right]
= \E_{{\mathsf{P}}} \left[ \frac{1}{R} \int_{{\Lambda}_R \times \R} |\El_{\eta}|^2 \right] < + \infty,$$ and the claim follows by applying Markov’s inequality.
### Good decay on the vertical axis {#good-decay-on-the-vertical-axis .unnumbered}
This control is related to the condition in the statement of the screening procedure.
\[claim:gooddecay\] For all $\delta > 0$, for all $s > 0$, for all $R$ large enough (depending on ${\mathsf{P}}, \delta, s$), the event $$\EventDecay(R,s) := \llbr \frac{1}{s^4 R} \int_{[-R, R] \times (\R \backslash (-s^2 R, s^2 R))} |\El|^2 \leq 1 \rrbr$$ satisfies $${\mathsf{P}}\left( \EventDecay(R,s) \right) \geq 1 - \delta.$$
For any $R$, for any $t > 0$, since ${\mathsf{P}}$ is stationary we have $$\E_{{\mathsf{P}}} \left[ \frac{1}{R} \int_{[-R, R] \times (\R \backslash (-t, t))} |\El|^2 \right] = \E_{{\mathsf{P}}} \left[ \int_{[-1, 1] \times (\R \backslash (-t, t))} |\El|^2 \right] < + \infty.$$ On the other hand, by dominated convergence we have $$\lim_{t \to + \infty} \E_{{\mathsf{P}}} \left[ \int_{[-1, 1] \times (\R \backslash (-t, t))} |\El|^2 \right] = 0.$$ Therefore, for any given $\delta, s$, we see that for $t \geq t_0$ large enough (depending on ${\mathsf{P}}, \delta, s$), it holds $$\E_{{\mathsf{P}}} \left[ \int_{[-1, 1] \times (\R \backslash (-t, t))} |\El|^2 \right] \leq \delta s^4\;.$$ Then for any $R \geq R_0 := \frac{t_0}{s^2}$ (depending on ${\mathsf{P}}, \delta, s$), we have $${\mathsf{P}}\left(\frac{1}{s^4 R} \int_{[-R, R] \times (\R \backslash (-s^2 R, s^2 R))} |\El|^2 > 1 \right) \leq \frac{1}{s^4} \E_{{\mathsf{P}}} \left[ \int_{[-1, 1] \times (\R \backslash (-s^2R, s^2R))} |\El|^2 \right] \leq \delta,$$ which proves the claim.
### The “good field” event {#the-good-field-event .unnumbered}
We introduce the event $\GoodField(R, s, M, \eta)$ as the set of all configurations ${\mathcal{C}}$ such that there exists a field $\El$ satisfying the following three conditions:
1. $\El$ is compatible with ${\mathcal{C}}$ in $\Old := {\Lambda}_{R(1-s)} = {\Lambda}_{R'}$.
2. $\El$ satisfies the “good energy” estimate $$\int_{\{-R', R'\} \times [-R, R]} |\El_{\eta}|^2 \leq M$$
3. $\El$ satisfies the “good decay” estimate $$\frac{1}{s^4 R} \int_{[-R, R] \times (\R \backslash (-s^2 R, s^2 R))} |\El|^2 \leq 1$$
\[claim:goodfield\] By construction, the event $\GoodField(R, s, M, \eta)$ depends only on the restriction of ${\mathcal{C}}$ to ${\Lambda}_{R'}$. Since the global field (compatible with ${\mathcal{C}}$ on the whole real line) is always a candidate, and in view of the definitions above, we have $${\mathsf{P}}\left(\GoodField(R, s, M, \eta) \right) \geq {\mathsf{P}}\left( \EventEnergy(R, s, M, \eta) \cap \EventDecay(R,s) \right),$$ so in particular, combining Claim \[claim:goodenergy\] and Claim \[claim:gooddecay\]. For all $\delta > 0$, for all $M$ large enough, for all $s > 0$, for all $R$ large enough, we have $${\mathsf{P}}\left(\GoodField(R, s, M, \eta) \right) \geq 1 - \delta.$$
### Good discrepancy estimates {#good-discrepancy-estimates .unnumbered}
The condition is related to the assumption in the screening procedure, and the two other conditions are useful estimates that we want to enforce.
\[claim:gooddiscr\] Let us introduce the following events: $$\begin{aligned}
\label{NoPoint} \NoPoint(R,s, \eta) & := \llbr \left| {\mathcal{C}}_{[-R' - 2\eta, -R' + 2\eta]} \right| = \left| {\mathcal{C}}_{[R' - 2\eta, R' + 2\eta]} \right| = 0 \rrbr \\
\label{LeftRight} \LeftRight(R, s) & := \llbr \left|\Discr_{[-R', 0]} \right| + \left| \Discr_{[0, R']} \right| \leq s R^{1/2} \rrbr, \\
\label{TotDiscr} \TotDiscr(R, s) & := \llbr \sum_{i=-R'}^{R'} \left| \Discr_{[-R', i]} \right| \leq s^2 R^{3/2}, \rrbr.\end{aligned}$$ and we let $\displaystyle{\EventDiscr(R, s, \eta) := \NoPoint(R, s, \eta) \cap \LeftRight(R,s) \cap \TotDiscr(R, s).}$
Then for all $\delta > 0$, for all $\eta > 0$ small enough (depending on $\delta$), for all $s \in \lp 0, \frac{1}{4} \rp$, for all $R$ large enough (depending on ${\mathsf{P}}, \delta, s$), we have $${\mathsf{P}}\left( \EventDiscr(R, s ,\eta) \right) \geq 1 - \delta.$$
For $\NoPoint$, we recall that ${\mathsf{P}}$ has intensity $1$, so for all $s, R$ and $\eta > 0$. $$\E_{{\mathsf{P}}} \left[ \left| {\mathcal{C}}_{[-R' - 2\eta, -R' + 2\eta]} \right| \right] = 4\eta.$$ Of course the number of points in a given interval is an integer valued random variable. We thus have $${\mathsf{P}}\left[ \left| {\mathcal{C}}_{[-R' - 2\eta, -R' + 2\eta]} \right| > 0 \right] = {\mathsf{P}}\left[ \left| {\mathcal{C}}_{[-R' - 2\eta, -R' + 2\eta]} \right| \geq 1 \right] \leq 4 \eta,$$ and we then take $\eta$ small enough (depending only on $\delta$).
For $\LeftRight$, we use Lemma \[lem:variancesublinear\] and Markov’s inequality. For $\TotDiscr$, we note that by Lemma \[lem:variancesublinear\] we have $$\begin{aligned}
\frac{1}{R^{3/2}}\E_{{\mathsf{P}}} \left[ \sum_{i = -R'}^{R'} |\Discr_{[-R',i]}| \right] \leq
\frac{1}{R}\sum_{i = -R'}^{R'}\sqrt{\frac{1}{R}\E_{{\mathsf{P}}} \left[ |\Discr_{[-R',i]}|^2 \right]}
\longrightarrow 0\;, \quad\text{as }R\to\infty\;,
\end{aligned}$$ and use again Markov’s inequality.
### Good truncation error {#good-truncation-error .unnumbered}
\[claim:goodtruncationerror\] For all $\delta, \epsilon > 0$, for all $\eta \in (0,1)$ small enough (depending on $P, \delta, \epsilon$), and for all $R > 1$, the event $$\label{def:EventTrunc}
\EventTrunc(R, \eta, \epsilon) := \llbr \iint_{\{x,y \in {\Lambda}_R \times {\Lambda}_R, 0 < |x-y| < 2\eta\}} - \log \left|x-y \right| d{\mathcal{C}}(x) d{\mathcal{C}}(y) \leq \frac{\epsilon}{100} R \rrbr$$ satisfies $${\mathsf{P}}\left(\EventTrunc(R, \eta, \epsilon)\right) \geq 1 - \delta.$$
By stationarity we have, for $R > 1$, $$\begin{gathered}
\E_{{\mathsf{P}}} \left[ \iint_{\{(x,y) \in {\Lambda}_R \times {\Lambda}_R, 0 < |x-y| < 2\eta\}} - \log \left|x-y \right| d{\mathcal{C}}(x) d{\mathcal{C}}(y) \right] \\
\preceq R \times \E_{{\mathsf{P}}} \left[ \iint_{\{(x,y) \in {\Lambda}_1 \times {\Lambda}_1, 0 < |x-y| < 2\eta\}} - \log \left|x-y \right| d{\mathcal{C}}(x) d{\mathcal{C}}(y) \right].\end{gathered}$$ From [@leble2017large Lemma 3.5.], we know that $$\lim_{\eta \to 0} \E_{{\mathsf{P}}} \left[ \iint_{\{(x,y) \in {\Lambda}_1 \times {\Lambda}_1, 0 < |x-y| < 2\eta\}} - \log \left|x-y \right| d{\mathcal{C}}(x) d{\mathcal{C}}(y) \right] = 0,$$ and we use Markov’s inequality to conclude.
Large box approximation
-----------------------
We now state the approximation result.
\[prop:largeboxapproximation\] Let ${\mathsf{P}}$ be a stationary process such that ${\mathcal{F}_{\beta}}({\mathsf{P}})$ is finite. Let $\epsilon > 0$ be fixed. There exists an “energy threshold” $M$, depending on ${\mathsf{P}}$ and $\epsilon$, such that for any $s > 0$ small enough, for all integer $R$ large enough (depending on ${\mathsf{P}}, \epsilon, s$), there exists an event $\EveR$ depending only on the restriction of configurations to ${\Lambda}_R$ such that $$\label{PEverbig}
{\mathsf{P}}(\EveR) \geq 1 - \epsilon,$$ and there exists a probability measure $\tPR$ on ${\mathrm{Conf}}({\Lambda}_R)$ such that:
- The configurations have exactly $2R$ points in ${\Lambda}_R$, $\tPR$-almost surely.
- The restrictions to $\Old = {\Lambda}_{R(1-s)}$ of the process $\tPR$ and of the process ${\mathsf{P}}$ conditioned on $\EveR$ coincide. More precisely, if $|_\Old$ denotes the restriction of a configuration to $\Old$, then we have $${|_\Old}_*\tPR = {|_\Old}_*{\mathsf{P}}[\cdot|\EveR]\;.$$
Moreover, the following inequalities hold: $$\begin{aligned}
\label{goodEner}
\frac{1}{|{\Lambda}_R|} \E_{\tPR} \left[ {\mathcal{W}^{\mathrm{int}}}({\mathcal{C}}; {\Lambda}_R) \right] & \leq {\mathcal{W}}({\mathsf{P}}) + \epsilon, \\
\label{goodEnt} \frac{1}{|{\Lambda}_R|} {\mathrm{Ent}}\left[\tPR | {\mathsf{\Pi}}_{|{\Lambda}_R} \right] & \leq {\mathcal{E}}({\mathsf{P}}) + \epsilon.\end{aligned}$$
For $\tPR$-a.e. configuration $\Cscr$:
- The points in $\New$ follow the conclusions of Claim \[claim:position\].
- Letting $S$ be the random variable such that $$\label{def:SC}
x_0(\Cscr) = \Xx_{R+S(\Cscr)}(\Cscr),$$ where $x_0(\Cscr)$ denotes the first point of $\Cscr$ to the right of $0$ as in and $\Xx_{k}(\Cscr)$ denotes the $k$-th point of $\Cscr$ to the right of $-R$, as in Claim \[claim:position\], we have $$\label{controlS}
|S| \leq s M^{1/2} R^{1/2}, \quad \tPR\text{-a.s.}$$
- We have $$\label{discrpostscr}
\sum_{i=-R}^{R} \left| \Discr_{[-R,i]}(\Cscr) \right | \preceq s^2 R^{3/2}.$$
Let us comment on , . For a given configuration with $2R$ points in ${\Lambda}_R$, we consider two different ways of enumerating points: with $\Xx_0$ starting from the left, as in Claim \[claim:position\], and with $x_0$ being the first positive point, as in . There should be roughly $R$ points in $[-R, 0]$, so $x_0$ should correspond to $\Xx_R$. The “shift” $S$ defined in quantifies the deviation to this guess, and affirms that this deviation is small.
The proof relies on the screening procedure of Proposition \[prop:screening\]. First, we need to show that we can apply this procedure with high probability.
### Finding a good event. {#finding-a-good-event. .unnumbered}
Let $\epsilon > 0$ be given.
**The truncation.** First, since ${\mathsf{P}}$ is stationary, we know from that, for any $R > 1$, $$\lim_{\eta \to 0} \E_{{\mathsf{P}}}\left[ \frac{1}{|{\Lambda}_R|} \frac{1}{\c} \int_{{\Lambda}_R \times \R} |\El_{\eta}|^2 + \log \eta \right] = {\mathcal{W}}({\mathsf{P}}),$$ and the limit is uniform in $R$ because the quantity (the expectation) whose limit is taken is independent of $R$, by stationarity. In particular for $\eta$ small enough (depending only on ${\mathsf{P}}$), for all $R > 1$, we have $$\begin{aligned}
\label{etatruncA}
& \left| \E_{{\mathsf{P}}}\left[ \frac{1}{|{\Lambda}_R|} \frac{1}{\c} \int_{{\Lambda}_R \times \R} |\El_{\eta}|^2 + \log \eta \right] - {\mathcal{W}}({\mathsf{P}})\right| \leq \frac{\epsilon}{100},\end{aligned}$$ Another important feature for choosing a suitable truncation is that the error terms in the monotonicity estimates of Lemma \[lem:monoto\] should be small. By Claim \[claim:goodtruncationerror\], we may take some $\eta > 0$ (depending on ${\mathsf{P}}, \epsilon$) such that $${\mathsf{P}}\lp \EventTrunc\lp R, \eta, \epsilon \rp \rp \geq 1 - \frac{\epsilon}{100}.$$ and by taking $\eta$ smaller if needed, we may also impose that $$\label{choiceeta}
\ErMon(\eta) < \frac{\epsilon}{100}, \quad \eta < \etac,$$ where $\ErMon$ is the error term in , , and $\etac$ is the constant in Proposition \[prop:screening\].
**The energy threshold.** We now fix an energy threshold $M$. We take it high enough such that most configurations have an energy at most $M$. By Claim \[claim:goodfield\], for $M$ large enough depending on ${\mathsf{P}}, \epsilon$, for any $s \in \left(0, \frac{1}{4} \right)$, for $R$ large enough (depending on ${\mathsf{P}}, s, M, \epsilon$) we have $$\label{goodfield}
{\mathsf{P}}\left( \GoodField(R, s, M, \eta) \right) \geq 1 - \frac{\epsilon}{100}.$$
**The size of the screening zone.** Next, we fix the size of the screening zone. The screening procedure affects a small region near the endpoints, and has an energy cost proportional to this size, controlled by $s$. We choose $s \in \lp 0, \frac{1}{4}\rp$ small enough such that for all $R$ $$\label{choiceofs}
\ErrScr \leq \frac{\epsilon R}{100},$$ where $\ErrScr$ is the error term in . This choice of $s$ depends only on $\eta, M$.
**Various controls.** Claim \[claim:gooddiscr\] ensures that, for $R$ large enough, depending on ${\mathsf{P}}, \epsilon, s$, we have $${\mathsf{P}}\left( \EventDiscr(R, s, \eta) \right) \geq 1 - \frac{\epsilon}{100}.$$
**The good event.** We can now define the event.
Let $\epsilon, \eta, M, s$ be as chosen in the previous paragraphs, and for all $R > 0$ we let $\EveR$ be the intersection of the events $$\EventTrunc(R, \eta, \epsilon), \GoodField(R, s, M, \eta), \EventDiscr(R, s, \eta),$$
By a union bound, for $R$ large enough depending only on ${\mathsf{P}}, \epsilon, s$ we obtain ${\mathsf{P}}\left(\EveR\right) \geq 1 - \epsilon$, which ensures that holds. By Claim \[claim:goodfield\], $\GoodField$ is measurable with respect to the restriction to ${\Lambda}_R$. Moreover $\EventTrunc, \EventDiscr$ clearly depend only on ${\mathcal{C}}$ in ${\Lambda}_R$. Therefore, $\EveR$ depends only on the restriction to ${\Lambda}_R$.
### Defining the modified process {#defining-the-modified-process .unnumbered}
Let ${\mathcal{C}}$ be in $\EveR$. By definition of $\GoodField(R, s, M, \eta)$, there exists an electric field $\El$ compatible with ${\mathcal{C}}$ on $\Old \subset {\Lambda}_R$ and satisfying $$\begin{aligned}
\Mec := \int_{\{-R', R'\} \times [-R, R]} |\El_{\eta}|^2 & \leq M, \\
\Hec := \frac{1}{s^4 R} \int_{[-R, R] \times \R \backslash (-s^2 R, s^2 R)} |\El|^2 & \leq 1.\end{aligned}$$ In the sequel, we choose a field $\El$ that satisfies these conditions *and has minimal energy* i.e. such that $\Mec$ is minimal. In particular, this electric field $\El$ depends only on ${\mathcal{C}}$ in $\Old$.
The assumption is also satisfied by definition of $\EventDiscr(R, s, \eta)$, which is included in $\NoPoint(R, s, \eta)$ as defined in .
Hence for all $R$ large enough (depending on ${\mathsf{P}}, \epsilon, s$) we may apply the screening procedure to ${\mathcal{C}}$. We let $\Phiscr_{s, \eta, R}({\mathcal{C}})$ be the resulting probability measure on ${\mathrm{Conf}}({\Lambda}_R)$.
We define $\tPR$ as the mixture of the $\Phiscr_{s, \eta, R}({\mathcal{C}})$ for ${\mathcal{C}}$ in $\EveR$, weighted by ${\mathsf{P}}({\mathcal{C}})$, i.e. $$\label{def:tPR}
\tPR := \frac{1}{{\mathsf{P}}(\EveR)} \int \Phiscr_{s, \eta, R}({\mathcal{C}}) \1_{{\mathcal{C}}\in \EveR} d{\mathsf{P}}({\mathcal{C}}).$$ Equivalently, if $F$ is a bounded function on ${\mathrm{Conf}}({\Lambda}_R)$, we let $$\label{def:tPRbis}
\E_{\tPR} [F] = \frac{1}{{\mathsf{P}}(\EveR)} \int \E_{\Phiscr_{s, \eta, R}({\mathcal{C}})} [F] \1_{{\mathcal{C}}\in \EveR} d{\mathsf{P}}({\mathcal{C}}).$$
We may already check that:
- By construction, the screened configurations all have $2R$ points in ${\Lambda}_R$.
- Still by construction, for a given ${\mathcal{C}}\in \EveR$, all the configurations $\Cscr$ in $\Phiscr_{s, \eta, R}({\mathcal{C}})$ coincide with ${\mathcal{C}}$ inside $\Old = {\Lambda}_{R'}$. Thus, from the definition , we see that the restrictions to ${\Lambda}_{R'}$ of the process $\tPR$ and of the process ${\mathsf{P}}$ conditioned to $\EveR$ coincide, in other words $${|_{{\Lambda}_{R'}}}_*\tPR = {|_{{\Lambda}_{R'}}}_*{\mathsf{P}}[\cdot|\EveR]\;.$$
### Energy estimate {#energy-estimate .unnumbered}
We prove . Let ${\mathcal{C}}$ be in $\EveR$. By the screening procedure, for all $\Cscr$ in the support of $\Phiscr_{s, \eta, R}({\mathcal{C}})$, there exists a screened electric field whose energy, in view of the controls , on the energy after screening, and by the choice of $s$ as in , is bounded as follows: $$\label{errorscreeningwellchosen}
\int_{{\Lambda}_R \times \R} |\Escr_{\eta}|^2 \leq \int_{{\Lambda}_R \times [-R, R]} |\El_{\eta}|^2 + \frac{\epsilon R}{100}.$$ Let $\Eloc$ be the local electric field generated by $\Cscr$ in ${\Lambda}_R$, as in definition . By “minimality of the local energy”[^3], see e.g. [@leble2017large Lemma 3.10], we have $$\label{minimalite}
\int_{\R \times \R} |\Eloc_{\eta}|^2 \leq \int_{{\Lambda}_R \times \R} |\Escr_{\eta}|^2.$$ Combining and , we get $$\label{ElocvsEeta}
\int_{\R \times \R} |\Eloc_{\eta}|^2 \leq \int_{{\Lambda}_R \times [-R, R]} |\El_{\eta}|^2 + \frac{\epsilon R}{100}.$$ Next, the monotonicity inequality reads $${\mathcal{W}^{\mathrm{int}}}(\Cscr; {\Lambda}_R) \leq \frac{1}{\c} \int_{\R \times \R} |\Eloc_{\eta}|^2 + |\Cscr| \log \eta + |\Cscr| \ErMon(\eta) + \ErTrun(\Cscr, \eta; {\Lambda}_R).$$ By construction we have $\ErTrun({\mathcal{C}}, \eta; {\Lambda}_R) \leq \frac{\epsilon R}{100}$, and we know by that the screening procedure does not augment the truncation error. Moreover, $\eta$ has been chosen such that $\ErMon(\eta) \leq \frac{\epsilon}{100}$. Using the fact that $|\Cscr| = |{\Lambda}_R| = 2R$, we obtain $$\label{WintscrElocscr}
{\mathcal{W}^{\mathrm{int}}}(\Cscr; {\Lambda}_R) \leq \frac{1}{\c} \int_{\R \times \R} |\Eloc_{\eta}|^2 + 2R \log \eta + \frac{4 R \epsilon}{100}.$$ Combining and yields $$\label{WintscrEeta}
{\mathcal{W}^{\mathrm{int}}}(\Cscr; {\Lambda}_R) \leq \frac{1}{\c} \int_{{\Lambda}_R \times [-R, R]} |\El_{\eta}|^2 + 2R \log \eta + \frac{5 R \epsilon}{100}.$$
Moreover, we obtain, in view of the control on ${\mathcal{W}^{\mathrm{int}}}$ and the definition of $\tPR$, $$\begin{gathered}
\label{calculenergy}
\E_{\tPR} \left[ {\mathcal{W}^{\mathrm{int}}}(\Cscr; {\Lambda}_R) \right] = \frac{1}{{\mathsf{P}}(\EveR)} \int \E_{\Phiscr_{s, \eta, R}({\mathcal{C}})} \left[{\mathcal{W}^{\mathrm{int}}}(\Cscr;\Lambda_R)\right] \1_{\EveR}({\mathcal{C}}) d{\mathsf{P}}({\mathcal{C}}) \\
\leq \frac{1}{{\mathsf{P}}(\EveR)} \int \left( \frac{1}{\c} \int_{{\Lambda}_R \times [-R, R]} |\El_{\eta}|^2 + 2R \log \eta \right) \1_{\EveR}({\mathcal{C}}) d{\mathsf{P}}({\mathcal{C}}) + \frac{5 R \epsilon}{100}.\end{gathered}$$ Then since ${\mathsf{P}}(\EveR) \geq 1 - \epsilon$, and holds, we write $$\frac{1}{{\mathsf{P}}(\EveR)} \int \left( \frac{1}{\c} \int_{{\Lambda}_R \times [-R, R]} |\El_{\eta}|^2 + 2R \log \eta \right) \1_{\EveR}({\mathcal{C}}) d{\mathsf{P}}({\mathcal{C}}) \leq |{\Lambda}_R| (1 + \epsilon) \left( {\mathcal{W}}({\mathsf{P}}) + \epsilon \right),$$ so we obtain $$\frac{1}{|{\Lambda}_R|} \E_{\tPR} \left[ {\mathcal{W}^{\mathrm{int}}}(\Cscr; {\Lambda}_R) \right] \leq {\mathcal{W}}({\mathsf{P}}) + \epsilon {\mathcal{W}}({\mathsf{P}}) + O(\epsilon).$$ Up to working with a smaller $\epsilon'$ instead of $\epsilon$, we may ensure that $\epsilon' {\mathcal{W}}({\mathsf{P}}) + O(\epsilon') \leq \epsilon$ and so holds.
### Entropy estimate {#entropy-estimate .unnumbered}
We prove . Let us first recall a general disintegration principle.
Let $X,Y$ be Polish spaces, let $\mu,\pi$ be two probability measures on $X$ and let $T:X\to Y$ be a measurable map. Let $\bar\mu=T_{*}\mu$ and $\bar\pi=T_{*}\pi$ and let $\mu(\cdot|T=y)$ and $\pi(\cdot|T=y)$ denote the regular conditional probabilities, i.e. we have the disintegration $$\begin{aligned}
\mu(A)= \int_Y\mu(A|T=y)d\bar\mu(y)\;,\qquad\pi(A)= \int_Y\pi(A|T=y)d\bar\pi(y)\quad \forall A\;. \end{aligned}$$ Then we have that $$\begin{aligned}
\label{eq:ent-desint}
{\mathrm{Ent}}[\mu|\pi]= {\mathrm{Ent}}[\bar\mu|\bar\pi] +\int_Y {\mathrm{Ent}}\big[\mu(\cdot|T=y)|\pi(\cdot|T=y)\big]d\bar\mu(y).\end{aligned}$$
This can be verified by a direct computation. Since the relative entropy w.r.t. a probability is non-negative, we have in particular $$\begin{aligned}
\label{eq:ent-push}
{\mathrm{Ent}}[\mu|\pi]\geq {\mathrm{Ent}}[\bar\mu|\bar\pi]\;.\end{aligned}$$
Let us briefly summarize some important properties from the construction of $\tPR$:
- First, we condition ${\mathsf{P}}$ to the “good event” $\EveR$, with ${\mathsf{P}}\lp \EveR \rp \geq 1 - \epsilon$, then we restrict to ${\Lambda}_R$. More precisely, we consider the process $$\label{def:chP}
\chP_{R}={|_{\Lambda_R}}_*{\mathsf{P}}[\cdot|\EveR],$$ namely the restriction to ${\Lambda}_R$ of the process ${\mathsf{P}}$ conditioned to $\EveR$.
- Then we apply the screening procedure. Given a configuration ${\mathcal{C}}$ in $\EveR$, we find (see Figure \[fig:screenstatement\]):
- A number $n = n({\mathcal{C}})$, such that $n({\mathcal{C}}) + |{\mathcal{C}}|_\Old| = 2R$.
- Points $p_1({\mathcal{C}}), \dots, p_n({\mathcal{C}})$ in $\New$ such that the segments $[p_i - \eta, p_i + \eta]$ are disjoint and do not intersect $[-R', R']$.
- We define $\Phiscr({\mathcal{C}})$ by adding $n({\mathcal{C}})$ points placed independently, one into each interval of the form $[p_i - \eta, p_i + \eta]$.
- We define $\tPR$ as the weighted average of $\Phiscr({\mathcal{C}})$ for ${\mathcal{C}}$ in $\EveR$, as in .
#### **Step 1:**
We first estimate the entropy of $\chP_{R}$ and claim that, for some constant $C$ (independent of ${\mathsf{P}}, R$) $$\label{eq:ent-restr}
{\mathrm{Ent}}[\chP_{R} |{\mathsf{\Pi}}_{{\Lambda}_R}] \leq (1+2\epsilon) {\mathrm{Ent}}[{\mathsf{P}}_{{\Lambda}_R}|{\mathsf{\Pi}}_{{\Lambda}_R}] + C.$$ Indeed, if $\rho$ is the density of ${\mathsf{P}}_{{\Lambda}_R}:={|_{{\Lambda}_R}}_*{\mathsf{P}}$ with respect to ${\mathsf{\Pi}}_{{\Lambda}_R}$, the density of $\chP_{R}$ is given by $\1_E \rho/{\mathsf{P}}(E)$, where we set $E=\EveR$ for brevity. Thus we have $${\mathrm{Ent}}[\chP_{R}|{\mathsf{\Pi}}_{{\Lambda}_R}] = \int_E\frac{\rho}{{\mathsf{P}}(E)}\log\lp\frac{\rho}{{\mathsf{P}}(E)}\rp d{\mathsf{\Pi}}_{{\Lambda}_R} =
- \log {\mathsf{P}}(E) + \frac{1}{{\mathsf{P}}(E)}\left( {\mathrm{Ent}}\left[{\mathsf{P}}_{{\Lambda}_R}|{\mathsf{\Pi}}_{{\Lambda}_R}\right] -\int_{E^c}\rho\log\rho d{\mathsf{\Pi}}_{{\Lambda}_R} \right).$$ By Jensen’s inequality, we have $$\begin{aligned}
\frac1{{\mathsf{\Pi}}_{{\Lambda}_R}(E^c)}\int_{E^c}\rho\log\rho d{\mathsf{\Pi}}_{{\Lambda}_R} \geq \frac{{\mathsf{P}}_{{\Lambda}_R}(E^c)}{{\mathsf{\Pi}}_{{\Lambda}_R}(E^c)}\log\frac{{\mathsf{P}}_{{\Lambda}_R}(E^c)}{{\mathsf{\Pi}}_{{\Lambda}_R}(E^c)} \geq -{e}^{-1}\;.\end{aligned}$$ since $r\log r\geq -{e}^{-1}$. Thus we obtain $${\mathrm{Ent}}[\chP_{R}|{\mathsf{\Pi}}_{{\Lambda}_R}] \leq {\mathrm{Ent}}[{\mathsf{P}}_{{\Lambda}_R}|{\mathsf{\Pi}}_{{\Lambda}_R}] + \frac{1-{\mathsf{P}}(E)}{{\mathsf{P}}(E)}{\mathrm{Ent}}[{\mathsf{P}}_{{\Lambda}_R}|{\mathsf{\Pi}}_{{\Lambda}_R}] - \log {\mathsf{P}}(E) + \frac{1}{e} \frac{{\mathsf{\Pi}}_{{\Lambda}_R}(E^c)}{{\mathsf{P}}_{{\Lambda}_R}(E)}.$$ Using the fact that ${\mathsf{P}}_{{\Lambda}_R}(E)={\mathsf{P}}(E)\geq 1-\epsilon$, and assuming e.g. $\epsilon \leq {\frac{1}{2}}$, it yields .
#### **Step 2:**
Let us now compare ${\mathrm{Ent}}[\chP_{R}|{\mathsf{\Pi}}_{{\Lambda}_R}]$ and ${\mathrm{Ent}}[\tPR|{\mathsf{\Pi}}_{{\Lambda}_R}]$.
Let ${\mathrm{Conf}}({\Lambda}_R)$ denote the set of point configurations in ${\Lambda}_R$ and ${\mathrm{Conf}}(\Old)$ denote the set of point configurations on $\Old = {\Lambda}_{R'}$. Using the disintegration formula for the map $T : {\mathrm{Conf}}(\Lambda_R) \to {\mathrm{Conf}}(\Old)$ given by the restriction ${\mathcal{C}}\mapsto {\mathcal{C}}\big|_\Old$ we have $$\label{eq:ent-est1}
\begin{split}
{\mathrm{Ent}}[\chP_{R}|{\mathsf{\Pi}}_{{\Lambda}_R}] &= {\mathrm{Ent}}[T_{*}\chP_{R}|T_{*} {\mathsf{\Pi}}_{{\Lambda}_R}] +
\int_{{\mathrm{Conf}}(\Old)}{\mathrm{Ent}}\big[\chP_R(\cdot|T=\mathcal
C)|{\mathsf{\Pi}}_{{\Lambda}_R} (\cdot|T=\mathcal C)\big]d(T_{*}\chP_R)(\mathcal C)\;,\\
{\mathrm{Ent}}[\tPR|{\mathsf{\Pi}}_{{\Lambda}_R}] &= {\mathrm{Ent}}[T_{*}\tPR|T_{*} {\mathsf{\Pi}}_{{\Lambda}_R}] +
\int_{{\mathrm{Conf}}(\Old)}{\mathrm{Ent}}\big[\tPR(\cdot|T=\mathcal
C)|{\mathsf{\Pi}}_{{\Lambda}_R}(\cdot|T=\mathcal C)\big]d(T_{*}\tPR)(\mathcal C)\;,\\
\end{split}$$ Note that $T_{*}{\mathsf{\Pi}}_{{\Lambda}_R}={\mathsf{\Pi}}_{\Old}$. Further, we have by construction that $T_{*}\tPR=T_{*}\chP_R$, hence the first terms in the decomposition of the entropy of $\chP_R$ and $\tPR$ coincide, and it remains to compare the conditional entropies.
By definition of the Poisson point process, a random configuration drawn from ${\mathsf{\Pi}}_{{\Lambda}_R}(\cdot|T=\mathcal C)$ consists of $\mathcal C$ in $\Old$ plus a sample of points in $\New =\Lambda_R \setminus \Old$ drawn from ${\mathsf{\Pi}}_{\New}$, the Poisson process in $\New$. Denoting by $\tPR^{\mathcal C}$ the image of $\tPR(\cdot|T=\mathcal C)$ under restriction to $\New$ we have $${\mathrm{Ent}}\left[\tPR(\cdot|T=\mathcal C)|{\mathsf{\Pi}}_{{\Lambda}_R}(\cdot|T=\mathcal C)\right] = {\mathrm{Ent}}\left[\tPR^{\mathcal C}|{\mathsf{\Pi}}_{\New}\right]\;.$$
By construction, a random configuration drawn from $\tPR(\cdot|T=\mathcal C)$ consists of $\mathcal C$ in $\Old$ plus $n(\mathcal C)$ points placed independently uniformly in the intervals $(p_i(\mathcal C)-\eta,p_i(\mathcal C)+\eta)$. To calculate the entropy with respect to the Poisson process in $\New$, let us introduce another disintegration concerning the number of points placed in $\New$. For a set $I\subset\R$ we set $${\mathrm{Conf}}(I)^{(n)}:= \{ {\mathcal{C}}\in {\mathrm{Conf}}(I), |{\mathcal{C}}| = n \}.$$ Given a probability $\sigma$ on ${\mathrm{Conf}}(I)$ we consider the restrictions $$\begin{aligned}
\label{eq:restrictions}
\sigma_n := \frac{1}{\sigma(n)}\sigma\big|_{{\mathrm{Conf}}(I)^{(n)}}, \quad \text{ where we denote } \sigma(n) :=\sigma\big({\mathrm{Conf}}(I)^{(n)}\big)\;.\end{aligned}$$ By definition of the Poisson point process, we have ${\mathsf{\Pi}}_{I}(n)= e^{-|I|}\frac{|I|^n}{n!}$, and the law of ${\mathsf{\Pi}}_{I, n}$ is that of $n$ independent points placed uniformly in $I$. Using again the disintegration formula , we get $$\begin{aligned}
\label{eq:ent-nb-points}
{\mathrm{Ent}}[\sigma|{\mathsf{\Pi}}_{I}] = \sum_{n=0}^\infty \sigma(n)
\Big[\log\Big(\frac{\sigma(n)}{{\mathsf{\Pi}}_{I}(n)}\Big)+{\mathrm{Ent}}[\sigma_n|{\mathsf{\Pi}}_{I, n}]\Big]\;.\end{aligned}$$
Recalling that $\tPR^{\mathcal C}$ is the law of $n(\mathcal C)$ independent points each placed uniformly into $n(\mathcal C)$ intervals of size $2\eta$ we find that for $n=n(\mathcal C)$ the entropy of $\left(\tPR^{\mathcal C}\right)_n$ with respect to ${\mathsf{\Pi}}_{\New, n}$ is given by the entropy of the uniform distribution on the set $$\begin{aligned}
B_{n,\eta}:=\bigcup_{\sigma\in S_n}(p_{\sigma(1)}-\eta,p_{\sigma(1)}+\eta)\times\cdots\times(p_{\sigma(n)}-\eta,p_{\sigma(n)}+\eta)\end{aligned}$$ relative to the uniform distribution on $\New^n$, i.e. $$\begin{aligned}
{\mathrm{Ent}}\left[\left(\tPR^{\mathcal C}\right)_n|{\mathsf{\Pi}}_{\New, n}\right] = \log\frac{|\New|^n}{|B_{n,\eta}|}
= \log \frac{|\New|^n}{n!(2\eta)^n} \;.\end{aligned}$$ By construction we have $\tPR^{\mathcal C}(m)=1$ for $m=n(\mathcal C)$ and $0$ otherwise. Thus we obtain from that for $n=n(\mathcal C)$ $$\begin{aligned}
\label{eq:ent-est3}
{\mathrm{Ent}}[\tPR^{\mathcal C}|\Pi_{\New}] &= \log \frac{1}{e^{-|\New|}|\New|^n/n!} + {\mathrm{Ent}}\left[\left(\tPR^{\mathcal C}\right)_n|{\mathsf{\Pi}}_{\New, n}\right]
\\\nonumber
&= \log\frac{e^{|\New|}n!}{|\New|^n} +\log \frac{|\New|^n}{n!(2\eta)^n} = |\New|-n(\mathcal C)\log 2\eta
\\\nonumber
&= 2sR -\left(2R-{\mathcal{C}}(\Old)\right)\log 2\eta.\end{aligned}$$ Finally, note that ${\mathrm{Ent}}[\chP_R(\cdot|T=\mathcal C)|{\mathsf{\Pi}}_{{\Lambda}_R}(\cdot|T=\mathcal C)]\geq 0$. Thus combining and we finally obtain $$\begin{aligned}
{\mathrm{Ent}}[\tPR|{\mathsf{\Pi}}_{{\Lambda}_R}] \leq {\mathrm{Ent}}[\chP_R|{\mathsf{\Pi}}_{{\Lambda}_R}] + 2sR -\log 2\eta \E_{\chP_R}\left[2R - |{\mathcal{C}}_{\Old} | \right]. \end{aligned}$$ By construction (see ), we have $$\left|2R - |{\mathcal{C}}_{\Old}| \right| \leq 2sR+sR^{1/2} \leq 3sR.$$ We thus obtain $$\label{eq:ent-est4}
{\mathrm{Ent}}[\tPR|{\mathsf{\Pi}}_{{\Lambda}_R}] \leq {\mathrm{Ent}}[\chP_R|{\mathsf{\Pi}}_{{\Lambda}_R}] + 5|\log \eta|sR.$$
Combining and , we obtain (for $R$ large, the constant error term in can be absorbed in the dominant error terms) $$\frac{1}{|{\Lambda}_R|} {\mathrm{Ent}}[\tPR|{\mathsf{\Pi}}_{{\Lambda}_R}] \leq \frac{1}{|{\Lambda}_R|} {\mathrm{Ent}}[{\mathsf{P}}_{{\Lambda}_R} | {\mathsf{\Pi}}_{{\Lambda}_R}] + \frac{2 \epsilon}{|{\Lambda}_R|} {\mathrm{Ent}}[{\mathsf{P}}_{{\Lambda}_R} | {\mathsf{\Pi}}_{{\Lambda}_R}] + \frac{5}{2}|\log \eta|s.$$ First, we observe that the error term $|\log \eta|s $ is of the same order as $\ErrScr$ and $s$ was chosen small enough so that $\ErrScr$ is small. On the other hand the limit defining the specific relative entropy is non-decreasing (it follows from a super-additivity argument, see e.g. [@friedli2017statistical Cor. 6.77] where $S$ there is the opposite of our entropy), so $$\frac{1}{|{\Lambda}_R|} {\mathrm{Ent}}[{\mathsf{P}}_{{\Lambda}_R} | {\mathsf{\Pi}}_{{\Lambda}_R}] + \frac{2 \epsilon}{|{\Lambda}_R|} {\mathrm{Ent}}\left[{\mathsf{P}}_{{\Lambda}_R} | {\mathsf{\Pi}}_{{\Lambda}_R}\right] \leq {\mathcal{E}}({\mathsf{P}}) + 2\epsilon {\mathcal{E}}({\mathsf{P}}),$$ and up to working with a smaller $\epsilon'$ we may ensure that $2 \epsilon' {\mathrm{Ent}}[{\mathsf{P}}_{{\Lambda}_R} | {\mathsf{\Pi}}_{{\Lambda}_R}] + 5 |\log \eta| s \leq \epsilon$, hence we obtain .
### Additional considerations
Since holds, and since Claim \[claim:position\] ensures that there are $sR \pm s M^{1/2} R^{1/2}$ points in $[-R, -R']$, we see that there are $R \pm s M^{1/2} R^{1/2}$ points of $\Cscr$ in $[-R,0]$ and thus holds.
Finally, follows from combining the definition of $\TotDiscr$ in $\Old$ with the discrepancy estimates in $\New$ , .
Displacement convexity {#sec:displacement}
======================
Labelling on the orthant
------------------------
In this section, it is more convenient to treat the laws of random $2R-$point configurations in ${\Lambda}_R$ as probability measures on $2R$-tuples. In order to identify a configuration and a $2R$-tuple, we introduce the orthant ${\mathcal{O}_R}$ $$\label{ON}
{\mathcal{O}_R}:= \left\lbrace (\Xx_1, \dots, \Xx_{2R}) \in \left({\Lambda}_N\right)^{2R} \ | \ \Xx_1 \leq \dots \leq \Xx_{2R}\right\rbrace.$$
- We denote by ${{\mathrm{Conf}}_R}$ the set of point configurations in $\Lambda_R$ with exactly $2R$ points.
- We let ${\mathsf{B}_R}$ be the Bernoulli point process with $2R$ points in ${\Lambda}_R$, which is a probability measure on ${{\mathrm{Conf}}_R}$.
- We let ${\mathbf{L}_R}$ be the normalized Lebesgue measure on ${\mathcal{O}_R}$.
We define the label map ${\pi_R}$ as $$\label{def:piN}
{\pi_R}: \begin{cases}
{{\mathrm{Conf}}_R}& \longrightarrow {\mathcal{O}_R}\subset \left(\Lambda_R\right)^{2R} \\
{\mathcal{C}}= \sum_{i=1}^{2R} \delta_{\Xx_i} & \mapsto (\Xx_1, \dots, \Xx_{2R}), \quad \Xx_1 \leq \dots \leq \Xx_{2R}.
\end{cases}$$
\[lem:piNbijec\] The label map ${\pi_R}$ is well-defined and is a bijection from ${{\mathrm{Conf}}_R}$ to ${\mathcal{O}_R}$, up to a subset of measure zero (for ${\mathsf{B}_R}$) in the source and a subset of measure zero (for ${\mathbf{L}_R}$) in the target.
The label map ${\pi_R}$ as in is well-defined and injective on the set of simple configurations, when all points are distinct. The image of this set in ${\mathcal{O}_R}$ is the set of strictly ordered $2R$-tuples $\Xx_1 < \Xx_2 < \dots < \Xx_{2R}$. It is clear that the first set has full measure in ${{\mathrm{Conf}}_R}$ (for ${\mathsf{B}_R}$), and the second set has full measure in ${\mathcal{O}_R}$ (for ${\mathbf{L}_R}$).
The following fact is easy to check.
\[lem:pushBNpiN\] ${\mathsf{B}_R}$ and ${\mathbf{L}_R}$ are images of each other under ${\pi_R}$ and its inverse, i.e. ${{\pi_R}}_*{\mathsf{B}_R}={\mathbf{L}_R}$ and $({\pi_R}^{-1})_*{\mathbf{L}_R}={\mathsf{B}_R}$.
\[prop:piNentro\] Let ${\mathsf{P}}$ be a point process on ${\Lambda}_R$ with almost surely $2R$ points, i.e. ${\mathsf{P}}({{\mathrm{Conf}}_R})=1$, and let ${\widehat{\mathsf{P}}}$ be its image under ${\pi_R}$, i.e. ${\widehat{\mathsf{P}}}=(\pi_R)_*{\mathsf{P}}$. Then, there exists a constant $c_R$ depending only on $R$ such that $$\label{LNversusPoisson}
{\mathrm{Ent}}[{\mathsf{P}}| {\mathsf{\Pi}}_{{\Lambda}_R}] = {\mathrm{Ent}}[{\widehat{\mathsf{P}}}| {\mathbf{L}_R}] + c_R\;.$$
Since ${\mathsf{P}}$ has almost surely $2R$ points and noting that the restriction of ${\mathsf{\Pi}}_{{\Lambda}_R}$ to ${{\mathrm{Conf}}_R}$ is ${\mathsf{B}_R}$, we infer from the desintegration formula that $$\begin{aligned}
{\mathrm{Ent}}[{\mathsf{P}}|{\mathsf{\Pi}}_{{\Lambda}_R}] &= \log \frac{e^{2R}(2R)!}{|2R|^{2R}} + {\mathrm{Ent}}[{\mathsf{P}}|{\mathsf{B}_R}]\end{aligned}$$ Since $\pi_R$ is essentially bijective, we have that ${\mathsf{P}}=(\pi^{-1}_R)_*{\widehat{\mathsf{P}}}$. By Lemma \[lem:pushBNpiN\], ${\mathbf{L}_R}=(\pi_R)_*{\mathsf{B}_R}$ and ${\mathsf{B}_R}=(\pi_R^{-1})_*{\mathbf{L}_R}$. Now, the claim follows immediately from .
The optimal transportation map for the quadratic cost
-----------------------------------------------------
### Definition of the transportation map {#definition-of-the-transportation-map .unnumbered}
Let $\hPz, \hPu$ be two probability measures on ${\mathcal{O}_R}$, with finite relative entropy with respect to ${\mathbf{L}_R}$. In particular, they are absolutely continuous with respect to ${\mathbf{L}_R}$.
There exists a map $\TN : {\mathcal{O}_R}\to \R^{2R}$ satisfying:
- The push-forward of $\hPz$ by $\TN$ is $\hPu$.
- There exists a convex function $\varphi : {\mathcal{O}_R}\to \R$ such that $\TN = \nabla \varphi$.
This follows from [@villani2003topics Theorem 2.12], because $\hPz$ and $\hPu$ are both compactly supported probability measures on $\R^{2R}$, absolutely continuous with respect to the Lebesgue measure.
The first item expresses the fact that $\TN$ transports $\hPz$ onto $\hPu$. The fact that $\TN$ is the gradient of a convex function expresses the *optimality* of $\TN$ for the quadratic cost. We will use the optimal character of $\TN$ only once, to argue that the relative entropy is displacement convex.
For any $t \in [0,1]$, we introduce the displacement interpolate ${\widehat{\mathsf{P}}}^t$ as $$\label{def:hPt}
{\widehat{\mathsf{P}}}^t := \left((1-t) \id + t \TN \right)_* \hPz,$$ which is consistent with the previously defined $\hPz, \hPu$ in the cases $t = 0$ and $t=1$.
\[defi:hip\] Let $\Xz = (\Xx_1^\z, \dots, \Xx_{2R}^\z)$ be a $2R$-tuple in ${\mathcal{O}_R}$, and let $\Xu = (\Xx_1^\u, \dots, \Xx^{\u}_{2N})$ in ${\mathcal{O}_R}$ be the image of $\Xz$ by the transportation map $\TN$. We introduce $\Xhal$ as $$\label{def:Xhal}
\Xhal := \frac{\id + \TN}{2}(\Xz) := \left( \frac{\Xx^\z_1 + \Xx^\u_1}{2}, \dots, \frac{\Xx_{2R}^\u + \Xx^\u_{2R}}{2} \right).$$
### Effect of the transport on the discrepancy {#effect-of-the-transport-on-the-discrepancy .unnumbered}
\[lem:transportondiscr\] Let $R > 0$, let $\Cz, \Cu$ be two point configurations with $2R$ points in ${\Lambda}_R$, and let $\Chal$ be the half-interpolate of $\Cz$ and $\Cu$. For any $r$ in $[-R,R]$, we have $$\label{discrpartransport}
\left| \Discr_{[-R,r]}(\Chal) \right| \leq \max \left( \left| \Discr_{[-R,r]}(\Cz)\right|, \left|\Discr_{[-R,r]}(\Cu)\right| \right)$$
The construction of $\Chal$ clearly implies, for any $r \in [-R,R]$, $$\min \left(\left|\Cz_{[-R,r]}\right|, \left|\Cu_{[-R,r]}\right|\right) \leq \left|\Ch_{[-R,r]}\right| \leq \max\left(\left|\Cz_{[-R,r]} \right|, \left|\Cu_{[-R,r]}\right|\right),$$ and the results follows from the definition of $\Discr$.
In particular, since the configurations that we construct all satisfy , this is also the case for all the configurations obtained by interpolation.
### Displacement convexity of the entropy {#displacement-convexity-of-the-entropy .unnumbered}
\[lem:disconventr\] The map $t \mapsto {\mathrm{Ent}}\left[{\widehat{\mathsf{P}}}^t| {\mathbf{L}_R}\right]$ is convex on $[0,1]$.
This is a well-known “displacement convexity” result, which was proven in the pioneering paper [@MCCANN1997153]\[Theorem 2.2\].
This is the only moment where we use the fact that $\TN$ is the *optimal* transport map for the quadratic cost.
In particular, we obtain $$\label{convexentropy1}
{\mathrm{Ent}}\left[{\widehat{\mathsf{P}}}^{{\frac{1}{2}}} | {\mathbf{L}_R}\right] \leq {\frac{1}{2}}\left( {\mathrm{Ent}}\left[{\widehat{\mathsf{P}}}^{1} | {\mathbf{L}_R}\right] + {\mathrm{Ent}}\left[{\widehat{\mathsf{P}}}^{2} | {\mathbf{L}_R}\right] \right).$$ If $\Pz_R, \Phal_R, \Pu_R$ are the push-forward of $\hPz, {\widehat{\mathsf{P}}}^{{\frac{1}{2}}}, \hPu$ by ${\pi_R}^{-1}$, using Lemma \[prop:piNentro\] we see that $$\label{convexentropy2}
{\mathrm{Ent}}\left[\Phal_R | {\mathsf{\Pi}}_{{\Lambda}_R} \right] \leq {\frac{1}{2}}\left( {\mathrm{Ent}}\left[\Pz_R | {\mathsf{\Pi}}_{{\Lambda}_R} \right] + {\mathrm{Ent}}\left[\Pu_R | {\mathsf{\Pi}}_{{\Lambda}_R} \right] \right).$$
Since $x \mapsto x\ln(x)$ is a strictly convex function on $(0, +\infty)$, the results of [@MCCANN1997153] imply that the inequality in , and thus in , is strict unless $\Pz_R = \Pu_R$. Inspecting the proof could possibly yield some quantitative bound, but, here, we rely instead on the energy term to get a tractable convexity inequality along the displacement interpolation.
Convexity inequality for the energy
-----------------------------------
Although [@MCCANN1997153]\[Section 3\] deals with the displacement convexity of energies similar to our, the setting is different and we cannot directly apply these results. Moreover, we crucially need a quantitative convexity estimate, which is the aim of this section.
### Convexity for the logarithmic interaction {#convexity-for-the-logarithmic-interaction .unnumbered}
We start by stating an elementary inequality:
Let $x,y > 0$. We have $$\label{convlog}
- \log \left( \frac{x+y}{2} \right) \leq \frac{- \log x - \log y}{2} - \frac{(x-y)^2}{8(x^2 + y^2)}.$$
The function $f : x \mapsto - \log (x)$ is convex on $(0, + \infty)$, since its second derivative is given by $f''(x) = \frac{1}{x^2} > 0$. The result follows by applying the following elementary inequality $$f\left( \frac{x+y}{2} \right) \leq \frac{f(x) + f(y)}{2} - \frac{1}{8} (x-y)^2 \inf_{[x,y]} f''(c).$$
As an immediate consequence, if $(\Xx^\z_1, \dots, \Xx^\z_{2R})$ and $(\Xx^\u_1, \dots, \Xx^\u_{2R})$ are two $2R$-tuples of points in ${\mathcal{O}_R}$, we have $$\begin{gathered}
\label{convexityenergyA}
\sum_{1 \leq i < j \leq 2R} \frac{ - \log (\Xx_j^\z - \Xx_i^\z) - \log (\Xx_j^\u - \Xx_i^\u) }{2} + \log \left( \frac{ \Xx_j^\z + \Xx_j^\u}{2} - \frac{\Xx_i^\z + \Xx_i^\u}{2} \right) \\
\geq \frac{1}{8} \sum_{1 \leq i < j \leq 2R} \frac{ \left( (\Xx_j^\z - \Xx_i^\z) - (\Xx_j^\u - \Xx_i^\u) \right)^2 }{ (\Xx_j^\z - \Xx_i^\z)^2 + (\Xx_j^\u - \Xx_i^\u)^2}.\end{gathered}$$
Let $\X = (\Xx_1, \dots, \Xx_{2R})$ be a $2R$-tuple in ${\mathcal{O}_R}$. We define the gaps of $\X$, “counted from the left” as $$\label{def:gaps2}
\hGap_{i} := \Xx_{i+1} - \Xx_{i} \ (1 \leq i \leq 2R-1)$$ We denote by $\hGap_i(\X)$ the $i$-th such gap of a given $\X \in {\mathcal{O}_R}$.
Let $\X = (\Xx_1, \dots, \Xx_{2R})$ be a $2R$-tuple of points in ${\mathcal{O}_R}$, and let ${\mathcal{C}}= \pi_{R}^{-1}(\X)$. We introduce the notation $$\label{intX}
{\mathrm{Int}}[\X] := \sum_{1 \leq i < j \leq 2R} - \log|\Xx_i - \Xx_j| = \iint_{x < y} - \log |x- y| d{\mathcal{C}}(x) d{\mathcal{C}}(y).$$
\[prop:strictXhal\] Let $\Xz, \Xu, \Xhal$ be as in Definition \[defi:hip\]. We have: $$\label{strictXhal}
{\mathrm{Int}}[\Xhal] \leq \frac{{\mathrm{Int}}[\Xz] + {\mathrm{Int}}[\Xu]}{2} - \frac{1}{8} \sum_{i = 1}^{2R - 1} \frac{\left|\hGap_i(\Xz) - \hGap_i(\Xu) \right|^2}{\left(\hGap_i(\Xz)\right)^2 + \left(\hGap_i(\Xu)\right)^2}.$$
It is a straightforward consequence of , by writing $$\sum_{1 \leq i < j \leq 2R} \frac{ \left( \left( (\Xx_j^\z - \Xx_i^\z \right) - \left(\Xx_j^\u - \Xx_i^\u \right) \right)^2 }{ \left(\Xx_j^\z - \Xx_i^\z \right)^2 + \left(\Xx_j^\u - \Xx_i^\u)\right)^2}
\geq \sum_{i=1}^{2R-1} \frac{ \left( \left(\Xx_{i+1}^\z - \Xx_i^\z \right) - \left(\Xx_{i+1}^\u - \Xx_i^\u\right) \right)^2 }{ \left(\Xx_{i+1}^\z - \Xx_i^\z\right)^2 + \left(\Xx_{i+1}^\u - \Xx_i^\u\right)^2},$$ and using the notation .
For $R > 0$, we introduce the “background potential” as $$\label{def:VR}
\V_R(t) := \int_{-R}^R \log |t-s| ds.$$
An explicit computation yields $$\begin{aligned}
\V_R(t)&=\int_{-R}^t\log(t-s)ds + \int_t^R\log(s-t)ds
=\Big[(R+t)\log(R+t)-(R+t)+(R-t)\log(R-t)-(R-t)\Big]\;.
\end{aligned}$$ and thus $$\label{VRtsec}
\V'_R(t)=\log(R+t)+\log(R-t), \quad \V''_R(t) = \frac{1}{R+t} + \frac{1}{R-t}.$$ As a straightforward consequence, we obtain:
\[lem:BFconvex\] $\V_R$ is a convex function on ${\Lambda}_R$.
Let ${\mathcal{C}}$ be in ${{\mathrm{Conf}}_R}$, and let $\X = {\pi_R}({\mathcal{C}})$. We have $$\label{defWintIntVN}
{\mathcal{W}^{\mathrm{int}}}({\mathcal{C}}; {\Lambda}_R) = 2 {\mathrm{Int}}[\X] + 2 \sum_{i=1}^{2R} \V_R(\Xx_i) + \const_R,$$ where ${\mathrm{Int}}$ is as in , $\V_R$ is as above, and $\const_R$ is a constant depending only on $R$.
It follows from the definition of ${\mathcal{W}^{\mathrm{int}}}$, by expanding the quadratic term $(d{\mathcal{C}}- dx)(d{\mathcal{C}}- dy)$, $$\begin{gathered}
{\mathcal{W}^{\mathrm{int}}}({\mathcal{C}}; {\Lambda}_R) = \iint_{{\Lambda}_R \times {\Lambda}_R \setminus \diamond} - \log |x-y| (d{\mathcal{C}}-dx) (d{\mathcal{C}}- dy) = 2 \iint_{x < y} - \log |x-y| d{\mathcal{C}}(x) d{\mathcal{C}}(y) \\ + 2 \iint_{{\Lambda}_R \times {\Lambda}_R \setminus \diamond} \log |x-y| d{\mathcal{C}}(x) dy
+ \iint_{{\Lambda}_R \times {\Lambda}_R \setminus \diamond} - \log |x-y| dx dy,\end{gathered}$$ and using definition .
### The convexity inequality {#the-convexity-inequality .unnumbered}
Let $\Cz, \Cu$ be in ${{\mathrm{Conf}}_R}$, $\Xz, \Xu$ be their image by ${\pi_R}$, let $\Xhal$ be as in .
- We define the “gain” term $\GGain(\Cz, \Cu; {\Lambda}_R)$ as: $$\label{def:GGain}
\GGain(\Cz, \Cu; {\Lambda}_R) := \sum_{i = 1}^{2R - 1} \frac{\left|\hGap_i(\Xz) - \hGap_i(\Xu) \right|^2}{\left(\hGap_i(\Xz)\right)^2 + \left(\hGap_i(\Xu)\right)^2}.$$
- We define the “background field contribution” term $\BF(\Cz, \Cu; {\Lambda}_R)$ as $$\label{def:BF}
\BF(\Cz, \Cu; {\Lambda}_R) := 2 \sum_{i=1}^{2R} \lp \V_R(\Xx_i^\hh) - {\frac{1}{2}}\left( \V_R(\Xx_i^\z) + \V_R(\Xx_i^\u) \right) \rp.$$
\[prop:Wintconvex\] Let $\Cz, \Cu$ be in ${{\mathrm{Conf}}_R}$, we have: $$\label{Wintconvex}
{\mathcal{W}^{\mathrm{int}}}(\Chal;\Lambda_R) \leq {\frac{1}{2}}\left({\mathcal{W}^{\mathrm{int}}}(\Cz;\Lambda_R) + {\mathcal{W}^{\mathrm{int}}}(\Cu;\Lambda_R)\right) - \frac{1}{4} \GGain(\Cz, \Cu; {\Lambda}_R).$$
We combine and , and we use the notation introduced in , . We obtain $${\mathcal{W}^{\mathrm{int}}}(\Chal;\Lambda_R) \leq {\frac{1}{2}}\left({\mathcal{W}^{\mathrm{int}}}(\Cz;\Lambda_R) + {\mathcal{W}^{\mathrm{int}}}(\Cu;\Lambda_R)\right) - \frac{1}{4} \GGain(\Cz, \Cu; {\Lambda}_R) + \BF(\Cz, \Cu; {\Lambda}_R).$$ Since $\V_R$ has been observed to be a convex function on ${\Lambda}_R$ (see Lemma \[lem:BFconvex\]), from the definition of $\BF$ in we see readily that $\BF(\Cz, \Cu; {\Lambda}_R) \leq 0$, so in particular[^4] holds.
The interpolate process {#sec:interpolateprocess}
=======================
Let $\Pz, \Pu$ be two minimisers of ${\mathcal{F}_{\beta}}$, and assume that $\Pz \neq \Pu$. Let $\g$ be the “gain” given by Proposition \[prop:gaps2\], i.e. $1/R$ times the right hand side of .
Definition of the interpolate process
-------------------------------------
For any $s >0$ and for $R$ large enough (depending on $s, \Pz, \Pu$), we apply the “large box approximation” of Proposition \[prop:largeboxapproximation\] to $\Pz, \Pu$ with $\epsilon = \frac{\g}{100}$. We let $\tPzR, \tPuR$ be the processes on ${\Lambda}_R$ obtained this way. Further, we let $\hPz_R=(\pi_R)_*\tPzR$, $\hPu_R=(\pi_R)_*\tPuR$ be the correponding measures on ${\mathcal{O}_R}$ obtained by pushforward via the labeling map $\pi_R$. We let $\hat Q_R$ be an optimal coupling of $\hPz_R$ and $\hPu_R$ given by the optimal transport map $\Trans_R$ and let $\hat \Phal_R$ be the half-interpolate measure on ${\mathcal{O}_R}$ obtained from displacement interpolation. Finally, we let $\Phal_R=(\pi_R^{-1})_*\tilde \Phal_R$ be the corresponding half-interpolate process on ${\Lambda}_R$ and let $Q_R=(\pi_R^{-1}\times\pi_R^{-1})_*\tilde Q_R$ be the corresponding coupling of $\tPzR, \tPuR$.
The energy gain is proportional to the size of the segment
----------------------------------------------------------
\[lem:gainisextensive\] Taking $R$ large enough, we have $$\label{gainisextensive}
\E_{\Q_R} \left[\GGain( \cdot, \cdot ; {\Lambda}_R) \right] \geq {\frac{1}{2}}\g R.$$
Let $\Cscrz, \Cscru$ be in the support of $\tPzR, \tPuR$. Let $\Sz, \Su$ be the quantities, as in , relating the two ways of enumerating points, i.e. such that $$x_0(\Cscrz) = \Xx^\z_{R + \Sz}, \quad x_0(\Cscru) = \Xx^\u_{R + \Su},$$ which means $\Cscrz$ (resp. $\Cscru$) has $R + \Sz$ (resp. $R + \Su$) points in $[-R,0]$.
From we know that (up to choosing $s$ small enough with respect to $M$) $\Sz, \Su$ are bounded by ${\frac{1}{2}}R^{1/2}$, so $S = \Su - \Sz$ is bounded by $R^{1/2}$. We may write, using the definition and switching indices: $$\GGain(\Cscrz, \Cscru; {\Lambda}_R) \geq \sum_{i=-R/2}^{R/2} \frac{\left|\Gap_i(\Cscrz) - \Gap_{i+S}(\Cscru) \right|^2}{\left(\Gap_i(\Cscrz)\right)^2 + \left(\Gap_{i+S}(\Cscru)\right)^2}.$$ Note that the right hand side only depends on the restrictions of $\Cscrz,\Cscru$ to $\Old={\Lambda}_{R'}$. Let us denote by $\bar \Q_R$ the image of the coupling $\Q_R$ under restriction to $\Old$. Since the screening procedure does not change the configurations in $\Old$ by construction, $\bar \Q_R$ is a coupling of the restrictions to $\Old$ of the stationary processes $\Pz$ and $\Pu$ conditioned on an event of probability $\geq 1 - \frac{\g}{100}$. $$\E_{\Q_R} \left[\GGain( \cdot, \cdot ; {\Lambda}_R) \right] \geq \E \left[ \sum_{i=-R/2}^{R/2} \frac{\left|\Gap_i(\Cz) - \Gap_{i+S}(\Cu) \right|^2}{\left(\Gap_i(\Cz)\right)^2 + \left(\Gap_{i+S}(\Cu)\right)^2} \right]\;,$$ where $(\Cz,\Cu)$ is distributed according to $\bar \Q_R$. In view of Corollary \[coro:gaps\], the last expression is bounded below by ${\frac{1}{2}}\g R$, which yields the result.
Energy of the interpolate process
---------------------------------
For $s$ small enough (depending on $\Pz, \Pu$), for $R$ large enough (depending on $\Pz, \Pu, s$) we have $$\label{EPhalWint}
\frac{1}{|{\Lambda}_R|} \E_{\Phal} \left[{\mathcal{W}^{\mathrm{int}}}(\Chal; {\Lambda}_R)\right] \leq {\frac{1}{2}}\left( {\mathcal{W}}(\Pz) + {\mathcal{W}}(\Pu) \right) - \frac{\g}{4}$$
We combine
- , from the approximation procedure, that controls the energies $\E_{\Pz}[{\mathcal{W}^{\mathrm{int}}}], \E_{\Pu}[{\mathcal{W}^{\mathrm{int}}}]$ in terms of ${\mathcal{W}}(\Pz)$, ${\mathcal{W}}(\Pu)$.
- , from the transportation argument, that expresses $\E_{\Phal}[{\mathcal{W}^{\mathrm{int}}}]$ in terms of $\E_{\Pz}[{\mathcal{W}^{\mathrm{int}}}], \E_{\Pu}[{\mathcal{W}^{\mathrm{int}}}]$ and the “gain” term.
- Lemma \[lem:gainisextensive\], saying that the “gain” is proportional to the volume.
Conclusion: proof of Theorem \[theo:main\] {#sec:conclusionproof}
==========================================
Constructing a better candidate
-------------------------------
We now turn the interpolate process, which is supported in a large segment $[-R,R]$, into a stationary point process on the whole real line. We proceed in two steps: first we paste independent copies of the interpolate process, and then we average over translations in the original segment. This construction was used in [@leble2016logarithmic] for similar purposes.
### Pasting copies on the line {#pasting-copies-on-the-line .unnumbered}
Let $\{K_i : = {\Lambda}_R - 2Ri \}_{i \in \Z}$ be a tiling of $\R$ by translates of the interval ${\Lambda}_R$, and let $\{\Chal_i\}_{i \in \Z}$ be independent random variables identically distributed according to $\Phal$. We let $\Paste_R$ be the map $$\Paste_R := \{ {\mathcal{C}}_i \}_{i \in \Z} \mapsto \sum_{i \in \Z} ({\mathcal{C}}_i - 2iR),$$ that creates a configuration on $\R$ from a collection of configurations on ${\Lambda}_R$ by copying one on each interval $K_i$. We let $\Ptot$ be the push-forward of $\{\Chal_i\}_{i \in \Z}$ by $\Paste_R$.
### The stationary candidate: averaging over an interval {#the-stationary-candidate-averaging-over-an-interval .unnumbered}
We let $\Pav$ be the law of the point process defined by “averaging $\Ptot$ over translations in ${\Lambda}_R$”. More precisely, we let $\Pav$ be the law of the point process defined by duality as follows: for any bounded measurable test function $F$, $$\label{def:Pav}
\E_{\Pav} [F] := \E_{\Ptot} \left[ \frac{1}{|{\Lambda}_R|} \int_{-R}^R F({\mathcal{C}}- t) dt \right].$$ Now, by construction, $\Pav$ is a stationary process of intensity $1$, and there are always $2L \pm 4R$ points in any interval of length $2L$.
Claims about the new process and conclusion
-------------------------------------------
The process $\Pav$ satisfies $$\label{fbetaPav} {\mathcal{F}_{\beta}}(\Pav) < {\frac{1}{2}}\lp {\mathcal{F}_{\beta}}(\Pz) + {\mathcal{F}_{\beta}}(\Pu) \rp.$$
Recall that ${\mathcal{F}_{\beta}}(\Pav) := \beta {\mathcal{W}}(\Pav) + {\mathcal{E}}(\Pav)$, we evaluate each term below.
### The specific relative entropy {#the-specific-relative-entropy-1 .unnumbered}
\[claim:SREPtot\] We have $\displaystyle{{\mathcal{E}}(\Pav) \leq {\frac{1}{2}}\left( {\mathcal{E}}(\Pz) + {\mathcal{E}}(\Pu) \right) + \frac{\g}{50}}$.
By construction, the process $\Ptot$ is made of independent copies of $\Phal$ on ${\Lambda}_R$, so we have $${\mathcal{E}}(\Pav) = \lim_{M \to \infty} \frac{1}{|{\Lambda}_M|} {\mathrm{Ent}}\left [\Pav_{{\Lambda}_M} | {\mathsf{\Pi}}_{{\Lambda}_M} \right] = \frac{1}{|{\Lambda}_R|} {\mathrm{Ent}}\left[ \Phal | {\mathsf{\Pi}}_{{\Lambda}_R} \right].$$ Using and (with $\epsilon = \frac{\g}{100}$) yields the claim.
### The intrinsic energy {#the-intrinsic-energy .unnumbered}
\[claim:intrinsicPtot\] We have $\displaystyle{\Wintg(\Pav) \leq {\frac{1}{2}}\left( {\mathcal{W}}(\Pz) + {\mathcal{W}}(\Pu) \right) - \frac{\g}{10}}$.
Since $\Wintg$ is defined in as a $\liminf$, it is enough to show that $$\liminf_{M \to \infty} \frac{1}{|{\Lambda}_{MR}|} \E_{\Phal} \left[ {\mathcal{W}^{\mathrm{int}}}\left({\mathcal{C}}; {\Lambda}_{MR} \right) \right] \leq {\frac{1}{2}}\left( {\mathcal{W}}(\Pz) + {\mathcal{W}}(\Pu) \right) - \frac{\g}{10}.$$ By the definition , $\Pav$ is a uniform mixture of processes that we will denote by $\Ptotz$ for $z \in {\Lambda}_R$, where $\Ptotz$ is the process $\Ptot$ translated by $z$. For $z \in [-R,R]$ and $M \geq 10$ fixed, a configuration of $\Ptotz$ in ${\Lambda}_{MR}$ consists of:
- $M-2$ “full” configurations ${\mathcal{C}}_1, \dots, {\mathcal{C}}_{M-2}$ obtained from independent copies of $\Phal$, and supported in the intervals ${\Lambda}_R - z - 2Ri$ for $i \approx - M/2 \dots M/2$,
- Two “partial” configurations $\Cleft$ and $\Cright$, where $\Cleft$ is supported in $[-MR, -MR + R - z]$ and $\Cright$ in $[MR - R - z, MR]$.
When we compute the interaction energy ${\mathcal{W}^{\mathrm{int}}}({\mathcal{C}}; {\Lambda}_{MR})$, we obtain: $$\label{ABCDEFG}
{\mathcal{W}^{\mathrm{int}}}({\mathcal{C}}; {\Lambda}_{MR}) = A + B + C + E + F + G,$$
- $A$ is the sum of the interactions of a configuration ${\mathcal{C}}_i$ with itself, for $i = 1, \dots, M-2$.
- $B$ is the interaction between $\Cleft$ and $\Cright$.
- $C$ is the interaction between $\Cleft$ and ${\mathcal{C}}_1$; the interaction between $\Cright$ and ${\mathcal{C}}_{M-2}$.
- $D$ is the interaction of $\Cleft$ with itself, and the interaction of $\Cright$ with itself.
- $E$ is the sum of the interactions between ${\mathcal{C}}_i$ and ${\mathcal{C}}_{i+k}$ for $i = 1, \dots M-2$ and $2 \leq k \leq M-2-i$, i.e. between *non-neighboring* “full” configurations.
- $F$ is all the interactions between two neighboring “full” configurations ${\mathcal{C}}_i, {\mathcal{C}}_{i+1}$.
- $G$ is the sum of the interactions between $\Cleft$ and all non-neighboring “full” configurations (${\mathcal{C}}_i$ for $i = 2, \dots M-2$), and the interactions between $\Cright$ and all non-neighboring “full” configurations (${\mathcal{C}}_i$ for $i = 1, \dots, M-3$).
**The term $A$ (self-interaction of full configurations).** Taking the expectation under $\Ptotz$, in view of , we have $$\frac{1}{M |{\Lambda}_R|} \E_{\Ptotz}(A) = \frac{M-2}{M} \frac{1}{|{\Lambda}_R|} \E_{\Phal} \left[ {\mathcal{W}^{\mathrm{int}}}\left({\mathcal{C}}; {\Lambda}_{R} \right) \right] \leq {\frac{1}{2}}\left( {\mathcal{W}}(\Pz) + {\mathcal{W}}(\Pu) \right) - \frac{\g}{4} + o_M(1).$$
It remains to study all the other terms and to show that they are either negligible with respect to $M$ as $M \to \infty$ or only yield a perturbation of order $MR$, that can be made arbitrarily small through the choice of $s$.
**The terms $B, C, D$.** We may already observe that the interaction between $\Cleft$ and $\Cright$ is bounded by $O(R^2 \log M)$, and is thus $o(M)$. So are the interactions between $\Cleft$ and ${\mathcal{C}}_1$ or between $\Cright$ and ${\mathcal{C}}_{M-2}$. We may also bound the self-interaction of $\Cleft$, $\Cright$ by a quantity independent of $M$. We thus have $$\frac{1}{M |{\Lambda}_R|} \E_{\Ptotz}(B + C+ D) = o_M(1).$$
**A priori bound on fluctuations.** To control the other “pairwise” interactions we rely on the following bound expressing fluctuations in terms of discrepancies.
\[lem:contrfluctgab\] Let $[a,b]$ be an interval of $\R$, let $g$ be a $C^1$ function on $[a,b]$ and let ${\mathcal{C}}$ be a point configuration on $[a,b]$. We have $$\label{fluctgab}
\int_a^b g(x) (d{\mathcal{C}}- dx) \preceq \sum_{k=a}^b \|g'\|_{\infty} \left( \left|\Discr_{[a,k]}({\mathcal{C}})\right| + \left|\Discr_{[k,k+1]}({\mathcal{C}})\right| + 1 \right) + \|g\|_{\infty} \left| \Discr_{[a,b]}({\mathcal{C}}) \right|.$$
In particular, if $[a,b] = {\Lambda}_R$ and ${\mathcal{C}}$ is a configuration with $2R$ points in ${\Lambda}_R$, the last term in the right-hand side of vanishes.
This follows from splitting $[a,b]$ into intervals of length $1$, using a Taylor’s expansion of $g$ on each interval, and using a summation by part. We refer e.g. to [@Leble:2018aa]\[Prop. 1.6\].
We use the notation $\tD_{R,k}({\mathcal{C}})$ for the summand in , $$\label{def:tDRk}
\tD_{R,k}({\mathcal{C}}) := \left| \Discr_{[-R,k]}({\mathcal{C}}) \right| + \left|\Discr_{[k, k+1]}({\mathcal{C}})\right| + 1.$$
Since the double integral defining ${\mathcal{W}^{\mathrm{int}}}$ (see ) involves the “fluctuation” terms $(d{\mathcal{C}}- dx)(d{\mathcal{C}}- dy)$, using we can derive the following control on the interaction between two configurations living in two non-neighboring copies of ${\Lambda}_R$:
\[prop:apriori\] Let ${\mathcal{C}}^a, {\mathcal{C}}^b$ be two configurations with $2R$ points supported on the intervals ${\Lambda}_R^a := {\Lambda}_R - 2Ra$, ${\Lambda}_R^b := {\Lambda}_R - 2Rb$ respectively, with $|a-b| \geq 2$. Then $$\begin{gathered}
\label{interactiondiscrepanci}
\iint_{{\Lambda}_R^a \times {\Lambda}_R^b} - \log |x -y| (d{\mathcal{C}}^a(x) - dx) (d{\mathcal{C}}^b(y) - dy) \\
\preceq \frac{1}{|a-b|^2 R^2} \sum_{k=0}^{2R} \sum_{j=0}^{2R} \tD_{R, k}({\mathcal{C}}^a +2Ra) \tD_{R,j}({\mathcal{C}}^b + 2Rb),\end{gathered}$$ where $\tD_{R,k}({\mathcal{C}})$ is as in .
We apply Lemma \[fluctgab\] twice (once for each variable) to $h(x,y) = - \log|x-y|$, bounding the second derivative of $h$, for $x$ in ${\Lambda}_R^a$ and $y$ in ${\Lambda}_R^b$, by $\frac{1}{|a-b|^2 R^2}$.
**The term $E$.** Combining the result of Proposition \[prop:apriori\] with the discrepancy bounds (which are still valid for the interpolate configurations, as observed in Lemma \[lem:transportondiscr\]), we may bound the interaction between ${\mathcal{C}}_i$ and ${\mathcal{C}}_{i+k}$, for $k \geq 2$, by $$\iint_{{\Lambda}_R^i \times {\Lambda}_R^{i+k}} - \log |x -y| (d{\mathcal{C}}_i(x) - dx) (d{\mathcal{C}}_{i+k}(y) - dy)
\preceq \frac{1}{k^2 R^2} \left(s^2 R^{3/2}\right)^2.$$ For a given $i$, we thus have $$\sum_{k \geq 2} \iint_{{\Lambda}_R^i \times {\Lambda}_R^{i+k}} - \log |x -y| (d{\mathcal{C}}_i(x) - dx) (d{\mathcal{C}}_{i+k}(y) - dy) \preceq s^4 R,$$ and summing again over $i = 1, \dots, M-2$, we bound the term $E$ in by $$E \preceq s^4 M R,$$ which is a contribution of order $MR$ that can be made arbitrarily small by taking $s$ small.
The remaining interactions, between all neighbors ${\mathcal{C}}_i,{\mathcal{C}}_{i+1}$; and between $\Cleft, \Cright$ and the ${\mathcal{C}}_i$’s, also yield arbitrarily small contributions. The argument is similar: we use Lemma \[lem:contrfluctgab\] and the discrepancy estimates, with two small modifications.
**The term $F$.** To treat neighbors, here is a sketch of the argument: take two configurations ${\mathcal{C}}^{l}$ (on the left) and ${\mathcal{C}}^{r}$ (on the right) living in $[-R, 0]$ and $[0,R]$ respectively. In view of Lemma \[lem:contrfluctgab\], we write their interaction as $$\begin{gathered}
\label{interaction}
\Interaction = \iint_{[-R, 0] \times [0, R]} - \log |x-y| (d{\mathcal{C}}^{l} - dx) (d{\mathcal{C}}^{r} - dy)
\\
\preceq \sum_{i,j=1}^R \frac{1}{(i+j)^2} |\Discr_{[0, -i]}({\mathcal{C}}^{l}) | \cdot |\Discr_{[0,j]}({\mathcal{C}}^{r})|,\end{gathered}$$ where we have only kept the leading order discrepancy in . The term $\frac{1}{(i+j)^2}$ comes from the fact that the second derivative $\partial_{xy} - \log|x-y|$ is controlled, for $x$ in $[-i-1, -i]$ and $y$ in $[j, j+1]$, by $\frac{1}{(i+j)^2}$. We are looking for a bound of the type $\Interaction \preceq o_s(1) R$, since then we have $O(M)$ pairs of neighbors, each yielding a contribution $o_s(1) R$, so the term $F$ would be of order $o_s(1) MR$, which is enough for our purposes.
Note that, to simplify, when writing we have assumed that the configurations were separated by a distance $1$ (there is no $i,j=0$ term), in reality the contribution of the terms at distance $\leq 1$ is bounded by $O(1)$ and we can forget about them. They key point is to use the fact that the second moment of the discrepancy in a segment is very small compared to the size of the segment, as expressed by Lemma \[lem:variancesublinear\]. In particular, we may write $$\label{discrisverysmall}
|\Discr_{[0, -i]}({\mathcal{C}}^{l}) | \cdot |\Discr_{[0,j]}({\mathcal{C}}^{r})| \ll \sqrt{i} \sqrt{j}.$$ Then, we estimate the double sum of as follows: $$\begin{gathered}
\sum_{i,j=1}^R \frac{1}{(i+j)^2} |\Discr_{[0, -i]}({\mathcal{C}}^{l}) | \cdot |\Discr_{[0,j]}({\mathcal{C}}^{r})| \preceq \sum_{i=1}^R \sum_{j = i}^R \frac{1}{(i+j)^2} |\Discr_{[0, -i}({\mathcal{C}}^{l}) | \cdot |\Discr_{[0,j]}({\mathcal{C}}^{r})|
\\
\ll \sum_{i=1}^R \sum_{j = i}^R \frac{1}{(i+j)^2} \sqrt{i} \sqrt{j}
\preceq \sum_{i=1}^R \sum_{j = i}^{2i} \frac{i}{i^2} + \sum_{i=1}^R \sqrt{i} \sum_{j = 2i}^{R} \frac{1}{j^2} \sqrt{j},\end{gathered}$$ and thus a direct computation yields $\Interaction \ll R.$
**The term $G$.** If we apply Lemma \[lem:contrfluctgab\] to $\Cleft$ (or $\Cright$), the boundary term in is not zero, but bounded by $2R \|g \|_{\infty}$. In particular, when estimating the interaction between $\Cleft$ and ${\mathcal{C}}_i$, for $i \geq 2$, we obtain a term similar to above, plus a boundary term: $$\begin{gathered}
\iint - \log |x -y| (d\Cleft(x) - dx) (d{\mathcal{C}}_i(y) - dy)
\preceq \frac{1}{i^2 R^2} \left( \sum_{k=0}^{2R} \tD_{R, k}(\Cleft) \right) \left( \sum_{j=0}^{2R} \tD_{R,j}({\mathcal{C}}_i)\right)\\
+ \frac{1}{i R} 2R \sum_{j=0}^{2R} \tD_{R,j}({\mathcal{C}}_i).\end{gathered}$$ Using the discrepancy bound $\sum_{j=0}^{2R} \tD_{R,j}({\mathcal{C}}_i) \preceq R^{3/2}$, the new term in the right-hand side is controlled by $\frac{1}{i} R^{3/2}$, and the sum of these terms over $i = 2, \dots, M-2$ is thus bounded by $$R^{3/2} \sum_{i=2}^{M-2} \frac{1}{i} \preceq R^{3/2} \log M,$$ which is negligible with respect to $M$.
### The electric energy {#the-electric-energy .unnumbered}
\[claim:electricPtot\] We have ${\mathcal{W}}(\Pav) < {\frac{1}{2}}\left( {\mathcal{W}}(\Pz) + {\mathcal{W}}(\Pu) \right) - \frac{\g}{10}$.
It follows from the previous Claim and the “electric-intrinsic” inequality , which applies here because, by construction, the discrepancy in any interval is bounded by $4R$ (see the remark following immediately ).
### Conclusion of the proof {#conclusion-of-the-proof .unnumbered}
Starting from the assumption that there are two distinct minimisers $\Pz, \Pu$ of ${\mathcal{F}_{\beta}}$, we have constructed a stationary point process $\Pav$ which satisfies $${\mathcal{F}_{\beta}}(\Pav) < {\frac{1}{2}}\lp {\mathcal{F}_{\beta}}(\Pz) + {\mathcal{F}_{\beta}}(\Pu) \rp = \min {\mathcal{F}_{\beta}},$$ which is absurd. Hence, the minimiser of ${\mathcal{F}_{\beta}}$ is unique, which proves the main theorem.
Miscellaneous proofs {#sec:annex}
====================
Proof of Lemma \[lem:variancesublinear\] {#sec:proofvariancesublinear}
----------------------------------------
We follow the argument developed in [@leble2017large Section 8.5], with parameters $\dd = 1, \ss = 0, \kk=1, \gamma = 0$, but the proof below is self-contained. In the sequel, equation numbers with a **bold typeface** refer to the corresponding equations in that paper.
We obtain (we could e.g. use ) with $\eta_0 = \frac{1}{4}$ (the precise value of $\eta_0$ is, in fact, irrelevant): $$\label{EPL1}
\frac{1}{2\pi} \E_{{\mathsf{P}}} \left[ \int_{\Lambda_1 \times \R} |\El_{\eta_0}|^2 \right] \leq {\mathcal{W}}(P) + C.$$
For any $T > 0$, we let $\HRT$ be the rectangle $\HRT := {\Lambda}_R \times [-T, T]$. Let us emphasize that here, in contrast to **(8.5)**, we do not yet fix $T$ with respect to $R$. The integration by parts as in **(8.6)** still holds, and we get $$\label{pHRT}
\int_{\partial \HRT} \El_{\eta_0} \cdot \vec{\nu} =- 2\pi \left( \Discr_{{\Lambda}_R}({\mathcal{C}}) + r_{\eta_0} \right),$$ where $r_{\eta_0}$ is an error term bounded by the number of points in a $\eta_0$-neighborhood of $\{-R, +R\}$. It is easy to see that $\E_{{\mathsf{P}}} \left[r_{\eta_0}^2\right]$ is bounded by a constant independent of $R$, and thus, since we are aiming for a $o(R)$ bound, this term is negligible - for simplicity we will forget it.
The main improvement on the existing proof is to observe that the choice $T \in (R, 2R)$ as in **(8.5)** is valid, but slightly sub-optimal. We replace it by the following claim.
\[claim:gutentail\] There exists $f : [0, + \infty) \to [0, + \infty)$ satisfying $$\label{fx}
\lim_{x \to \infty} f(x) = + \infty, \quad \lim_{x \to \infty} \frac{f(x)}{x} = 0$$ and such that $$\label{gutentail}
\lim_{x \to \infty} x \E_{{\mathsf{P}}} \left[ \int_{{\Lambda}_1 \times \{-f(x), f(x)\}} | \El_{\eta_0} |^2 \right] = 0.$$
Let $\tail(x)$ be the quantity $$\label{def:tail}
\tail(x) = \E_{{\mathsf{P}}} \left[ \int_{{\Lambda}_1 \times (\R \backslash (-x, x))} |\El_{\eta_0}|^2 \right].$$ The map $x \mapsto \tail(x)$ is continuous, positive, non-increasing, with $\lim_{x \to + \infty} \tail(x) = 0$. So, introducing the map $$u \mapsto \frac{u}{\sqrt{\tail(u)}},$$ it is continuous on $[0, + \infty)$, increasing, is equal to $0$ at $0$ and tends to $+ \infty$ at $+ \infty$. Thus, for any given $x > 0$ there exists (a unique) $u > 0$ such that $$\label{utailueqx}
\frac{u}{\sqrt{\tail(u)}} = x.$$ Now, by a mean value argument, we may find $v \in [u, 2u]$ such that $$\E_{{\mathsf{P}}} \left[ \int_{{\Lambda}_1 \times \{-v, v\}} | \El_{\eta_0} |^2 \right] \leq \frac{1}{u} \tail(u).$$ We define $f(x)$ as the smallest such real number $v$. It is easy to check that the properties are satisfied, and moreover we have $$x \E_{{\mathsf{P}}} \left[ \int_{{\Lambda}_1 \times \{-f(x), f(x)\}} | \El_{\eta_0} |^2 \right] \leq \frac{x}{u} \tail(u) = \sqrt{\tail(u)},$$ where we have used . This quantity (seen as depending on $x$) tends to $0$ as $x \to \infty$, which proves .
We now take $T = f(R)$ in , and we apply the Cauchy-Schwarz inequality to the left-hand side. Since we are dealing with two (slightly) different lengths $R, f(R)$, it is important to be more precise than in **(8.8)** and to split $\partial H_{R, f(R)}$ as $$\partial \HRT = {\Lambda}_R \times \{-f(R), f(R)\} \cup \{-R, R\} \times [-f(R),f(R)].$$ We obtain $$\left(\int_{\partial H_{R, f(R)}} \El_{\eta_0} \cdot \vec{\nu}\right)^2 \preceq R \int_{{\Lambda}_R \times \{-f(R), f(R)\}} |\El_{\eta_0}|^2 + f(R) \int_{\{-R, R\} \times [-f(R),f(R)]} |\El_{\eta_0}|^2.$$
By stationarity, we have, in view of $$\frac{1}{2\pi} \E_{{\mathsf{P}}} \left[ \int_{\{-R, R\} \times [-f(R),f(R)]} |E_{\eta_0}|^2 \right] = \frac{1}{2\pi} \E_{P} \left[ \int_{{\Lambda}_1 \times [-f(R),f(R)]} |E_{\eta_0}|^2 \right] \leq {\mathcal{W}}(P) + C,$$ and also $$\label{La1f(R)}
\E_{{\mathsf{P}}} \left[ \int_{{\Lambda}_R \times \{-f(R), f(R)\}} |E_{\eta_0}|^2 \right] = R \E_{{\mathsf{P}}} \left[ \int_{{\Lambda}_1 \times \{-f(R), f(R)\}} |E_{\eta_0}|^2 \right].$$ According to , $f(R) = o(R)$, and from we see that the right-hand side of is $o_R(1)$. We thus obtain $$\E_{{\mathsf{P}}} \left[ \left(\int_{\partial H_{R, f(R)}} E_{\eta_0} \cdot \vec{\nu} \right)^2 \right] = R o_R(1) + o(R) \left({\mathcal{W}}(P) + C\right),$$ and thus, in view of , we obtain $\E_{{\mathsf{P}}} \left[ \Discr^2_R({\mathcal{C}}) \right] = o(R)$, which proves .
Proof of Lemma \[lem:gapL2\] {#sec:proofGapL2}
----------------------------
The argument is essentially a re-interpretation of the discrepancy controls as given e.g. in [@petrache2017next Lemma 2.2].
It is enough to show that $$\sum_{i=-\frac{R}{2}}^{\frac{R}{2}} \1_{\Gap_i \geq 10} \left(\Gap_i \right)^2 \preceq \int_{[-R, R] \times \R} |\El_{\eta}|^2.$$
Let $i$ such that $\Gap_i \geq 10$. Let $m = \frac{x_i + x_{i+1}}{2}$ and for $\ell > 0$ let $H_\ell$ be the rectangle $[m-\ell, m+\ell] \times [-\ell, \ell]$. By a mean value argument, we can find $\ell \in \left[\frac{1}{4} \Gap_i, \frac{1}{2} \Gap_i\right]$ such that $$\label{mvgapL2}
\int_{\partial H_{\ell}} |\El_{\eta}|^2 \preceq \frac{1}{\Gap_i} \int_{[x_i, x_{i+1}] \times \R} |\El_{\eta}|^2.$$
On the other hand, using and an integration by parts, we have: $$\int_{\partial H_{\ell}} \El_{\eta} \cdot \vec{n} = + \int_{[m-\ell, m+\ell] \times [-\ell, \ell]} \div(\El_{\eta}) = 4\pi \ell.$$ Indeed, by construction there is no point of ${\mathcal{C}}$ between $m-\ell$ and $m+\ell$.
Let us observe that $\Gap_i \preceq \ell$. Thus, using Cauchy-Schwarz’s inequality and , we obtain $$|\Gap_i|^2 \leq \Gap_i \int_{\partial H_{\ell}} |\El_{\eta}|^2 \preceq \int_{[x_i, x_{i+1}] \times \R} |\El_{\eta}|^2.$$
The results follows by summing on $i$, and observing that $$\sum_{i = -R/2}^{R/2} \int_{[x_i, x_{i+1}] \times \R} |\El_{\eta}|^2 \leq \int_{{\Lambda}_R \times \R} |\El_{\eta}|^2,$$ since by assumption there are at least $R/2$ points on both sides of ${\Lambda}_R$.
Proof of Lemma \[prop:gapsdiff\] {#sec:proofgapsdiff}
--------------------------------
Since $\Pz \neq \Pu$, there exists a continuous function $F : {\mathrm{Conf}}\to \R$ with $\|F\|_{\infty} = 1$, and $c > 0$ such that $$\label{EPFEQF}
\E_\Pz [ F ] - \E_\Pu [F] = c.$$ Furthermore, without loss of generality we may assume that $F$ is *local* in the sense that there exists $N > 0$, such that for any ${\mathcal{C}}$ in ${\mathrm{Conf}}$, $$\label{Fislocal}
F({\mathcal{C}}) = F({\mathcal{C}}\cap {\Lambda}_N).$$ Indeed, by dominated convergence, we have $$\lim_{R \to \infty} \left( \E_\Pz \left[ F( \cdot \cap {\Lambda}_R ) \right] - \E_\Pu \left[ F( \cdot \cap {\Lambda}_R ) \right] \right) = \E_\Pz [ F ] - \E_\Pu [F].$$
For any $M > 0$, we define the function $F_M : {\mathrm{Conf}}\to \R$ as $$F_M({\mathcal{C}}) := \frac{1}{2M} \int_{-M}^M F({\mathcal{C}}+ t) dt.$$ Since, $\Pz, \Pu$ are stationary, we have, for $M$ arbitrary, in view of , $$\label{EPFM}
\E_\Pz [ F_M ] - \E_\Pu [F_M] = c > 0.$$ Also, since $F$ is local and satisfies , we have $$\label{FMislocal}
F_M({\mathcal{C}}) = F_M\left( {\mathcal{C}}\cap {\Lambda}_{M+N} \right).$$
Recall that, in Definition \[def:gaps\], we denote by $x_0({\mathcal{C}})$ the first non-negative point of ${\mathcal{C}}$. Since $\Pz, \Pu$ are assumed to have finite energy, the discrepancy estimate holds and we have $$\lim_{T \to \infty} \Pz\left(x_0({\mathcal{C}}) > T\right) + \Pu\left(x_0({\mathcal{C}}) > T\right) = 0,$$ so we may choose $T$ (depending on $c, \Pz, \Pu$) such that $$\label{conditionT}
\Pz\left(x_0({\mathcal{C}}) > T\right) + \Pu\left(x_0({\mathcal{C}}) > T\right) \leq \frac{c}{100}.$$
Once $T$ is fixed, we choose $M$ such that $$\label{conditionM}
\frac{T}{M} \leq \frac{c}{100}.$$ We may also impose that $M > N$, where $N$ is as in , so in view of we have for any ${\mathcal{C}}$ in ${\mathrm{Conf}}$, $$\label{FMislocalbis}
F_M({\mathcal{C}}) = F_M\left( {\mathcal{C}}\cap {\Lambda}_{2M} \right).$$
\[claim:recenter\] If $x_0({\mathcal{C}}) \leq T$, and holds, we have $$\label{recenter}
|F_M({\mathcal{C}}) - F_M\left({\mathcal{C}}- x_0({\mathcal{C}})\right)| \leq \frac{c}{100}$$
We use the definition of $F_M$, the fact that $F$ is bounded by $1$, and the choice of $M$ with respect to $T$ as in , and compute $$\begin{gathered}
|F_M({\mathcal{C}}) - F_M({\mathcal{C}}- x_0({\mathcal{C}}))| \leq \frac{1}{2M} \left( \int_{M - x_0({\mathcal{C}})}^M |F({\mathcal{C}}+ t)| dt + \int_{-M - x_0({\mathcal{C}})}^{-M} |F({\mathcal{C}}+t)| dt \right) \\
\leq \frac{2 x_0({\mathcal{C}})}{2M} \|F\|_{\infty} \leq \frac{T}{M} \leq \frac{c}{100}.\end{gathered}$$
Since $\Pz, \Pu$ have intensity $1$, for $r \geq 1$ large enough (depending on $M, c, \Pz, \Pu$), we have $$\label{choicer}
\Pz\left(|{\mathcal{C}}\cap {\Lambda}_{8M}| > r \right) + \Pu\left(|{\mathcal{C}}\cap {\Lambda}_{8M}| > r\right) \leq \frac{c}{100}.$$ For $r$ fixed such that holds, we consider the map on $\R^{2r+1}$ defined by $$\label{def:H}
H\left(a_{-r}, \dots, a_{r}\right) := F_M\left( \delta_0 + \sum_{i=1}^r (\delta_{p_i} + \delta_{p_{-i}}) \right),$$ where we let the $p_i$’s be $$\label{def:pi}
p_i := \sum_{k = 1}^i a_k \ (i \geq 1), \quad p_{-i} : = \sum_{k=1}^i - a_{-k} \ (i \geq 1).$$ In other words $H\left(a_{-r}, \dots, a_{r}\right)$ is obtained by applying $F_M$ to the configuration made of one point at $0$ and $2r$ points located at $p_i$ for $|i| = 1, \dots, r$. The idea is that if $a_{-r}, \dots, a_r$ are the first gaps of ${\mathcal{C}}$, then this configuration is made of the first points of ${\mathcal{C}}- x_0({\mathcal{C}})$. Since $F_M$ is bounded by $1$, clearly so is $H$.
\[claim:EPHEQH\] We have $$\label{EPHEQH}
\E_{\Pz} \left[H\left(\Gap_{-r}({\mathcal{C}}), \dots, \Gap_{r}({\mathcal{C}}) \right) \right] - \E_{\Pu} \left[H\left(\Gap_{-r}({\mathcal{C}}), \dots, \Gap_{r}({\mathcal{C}}) \right) \right] > 0.$$
First of all, since $\Pz, \Pu$ have finite energy, the configurations have almost surely infinitely many points in $\R_+$ and $\R_-$, hence the gaps are almost surely all finite and $H\left(\Gap_{-r}({\mathcal{C}}), \dots, \Gap_{r}({\mathcal{C}}) \right)$ is well-defined almost surely.
Since $H$ is bounded by $1$, and since , hold, we have $$\label{EPHEPHcond}
\left| \E_{\Pz} \left[ H\left(\Gap_{-r}, \dots, \Gap_{r} \right) \right] - \E_{\Pz} \left[\1_{x_0({\mathcal{C}}) \leq T} \1_{|{\mathcal{C}}\cap {\Lambda}_{8M}| \leq r} H\left(\Gap_{-r}, \dots, \Gap_{r} \right) \right] \right| \leq \frac{c}{100} + \frac{c}{100},$$ and the same holds for $\Pu$. Knowing that $x_0({\mathcal{C}}) \leq T$, we have by $$\left|F_M ({\mathcal{C}}- x_0({\mathcal{C}})) - F_M({\mathcal{C}}) \right| \leq \frac{c}{100},$$ and since $F_M$ satisfies , we have $$F_M ({\mathcal{C}}- x_0({\mathcal{C}})) = F_M \left( \left({\mathcal{C}}- x_0({\mathcal{C}})\right) \cap {\Lambda}_{2M} \right).$$ Of course, the point configuration ${\mathcal{C}}- x_0({\mathcal{C}})$ can be written in terms of the gaps of ${\mathcal{C}}$ as $${\mathcal{C}}- x_0({\mathcal{C}}) = \delta_0 + \sum_{i=1}^{+\infty} \left(\delta_{p_i} + \delta_{p_{-i}} \right),$$ where the $p_i$’s are as in . Knowing that, moreover, $|{\mathcal{C}}\cap {\Lambda}_{8M}| \leq r$, we see that $$\left({\mathcal{C}}- x_0({\mathcal{C}}) \right) \cap {\Lambda}_{2M} = \left(\delta_0 + \sum_{i=1}^r \delta_{p_i} + \delta_{p_{-i}} \right) \cap {\Lambda}_{2M},$$ and thus, using again , $$\begin{gathered}
F_M \left( \left({\mathcal{C}}- x_0({\mathcal{C}})\right) \cap {\Lambda}_{2M} \right) = F_M \left( \left(\delta_0 + \sum_{i=1}^r (\delta_{p_i} + \delta_{p_{-i}}) \right) \cap {\Lambda}_{2M} \right) = F_M \left(\delta_0 + \sum_{i=1}^r (\delta_{p_i} + \delta_{p_{-i}}) \right)
\\
= H\left(\Gap_{-r}, \dots, \Gap_{r}\right).\end{gathered}$$ We thus obtain by $$\left| \E_{\Pz} \left[\1_{x_0({\mathcal{C}}) \leq T} \1_{|{\mathcal{C}}\cap {\Lambda}_{8M}| \leq r} H\left(\Gap_{-r}, \dots, \Gap_{r} \right) \right] - \E_{\Pz} \left[ \1_{x_0({\mathcal{C}}) \leq T} \1_{|{\mathcal{C}}\cap {\Lambda}_{8M}| \leq r} F_M({\mathcal{C}}) \right] \right| \leq \frac{c}{100},$$ which easily yields, in view of , $$\label{EPHcondEPFM}
\left| \E_{\Pz} \left[ H\left(\Gap_{-r}, \dots, \Gap_{r} \right) \right] - \E_{\Pz} \left[ F_M({\mathcal{C}}) \right] \right| \leq \frac{5c}{100},$$ and the same goes for $\Pu$, which proves the claim.
Finally, we use a density argument in $L^1(\R_{+}^{2r+1})$ to find a test function on $\R^{2r+1}_{+}$ that satisfies *and* is compactly supported and Lipschitz with respect to the $\| \cdot \|_{1}$ norm. By possibly reducing the lower bound in , we can assume the test function to be $1-$Lipschitz.
Proof of Proposition \[prop:gaps2\] {#sec:proofgaps2}
-----------------------------------
We recall that ${\mathrm{Conf}}(\R)$ denotes the space of point configuration on $\R$.
### Detecting the local difference {#detecting-the-local-difference .unnumbered}
Let $c > 0, r \geq 1$, and a function $H : \R_{+}^{2r+1} \to \R$ as given by Lemma \[prop:gapsdiff\], such that $$\label{EPH}
\E_{\Pz} \left[ H(\Gap_{-r}, \dots, \Gap_r) \right] -
\E_{\Pu} \left[ H(\Gap_{-r}, \dots, \Gap_r) \right] \geq c.$$ The function $H$ is compactly supported, so let $L$ be such that $H$ is supported in $[0, L]^{2r +1}$. Without loss of generality, we can take $L > 10$. To clarify notation, let us define $\tH({\mathcal{C}})$ for a configuration ${\mathcal{C}}$ in ${\mathrm{Conf}}(\R)$ as $$\label{def:tH}
\tH({\mathcal{C}}) := H(\Gap_{-r}({\mathcal{C}}), \dots, \Gap_{r}({\mathcal{C}})).$$ Strictly speaking, $\tH$ is not defined everywhere on ${\mathrm{Conf}}(\R)$, but it is well-defined on the set of configurations with at least $r+1$ points on $\R_+$ and $\R_-$, because it ensures that the gaps $\Gap_{-r}({\mathcal{C}}), \dots, \Gap_{r}({\mathcal{C}})$ are all finite.
### A test function to detect the global difference {#a-test-function-to-detect-the-global-difference .unnumbered}
For $R > 1$, we define the function $\hHR$ on ${\mathrm{Conf}}(\R)$ as $$\label{def:hHR} \hHR : {\mathcal{C}}\mapsto \int_{0}^{\Rdix} \tH({\mathcal{C}}- t) dt.$$ Clearly, $\hHR$ is bounded by $\frac{R}{10}$. Since $\Pz, \Pu$ are stationary, we have of course, using , $$\label{EPQhHR}
\E_{\Pz}[ \hHR ] - \E_{\Pu}[ \hHR] \geq \frac{cR}{10}.$$ Strictly speaking, $\hHR$ is not defined everywhere on ${\mathrm{Conf}}(\R)$. It is well-defined on the set configurations with at least $r+1$ points on $\left[\frac{R}{10}, + \infty\right)$ and on $\left(-\infty, 0 \right]$, because it guarantees that the gaps $\Gap_{-r}({\mathcal{C}}-t), \dots, \Gap_{r}({\mathcal{C}}-t)$ are finite for all $t \in \left[0, \frac{R}{10}\right]$.
Now, if $\Q$ is a coupling of $\Pz$ and $\Pu$, we may re-write as $$\label{usingcoupling1}
\E_{\Q} \left[ \int_{0}^{\Rdix} \tH(\Cz - t) dt - \int_{0}^{\Rdix} \tH(\Cu - t) dt \right] \geq \frac{cR}{10}.$$ Since, we work only with a coupling of the restrictions of $\Pz$ and $\Pu$ to ${\Lambda}_R$, we have to restrict the argument above to an event of high probability ensuring in particular that there are at least $r+1$ points in $[-R,0]$ and $[R/10,R]$, so that $\hHR$ depends only on the configuration in ${\Lambda}_R$, see for a precise formulation.
### Strategy of the proof {#strategy-of-the-proof-1 .unnumbered}
To simplify, let us assume that $\tH({\mathcal{C}})$ depends only on the first gap of ${\mathcal{C}}$ (the one with index $0$, between $x_0({\mathcal{C}})$ and $x_1({\mathcal{C}})$, see Definition \[def:gaps\]). Since $H$ is Lipschitz, for any two configurations $\Cz, \Cu$ and any $t$, we thus have $$|\tH(\Cz - t) - \tH(\Cu - t)| \leq \left|\Gap_0(\Cz-t) - \Gap_0(\Cu-t)\right|.$$ As $t$ goes from $0$ to $\frac{R}{10}$, the first gap of $\Cz-t$ will successively correspond to the first gap of $\Cz$, then the second one, etc. up to a gap of order $\approx \frac{R}{10}$ in $\Cz$, and similarly for $\Cu -t$. If the gaps of $\Cz -t$ and $\Cu-t$ were always “aligned”, i.e. for any $t$, the index $k_0$ such that $\Gap_0(\Cz-t) = \Gap_{k_0}(\Cz)$ and the index $k_1$ (defined similarly for $\Cu$) are equal, we would write, using a Fubini-type argument $$\label{preFubiniA}
\int_0^{\frac{R}{10}} |\tH(\Cz - t) - \tH(\Cu - t)| dt \precapprox \sum_{k=0}^{\frac{R}{10}} |\Gap_{k}(\Cz) - \Gap_{k}(\Cu)|,$$ and thus, in view of and , we would get a lower bound of order $R$ on a certain “gap difference”. Getting a lower bound of the type $$\label{nosquareisenough}
\sum_{k=0}^{\frac{R}{10}} |\Gap_{k}(\Cz) - \Gap_{k}(\Cu)| \geq c R$$ would be enough for our purposes.
Compared to this situation, the quantity $\tH({\mathcal{C}})$ depends on more than one gap. However, it only depends on a finite number of gaps (here, at most $2r +1$), and the strategy can be easily adapted. The major complication comes from aligning gaps, we describe it below.
We want to detect a difference between the gaps of $\Cz$ and $\Cu$. In the statement of Proposition \[prop:gaps2\], we include the possibility of a fixed shift $S$, but let us take $S = 0$, and let us try to transform into a bound of the type . Let us consider a typical[^5] situation where $x_0(\Cz) > 0$ and $x_0(\Cu) > 0$. Taking the time $t = 0$ in the integrals, we can bound $$\label{arfarf}
| \tH(\Cz - 0) - \tH(\Cu - 0) | \leq \sum_{i=-r}^{r} |\Gap_i(\Cz - 0) - \Gap_i(\Cu - 0)| = \sum_{i=-r}^{r} |\Gap_i(\Cz) - \Gap_i(\Cu)|.$$ The right-hand side of appears in the sum in , and the right-hand side of appears in the integral of , so can be used to transform the lower bound in into a lower bound of the type . Now let us increase $t$, we have again, using the assumption on $H$, $$| \tH(\Cz - t) - \tH(\Cu - t) | \leq \sum_{i=-r}^{r} |\Gap_i(\Cz - t) - \Gap_i(\Cu - t)|,$$ and for $t$ small we still have $$\Gap_i(\Cz - t) = \Gap_i(\Cz), \quad \Gap_i(\Cu - t) = \Gap_i(\Cu).$$ However, these identities cease to hold as soon as we encounter a point of $\Cz$ or $\Cu$, i.e. as soon as $t = T_0 := \min\left(x_0(\Cz), x_0(\Cu)\right)$. Indeed, assuming e.g. that the first point encountered is $x_0(\Cz)$, we have, for $t$ slightly larger than $T_0$, but smaller than $x_0(\Cu)$, $$\Gap_i(\Cz - t) = \Gap_{i+1}(\Cz), \quad \text{ but still } \Gap_i(\Cu - t) = \Gap_i(\Cu),$$ so comparing naively the two integrands with the Lipschitz control of $H$ gives us a lower bound $$\label{arfarf2}
| \tH(\Cz - t) - \tH(\Cu - t) | \leq \sum_{i=-r}^{r} |\Gap_{i+1}(\Cz) - \Gap_i(\Cu )|.$$ The left-hand side of is still present in the integrals of , however the right-hand side of does *not* appear in the sum , and thus is useless for us. To remedy this misalignment, we need to shift the configuration $\Cu$ by $x_0(\Cu) - x_0(\Cz)$, i.e. to add a quantity $x_0(\Cu) - x_0(\Cz)$ to the “proper time” $t$ of $\Cu$. Indeed, we have, for $t$ slightly larger than $T_0 + x_0(\Cu) - x_0(\Cz) = x_0(\Cu)$, $$\Gap_i(\Cu - t) = \Gap_{i+1}(\Cu),$$ and thus comparing $\Cz - t$ and $\Cu -t - (x_0(\Cu) - x_0(\Cz))$ yields again a summand from . Each time that we encounter a “$k$-th point”, no matter whether it comes from $\Cz$ or from $\Cu$, we need to shift the “proper time” of the other configuration in order to “align the gaps”. Doing this, we effectively “lose” portions of the interval $[0, \Rdix]$ on which we integrate, which possibly deteriorates the lower bound of . In fact, it turns out that the loss can be expressed in terms of the “gap difference” itself.
Another technicality occurs when one tries to make the Fubini argument yielding rigorously, and we will need to estimate the time “spent” on each gap $\Gap_i$ for $i \in [0, \frac{R}{10}]$.
### A good event {#a-good-event .unnumbered}
Let us introduce the following events $$\begin{aligned}
\EG_1 := & \llbr \left|{\mathcal{C}}_{\left[\frac{R}{10}, R\right]}\right| \geq r + 1 \rrbr \cap \llbr \left| {\mathcal{C}}_{\left[-R, 0\right]}\right| \geq r + 1 \rrbr \\
\EG_2 := & \llbr x_{\Pos({\mathcal{C}};\frac{R}{10}) + r} \leq \frac{R}{4} \rrbr \\
\EG_3 := &
\llbr x_{R^{1/2}}({\mathcal{C}}) \leq 2 R^{1/2} \rrbr \cap \llbr x_{-R^{1/2}}({\mathcal{C}}) \geq -2 R^{1/2} \rrbr \\
\EG_4 := & \llbr \left|{\mathcal{C}}_{\left[-\Rdix, \Rdix\right]}\right| \leq \frac{R}{3} \rrbr,\end{aligned}$$ and $\EveGap_R$ as the intersection $$\label{EventR}
\EveGap_R := \EG_1 \cap \EG_2 \cap \EG_3 \cap \EG_4.$$ By construction, if ${\mathcal{C}}$ is in $\EveGap_R$, we know that:
- (Event $\EG_1$.) There are at least $r+1$ points in $[\frac{R}{10}, R]$ and in $[-R, 0]$, so $\hHR({\mathcal{C}}|_{{\Lambda}_R})$ is well-defined.
- (Event $\EG_2$.) The $r$-th gap of ${\mathcal{C}}- \frac{R}{10}$ (which is the “right-most” gap that we consider when applying $\hHR$ to a configuration ${\mathcal{C}}$) corresponds to points in $\left[0, \frac{R}{4}\right]$. It implies that $\hHR({\mathcal{C}})$ depends only on ${\mathcal{C}}|_{{\Lambda}_{\frac{R}{4}}}$, namely $$\label{hHRbecomeslocal}
\hHR({\mathcal{C}}) = \hHR({\mathcal{C}}|_{{\Lambda}_{\frac{R}{4}}}).$$
- (Event $\EG_3$.) The $R^{1/2}$-th point on each side is at distance at most $2R^{1/2}$ . Since by assumption $S$ is a random variable bounded by $R^{1/2}$, it yields $$\label{bornexS}
|x_{S}({\mathcal{C}})| \preceq R^{1/2}.$$ We may note that we also have the very rough bound $$\label{bornefirstidiot}
x_0({\mathcal{C}}) \preceq R^{1/2}.$$
- (Event $\EG_4$.) There are at most $\frac{R}{3}$ points in $\left[-\Rdix, \Rdix\right]$, and in particular $$\label{xkRdix}
|x_k({\mathcal{C}})| \leq \Rdix \implies |k| \leq \frac{R}{3}.$$
Since $r$ is fixed, the discrepancy estimates and guarantee that, $$\lim_{R \to \infty} \Pz \left( \EveGap_R \right) = \lim_{R \to \infty} \Pu \left( \EveGap_R \right) = 1,$$ so for $R$ large enough, we have $$\label{choiceR0}
\Pz \left( \EveGap_{R} \right) \geq 1 - \frac{c}{100}, \quad \Pu \left( \EveGap_R \right) \geq 1 - \frac{c}{100}.$$
### The quantity to compute {#the-quantity-to-compute .unnumbered}
For $R$ large enough, we may write, since $\hHR$ is bounded by $\frac{R}{10}$ $$\label{hHReveGap}
\E_{\Pz}[ \hHR ] - \E_{\Pu}[ \hHR] \leq \E_{\Pz}[ \hHR \1_{\EveGap_R} ] - \E_{\Pu}[ \hHR \1_{\EveGap_R}] + \frac{2c}{100} \times \frac{R}{10}.$$ On the other hand, since $\Q$ is a coupling of the restrictions of $\Pz, \Pu$ to ${\Lambda}_R$, in view of we have $$\E_{\Pz}[ \hHR \1_{\EveGap_R}] - \E_{\Pu}[ \hHR \1_{\EveGap_R}] = \E_{\Q} \left[ \hHR(\Cz) \1_{\EveGap_R}(\Cz) - \hHR(\Cu)\1_{\EveGap_R}(\Cu) \right].$$ We now turn to compute the quantity $$\begin{gathered}
\label{GapsunderPi}
\E_{\Q} \left[ \1_{\EveGap_R}(\Cz) \hHR(\Cz) - \1_{\EveGap_R}(\Cu) \hHR(\Cu) \right] \\
= \E_{\Q} \left[ \1_{\EveGap_R}(\Cz) \int_{0}^{\Rdix} \tH(\Cz - t) dt - \1_{\EveGap_R}(\Cu) \int_{0}^{\Rdix} \tH(\Cu - t) dt \right].\end{gathered}$$ Let $\Cz, \Cu$ be fixed, and assume both belong to $\EveGap_R$.
### The initial shift {#sec:initialshift .unnumbered}
Let $v_0, v_1$ be defined as $$\label{choixv}
v_0 = x_0(\Cz), \quad v_1 = x_{S}(\Cu).$$ In view of , , we have $$\label{bornev}
|v_0| + |v_1| \preceq R^{1/2}.$$
By the definition , of the enumeration of points and gaps, the choice of $v_0, v_1$ as in ensures that for all $i$ $$\label{translationGaps}
x_{i}(\Cz - v_0) = x_i(\Cz), \quad x_{i}(\Cu - v_1) = x_{i+S}(\Cu), \quad \Gap_i(\Cz - v_0) = \Gap_{i}(\Cz), \quad \Gap_i(\Cu - v) = \Gap_{i+S}(\Cu).$$
In the sequel, we will write, for simplicity, $\Cz = \Cz - v_0$ and $\Cu = \Cu - v_1$. This amounts to doing a translation in each integral of , and the error is of order $|v_0| + |v_1|$ which is bounded as in , and thus negligible for our purposes (we pursue a lower bound of order $R$). It places us in an “ideal” situation where $S = 0$, and where, at time $t=0$, both configurations have a point at $0$.
### The proper times {#the-proper-times .unnumbered}
- We let $T_0 = 0$, and recall (see previous paragraph) that $x_0(\Cz) = x_0(\Cu) = 0$. We define the functions $$\tz_0(t) = t, \quad \tu_0(t) = t.$$
- We let $T_1$ be the first time at which we encounter a new point, more precisely:
$$\label{def:T1}
T_1 := \min \left\lbrace t: \ \ \tz_0(t) - \tz_0(T_0) = \Gap_0(\Cz) \text{ or } \tu_0(t) - \tu_0(T_0) = \Gap_0(\Cu) \right\rbrace.$$
- We then define the functions $$\label{def:tz1}
\tz_1(t) = \tz_0(t) + \left(\Gap_0(\Cz) - \left(\tz_0(T_1) - \tz_0(T_0) \right) \right), \quad \tu_1(t) = \tu_0(t) + \left(\Gap_0(\Cu) - \left(\tu_0(T_1) - \tu_0(T_0) \right) \right).$$ Of course, we have in fact $\tz_0(T_0) = \tu_0(T_0) = 0$, $T_1 = \min(\Gap_0(\Cz), \Gap_0(\Cu))$, and one of the quantities $$\Gap_0(\Cz) - T_1, \quad \Gap_0(\Cu) - T_1$$ is equal to zero, while the other one is the “shift” $$\Gap_0(\Cz) - \Gap_0(\Cu), \text{ or } \Gap_0(\Cu) - \Gap_0(\Cz),$$ that we must apply to the configuration with a larger first gap.
- Assume that $T_{k}, \tz_k, \tu_k$ have been defined for some $k \geq 1$. We let the time $T_{k+1}$ be given by $$\label{def:Tk}
T_{k+1} := \min \left\lbrace t, \ \tz_k(t) - \tz_k(T_k) = \Gap_{k}(\Cz) \text{ or } \tu_k(t) - \tu_k(T_k) = \Gap_k(\Cu) \right\rbrace,$$ and we introduce the functions $$\label{def:tzk}
\begin{cases}
\tz_{k+1}(t) & := \tz_k(t) + \left(\Gap_k(\Cz) - \left(\tz_k(T_{k+1}) - \tz_k(T_k) \right)\right), \\
\tu_{k+1}(t) & := \tu_k(t) + \left(\Gap_k(\Cu) - \left(\tu_k(T_{k+1}) - \tu_k(T_k)\right) \right).
\end{cases}$$ Once again, by definition one of the “shifts” is equal to zero, and the other one is given by $$\label{shift}
\Gap_k(\Cz) - \Gap_k(\Cu) \text{ or } \Gap_k(\Cu) - \Gap_k(\Cz).$$
- Let $\Kmax$ be defined as $$\label{def:Kaymax}
\Kmax:= \min \left\lbrace k: T_k \geq \Rdix\right\rbrace$$ We will only consider $k \leq \Kmax$, in other words we stop the definition when $T_{k} \geq \Rdix$ and let $\Kmax$ be the number of steps. Since holds, we have $
\Kmax \leq \frac{R}{3}.
$
\[claim:shifts\] For all $k$, we have $$\label{postkbis}
x_0(\Cz - \tz_k(T_k)) = x_k(\Cz) - \tz_k(T_k) = 0, \quad x_0(\Cu - \tu_k(T_k)) = x_k(\Cu) - \tu_k(T_k) = 0,$$ and for $t$ in $(T_k, T_{k+1})$, we have, using the notation of , $$\label{postk}
\Pos(\Cz, \tz_k(t)) = \Pos(\Cu, \tu_k(t)) = k + 1,$$ which in particular implies the following “gap alignment” identity for $t\in(T_k, T_{k+1})$: $$\label{gapsarealigned}
\Gap_0\left(\Cz - \tz_k(t) \right) = \Gap_{k+1}(\Cz), \quad \Gap_0\left(\Cu - \tu_k(t) \right) = \Gap_{k+1}(\Cu).$$
We recall that we enforced $x_0(\Cz) = x_1(\Cz) = 0$, which is for $k=0$. It allows us to re-write the definition as $$T_1 = \min\left(t, \ t = \Gap_0(\Cz) \text{ or } t = \Gap_0(\Cu) \right),$$ thus in fact $\Gap_0(\Cz) = x_1(\Cz)$ and $\Gap_0(\Cu) = x_1(\Cu)$, and $T_1$ is *the first positive time at which a point of $\Cz$ or $\Cu$ is encountered*.
For $0 < t < T_1$, the configuration $\Cz -t$ has a “first negative point” given by $x_0(\Cz) -t = -t$ and a “first positive point” given by $x_1(\Cz) -t$, and the same holds for $\Cu - t$. Thus holds for $k = 0$.
Assume that holds for some $k \geq 0$. And without loss of generality, assume that $$T_{k+1} = \min\left( t, \tz_k(t) - \tz_k(T_k) = \Gap_k(\Cz) \right).$$ Since we know, by induction hypothesis, that $\tz_k(T_k) = x_k(\Cz)$, we see that $\tz_{k+1}(T_{k+1})$ must be given by $x_k(\Cz) + \Gap_k(\Cz)$, which is equal to $x_{k+1}(\Cz)$. We obtain also, following the definition, $$\tu_{k+1}(T_{k+1}) = \tu_{k}(T_{k+1}) + \left(\Gap_{k}(\Cu) - (\tu_k(T_{k+1}) - \tu_k(T_k) )\right),$$ and using the induction hypothesis $\tu_k(T_k) = x_k(\Cu)$, we obtain $$\tu_{k+1}(T_{k+1}) = x_{k}(\Cu) + \Gap_{k}(\Cu) = x_{k+1}(\Cu),$$ which proves at rank $k+1$. We also deduce that holds (at rank $k$). The claim is thus proven by induction.
\[claim:timeloss\] For a bounded function $F(t)$, we have $$\label{timelossA}
\left| \int_{0}^{\frac{R}{10}} F(t) dt - \sum_{k=0}^{\Kmax-1} \int_{T_k}^{T_{k+1}} F(\tz_k(t)) dt \right| \leq \|F\|_{\infty} \sum_{k=0}^{\Kmax} \left |\Gap_{k}(\Cz) - \Gap_{k}(\Cu)\right|,$$ and similarly when replacing $\tz$ by $\tu$.
The intervals $(T_k, T_{k+1})$ are disjoint, and the change of variable $t \mapsto \tz_k(t)$ has speed $1$. The “time loss” $$\left[0, \frac{R}{10} \right] \setminus \bigcup_{k=0}^{\Kmax-1} \llbr \tz_k\left((T_k, T_{k+1})\right) \rrbr$$ comes from the time shifts , and its length is thus bounded by the sum of all possible shifts, which yields the right-hand side of .
### Using the Lipschitz bound on shifted intervals {#using-the-lipschitz-bound-on-shifted-intervals .unnumbered}
\[claim:Lipschitz\] For all $k \leq \Kmax$, for $t \in (T_k, T_{k+1})$, we have $$\label{LipschitzBound2}
\left| \tH(\Cz - \tz_k(t)) - \tH(\Cu - \tu_k(t)) \right| \preceq L \sum_{i=-r}^r \frac{\left|\Gap_{i+1+k}(\Cz) - \Gap_{i+1+k}(\Cu) \right|}{\Gap_{i+1+k}(\Cz) + \Gap_{i+1+k}(\Cu)}.$$
We first use the fact that $H$ is $1$-Lipschitz with respect to the $\| \cdot \|_{1}$ norm on $\R_+^{2r+1}$, and the identity to obtain $$\label{LipschitzBound1}
\left| \tH(\Cz - \tz_k(t)) - \tH(\Cu - \tu_k(t)) \right| \leq \sum_{i=-r}^r \left|\Gap_{i+k+1}(\Cz) - \Gap_{i+k+1}(\Cu) \right|.$$ To improve into , we recall that the function $H$ is bounded by $1$ and supported in $[0,L]^{2r+1}$ for some $L > 10$.
For any $i$ between $-r$ and $r$, let us distinguish cases:
- The gaps $\Gap_{i+k+1}(\Cz)$ and $\Gap_{i+k+1}(\Cu)$ are both smaller than $2L$, in which case we certainly have $$\left| \Gap_{i+k+1}(\Cz) - \Gap_{i+k+1}(\Cu) ) \right|
\leq 4L \frac{\left| \Gap_{i+k+1}(\Cz) - \Gap_{i+k+1}(\Cu) \right|}{ \Gap_{i+k+1}(\Cz) + \Gap_{i+k+1}(\Cu) )}.$$
- Both are larger than $L$, in which case the left-hand side of is zero.
- One of these quantities is smaller than $L$ (say, without loss of generality, the first one) and the other one is larger than $2L$, in which case the left-hand side of is bounded by $1$, and the right-hand side contains the term $$4L \frac{\left| \Gap_{i+k+1}(\Cz) - \Gap_{i+k+1}(\Cu) \right|}{ \Gap_{i+k+1}(\Cz) + \Gap_{i+k+1}(\Cu)} \geq 4L \frac{\Gap_{i+k+1}(\Cu) - L}{L + \Gap_{i+k+1}(\Cu)},$$ but clearly $\frac{4L (x-L)}{x+L} \geq 2$ for $x \geq 2L$ and $L > 10$, so the right-hand side of is bounded below by $2$, and the inequality holds.
### Shifting integrals and a Fubini argument {#shifting-integrals-and-a-fubini-argument .unnumbered}
\[claim:shiftandFubini\] $$\label{shiftFubini}
\left| \int_{0}^{\Rdix} \tH(\Cz -t) dt - \int_{0}^{\Rdix} \tH(\Cu - t) dt \right|
\leq \sum_{i = -\infty}^{+\infty} L \Weight_i \frac{\left|\Gap_{i}(\Cz) - \Gap_{i}(\Cu) \right|}{\Gap_i(\Cz) + \Gap_i(\Cu)} + \Error,$$ where $\Weight_i$ satisfies $$\label{boundWi}
\Weight_i \preceq \sum_{j = i-r-2}^{i+r} \min\left(\Gap_j(\Cz), \Gap_j(\Cu) \right),$$ and the range condition $$\label{weightzero}
\Weight_i = 0 \text{ if } |i| \geq R/2,$$ and $\Error$ is bounded by $$\label{ErrorFub}
\Error \preceq \sum_{i = 0}^{R/2} \left|\Gap_{i}(\Cz) - \Gap_{i}(\Cu) \right|.$$
We use Claim \[claim:timeloss\] with $F = \tH(\Cz - \cdot)$ and $\tH(\Cu - \cdot)$, and write $$\begin{gathered}
\label{PFA}
\int_{0}^{\frac{R}{10}} \tH(\Cz -t) dt - \int_{0}^{\Rdix} \tH(\Cu - t) dt
\\
= \sum_{k=0}^{\Kmax-1} \int_{T_k}^{T_{k+1}} \left( \tH(\Cz - \tz_k(t)) - \tH(\Cu - \tu_k(t)) \right) dt + \Error, \end{gathered}$$ with $\Error$ bounded as in .
Next, we use from Claim \[claim:Lipschitz\], and write, for every $k$ $$\label{PFB}
\left|\int_{T_k}^{T_{k+1}} \tH(\Cz - \tz_k(t)) - \tH(\Cu - \tu_k(t)) dt\right| \preceq L (T_{k+1} - T_k) \sum_{i=-r}^r \frac{\left|\Gap_{i+k+1}(\Cz) - \Gap_{i+k+1}(\Cu) \right|}{\Gap_{i+k+1}(\Cz) + \Gap_{i+k+1}(\Cu)}.$$ Combining , , a Fubini argument yields , with weights $$\Weight_i \preceq \sum_{k} \1_{k \in [i-r-1, i+r-1], 0 \leq k \leq \Kmax} \left(T_{k+1} - T_k\right).$$ This is a telescopic sum, and we may write $$\Weight_i \preceq T_{i+r} - T_{\min\{i-r-1,0\}}.$$ By definition of the times $T_k$, we see that $$\label{boundFubini}
T_{i+r} - T_{\min\{i-r-1,0\}} \leq \sum_{j = i-r-1}^{i+r} \min\left(\Gap_j(\Cz), \Gap_j(\Cu) \right),$$ which yields . Moreover, we see that $\Weight_i$ is $0$ if $i+r-1<0$ or $i+r-1>\Kmax$, which in particular (since $\Kmax \leq \frac{R}{3}$, taking $R$ large enough with respect to $r$) implies .
### Final computation, step 1. {#final-computation-step-1. .unnumbered}
Let us introduce the quantity $$\label{def:GCzCu}
\Gaingain := \sum_{i = -\frac{R}{2}}^{\frac{R}{2}} \frac{\left(\Gap_i(\Cz) - \Gap_{i}(\Cu)\right)^2}{\left(\Gap_i(\Cz)\right)^2 + \left(\Gap_{i}(\Cu)\right)^2},$$ and let us recall that the goal of this proof is to obtain a lower bound on the expectation, under the coupling $\Q$, of $\Gaingain$ that would be proportional to $R$.
\[claim:finalcomputm1\] There exists a constant $\CLr$ depending only on $L, r$ such that $$\begin{gathered}
\label{finalcomputm1}
\left| \int_{0}^{\Rdix} \tH(\Cz -t) dt - \int_{0}^{\Rdix} \tH(\Cu - t) dt\right| \\
\leq \CLr \left(\sum_{i=-\frac{R}{2}}^{\frac{R}{2}} \Gap_k(\Cz)^2 + \Gap_k(\Cu)^2 \right)^{1/2} \left(\Gaingain \right)^{1/2} .\end{gathered}$$
We write the result of Claim \[claim:shiftandFubini\] as follows: $$\begin{gathered}
\label{beforefine}
\left| \int_{0}^{\Rdix} \tH(\Cz -t) dt - \int_{0}^{\Rdix} \tH(\Cu - t) dt \right|
\preceq \sum_{i = -R/2}^{R/2} L \left( \sum_{j = i-r-2}^{i+r} \min\left(\Gap_j(\Cz), \Gap_j(\Cu) \right) \right) \frac{\left|\Gap_{i}(\Cz) - \Gap_{i}(\Cu) \right|}{\Gap_i(\Cz) + \Gap_i(\Cu)} \\ + \sum_{i = 0}^{R/2} \left|\Gap_{i}(\Cz) - \Gap_{i}(\Cu) \right|.\end{gathered}$$
For the first term in the right-hand side of , we use Cauchy-Schwarz’s inequality and the trivial bound $$\min\left(\Gap_j(\Cz), \Gap_j(\Cu) \right)^2 \leq \Gap_j(\Cz)^2 + \Gap_j(\Cu)^2,$$ while for the last term in the right-hand side, we write $$\left|\Gap_{i}(\Cz) - \Gap_{i}(\Cu) \right| = \frac{\left|\Gap_{i}(\Cz) - \Gap_{i}(\Cu) \right|}{\Gap_{i}(\Cz) + \Gap_{i}(\Cu)} \left(\Gap_{i}(\Cz) + \Gap_{i}(\Cu)\right),$$ and use Cauchy-Schwarz’s inequality. In both cases, we obtain a term bounded by the right-hand side of , up to a multiplicative constant depending only on $L, r$.
### Final computation, step 2. {#final-computation-step-2. .unnumbered}
Taking the expectation of under the coupling $\Q$ (to simplify, we forget about the conditioning on the event $\EveGap_R$, which does not affect the estimate see e.g. . We obtain, replacing the left-hand side of as in , $$\label{takingcoupling}
\E_{\Pz}[ \hHR ] - \E_{\Pu}[ \hHR] \preceq \E_{\Q} \left[ \left(\sum_{i=-\frac{R}{2}}^{\frac{R}{2}} \Gap_k(\Cz)^2 + \Gap_k(\Cu)^2 \right)^{1/2} \left(\Gaingain \right)^{1/2} \right].$$ Using Cauchy-Schwarz’s inequality, Lemma \[lem:gapL2\], and the fact that $\Pz, \Pu$ have finite energy, the right-hand side of can be bounded in order to yield $$\E_{\Pz}[ \hHR ] - \E_{\Pu}[ \hHR] \preceq R^{1/2} \E_{\Q} \left[ \Gaingain \right]^{{\frac{1}{2}}},$$ and if we use to bound below $\E_{\Pz}[ \hHR ] - \E_{\Pu}[ \hHR]$ we deduce that $$R \preceq R^{1/2} \E_{\Q} \left[ \Gaingain \right]^{{\frac{1}{2}}},$$ so $R \preceq \E_{\Q} \left[ \Gaingain \right]$, which concludes the proof.
The screening procedure {#sec:proofscreening}
=======================
We present a sketch of the screening argument as developed in [@petrache2017next Section 6]. For the notation of that paper, our setting is $d = 1,\, s = 0,\, k=1, \gamma = 0,\, g(x) = - \log |x|,\, c_{d,s} = 2\pi$, and our $s$ is their $\epsilon$. We also use the fact that, in the present case, the background measure has a constant intensity. We do not claim to make any serious improvement on the procedure.
We recall that $R' = R(1-s)$, that by assumption , we have $$\begin{aligned}
\label{goodboundary1}
\int_{\{-R', R'\} \times [-R, R]} |\El_{\eta}|^2 & = \Mec \leq M, \\
\label{gooddecay1} \frac{1}{s^4 R} \int_{{\Lambda}_R \times \R \backslash (-s^2 R, s^2 R)} |E|^2 & = \Hec \leq 1,\end{aligned}$$ and, that by assumption , the smeared out charges $\sigma_{x,\eta}$, for $x$ in ${\mathcal{C}}$, do not intersect the boundary ${\Lambda}_{R'} \times \R$.
By a mean value argument, there exists $\l$ in $[s^2 R, 2 s^2 R]$ such that $$\label{goodboundary2}
\int_{{\Lambda}_R \times \{-\l, \l\} } |E|^2 \leq s^2 \Hec \leq s^2.$$
#### **Subdividing the domain.**
As depicted on Figure \[fig:screening\], we decompose ${\Lambda}_R \times [-R,R]$ in three parts: $$D_0 = {\Lambda}_{R'} \times [-\l, \l], \quad D_{\partial} = \left( {\Lambda}_R \times [-\l, \l] \right) \backslash D_0, \quad D_1 = ({\Lambda}_R \times [-R, R]) \backslash (D_0 \cup D_{\partial}).$$
![Sketch of the situation.[]{data-label="fig:screening"}](screening3.pdf)
Roughly speaking, here is what the screening procedure entails:
- The point configuration will be kept in $\Old$ and the existing electric field will be kept in $D_0$.
- In $D_{\partial}$, we throw away the field and configuration. We will place a correct number of new points and define an electric field whose normal component coincides with the existing one on the vertical dotted lines and reaches $0$ on the vertical full line. There will be, however, a non-vanishing component of the field at the level of the dashed line.
- In $D_1$, we manipulate the electric field, starting from the horizontal dashed line, in order to reach a zero normal component on the exterior (full line).
We let $\Nint = |{\mathcal{C}}\cap {\Lambda}_{R'}| = |{\mathcal{C}}\cap D_0|$, be the number of points of ${\mathcal{C}}$ inside $D_0$ - those will not be touched, and we will place $|{\Lambda}_R| - \Nint$ points in $\New$ in order for the final configuration to have $|{\Lambda}_R|$ points.
We also define $\UZ$ as the quantity $$\label{def:UZ}
\UZ := \frac{1}{2 (|{\Lambda}_R| - |{\Lambda}_{R'}|)} \int_{[-R',R'] \times \{-\l, \l\}} \El \cdot \vec{\nu}.$$ The integral in corresponds to the integral of the normal component of the existing field on the horizontal dotted lines in Figure \[fig:screening\].
The screening procedure enforces that the normal component of the constructed field on the dashed line part of the boundary of $D_{\partial}$ exactly compensates $\UZ$, so that the total flux on the boundary of both rectangles that form $D_1$ is $0$.
We have $$\label{sizeUZ}
\UZ \preceq R^{-1/2}.$$
Since $R' = R(1-s)$ and by definition of $\UZ$, we have $$\UZ \preceq \frac{1}{s R} \int_{ [-R',R'] \times \{-\l, \l\} } |E|,$$ and combining the Cauchy-Schwarz’s inequality with we obtain .
We split ${\Lambda}_R \backslash {\Lambda}_{R'}$ into intervals $H_i$ whose lengths belong to $[\l/2, 2\l]$, and we let $\tH_i = H_i \times [-\l, \l]$. We denote by $\Hleft$ (resp. $\Hright$ the interval exactly to the left (resp. to the right) or $-R'$ (resp. $R'$).
For any interval in this decomposition, we let $m_i$ be such that $$\label{def:mi}
2\pi (m_i - 1) |H_i| = \int_{ \left(\{-R', R'\} \times [-\l, \l] \right) \cap \partial \tH_i} \El_{\eta} \cdot \vec{\nu} - 2 \UZ |H_i|.$$ The first term in the right-hand side of is only present if $H_i$ is $\Hleft$ or $\Hright$, the second term is always present.
\[claim:onmi\] We quantify how close to $1$ the number $m_i$ defined in is.
- If $H_i$ is not $\Hleft$ or $\Hright$, we have then $$\label{sizemi1}
|m_i - 1| \preceq R^{-1/2}$$
- If $H_i$ is one of the two intervals $\Hleft, \Hright$ that have an intersection with $\partial {\Lambda}_{R'}$, we have $$\label{sizemi}
|m_i - 1| \preceq R^{-1/2} + \Mec^{1/2} R^{-1/2} s^{-1}.$$
In particular, in both cases, for $R$ large enough, we have $$\label{mimoins1hal}
|m_i - 1| \preceq {\frac{1}{2}}.$$
In the first case, we have $|m_i - 1| \preceq \UZ$ and we apply . In the second case, we add the contribution of the integral $$\frac{1}{|H_i|} \int_{\partial D_0 \cap \partial \tH_i} \El_{\eta} \cdot \vec{\nu} \preceq \l^{-1} \l^{1/2} \left( \int_{\partial D_0} |\El_{\eta}|^2 \right)^{1/2} \preceq (s^2 R)^{-1/2} \Mec^{1/2},$$ where we have used the Cauchy-Schwarz inequality and .
Then each interval $H_i$ is divided into sub-intervals of length $\frac{1}{m_i}$, and in each of these intervals, exactly one point of the new screened configuration is placed. More precisely, a point is placed randomly at distance less than $\frac{\eta}{4}$ from the center of the interval. This randomness is important, because it “creates volume” and yields the description of the “new points” as in .
For the end of the screening procedure (constructing the screened electric field and estimating its energy), we refer to the proof of [@petrache2017next]. Let us emphasize that, in our setting, some technicalities become irrelevant and that the construction and estimates could be written in a much more concise way. Here is a short sketch thereof:
1. First, as the $k$-th point $\Xx_k$ is being “placed”, in the sub-interval $I_k$ we define the electric field by solving $$- \div\ \El = 2\pi \left( \delta_{\Xx_k} - dx \right)$$ in $I_k \times [-\l, \l]$, with some boundary condition. These boundary conditions are chosen in a compatible fashion for two neighboring sub-intervals. We also impose a zero boundary condition on the left for the sub-interval that contains $-R$ (and similarly for the rightmost sub-interval), and to match the existing boundary condition given by the pre-existing field $\El$ for the two sub-intervals that share an endpoint with $\Old$. Estimating the “energy” created this way is an additional task, but we find that it only yields a small error compared to the total energy.
2. At this stage, the electric field is defined on $\Old \times [-\l, \l]$, the region denoted by $D_0$ in Figure \[fig:screening\], by keeping the pre-existing field; and on $\New \times [-\l, \l]$, the region denoted by $D_{\partial}$ in Figure \[fig:screening\], by defining it on each sub-interval, as in the previous step. It remains to define it on the region denoted by $D_1$ in Figure \[fig:screening\]. Of course, we do not place any points here, but we tile this region by small rectangles of side-length $\approx \l$, and on each of these we solve $$- \div\ \El = 0,$$ with an appropriate choice of (mutually compatible) boundary conditions, which allow us to pass from whatever boundary condition exists at the frontier of $D_{\partial}, D_0$, to the desired Neumann condition $\El \cdot \vec{\nu} = 0$ on the boundary on the big rectangle in Figure \[fig:screening\]. There again, one must estimate the energy of these “patching” fields.
3. We then obtain the desired screened electric field.
By construction, for any $H_i$ we place exactly one point at the center of each sub-interval of length $\frac{1}{m_i}$. Since the length of $H_i$ is in $[\l/2, 2\l]$, with $\l \preceq s^2 R$, and since $|m_i - 1| \leq {\frac{1}{2}}$ (see ) the number of sub-intervals in each interval is $\preceq s^2 R$. In particular $$\label{pointsinHleft}
\# \text{ points in $\Hleft$ } \preceq s^2R.$$
In view of Claim \[claim:onmi\], the distance between the position of $k$-th point (starting from the leftmost one) and $-R + k - {\frac{1}{2}}$ is bounded, up to a universal multiplicative constant, by
- $k R^{-1/2}$ as long as the point does not belong to $\Hleft$.
- $k R^{-1/2} + (k-k_0) \Mec^{1/2} R^{-1/2} s^{-1}$ if the point belongs to $\Hleft$, where $k_0$ is the index of the last point outside $\Hleft$. By , we may write $$k R^{-1/2} + (k-k_0) \Mec^{1/2} R^{-1/2} s^{-1} \preceq \left(1+\Mec^{1/2}\right) s R^{1/2}.$$
We thus have, as claimed in , $$\left| \Xx_k - \bXx_k \right| \preceq k R^{-1/2},$$ as long as $\Xx_k$ is not in $\Hleft$, that is for $k$ such that $$s^2 R \preceq |sR - k|.$$ When $\Xx_k$ is in $\Hleft$, that is for $k$ such that $$|sR - k| \preceq s^2 R,$$ we have, as claimed in , $$\left| \Xx_k - \bXx_k \right| \preceq \Mec^{1/2} s R^{1/2}.$$
Let us recall that $\kkmax$ is the index of the first point such that $\Xx_k \geq -R + sR$. We deduce that $\kkmax$ is equal to $sR$, up to an error of order $\Mec^{1/2} sR^{1/2}$, which yields .
The inequalities , can then be converted into discrepancy estimates and .
[^1]: ME gratefully acknowledges support by the German Research Foundation through the Hausdorff Center for Mathematics and the Collaborative Research Center 1060 *Mathematics of Emergent Effects*. MH has been funded by the Vienna Science and Technology Fund (WWTF) through project VRG17-005.
[^2]: One could e.g. enforce that $R/2$ is an integer, but more generally for $x$ real, having “at least $x$ points” means “having at least $\ceil{x}$ points”, $\sum_{k=0}^{x}$, means $\sum_{k=0}^{\floor{x}}$ etc.
[^3]: Remark: this is the only moment where we really use the fact that the new electric field that we have produced is *screened*.
[^4]: In fact, the contribution of the backgroung potential can be shown to be negligible with respect to $R$.
[^5]: For technical reasons, when writing the formal proof below, we actually enforce that $x_0(\Cz) = x_0(\Cu) = 0$, which twists the enumeration a little bit, but does not modify the general strategy.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Extensive work on single molecule magnets has identified a fundamental mode of relaxation arising from the nuclear-spin assisted quantum tunnelling of nearly independent and quasi-classical magnetic dipoles. Here we show that nuclear-spin assisted quantum tunnelling can also control the dynamics of purely emergent excitations: magnetic monopoles in spin ice. Our low temperature experiments were conducted on canonical spin ice materials with a broad range of nuclear spin values. By measuring the magnetic relaxation, or monopole current, we demonstrate strong evidence that dynamical coupling with the hyperfine fields bring the electronic spins associated with magnetic monopoles to resonance, allowing the monopoles to hop and transport magnetic charge. Our result shows how the coupling of electronic spins with nuclear spins may be used to control the monopole current. It broadens the relevance of the assisted quantum tunnelling mechanism from single molecular spins to emergent excitations in a strongly correlated system.'
author:
- 'C. Paulsen'
- 'S. R. Giblin'
- 'E. Lhotel'
- 'D. Prabhakaran'
- 'K. Matsuhira'
- 'G. Balakrishnan'
- 'S. T. Bramwell'
title: Nuclear spin assisted quantum tunnelling of magnetic monopoles in spin ice
---
Introduction {#introduction .unnumbered}
============
In the canonical dipolar spin ice materials ([Dy$_{2}$Ti$_{2}$O$_{7}$]{}, [Ho$_{2}$Ti$_{2}$O$_{7}$]{}) [@Harris1997; @BramwellHarris98; @Ramirez; @BramwellGingras], rare earth ions with total angular momentum $J = 15/2$ (Dy$^{3+}$) and $J = 8$ (Ho$^{3+}$) are densely packed on a cubic pyrochlore lattice of corner-linked tetrahedra. The ions experience a very strong $\langle111\rangle$ crystal field, resulting in two effective spin states ($M_{\rm J} = \pm J$) that define a local Ising-like anisotropy. At the millikelvin temperatures discussed here ($0.08~{\rm K}<T<0.2$ K), a lattice array of such large and closely-spaced spins would normally be ordered by the dipole-dipole interaction [@LuttingerTisza], but the pyrochlore geometry of spin ice frustrates the dipole interaction and suppresses long-range order. Instead, the system is controlled by an ice-rule, that maps to the Pauling model of water ice [@Harris1997; @BramwellHarris98; @Ramirez; @BramwellGingras]. In the effective ground state, the spins describe a flux with closed-loop topology and critical correlations, that may be described by a local gauge symmetry rather than by a traditional broken symmetry [@CMS]. This strongly correlated spin ice state is stabilised by a remarkable self-screening of the dipole interaction [@Isakov; @Gingras]. Excitations out of the spin ice state fractionalise to form effective magnetic monopoles [@CMS; @Ryzhkin], but the excited states are no longer self-screened and this manifests as an effective Coulomb interaction between monopoles. The static properties of spin ice are accurately described by the monopole model [@KaiserDH]. The dynamic properties can also be described by assuming an effective monopole mobility [@JaubertHoldsworth; @Bovo; @Paulsen_Wien], but there have been few studies of the microscopic origin of the monopole motion [@Tomasello].
{width="10cm"}
The field and energy scales involved in monopole motion are illustrated in Fig. 1, a-e. When a monopole hops to a neighbouring site a spin is flipped (Fig. 1a). For an isolated monopole (far from any others) this spin flip takes place at nominally zero energy cost (Fig. 1b) because contributions from near-neighbour antiferromagnetic superexchange and ferromagnetic dipole–dipole coupling individually cancel. The cancellation of the field contribution relies on the dipolar self-screening [@Isakov] that maps the long-range interacting system [@Gingras] to the degenerate Pauling manifold of the near neighbour spin ice model [@BramwellHarris98]. This surprising cancellation is a key result of the many–body physics of spin ice. In practice, a monopole hop may also involve a finite energy change arising from longitudinal fields at the spin site: the main source of fields is nearby monopoles [@CMS] (Fig. 1b), while further contributions arise from corrections to the mapping, which give a finite energy spread to the Pauling manifold [@Vedmedenko] (here of order $\sim$0.1 K [@Melko]). The mechanism of the hop is believed to be quantum tunnelling and several key signatures of this have been observed in the high temperature regime between 2 K and 10 K [@Ehlers; @Snyder; @JaubertHoldsworth; @Bovo; @Tomasello].
At lower temperatures ($T < 0.6$ K), spin ice starts to freeze [@Snyder]. This is due in part to the rarefaction of the monopole gas whose density $n(T)$ varies as $\sim e^{-|\mu|/T}$ where the chemical potential $|\mu| = 4.35 $ and 5.7 K for [Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} respectively [@CMS], and also in part to geometrical constraints that create noncontractable, monopole-antimonopole pairs that cannot easily annihilate [@noncontractable]. These factors, which are independent of the monopole hopping mechanism, suggest that the relaxation rate $\nu(T) \propto n(T)$ will fall to exponentially small values at low temperature ($T~<~ 0.35$ K).
Previous thermal quenching experiments have demonstrated monopole populations well below the nominal freezing temperature that are both long lived and able to mediate magnetic relaxation [@PaulsenAQP]. This paradoxical frozen but dynamical character of the system suggests the relevance of resonant magnetic tunnelling, where magnetisation reversal can only occur when the longitudinal field is smaller than the tunnelling matrix element $\Delta E$. The monopolar fields may add a longitudinal component that takes the spin off the resonance condition (Fig. 1c,d) but in addition may add a transverse component that amplifies $\Delta E$: together these lead to a suppression and dispersion of the monopole mobility.
In the following, we will demonstrate experimentally that hyperfine interactions (Fig. 1e) play a significant role in bringing monopoles back to their resonance condition, enabling dynamics at very low temperatures ($T~<~0.35$ K).
Results {#results .unnumbered}
=======
[**Samples**]{}
$I$
-- ------------------------------------- ------ -------- -------- --------- ---------------------- --
7/2 4.17 0.3 0.034 5.7 1$\times$10$^{-5} $
$\approx$ 19 $\%$, $^{161}$Dy 5/2 -0.48 -0.0039
$\approx$ 25 $\%$, $^{163}$Dy 5/2 0.67 0.0054
$\approx$ 66 $\%$, $^{\rm other}$Dy 0 0 0
0 0 0 0
5/2 0.67 0.0599 0.0054
To investigate the effect of nuclear spins on the magnetic relaxation in spin ice, we studied four spin ice samples: [Ho$_{2}$Ti$_{2}$O$_{7}$]{}, with $I = 7/2$ and three [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples spanning a range of nuclear spin composition from $I=0$ to $I=5/2$. Details of nuclear spins and hyperfine parameters are given in Table 1. Ho$^{3+}$ is a non-Kramers ion with intrinsically fast dynamics owing to the possibility of transverse terms in the single-ion spin Hamiltonian, while Dy$^{3+}$, being a Kramers ion, has intrinsically much slower dynamics. However, it should be noted that, at low temperature, bulk relaxation is slower in [Ho$_{2}$Ti$_{2}$O$_{7}$]{} than in [Dy$_{2}$Ti$_{2}$O$_{7}$]{}, owing to its larger $|\mu|$ and hence much smaller monopole density (See Supplementary Fig. 1).
[**Thermal protocol**]{}
![Controlled cooling of spin ice below its freezing temperature. How the temperature of the samples varies during and after the AQP: (a) The applied field (black) during an AQP, and the temperatures measured by a small thermometer glued directly on top of the samples (schematically shown in (b)) vs log time for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} (HTO, red) and $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} (DTO, blue). The inset shows a zoom of the first 6 seconds vs time. (b) Comparison of the sample cooling rates $dT/dt$ as a function of temperature after the AQP for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} and $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} from the data in (a) to the Òequilibrium cooling rateÓ $dT/d \tau$ extracted from ac susceptibility data for the two samples (see Supplementary Fig. 6). The cooling rate for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} crosses the equilibrium rate at $\sim 0.9$ K, and $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} at 0.72 K.[]{data-label="two"}](figure2.pdf){width="7.5cm"}
In previous experiments we have accurately manipulated the monopole density in [Dy$_{2}$Ti$_{2}$O$_{7}$]{} by rapid magnetothermal cooling (Avalanche Quench Protocol, AQP) the sample through the freezing transition, allowing the controlled creation of a non-equilibrium population of monopoles in the frozen regime [@PaulsenAQP]. However it is more problematic to cool samples containing Ho, due to the large Ho nuclear spin which results in a Schottky heat capacity of 7 J mol$^{-1}$ K$^{-1}$ at 300 mK. Indeed this anomaly has been exploited by the Planck telescope where the bolometers are attached to the cold plate by yttrium-holmium feet thus allowing passive filtering with a several hour time constant that was crucial to the operation of the system [@PrivateComm]. For [Ho$_{2}$Ti$_{2}$O$_{7}$]{} this means difficulty in cooling. Therefore during some of the runs the sample temperature was recorded via a thermometer directly mounted on the sample face. Fig. 2a shows the monitoring of the sample temperature as it approaches equilibrium for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} and [Dy$_{2}$Ti$_{2}$O$_{7}$]{} during and after the AQP. The inset of Fig. 2a shows that only a few seconds are required to cool the samples from 0.9 K to 0.2 K, which is well below the freezing transition. Whereas [Dy$_{2}$Ti$_{2}$O$_{7}$]{} continues to cool, reaching 80 mK after only 10 s, [Ho$_{2}$Ti$_{2}$O$_{7}$]{} takes nearly 2000 s to reach the same temperature. Hence the data shown here were taken at 80 mK for [Dy$_{2}$Ti$_{2}$O$_{7}$]{} and 200 mK (and 80 mK when possible) for [Ho$_{2}$Ti$_{2}$O$_{7}$]{}.
[**Monopole density**]{}
We have phenomenologically estimated how the monopole density depends upon the rate of sample cooling, $dT/dt$ and the spin relaxation time $\tau(T) = 1/\nu(T)$, which is derived from the peaks in the imaginary component of the ac susceptibility. Differentiation of $\tau(T)$ to give $d\tau/dT$ and hence $dT/d\tau$, allows definition of an equilibrium cooling rate $dT/d\tau$, that gives the maximum cooling rate that may still maintain equilibrium. Fig. 2b compares $dT/dt$ and $dT/d \tau$ for both [Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{}. It can be seen that after the AQP, $dT/dt$ for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} crosses the equilibrium curve and goes out of equilibrium at $\approx 0.9$ K, and for [Dy$_{2}$Ti$_{2}$O$_{7}$]{} at $\approx 0.72$ K. The upper limit of the monopole density at low temperature can be estimated by equating it to the theoretical value at the crossing temperature: thus we find one monopole on approximately every $10^3$ tetrahedra for both [Ho$_{2}$Ti$_{2}$O$_{7}$]{} and [Dy$_{2}$Ti$_{2}$O$_{7}$]{}.
[**Spontaneous relaxation**]{}
We studied the effect of wait time $t_{\rm w}$ between the end of the avalanche quench and the application of the field with the aim to determine the effect of nuclear spins on the monopole dynamics. Varying the wait time deep in the frozen regime allowed us to gauge the spontaneous evolution of the zero-field monopole density as a function of time: that is, if monopoles recombine in a time $t_{\rm w}$, then the observed monopole current will be smaller, the longer the wait time. Two separate experiments were designed to study these effects. In the first experiment (Fig. 3) after waiting we applied a constant field and measured the magnetization $M$ as a function of time. In the second experiment (Fig. 4) we investigated the effect of wait time on the magnetothermal avalanches [@Slobinsky; @Jackson; @Krey2012] that occur on ramping the field to high values. Both of these allowed access to the magnetic current density $J_{\rm m} = dM / dt$. Full details of the experimental conditions are given in Supplementary Note 1 and Supplementary Fig. 2.
The monopole current is controlled by multiple factors. In the simplest model [@Ryzhkin] there are three of these: the monopole density $n$, the monopole mobility $u$ (related to the spin tunnelling rate) and the bulk susceptibility $\chi$. Thus $J_{\rm m} = dM/dt = \nu (M_{\rm eq} - M)$ where $M_{\rm eq} = \chi H$ is the equilibrium magnetisation and $\nu \propto u n$. In general it is difficult to deconvolve these various factors. In Ref. it was achieved by independent measurement of $n(T)$ and $\chi(T)$ to reveal $u(T)$. In the present time-dependent experiments we cannot perform such a direct separation, but by studying [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples with different isotopes, it seems reasonable to assume that the susceptibility and starting density are roughly the same, so the variation in mobility (hop rate) will dominate differences between the samples. Inclusion of [Ho$_{2}$Ti$_{2}$O$_{7}$]{} in the comparison gives a further point of reference: the starting monopole densities (see above) and susceptibilities for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} are expected to be comparable to those of [Dy$_{2}$Ti$_{2}$O$_{7}$]{}, while the the tunnel splitting (which controls the intrinsic mobility) is also estimated to be of the same order [@Tomasello] in the appropriate range of internal fields (see Fig. 1 and Ref. , Fig. 5).
{width="12cm"}
Fig. 3 summarizes results for the relaxation of the magnetization $M(t)$ for the different samples, as well as the value of $M(t = 400~{\rm s})$ and the monopole current $J_{\rm m}(t = 0)$ as a function of wait time, for a constant applied field of 0.08 T. The [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples show a clear progression in wait time effect that correlates strongly with their relative densities of nuclear spin states. Thus the monopoles recombine during the wait period much more effectively the larger the nuclear spin: that is, the larger the nuclear spin the higher the monopole mobility, the faster the recombination, and the fewer the monopoles at the start of the measurement. In Fig. 3e, f, higher mobility means the relaxation curves ($M(t = 400~{\rm s})$ and $J_{\rm m}(t = 0)$) shift both up and to the left, so a crossover in curves is expected – and this is indeed observed at the longer times. Near to equilibrium a second crossover would be expected (i.e. the equilibrium current density is higher for the highest mobility), but this crossover is clearly very far outside our time window. Hence our [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples are always far from equilibrium.
The effects observed for [Dy$_{2}$Ti$_{2}$O$_{7}$]{} are yet more dramatic in [Ho$_{2}$Ti$_{2}$O$_{7}$]{}, consistent with the Ho$^{3+}$ non-Kramers character, large nuclear spin, and large hyperfine coupling. Relaxation at 200 mK covers more than two orders of magnitude but is practically extinguished for long wait times, showing that excess monopoles spontaneously recombine to eliminate themselves from the sample. The plots indicate that the half life for monopole recombination in [Ho$_{2}$Ti$_{2}$O$_{7}$]{} would be approximately 150 seconds (much shorter than the equilibrium relaxation time) and suggests that equilibrium in the monopole density is reached at long times. Using the above estimate for the initial monopole density $n(t = 0) \sim 10^{-3}$, we recover a nominal equilibrium density of $n_{\rm eq} = 10^{-5}$ (per rare earth atom). Although this estimate is an upper limit it is nevertheless far from the expected equilibrium density, $n_{\rm eq}^{\rm 200~mK} \sim 10^{-13}$ (calculated by the method of Ref. , see Supplementary Fig. 1). It continues to evolve with temperature, being lower by a further order of magnitude at $T = 80$ mK (Fig. 3f). Most likely, the actual equilibrium monopole density is amplified by defects and disorder in the sample.
[**Magnetothermal avalanches**]{}
{width="12cm"}
Fig. 4 illustrates the effect of $t_{\rm w}$ on the magnetothermal avalanches. These occur when the injected power ($\mu_0 H \cdot J_{\rm m}$) overwhelms the extraction of thermal energy from the sample to the heat bath [@Jackson] such that monopoles are excited in great excess as the temperature steeply rises. The faster and more abundant the monopoles, the lower the avalanche field. To obtain the data in Fig. 4, after the AQP and $t_{\rm w}$, the applied field was swept at a constant rate, 0.02 T.s$^{-1}$ up to 0.4 T. If the avalanche field $H_{\rm ava}(t_{\rm w})$, is defined as the field where the magnetization crosses 1 $\mu_{\rm B}$ per rare earth ion, (0.5 $\mu_{\rm B}$ for the $^{163}$Dy sample) then the difference in avalanche field $\Delta H_{\rm ava}= H_{\rm ava}(t_{\rm w}) - H_{\rm ava} (t_{\rm w}={\rm minimum})$ allows us to compare the spread of fields for all samples.
Fig. 4a and b show the experimental results for the isotopically enriched [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples at 80 mK, demonstrating a very clear pattern. In general the spread of $H_{\rm ava}(t_{\rm w})$ becomes larger, the larger the nuclear spin, showing again that the nuclear spins strongly enhance the monopole mobility. Thus, the $^{162}$Dy sample (no nuclear spin) shows negligible evolution of the position of the avalanche field. For $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} (shown in Supplementary Fig. 4) the effect is small, while for the $^{163}$Dy sample (maximum nuclear spin) the effect of $t_{\rm w}$ can be clearly seen as a steady progression of $H_{\rm ava}(t_{\rm w})$ to higher fields for increasing $t_{\rm w}$ due to the smaller initial monopole density at the start of the field ramp. Also shown in the figure are the curves that result from slow conventional zero field cooling (CC) from 900 mK to 80 mK (at 1 mK.s$^{-1}$) followed by a 1000 s wait period. For the $^{162}$Dy and $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples the CC avalanche field is offset to higher fields, well outside the distribution of $H_{\rm ava}(t_{\rm w})$. For the $^{163}$Dy sample the CC curves falls within the distribution but near the long wait time curves. Also, we note for $^{163}$Dy, which happens to have better thermal contact, and thus faster cooling during the AQP, the CC curve again falls outside the distribution (shown in Supplementary Fig. 10 b). Thus slow cooling is more efficient at approaching equilibrium in [Dy$_{2}$Ti$_{2}$O$_{7}$]{} than is the AQP cooling followed by a long $t_{\rm w}$, especially for the low nuclear moment samples. This is typical behaviour for frustrated or disordered systems because slow cooling allows the system time to explore all available phase space.
Fig. 4c shows a much greater effect of $t_{\rm w}$ for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} with a larger spread of fields, saturating near 0.32 T for the the longest $t_{\rm w}$. This is again consistent with the conclusion that that the larger the nuclear spin moment, the more effective the spontaneous monopole recombination. The measurements were performed primarily at 200 mK, but the same conclusion follows from measurements at 80 mK. [Ho$_{2}$Ti$_{2}$O$_{7}$]{} also exhibits some unusual behaviour suggesting that the monopole density and magnetization do not approach equilibrium in a simple way. First, the magnetization jumps fall short of the $M$ vs $H$ equilibrium curve taken at 900 mK, even though thermometers placed on the sample indicate that the sample does indeed heat above 900 mK (see Supplementary Note 2 and Supplementary Fig. 3 for more details). Secondly, in contrast to the behaviour of [Dy$_{2}$Ti$_{2}$O$_{7}$]{} discussed above, the CC curve of [Ho$_{2}$Ti$_{2}$O$_{7}$]{} falls in the middle of the distribution of $H_{\rm ava}$($t_{\rm w}$) indicating, unusually, that waiting long enough at low temperature is an equally efficient way of approaching equilibrium as slow cooling.
Discussion {#discussion .unnumbered}
==========
The experimental result demonstrated here is that magnetic monopole dynamics in the frozen regime of spin ice are greatly enhanced by the hyperfine coupling of the electronic and nuclear moments. We now argue that this observation finds a natural – albeit surprising – explanation by analogy with the properties of single-molecule magnets [@review]. These are metal-organic clusters with large composite spins: some of the most studied include the so called Mn$_{12}$ and Fe$_8$ systems, both of which can be thought of as an ensemble of identical, weakly interacting nanomagnets of net spin S = 10 with an Ising-like anisotropy. The degenerate $M_{\rm s} = \pm S$ states are split by the ligand electric field into a series of doublets. At temperatures smaller than the level separation, the spins flip by resonant tunnelling through a quasi-classical barrier. The signature of a resonant tunnelling effect in Fe$_8$ is a peak in the low temperature relaxation rate around $H = 0$ [@Sa]. It quickly became clear that to understand the resonant tunnelling both dipolar and dynamic nuclear spin contributions to the interactions need to be accounted for. The typical dipolar field in such a system is $\approx 0.5$ K, and the relevant tunnel splitting $\Delta E$ of the order 10$^{-8}$ K, meaning that a broad distribution of dipolar field and a static hyperfine contribution would force all the spins off resonance. Prokof’ev and Stamp [@PS] proposed that dynamic nuclear fluctuations can drive the system to resonance, and the gradual adjustment of the dipole fields in the sample caused by tunnelling, brings other clusters into resonance and allows a continuous relaxation. Hence the observation of relaxation in single molecule magnets is fundamentally dependent on the hyperfine coupling with the fields of nuclear spins [@Wernsdorfer00].
The Prokof’ev and Stamp model [@PS] certainly does not apply in detail to spin ice at low temperatures. First, in single molecule magnets the spin of any particular complex in the system is available to be brought to resonance, whereas in spin ice, only those spins that are instantaneously associated with a diffusing monopole are available to tunnel (and this presumes that more extended excitations can be neglected). The remaining spins – the vast majority – are, in contrast, static and instantaneously ordered by the ice rules. The rate of flipping of these quasi-ordered spins, which corresponds to monopole pair creation, is negligible at the temperatures studied and the process is not relevant to our experiments. Thus, even at equilibrium, spin ice has an effective number of flippable spins that depends on temperature (see Supplementary Fig. 1). Away from equilibrium, where our experiments are performed, the number of flippable spins in spin ice further depends on time, with monopole recombination depleting their number. In addition, it seems reasonable to assume that the reduction of the density of monopoles is even more important during the relaxation process; as monopoles move through the matrix magnetizing the sample they will annihilate when they encounter a monopole of opposite charge, or become trapped on a defect or on the sample surface. This feature of spin ice is a second important difference with single molecule magnets, as modelled in Ref. 27.
A third difference relates to the distribution of internal fields in the system. In spin ice only, the actual field associated with a flippable spin, both before and after a flip, is a monopolar field. Flipping a spin transfers a monopole from site to site (Fig. 1a), dragging the monopolar field with it: a field that is much stronger and of longer range than any conventional dipole field. However, the change in field on a spin flip is dipolar, as in single molecule magnets.
In short, the flippable spins in spin ice are really an aspect of the emergent monopole excitation rather than a perturbed version of an isolated (composite) spin as assumed for the single molecule magnets in Ref. . Yet despite this difference, it seems reasonable to suggest that the basic idea of Ref. does apply to spin ice. The longitudinal monopolar fields will take flippable spins off resonance (Fig. 1c-e), while the transverse ones will tend to broaden the resonance well beyond the tunnel splitting calculated for an isolated spin. i.e. $\Delta E= 10^{-5}$ K [@Tomasello]. An applied field can also take flippable spins on or off resonance or broaden the resonance, depending on its direction. Nevertheless, in zero applied field, at very low temperatures we would expect all flippable spins associated with isolated monopoles to be off resonance and hence unable to relax, unless they are brought back to resonance by a combination of the monopole fields and the fluctuating nuclear spins: nuclear assisted flipping of spins will then bring further spins to resonance via the change in dipolar fields, as in the Prokof’ev-Stamp picture [@PS]. Our experimental results for the wait time dependence of various properties clearly support this proposition: in zero field (during $t_{\rm w}$) the sample with no nuclear spin is scarcely able to relax its monopole density, while the larger the nuclear spin, the quicker the relaxation. For flippable spins associated with closely–spaced monopole-antimonopole pairs the situation is slightly different. Although they are strongly off-resonance (Fig. 1b), the decreasing transition matrix elements will be compensated by the increasing Boltzmann factors required for detailed balance. Also, for the final recombination, a favourable change in exchange energy will reduce the field required to bring spins to resonance (see caption, Fig. 1).
We note in passing that the differences between single molecule magnets and spin ice are also evident in our data. Specifically, a $t^{1/2}$ initial relaxation of the magnetisation is a property of single molecule magnets, with the $t^{1/2}$ form arising from the dipole interactions [@PS; @Pauling-t-half]. Given the very unusual field distribution in spin ice, and the complicating factor of monopole recombination, as described above, it is hardly likely that this functional form will apply. We test for a $t^{1/2}$ decay in the Supplementary Fig. 5 and confirm that it can only be fitted over a narrow time range: to calculate the true time dependence in spin ice poses a theoretical problem.
Our main result has implications for both the theory of spin ice and the theory of nuclear spin assisted quantum tunnelling. First, in previous work [@Paulsen_Wien] we have shown how the low-temperature quenched monopole populations of [Dy$_{2}$Ti$_{2}$O$_{7}$]{} obey the nonlinear and non-equlibrium response of monopole theory [@Kaiser2] that was developed assuming a single hop rate. In view of our findings, the theory should apply most accurately to the [Dy$_{2}$Ti$_{2}$O$_{7}$]{} sample with no nuclear spins and least accurately to [Ho$_{2}$Ti$_{2}$O$_{7}$]{} where the hyperfine splitting energies are of a similar order to the Coulomb energies. In other measurements, presented in Supplementary Figs. 7 and 8, we confirm that this is the case; hence a generalisation of the theory of Ref. to include the effect of nuclear spins seems an attainable goal. We also note that [Ho$_{2}$Ti$_{2}$O$_{7}$]{} offers the unusual situation that, at low temperatures ($<$ 0.35 K) and sufficient wait times, the nuclear spins are ice-rule ordering antiparallel to their electronic counterparts; hence spin ice offers a rare chance to investigate the effect of correlation on nuclear spin assisted quantum tunnelling in a controlled environment. Perhaps this will shed light on some of the unusual properties particular to [Ho$_{2}$Ti$_{2}$O$_{7}$]{}, as noted above.
Spin ice thus exemplifies a remarkable extension of the concept of nuclear spin assisted quantum tunnelling [@PS] to the motion of fractionalised topological excitations [@CMS]. This is made possible by the fact that the emergent excitations of the system – the monopoles – are objects localised in direct space that move through flipping spins. As well as illustrating this generic point, our result may also have practical consequences. We have established how coupling with nuclear spins controls the magnetic monopole current and the spectacular magnetothermal avalanches: hence any experimental handle on the nuclear spins of the system would also be a rare experimental handle on the monopole current. Any future application of magnetic monopoles in spin ice will surely rely on the existence of such experimental handles.
Methods {#methods .unnumbered}
=======
[*Samples.* ]{}Single crystals were grown by the floating zone method for all samples, the natural $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} samples (DTO, HTO) were prepared at the Institute of Solid State Physics, University of Tokyo, Japan, and $^{162}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}, $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} at Warwick University and Oxford University respectively.
[*Measurements.* ]{} Measurements were made using a low temperature SQUID magnetometer developed at the Institut Néel in Grenoble. The magnetometer is equipped with a miniature dilution refrigerator with a base temperature of 65 mK. The fast dynamics after a field change were measured in a relative mode, the slower measurements were made by the extraction method, and the initial relative measurements were adjusted to the absolute value extraction points. The field could be rapidly changed at a rate up to 2.2 T s$^{-1}$.
For all the data shown here the field was applied along the \[111\] crystallographic direction. Measurements were also performed perpendicular to the \[111\] direction, as well as along the \[001\] and \[011\] directions and on a polycrystalline sample, examples of which are discussed in Supplementary Note 3. In total ten different samples were studied. The direction of the applied field as well as differences in the sample shapes and thermal contact with the sample holder can effect some of the details of the measurements. However this does not change the main conclusion of the paper: the demonstration of the importance of nuclear assisted quantum tunnelling to the relaxation.
The measurements of temperature vs time shown in Fig. 2, a bare-chip Cernox 1010-BC resistance thermometer from LakeShore Cryogenics was wrapped in Cu foil and glued on top of the sample as shown in the inset of Fig. 2b.
Cooling [Ho$_{2}$Ti$_{2}$O$_{7}$]{} was difficult and warming was also tricky using the AQP, depending on the initial temperatures and wait times. Therefore to ensure the sample was heated above 900 mK, two AQP were used, separated by 300 seconds, which explains why the starting temperature for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} was higher in Fig. 2 (see Supplementary Note 2 and Supplementary Fig. 3 for further discussions).
A schematic of the AQP used for the preparation of the samples is shown in Fig. 3d. First a field of $-0.3$ T was applied and the sample was allowed to cool to base temperature for 20 minutes. The field was then reversed at 2.2 T s$^{-1}$ to $+ 0.3$ T for 4 s then reduced to zero. After a wait period ranging from 10 to 50,000 s, a field of 0.08 T was applied and the relaxation of the magnetization was recorded. The field $B= 0.08$ T was chosen because it is large enough to get sizeable relaxation, but small compared to the avalanche fields shown in Fig. 4. In this way, when applying the magnetic fields, the relaxation is well behaved and the sample does not heat.
The AQP used for the data of Fig. 4 was similar to the above, except the avalanche field was $\pm 0.4$ T. After the wait period the field was ramped at 0.02 T s$^{-1}$ while the magnetization and temperature of the sample were continuously recorded. For the slow CC protocol measurements shown in Fig. 3, the samples were first heated to 900 mK for 10 s, then cooled at a rate of approximately 0.01 K s$^{-1}$, followed by a waiting period of 1000 s.
Data availability {#data-availability .unnumbered}
=================
Information on the data underpinning the results presented here, including how to access them, can be found in the Cardiff University data catalogue at http://doi.org/10.17035/d.2019.0069144874.\
The datasets obtained and/or analyzed in this study are also available from the corresponding author on reasonable request.
[99]{} Harris, M. J., Bramwell, S. T., McMorrow, D. F., Zeiske T. & Godfrey, K. W. Geometrical frustration in the ferromagnetic pyrochlore Ho$_2$Ti$_2$O$_7$. [*Phys. Rev. Lett.*]{} [**79**]{}, 2554 – 2557 (1997). Bramwell, S. T. & Harris, M. J. Frustration in Ising-type spin models on the pyrochlore lattice. [*J. Phys.: Condens. Matter*]{} [**10**]{}, L215 - L220 (1998). Ramirez, A. P., Hayashi, A., Cava, R. J., Siddharthan, R. B. & Shastry, S. Zero-point entropy in Ôspin iceÕ. [*Nature*]{} [**399**]{}, 333 – 335 (1999). Bramwell, S. T. & Gingras, M. J. P. Spin ice state in frustrated magnetic pyrochlore materials. [*Science*]{} [**294**]{}, 1495 – 1501 (2001). Luttinger, J. M & Tisza, L. Theory of dipole interactions in crystals. [*Phys. Rev.*]{} [**70**]{}, 954 (1946). Castelnovo, C., Moessner, R. & Sondhi, S. L. Magnetic monopoles in spin ice. [*Nature*]{} [**451**]{}, 42–45 (2008). den Hertog, B. C. & Gingras, M. J. P. Dipolar interactions and origin of spin ice in Ising pyrochlore magnets. [*Phys. Rev. Lett.*]{} [**84**]{}, 3430 (2000). Isakov, S. V., Moessner, R. & Sondhi, S. L. Why spin ice obeys the ice rules [*Phys. Rev. Lett.*]{} [**95**]{}, 217201 (2005). Ryzhkin, I. A. Magnetic relaxation in rare-earth pyrochlores. [*J. Exp. and Theor. Phys.*]{} [**101**]{}, 481–486 (2005).
Kaiser, V., Bloxsom, J. A., Bovo, L., Bramwell, S. T., Holdsworth, P. C. W. & Moessner, R. Emergent Electrochemistry in Spin Ice: Debye–H[ü]{}ckel Theory and Beyond. [*Phys. Rev. B*]{} [**98**]{}, 144413 (2018). Paulsen, C., Giblin, S. R., Lhotel, E., Prabhakaran, D., Balakrishnan, G., Matsuhira, K. & Bramwell, S. T. Experimental signature of the attractive Coulomb force between positive and negative magnetic monopoles in spin ice. [*Nature Phys.*]{} [**12**]{}, 661 (2016). Jaubert, L. D. C. & Holdsworth, P. C. W. Signature of magnetic monopole and Dirac string dynamics in spin ice. [*Nature Phys.*]{} [**5**]{}, 258 - 261 (2009). Bovo, L., Bloxsom, J. A., Prabhakaran, D., Aeppli, G. & Bramwell, S. T. Brownian motion and quantum dynamics of magnetic monopoles in spin ice. [*Nature Comms.*]{} [**4**]{}, 1535 (2013). Tomasello, B., Castelnovo, C., Messier, R. & Quintanilla, J. Single-ion anisotropy and magnetic field response in the spin-ice materials Ho$_{2}$Ti$_{2}$O$_{7}$ and Dy$_{2}$Ti$_{2}$O$_{7}$. [*Phys. Rev. B*]{} [**92**]{}, 155120 (2015). Vedmedenko, E. Y. Dynamics of bound monopoles in artificial spin ice: How to store energy in Dirac strings. [*Phys. Rev. Lett.*]{} [**116**]{}, 077202 (2016). Melko, R. G., den Hertog, B. C. & Gingras, M. J. P. Long range order at low temperatures in dipolar spin ice. [*Phys. Rev. Lett.*]{} [**87**]{}, 067203 (2001). Ehlers, G., Cornelius, A. L., Fennell, T. Koza, M., Bramwell, S. T. & Gardner, J. S. Evidence for two distinct spin relaxation mechanisms in ‘hot’ spin ice Ho$_{2}$Ti$_{2}$O$_{7}$. [*J. Phys.: Condens. Matter*]{} [**16**]{}, S635–S642 (2004). Snyder, J., Ueland, B. G., Slusky, J. S., Karunadasa, H. Cava, R. J. & Schiffer, P. Low-temperature spin freezing in the Dy$_2$Ti$_2$O$_7$ spin ice. [*Phys. Rev. B*]{} [**69**]{}, 064414 (2004). Castelnovo, C., Moessner, R. & Sondhi, S. L. Thermal quenches in spin ice. [*Phys. Rev. Lett.*]{} [**104**]{}, 107201 (2010). Paulsen, C., Jackson, M. J., Lhotel, E., Canals, B., Prabhakaran, D., Matsuhira, K., Giblin, S. R. & Bramwell, S. T. Far-from-equilibrium monopole dynamics in spin ice. [*Nature Physics*]{} [**10**]{}, 135–139 (2014).
Planck Collaboration. Planck Early Results. II. The thermal performance of Planck, [*Astronomy & Astrophysics*]{} [**536**]{} (2011). doi: 10.1051/0004-6361/201116486. Slobinsky, D., Castelnovo, C., Borzi, R. A., Gibbs, A. S., Mackenzie, A. P., Moessner, R. & Grigera, S. A. Unconventional Magnetization Processes and Thermal Runaway in Spin-Ice Dy$_2$Ti$_2$O$_7$. [*Phys. Rev. Lett.*]{} [**105**]{}, 267205 (2010). Jackson, M. J., Lhotel, E., Giblin, S. R., Bramwell, S. T., Prabhakaran, D., Matsuhira, K., Hiroi, Z., Yu, Q. & Paulsen, C. Dynamic behavior of magnetic avalanches in the spin-ice compound Dy$_{2}$Ti$_{2}$O$_{7}$. [*Phys. Rev. B*]{} [**90**]{}, 064427 (2014). Krey, C., Legl, S., Dunsiger, S. R., Meven, M., Gardner, J. S., Roper, S. R. & Pfleiderer, C. First Order Metamagnetic Transition in Ho$_2$Ti$_2$O$_7$ Observed by Vibrating Coil Magnetometry at Milli-Kelvin Temperatures. [*Phys. Rev. Lett.*]{} [**108**]{} 257204.
Gatteschi, D. & Sessoli, R. Quantum Tunneling of Magnetization and Related Phenomena in Molecular Materials. [*Angewandte Chemie*]{} [**42**]{}, 268 – 297 (2003).
Sangregorio, C., Ohm, T., Paulsen, C., Sessoli, R. and Gatteschi, D. Quantum tunneling of the magnetization in an Iron cluster nanomagnet. [*Phys. Rev. Lett.*]{} [**78**]{}, 4645 – 4648 (1997). Prokof’ev, N. V. & Stamp, P. C. E. Low-temperature quantum relaxation in a system of magnetic nano molecules. [*Phys. Rev. Lett.*]{} [**80**]{}, 5794 (1998).
Wernsdorfer, W., Caneschi, A., Sessoli, R., Gatteschi, D., Cornia, A., Villar, V., & Paulsen, C. Effects of Nuclear Spins on the Quantum Relaxation of the Magnetization for the Molecular Nanomagnet Fe$_8$. [*Phys. Rev. Lett.*]{} [**84**]{}, 2965 – 2968 (2000).
Ohm, T., Sangregorio, C. & Paulsen, C. [*J. Low Temp. Physics*]{} [**113**]{}, 1141 – 1146 (1998). Kaiser, V., Bramwell, S. T., Holdsworth, P. C. W. & Moessner, R. ac Wien Effect in Spin Ice, Manifest in Nonlinear, Nonequilibrium Susceptibility. [*Phys. Rev. Lett.*]{} [**115**]{}, 037201 (2015).
[**Acknowledgements:**]{} S.R.G. thanks Cardiff University for ‘seedcorn’ funding, and acknowledges the EPSRC for EP/L019760/1. S.T.B. thanks Patrik Henelius for communicating his independent ideas on nuclear assisted quantum tunnelling in spin ice, and acknowledges the EPSRC for EP/S016465/1. E.L. and C.P. acknowledge financial support from ANR, France, Grant No. ANR-15-CE30-0004. G.B. wishes to thank financial support from EPSRC, UK, through grant EP/M028771/1.\
[**Author Contributions:**]{} The experiments were designed and performed by C.P. with inputs and discussions from E.L. and S.R.G. The data were analyzed by C.P., E.L., S.R.G., and S.T.B. Contributed materials were fabricated by K.M., D.P. and G.B. The paper was written by C.P., E.L., S.R.G., and S.T.B.
[**Nuclear spin assisted quantum tunnelling of magnetic monopoles in spin ice\
Supplementary Information**]{}
![\[fig1SI\] [**Density of single-charge monopoles (equilibrium number per diamond lattice site in zero field) versus temperature**]{}, calculated by the Debye-Hückel theory of Kaiser et al. [@KaiserDH]. The analytic calculation is very accurate for the monopole model of spin ice: it includes both single and double-charge monopoles, but only the single-charge monopoles are relevant at the temperatures we study. The density of ‘flippable’ spins per spin site is $3/2$ times the monopole density.](Plot_SI_nvsT.pdf){width="8cm"}
Experimental details
--------------------
[**Creating a large density of monopoles using the avalanche quench protocol in [Ho$_{2}$Ti$_{2}$O$_{7}$]{}.**]{} We have previously described the avalanche quench protocol (AQP) in detail for [Dy$_{2}$Ti$_{2}$O$_{7}$]{} (see Ref. and its Supplementary Information). From magnetisation measurements recorded during and after the AQP, we inferred that samples of [Dy$_{2}$Ti$_{2}$O$_{7}$]{} heat systematically to temperatures above 900 mK, even though the reference thermometer on the sample holder only registered a small jump [@Jackson14].
However from the onset, it was clear that [Ho$_{2}$Ti$_{2}$O$_{7}$]{} was different. For example, when performing measurements where the field is ramped at a steady rate, the magnetic avalanches of [Ho$_{2}$Ti$_{2}$O$_{7}$]{} never reach the 900 mK equilibrium value as seen in Fig. 4c. Sometimes depending on previous measurements, the AQP worked very poorly, or did not seem to work at all.
This was the motivation for measuring the sample temperature directly by mounting a thermometer on the samples during some of the runs.
[**Temperature measurements during the AQP.**]{} We attempted direct temperature measurements with 4 different thermometers; 2 homemade RuO$_2$ resistance thermometers (filed down to reduce mass with wires attached with silver epoxy), a Cernox 1010-SD thermometer with platinum leads, and a bare-chip Cernox 1010-BC, both from LakeShore Cryogenics. All had short comings but in the end most of the measurements shown here were made using the bare chip thermometer. This was the lightest of the four, but had a small but noticeable magnetoresistance that we have corrected for. A constant current source delivered 10 nA, and the voltage was measured with a Stanford Instruments model 830 lock-in amplifier running at 1100 Hz. This setup was a compromise, the measurements of the temperature were fast, but prone to some drift and noise.
During normal measurements samples are sandwiched between two long narrow pieces of Cu that are anchored to the mixing chamber of a miniature dilution refrigerator. The samples are glued in place, then teflon tape is tightly wrapped around the Cu strips, clamping the samples to the Cu. For measurements with the thermometer glued to the sample, only one Cu strip was used, the second was suspended away from the thermometer, as shown in Supplementary Figure \[fig1SI\].
![\[fig1SI\] [**Pictures of the sample mounting**]{}. Upper left: Sample of [Dy$_{2}$Ti$_{2}$O$_{7}$]{} and Cernox bare chip resistor with leads protected by kapton tape. Upper right: [Ho$_{2}$Ti$_{2}$O$_{7}$]{} glued onto the bare chip resistor with thin Cu foil protruding. Lower left: the sample + thermometer have been turned over, and the sample surface has been glued onto the Cu sample holder using GE varnish. The Cu foil has been folded back to cover the upper half of the sample and bare chip resistor. Lower right: standoffs hold upper part of sample holder away, teflon tape has been wrapped around the sample and thermometer. The thermometer is isolated from the Cu sample holder by the sample.](figure1_SI.pdf){width="8.5cm"}
Results
-------
As already discussed, cooling samples that contain Ho can be problematic, and this is also true for heating samples with the AQP when the starting temperature was well below 200 mK. This is shown in Supplementary Figure \[fig2SI\] for a series of AQP taken on [Ho$_{2}$Ti$_{2}$O$_{7}$]{} when the sample was first cooled to 65 mK after waiting 4 hours. Point (a) is the beginning of sequence when we applied a field of $-0.3$ T on the sample followed by a wait period of 180 s. For [Dy$_{2}$Ti$_{2}$O$_{7}$]{}, already a large $>1$ K spike in temperature would be seen at (a) as the sample rapidly magnetizes in the field, but for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} only a small jump of about 0.3 K was recorded. The jump in temperature at the first AQP is also small, only reaching about 0.6 K, compared to $>1.4$ K for [Dy$_{2}$Ti$_{2}$O$_{7}$]{} under similar conditions. At (c) we begin a second AQP, again setting $\mu_0 H=-0.3$ T, and waiting 180 s. But this time the jump in temperature reaches nearly 0.8 K, and the second AQP at (d) shows that now, the sample has warmed $>1.3$ K.
During the AQP, heat from the flipping of the electronic spins is absorbed by the sample. But because of the large heat capacity of the Ho nuclei, much of the energy is absorbed by the nuclear spin bath, raising its temperature but resulting in a small overall jump in sample temperature. However for the second AQP, the starting sample temperature is now greater, nearly 0.16 K, and this is enough to heat the spins above one Kelvin.
Thus [Ho$_{2}$Ti$_{2}$O$_{7}$]{} measurements were systematically made with 2 or 3 AQP in succession in order to ensure the sample is warmed above 1 K.
Note that the need for several AQP also suggests that at low temperature, well below 300 mK, the nuclear spins begin to freeze out, and anti-align with their respective electronic spin, thus 2 in – 2 out for the electronic spin becomes 2 out – 2 in for its nuclear counterpart.
![\[fig2SI\] [**The applied field and the sample temperature for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} as a function of running time for a double AQP.**]{} The sample was cooled for 4 hours to base temperature of approximately 65 mK. (a) At $t=0$, the field was changed from 0 to $-0.3$ T. A relatively weak jump in the temperature can be seen. (b) At $t=180$ s, the first AQP is performed: the field goes from $-0.3$ to $+0.3$ T, then after 4 s from $+0.3$ T to zero. The temperature on the sample reaches about 0.55 K, not sufficient to randomize the spins. (c) At $t=200$ s the field is again put at $-0.3$ T in preparation for the next AQP. This time the jump in sample temperature is larger, but still less than required. (d) At $t=380$ s the second AQP takes place, warming the sample above 1 K. ](figure2_SI.pdf){width="8cm"}
Another nagging problem was that the samples of [Ho$_{2}$Ti$_{2}$O$_{7}$]{} did not reach $M=0$ after the AQP. The origin of this is not clear, but our data suggests that [Ho$_{2}$Ti$_{2}$O$_{7}$]{} cools too fast. When the field is switched off, the applied field $H$ goes to zero before the sample even starts to change its magnetisation. The sample then feels the internal field $H_{\rm internal}=-D\cdot M$ where D is the demagnetisation factor and avalanches against this. As the magnetisation decreases, $H_{\rm internal}$ also decreases, the sample heats, but then cools so rapidly that the magnetisation gets ‘stuck’ at a small positive value of the order 1 emu/g. For convenience, the solution was to add a small overshoot for the field of about $-0.004$ T for 1 s, then switch back to $H=0$. This resulted in a starting $M$ closest to zero. Note that measurements without the overshoot gave the same results, but with an offset. This ultra rapid cooling may also explain why the avalanches shown in Fig. 4c for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} (while ramping of the field) fall below the equilibrium value expected for 900 mK, in contrast to [Dy$_{2}$Ti$_{2}$O$_{7}$]{}.
![\[fig3SI\] [**Plots for natural [Dy$_{2}$Ti$_{2}$O$_{7}$]{} corresponding to Figs. 3 and 4 of the main manuscript.**]{} (a) Avalanches of the magnetisation recorded while the field was ramped at 0.02 T/s for $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} (DTO). The samples were first prepared using the AQP and then followed by various wait times except for the curve marked ‘ZFC’, where the sample was first prepared using the conventional zero field cooled (CC) protocol (red circles). Also shown is the equilibrium $M$ vs $\mu_0 H$ taken at 900 mK (solid black dots). (b) Magnification, showing the spread in avalanche fields as a function of the wait time, and the ZFC far outside the pack. (c) The effect of wait time on the relaxation of the magnetisation $M$ vs time for $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} measured at 80 mK. The samples were again first prepared using the AQP. After the specified wait periods, a field of 0.08 T was applied and the magnetisation as a function of time was recorded.](figure3_SI.pdf){width="\textwidth"}
![\[fig5SI\] [**Time dependence of the magnetisation at short times**]{}. The same data of Fig. 3 in the main text as well as data for the $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} sample plotted against $\sqrt{time}$. An important prediction from ProkofÕev and Stamp [@PS] was that the initial relaxation of the magnetisation should follow a square root time dependence. This worked well for the SMM Fe$_8$ up to 1000 s or more. The situation for spin ice is quite different, the right hand side of the figure shows the data can at best be fitted over a very restricted range in time only up to 1 s. ](figure5_SI.pdf){width="13cm"}
![\[fig6SI\] [**Experimental values of relaxation time $\tau$ from susceptibility and magnetisation measurements.**]{} Relaxation time $\tau$ vs temperature for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} and $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}. The green and blue data points ($\tau$ less than 100 s) were taken from the peaks in the imaginary susceptibility. The red points for [Ho$_{2}$Ti$_{2}$O$_{7}$]{} come from analyzing dc relaxation (all raw data was first corrected for demagnetisation effects). The slope $dT/d\tau$ defining the equilibrium cooling rate shown in Fig. 2 are taken from fits to these curves. ](figure6_SI.pdf){width="7.2cm"}
![\[fig7SI\] [**Effect of monopole current $J_m=dM/dt$ on wait time for [Ho$_{2}$Ti$_{2}$O$_{7}$]{}**]{} left: Monopole current $J_m=dM/dt$ obtained from the relaxation of the magnetisation of [Ho$_{2}$Ti$_{2}$O$_{7}$]{} measured after different waiting times, and measured at 800 Oe. $J_m$ is obtained by extrapolating the derivative of the magnetisation with respect to time at $t=0$ (See Ref. for the detailed procedure). right: $J_m$ vs $1/T$ obtained from the saturation value of the left figure, i.e. when the current value does not depend anymore on the waiting time. ](i_800_Ho.pdf "fig:"){width="7.2cm"} ![\[fig7SI\] [**Effect of monopole current $J_m=dM/dt$ on wait time for [Ho$_{2}$Ti$_{2}$O$_{7}$]{}**]{} left: Monopole current $J_m=dM/dt$ obtained from the relaxation of the magnetisation of [Ho$_{2}$Ti$_{2}$O$_{7}$]{} measured after different waiting times, and measured at 800 Oe. $J_m$ is obtained by extrapolating the derivative of the magnetisation with respect to time at $t=0$ (See Ref. for the detailed procedure). right: $J_m$ vs $1/T$ obtained from the saturation value of the left figure, i.e. when the current value does not depend anymore on the waiting time. ](i_800sat_vs1-T.pdf "fig:"){width="7.5cm"}
![\[fig8SI\] [**Effect of the nuclear spins on the monopole current**]{}. Monopole current $J_m=dM/dt$ vs $\sqrt{H}$ determined for $^{\rm nat}$Dy and $^{162}$ [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples (left), $^{163}$ [Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} samples (middle), and in different cooling conditions for the [Ho$_{2}$Ti$_{2}$O$_{7}$]{} sample. $J_m$ is obtained by extrapolating the derivative of the magnetisation with respect to time at $t=0$ (See Ref. for the detailed procedure). Natural and $^{162}$Dy (no nuclear spin) [Dy$_{2}$Ti$_{2}$O$_{7}$]{} follows the $\sqrt{H}$ behavior expected for magnetic monopoles interacting through the Coulomb force, $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} samples do not, whatever the cooling process and so the initial density of monopoles. This result shows, as suggested in the main text, that the idealised emergent chemical kinetics of monopole theory does not apply in $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{}, where the dynamics is strongly affected by the nuclear spin effects, because the hyperfine splitting energies are of a similar order to the Coulomb energies. ](ivsh_DTO_no_N.pdf "fig:"){width="5.5cm"} ![\[fig8SI\] [**Effect of the nuclear spins on the monopole current**]{}. Monopole current $J_m=dM/dt$ vs $\sqrt{H}$ determined for $^{\rm nat}$Dy and $^{162}$ [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples (left), $^{163}$ [Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} samples (middle), and in different cooling conditions for the [Ho$_{2}$Ti$_{2}$O$_{7}$]{} sample. $J_m$ is obtained by extrapolating the derivative of the magnetisation with respect to time at $t=0$ (See Ref. for the detailed procedure). Natural and $^{162}$Dy (no nuclear spin) [Dy$_{2}$Ti$_{2}$O$_{7}$]{} follows the $\sqrt{H}$ behavior expected for magnetic monopoles interacting through the Coulomb force, $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} samples do not, whatever the cooling process and so the initial density of monopoles. This result shows, as suggested in the main text, that the idealised emergent chemical kinetics of monopole theory does not apply in $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{}, where the dynamics is strongly affected by the nuclear spin effects, because the hyperfine splitting energies are of a similar order to the Coulomb energies. ](ivsh_nuk_Hob4.pdf "fig:"){width="5.5cm"} ![\[fig8SI\] [**Effect of the nuclear spins on the monopole current**]{}. Monopole current $J_m=dM/dt$ vs $\sqrt{H}$ determined for $^{\rm nat}$Dy and $^{162}$ [Dy$_{2}$Ti$_{2}$O$_{7}$]{} samples (left), $^{163}$ [Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} samples (middle), and in different cooling conditions for the [Ho$_{2}$Ti$_{2}$O$_{7}$]{} sample. $J_m$ is obtained by extrapolating the derivative of the magnetisation with respect to time at $t=0$ (See Ref. for the detailed procedure). Natural and $^{162}$Dy (no nuclear spin) [Dy$_{2}$Ti$_{2}$O$_{7}$]{} follows the $\sqrt{H}$ behavior expected for magnetic monopoles interacting through the Coulomb force, $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{} samples do not, whatever the cooling process and so the initial density of monopoles. This result shows, as suggested in the main text, that the idealised emergent chemical kinetics of monopole theory does not apply in $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} and [Ho$_{2}$Ti$_{2}$O$_{7}$]{}, where the dynamics is strongly affected by the nuclear spin effects, because the hyperfine splitting energies are of a similar order to the Coulomb energies. ](ivsh_Hob4.pdf "fig:"){width="5.5cm"}
![\[fig9SI\] [**The effect of wait time $t_{\rm w}$ on the magneto-thermal avalanches for two different samples of natural DTO, and for two different runs**]{} (a) shows the value of the avalanche field $H_{\rm ava}$($t_{\rm w}$), defined as the field where the magnetization crosses 1 $\mu_{\rm B}$ per rare earth ion. (b) is a plot of the difference in avalanche field $\Delta H_{\rm ava}= H_{\rm ava}(t_{\rm w}) - H_{\rm ava} (t_{\rm w}=minimum)$. ](figure9_SI.pdf){width="16cm"}
Different Samples and Measuring Directions
------------------------------------------
During the course of this study, 10 different samples were measured with some samples measured along multiple axis. The mass of the samples ranged between 3 to 40mg, and they had various shapes. No corrections for demagnetization effects have been taken into account for the results presented in the main text. This is because for most of the measurements shown in the main text, the sample was far from equilibrium, and the magnetization was very small, and thus the demagnetizing field -NM was small. Nevertheless the sample shape and field direction do effect the observed relaxation curves and the avalanche fields.
In this section we show that although there are some variations between samples, between cooling runs, and for different directions, these differences do not change the main conclusion of the paper; the demonstration that nuclear assisted quantum tunneling is operative regardless of field direction.
Supplementary Figure 9(a) shows the effect of $t_{\rm w}$ on the magneto-thermal avalanches for two different samples of natural DTO, and for two different runs. Sample 1 (also shown in the main text) was rectangular shaped parallelepiped and sample 2 was a square thin platelet shaped sample. The measurements shown in the figures were taken with the field along the \[111\] axis for both samples. The left panel shows the value of the avalanche field $H_{\rm ava}$($t_{\rm w}$), defined as the field where the magnetization crosses 1 $\mu_{\rm B}$ per rare earth ion. The curves are clearly offset from one another, even the two curves taken on the same sample, but during different runs. The initial position of the avalanche field is very sensitive to thermal contact with the sample holder. For sample 2 run 1, the thermal contact was made using two Cu bands with the sample sandwiched between the two. For sample 2 run 2 only one Cu band was used, thus the thermal contact was worse. The better thermalized sample has a higher avalanche field, because leading up to the avalanche heat could be more efficiently evacuated from the sample. Supplementary Figure 9(b) is a plot of the difference in avalanche field $\Delta H_{\rm ava}= H_{\rm ava}(t_{\rm w}) - H_{\rm ava} (t_{\rm w}=minimum)$. As can be seen, for sample 2, the two runs collapse onto one another, but the shape of the curve for sample 1 is slightly different. A more systematic study needs to be made to understand if this is a shape dependent effect, or sample dependent.
![\[fig10SI\] [**Measurements made on a polycrystalline sample of [$^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}]{} (sample 2)**]{} The sample was first prepared using the same avalanche quench protocols (AQP) outlined in the main text and methods section, followed by various waiting times. (a) shows the effects of wait time on the relaxation of the magnetization in a field of 0.8 T at 80mK. (b) shows the effects of wait time on the position of the avalanche field when the field is ramped from 0 to 0.4 T at a constant rate of 0.02 T/s at 80mK. Also shown in the figure is the conventional zero field cooling (CC) curve were the sample was slowly cooled from 900 mK to 80 mK (at 1 mK/s) followed by a 1000 s wait period. For this sample the CC avalanche field is offset to higher fields, and is well outside the distribution of $H_{\rm ava}$($t_{\rm w}$). ](figure10_SI.pdf){width="16cm"}
For $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}, two samples were studied. Sample 1 was an odd shaped disk. The \[111\] direction was perpendicular to the surface of the disk, and resulted in a very large demagnetization factor for this direction. Importantly this resulted in difficulty thermalizing the sample along this direction to our Cu sample holder. This resulted in a much less efficient AQP cooling. We estimate that the sample took about 5 seconds to cool below 500mK, and about 40 seconds to cool below 100mK. This is much slower than the usual AQP as described in Fig. 2 of the main text, but still faster than the CC method. Sample 1 was measured along the \[111\] direction (shown in the main text) and perpendicular to the \[111\] direction. Sample 2 was a polycrystalline sample and the effects of wait time on the relaxation of the magnetization and position of the avalanche field are shown in Supplementary Figure 10 (a) and (b). These data sets are very similar to those presented in the main text in terms of the strength of the effect of wait time for $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}. (see Figures 3 and 4)
However there are two interesting differences.
Firstly, Supplementary Figure 11 (a) shows the monopole current $J_{\rm m}=dM/dt$ at $t=0$ vs log wait time for sample 1 \[111\] and perpendicular to \[111\] as well as for polycrystalline sample 2. As can be seen in the figure, although the slopes of the three curves are roughly the same, the \[111\] data fall significantly below the two perpendicular curves. Most likely this is not an intrinsic effect, but comes from the poor thermalization for the \[111\] sample run: as mentioned above, for this direction after the AQP the sample cooled much slower, therefore the initial monopole density at the beginning of the wait period was much reduced, so the initial monopole current was less, shifting the \[111\] curve down in the plot. A second difference that can be seen in Supplementary Figure 10 (b) is that the avalanche field for the CC method occurs at much higher fields and is well outside the distribution of curves obtained by the AQP method. The same result was found for sample 1 perpendicular to \[111\]. This can be contrast to the data shown for the \[111\] sample in the main text, and again can be explain by the slower cooling for the \[111\] sample run.
![\[fig11SI\] [**comparison of different samples and different measuring directions for HTO, [$^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}]{} and $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} (DTO)**]{} (a) the monopole current $J_{\rm m}=dM/dt$ at $t=0$ vs log wait time for two samples of [$^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}]{} and three samples of HTO. (b) Plot of difference in avalanche field $\Delta H_{\rm ava}= H_{\rm ava}(t_{\rm w}) - H_{\rm ava}(t_{\rm w}=minimum)$ against wait time for various directions and various samples of HTO, [$^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}]{} and $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} (DTO). The top 4 curves in the figure are measurements for 3 different samples of HTO (squares). Sample 1 was a needle shaped sample measured along the \[111\] (long) direction. Sample 2 was also needle shaped and measured along the \[001\] direction. Sample 3 was a square platelet, and was measure along the \[001\] and \[110\] axis. The middle 3 curves are for two different samples of $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}(solid dots) measured at 80mK. Sample 1 was measured along the \[111\] direction and perpendicular to the \[111\] direction, and sample 2 was a poly-crystal. The bottom two curves are for natural DTO (triangles) taken on two different samples along the \[111\] and \[001\] directions. (same data as shown in Supplementary Figure 9 (b)) ](figure11_SI.pdf){width="16cm"}
The results for measurements on 3 different samples of HTO are also shown in Supplementary Figure 11. Sample 1 was a needle shaped sample measured along the \[111\] (long) direction. Sample 2 was also needle shaped and measured along the \[001\] direction. Sample 3 was a square platelet, and was measure along the \[001\] and \[110\] axis. Supplementary Figure 11 (a) shows the monopole current $J_{\rm m}=dM/dt$ at $t=0$ vs log wait time for sample 1 \[111\] (also in the main text) compared to sample 2 \[110\]. The effect of wait time on the currents for these two samples are very similar; the rate at which monopoles recombine is seen to be much faster than that of $^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}, and both seem to saturate at very long wait times. Supplementary Figure 11 (b) are plots of difference in avalanche field $\Delta H_{\rm ava}= H_{\rm ava}(t_{\rm w}) - H_{\rm ava}(t_{\rm w}=minimum)$ against log wait time for the three samples of HTO, as well as two samples of [$^{163}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{}]{} and for two samples of $^{\rm nat}$[Dy$_{2}$Ti$_{2}$O$_{7}$]{} (DTO) (same data as shown in Supplementary Figure 9 (b)).
[99]{} Kaiser, V., Bloxsom, J. A., Bovo, L., Bramwell, S. T., Holdsworth, P. C. W. & Moessner, R. Emergent Electrochemistry in Spin Ice: Debye–H[ü]{}ckel Theory and Beyond. [*Phys. Rev. B*]{} [**98**]{}, 144413 (2018).
Paulsen, C., Giblin, S. R., Lhotel, E., Prabhakaran, D., Balakrishnan, G., Matsuhira, K., and Bramwell S. T. Experimental signature of the attractive Coulomb force between positive and negative magnetic monopoles in spin ice. [*Nature Physics*]{} [**12**]{}, 661 (2016).
Jackson, M. J., Lhotel, E., Giblin, S. R., Bramwell, S. T., Prabhakaran, D., Matsuhira, K., Hiroi, Z., Yu, Q., and Paulsen, C. Dynamic behavior of magnetic avalanches in the spin-ice compound [Dy$_{2}$Ti$_{2}$O$_{7}$]{}. [*Phys. Rev. B*]{} [**90**]{}, 064427 (2014).
Prokof’ev, N. V. & Stamp, P. C. E. Low-temperature quantum relaxation in a system of magnetic nano molecules. [*Phys. Rev. Lett.*]{} [**80**]{}, 5794 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Construction of ambiguity set in robust optimization relies on the choice of divergences between probability distributions. In distribution learning, choosing appropriate probability distributions based on observed data is critical for approximating the true distribution. To improve the performance of machine learning models, there has recently been interest in designing objective functions based on $L_p$-Wasserstein distance rather than the classical Kullback-Leibler (KL) divergence. In this paper, we derive concentration and asymptotic results using Bregman divergence. We propose a novel asymmetric statistical divergence called Wasserstein-Bregman divergence as a generalization of $L_2$-Wasserstein distance. We discuss how these results can be applied to the construction of ambiguity set in robust optimization.'
author:
- 'Xin Guo, Johnny Hong, Nan Yang'
bibliography:
- 'paper.bib'
title: Ambiguity set and learning via Bregman and Wasserstein
---
Introduction
============
Comparing probability distributions has been a recurring theme in many research areas of machine learning. In distribution learning, for example, one is interested in approximating the true distribution by an element in a predetermined class of probability distributions, and this element is chosen based on the observed data. Such choices rely on the divergence used in comparing distributions. While there is an abundance in statistical divergences, there is no consensus about the “ideal" way to measure the difference between distributions. In the theory of robust optimization, optimization problems are formulated under appropriate uncertainty sets for the model parameters and/or for the solutions against a certain measure of robustness. For instance, tractable uncertainty sets can be formulated in terms of chance constraints and expectation constraints under a given distribution ${\mathbb{P}}$ [@Guan:2012ec]. However, when the distribution ${\mathbb{P}}$ itself is unknown, which is the usual scenario in most data-driven research, the concept of ambiguity set is introduced [@Bayraksan:2015ge]. Thus, instead of optimizing under one particular distribution and under a deterministic set, distributionally robust stochastic optimization, aka DRSO, formulates optimization problems with a set of possible distributions, under the concept of ambiguity set. Specifically, one could consider minimizing the expected loss as follows, $$\min_{X\in{\mathcal{X}}}\max_{{\mathbb{P}}\in\mathcal{P}}{\mathbb{E}}_{\mathbb{P}}[l(X;\xi)],$$ where $X$ is the decision variable, allowed to vary inside the feasible region ${\mathcal{X}}$, and the random element $\xi$ follows distribution ${\mathbb{P}}\in \mathcal{P}$, with $\mathcal{P}$ the ambiguity set and $l$ the loss function.
In the data driven setting where we have iid samples $\{\xi_i\}_{i=1}^n$ drawn from ${\mathbb{P}}$, the ambiguity set $\mathcal{P}$ can be constructed so that it contains all distributions that are within a certain divergence from the empirical distribution, where the radius of the ambiguity set is large enough so that it contains ${\mathbb{P}}$ with high probability. Alternative methods to construct ambiguity sets use moment constraints under ${\mathbb{P}}\in \mathcal{P}$, where $\mathcal{P}$ consists of all probability distributions with first order and second order moments matching the sample moments. Again, the key is to define and measure the difference between various distributions.
In both the literature of learning and robust optimization, one popular choice to measure the difference between two distributions is the Kullback-Leibler divergence, which has strong theoretical foundation in information theory and large deviations [@Pardo:1999]. However, there are two issues in using the KL divergence. The first one is that the KL divergence between a continuous distribution and its empirical version, which is always a discrete distribution, is undefined (or infinite). The second issue is that KL divergence does not take into consideration the relative position of probability mass. As an example, consider the discrete distribution ${\mathbb{P}}$ which puts $1/2$ mass on 0 and $1/2$ mass on 1, and the discrete distribution ${\mathbb{Q}}$ which puts $1/2$ mass on $\epsilon$ and $1/2$ mass on $1-\epsilon$. The KL divergence does not reflect the convergence of $Q$ to $P$ as $\epsilon \downarrow 0$, hence it is too restrictive. It is therefore natural to use alternative measures for distributions, such as $f$-divergence, $L_p$-Wasserstein distance, and Prohorov metric. (See Section \[RelatedWork\] for more details).
On the other hand, KL divergence belongs to a class of divergences known as *Bregman divergences*. Bregman divergences [@Bregman:1967ab] are introduced by Lev Bregman in 1967 in solving a problem in convex optimization. Since its inception, Bregman divergences have found applications not only in convex optimization but also in statistics and machine learning, for example, clustering [@Lucic:2016uv] [@Banerjee:2005vsa], inverse problems [@LeBesenerais:1999] [@Jones:1990], classification [@Srivastava:2007], logistic regression and AdaBoost [@Collins:2002wc] [@Murata:2004] [@Lafferty:1999], regression [@Kivinen:2001], mirror descent ([@Nemirovski:1983ab]), and generalized accelerated descent algorithms [@Wibisono:2016gx] [@Taskar:2006]. Bregman divergences are asymmetric in general, which can potentially be more desirable in the setting of comparing distributions, compared to a symmetric measure such as $L_p$-Wasserstein distance.
Our goal is to address the following questions:
- How can we define appropriate divergences in the general setting of comparing distributions?
- How can we define appropriate divergences, in the particular context of robust optimization and distribution learning?
In this paper, we report some progress toward our goal. Our main contributions are as follows:
- We derive a weak convergence result using Bregman divergence in parametric distributions. The result describes precisely how the Hessian of the underlying convex function in Bregman divergence impacts the statistical properties of the divergence measure in the asymptotic setting.
- In the non-asymptotic setting, we prove concentration results using Bregman divergence between the true discrete distribution and the empirical distributions. This allows the construction of ambiguity set in robust optimization.
- We propose a novel statistical divergence called *Wasserstein-Bregman divergence*, which is essentially a marriage between Wasserstein distance and Bregman divergence. We find that this divergence has the ability to capture the asymmetry in comparing distributions, while retains nice analytical properties of Wasserstein distance for the purpose of optimization.
Related Work {#RelatedWork}
------------
#### DRSO with KL Divergence.
In [@Hu:2013tc], they formulate a robust optimization problem in terms of a KL divergence constraint and show that the problem can be converted into a convex optimization problem which can be solved analytically. In [@Guan:2012ec], they show that chance constraints with KL divergence ambiguity sets can be reformulated into a traditional chance constraint problem with different risk levels.
#### DRSO with $L_p$-Wasserstein Distance.
In [@Esfahani:2015wh], they propose the use of $L_1$-Wasserstein ambiguity set. They show that Wassstein ambiguity sets provide a better out-of-sample guarantee than the KL divergence, because a continuous ${\mathbb{P}}$ will always be outside the KL divergence ball centered at the empirical distribution $\hat{{\mathbb{P}}}_n$, which is discrete, whereas the Wasserstein ball contains continuous as well as discrete distributions. They also show that the robust optimization problem, under some mild conditions, can be converted into a finite-dimensional convex programming problem, solvable in polynomial time. In [@Shafieezadeh-Abadeh:2015wr], they use Wasserstein ambiguity set for distributionally robust logistic regression. Specifically they study $\inf_\beta \sup_{{\mathbb{P}}\in \mathcal{P}} {\mathbb{E}}_{{\mathbb{P}}}[l_\beta(x,y)]$, where $l_\beta(x,y)$ is the logloss function with parameter $\beta$. They show that this problem has a tractable convex reformulation and provide confidence interval for the objective function, which is the out of sample performance guarantee. In [@Anonymous:W0HO4cob], they use the $L_1$-Wasserstein ball as the ambiguity set. They show that the candidate probability distributions in the ball can be reduced to a subset whose elements can be described using extreme/exposed points of the set, hence a tractable reformulation of the original problem becomes possible. In [@Gao:2016vo], they consider the $L_p$-Wasserstein ball for $p \geq 1$, and give necessary and sufficient conditions for the worst-case distributions to exist. In [@Fournier:2015kk], they inspect the convergence rate of the empirical distribution to the true distribution under Wasserstein distance.
#### Distribution Learning with $L_2$-Wasserstein Distance.
In [@Arjovsky:2017vh], they use neural network to learn probability density and define the objective function for optimization to be the $L_2$-Wasserstein distance. They have shown promising results on a numerical experiments in image generation.
Background
==========
In this section, we will review definitions and relevant properties of Bregman divergence and Wasserstein distance.
Bregman Divergence
------------------
For two vectors $x$ and $y$ in ${\mathbb{R}}^d$ and a strictly convex function $\phi(x): {\mathbb{R}}^d \to {\mathbb{R}}$, the *Bregman divergence* is defined as $$D_\phi(x,y) = \phi(x) - \phi(y) - \langle \nabla \phi(y), x-y\rangle.$$
For two continuous distributions ${\mathbb{P}}$ and ${\mathbb{Q}}$, one can define Bregman divergence as in [@Jones:1990tf], $$\begin{aligned}
&D_\phi({\mathbb{P}},{\mathbb{Q}}) \\
&= \int\left[\phi(p(x)) - \phi(q(x)) - \phi'(q(x))(p(x)-q(x))\right] d\mu(x),\end{aligned}$$ where $p(x)$ and $q(x)$ are probability density functions of ${\mathbb{P}}$ and ${\mathbb{Q}}$ respectively, $\mu$ is the base measure, and $\phi: {\mathbb{R}}\to {\mathbb{R}}$ is a strictly convex function.
Examples of Bregman divergences include
- $L^2$ loss: $D_\phi(x,y) = \|x-y\|_2^2$, where $\phi(x) = \|x\|_2^2$,
- Itakura-Saito divergence: $D_\phi(x,y) = x/y - \log(x/y) - 1$, where $\phi(x) = -\log x$,
- KL divergence: $D_\phi(x,y) = \sum_{i = 1}^d x_i \log (x_i/y_i)$, where $\phi(x) = \sum_{i=1}^d x_i \log x_i$,
- Mahalanobis distance: $D_\phi(x,y) = (x-y)^T A (x-y)$, where $\phi(x) = x^T A x$, $A$ is a strictly positive definite matrix.
As a divergence function, $D_\phi(x,y)$ is always nonnegative by the convexity of $\phi$. $D_\phi(x,y) = 0$ if and only if $x = y$. However, it is not a metric because it is not symmetric, and it does not satisfy the triangle inequality. In [@Pardo:2003fr], they show an asymptotic equivalence between $f$-divergences (in particular, $\chi^2$-divergence) and Bregman divergences under some conditions.
#### $k$-means Using Bregman.
In [@Banerjee:2005jd], they show that conditional expectation is the optimal predictor for all Bregman divergences. Moreover, Bregman divergences are the only class of such loss functions. This property ensures the convergence of $k$-means algorithm when Bregman divergence is used as a loss function.
#### Connections with Exponential Family.
In [@Banerjee:2005vsa], they show that there is a one-to-one correspondence between Bregman divergences and exponential family. That is, take an exponential family in a canonical form of: $$p_\theta(x) = \exp(\theta^Tx - \psi(\theta))h(x),$$ where $\theta, x\in{\mathbb{R}}^d$. $\psi$ is the cumulant function with its Legendre convex conjugate $\phi$ defined as $$\phi(x) = \sup_t [\langle x, t \rangle - \psi(t)].$$ Then $$p_\theta(x) = \exp(-D_\phi(x,\mu(\theta)) - g_\phi(x)),$$ with $\mu(\theta) = \nabla \psi(\theta)$. This one-to-one correspondence comes from the duality property of Bregman divergence, which states that $$D_\phi(p,q) = D_\phi(q^*, p^*),$$ with $p^* = \nabla \phi(p)$ and $q^* = \nabla \phi(q)$.
#### Connections to Fisher Information.
In the case where $X\sim p_\theta$ and $p_\theta$ belongs to a regular exponential family, the Fisher information of $\mu = {\mathbb{E}}X$ has a nice representation. For notation simplicity, we present the result for the one-dimensional case. This result can be easily extended to higher dimensions.
\[FisherInfo\] Suppose $X\sim p_{\theta}$ belongs to a regular exponential family. Let $\mu = {\mathbb{E}}(X)$, $\psi$ be the cumulant function and $\phi$ be the convex conjugate of $\psi$. Assume that $\psi$ is three-time differentiable. Then $$I(\mu) = {\mathbb{E}}\left[\frac{\partial^2}{\partial \mu^2} D_{\phi}(x, \mu)\right] = \phi''(\mu).$$
The first equality follows directly from the representation $p_\theta(x) = \exp(-D_{\phi}(x, \mu) - g_{\phi}(x))$. The second equality follows from a straightforward calculation, $$\begin{split}
{\mathbb{E}}\left[\frac{\partial^2}{\partial \mu^2} D_{\phi}(x, \mu)\right] &= {\mathbb{E}}\left[\frac{\partial^2}{\partial \mu^2} [\phi(x) - \phi(\mu) - \phi'(\mu)(x - \mu)]\right] \\
&= {\mathbb{E}}\left[\frac{\partial}{\partial \mu}[-\phi''(\mu)(x - \mu)]\right]\\
&= {\mathbb{E}}[-\phi'''(\mu)(x-\mu) + \phi''(\mu)]\\
&=\phi''(\mu).
\end{split}$$
#### Bias-Variance Decomposition.
In [@Buja05lossfunctions], they show that expected Bregman divergence has a bias-variance decomposition $${\mathbb{E}}D_\phi (\hat{\theta}, \theta) = D_\phi({\mathbb{E}}\hat{\theta}, \theta) + {\mathbb{E}}D_\phi(\hat{\theta}, {\mathbb{E}}\hat{\theta}).$$
Setting $\phi(x) = \| x \|_2^2$ recovers the usual bias-variance decomposition for squared-error loss,
$$\begin{aligned}
&{\mathbb{E}}D_\phi (\hat{\theta}, \theta) \\
&= {\mathbb{E}}[(\hat{\theta} - \theta)^2] \\
&= ({\mathbb{E}}\hat{\theta} - \theta)^2 + {\mathbb{E}}[(\hat{\theta} - \theta)^2] \\
&= D_\phi({\mathbb{E}}\hat{\theta}, \theta) + {\mathbb{E}}D_\phi(\hat{\theta}, {\mathbb{E}}\hat{\theta}).\end{aligned}$$
Figure \[fig:kNN\_tradeoff\] shows how various choices can lead to different measures of bias-variance tradeoff in selecting the number of neighbors for $k$-nearest neighbor ($k$-NN) algorithm.
![The bias-variance tradeoff of $k$-NN algorithm based on the loss functions $D_\phi(x, y) = e^x - e^y - e^y(x - y)$ and $D_\phi(x, y) = x \log \left(\frac{x}{y}\right) + (1 - x)\log\left(\frac{1 - x}{1 - y}\right)$. For each plot, the solid dot indicates the parameter that minimizes the corresponding loss function. The data used for this illustration is the `spam` dataset collected at Hewlett-Packard Labs, readily available in the R package *kernlab*. []{data-label="fig:kNN_tradeoff"}](kNN_tradeoff_a "fig:"){width="37.00000%"} ![The bias-variance tradeoff of $k$-NN algorithm based on the loss functions $D_\phi(x, y) = e^x - e^y - e^y(x - y)$ and $D_\phi(x, y) = x \log \left(\frac{x}{y}\right) + (1 - x)\log\left(\frac{1 - x}{1 - y}\right)$. For each plot, the solid dot indicates the parameter that minimizes the corresponding loss function. The data used for this illustration is the `spam` dataset collected at Hewlett-Packard Labs, readily available in the R package *kernlab*. []{data-label="fig:kNN_tradeoff"}](kNN_tradeoff_c "fig:"){width="37.00000%"}
Wasserstein Distance
--------------------
Wasserstein distance is a divergence defined between probability distributions on a given metric space. It is also known as Kantorovich-Monge-Rubinstein metric. Wasserstein distances are vastly used in optimal transport [@Villani:2009ab], and they have found applications in many areas, such as the study of mixing for Markov chains in probability theory [@Dobrusin310:1996] [@Peres679:un], rates of fluctuations for empirical measures in statistics [@Rachev695:1991] [@Rachev696:1998] [@Dobric307:1995], and propagation of chaos in statistical mechanics [@Dobrusin308:1970] [@Spohn757:1991].
For any two probability distributions ${\mathbb{P}}$ and ${\mathbb{Q}}$ defined on a compact metric space $(\mathcal{X}, d)$, the *Wasserstein distance* between ${\mathbb{P}}$ and ${\mathbb{Q}}$ is defined as $$\begin{aligned}
W_p({\mathbb{P}}, {\mathbb{Q}}) &= \left( \inf_{\gamma \in \Pi({\mathbb{P}},{\mathbb{Q}})} \int_{\mathcal{X} \times \mathcal{X}} d(x, y)^p d\gamma(x,y)\right)^{1/p},\end{aligned}$$ where $\Pi({\mathbb{P}},{\mathbb{Q}})$ denotes the set of all couplings of ${\mathbb{P}}$ and ${\mathbb{Q}}$, i.e., all joint distributions defined on ${\mathcal{X}}\times {\mathcal{X}}$ with marginal distributions being ${\mathbb{P}}$ and ${\mathbb{Q}}$.
For example, if $\mathcal{X} \subset {\mathbb{R}}^d$, a natural choice of metric is $d(x, y) = \| x - y \|_p$. This leads to the definition of $L^p$-Wasserstein distance.
For any two probability distributions ${\mathbb{P}}$ and ${\mathbb{Q}}$ defined on a compact metric space $\mathcal{X} \subset {\mathbb{R}}^d$, the *Wasserstein distance of order $p\in[1,\infty]$* (or $L_p$-Wasserstein distance) between ${\mathbb{P}}$ and ${\mathbb{Q}}$ is defined as $$\begin{aligned}
W_p({\mathbb{P}}, {\mathbb{Q}}) &= \left( \inf_{\gamma \in \Pi({\mathbb{P}},{\mathbb{Q}})} \int_{\mathcal{X} \times \mathcal{X}} \|x - y\|_p^p d\gamma(x,y)\right)^{1/p},\end{aligned}$$ where $\Pi({\mathbb{P}},{\mathbb{Q}})$ denotes the set of all couplings of ${\mathbb{P}}$ and ${\mathbb{Q}}$, i.e., all joint distributions defined on ${\mathcal{X}}\times {\mathcal{X}}$ with marginal distributions being ${\mathbb{P}}$ and ${\mathbb{Q}}$.
For probability distributions, convergence under Wasserstein distance of order $p$ is equivalent to weak convergence plus convergence of the first $p$ moments.
For example, the Euclidean metric leads to the $L_2$-Wasserstein distance. The squared $L_2$-Wasserstein distance is defined as $$\begin{aligned}
W_2({\mathbb{P}}, {\mathbb{Q}})^2 &= \inf_{\gamma \in \Pi({\mathbb{P}},{\mathbb{Q}})} \int_{\mathcal{X} \times \mathcal{X}} \|x - y\|^2 d\gamma(x,y) \\
&= \inf_{\gamma \in \Pi({\mathbb{P}},{\mathbb{Q}})} {\mathbb{E}}_{X,Y \sim \gamma} [\|X - Y\|_{L^2}] \\
&= {\mathbb{E}}_{X \sim {\mathbb{P}}} [\|X\|_2^2] + {\mathbb{E}}_{Y \sim {\mathbb{Q}}} [\|Y\|_2^2] \\
&\qquad+ \inf_{\gamma \in \Pi({\mathbb{P}},{\mathbb{Q}}} [\langle -2Y, X \rangle].\end{aligned}$$ Wasserstein distance is a metric, so it is always nonnegative, it is equal to 0 if and only if ${\mathbb{P}}= {\mathbb{Q}}$. It is symmetric, and it satisfies the triangle inequality.
Main Results
============
Let $p = (p_1, \ldots, p_d) \in {\mathbb{R}}^d$ be the probability distribution of a discrete random variable $X$, where $p_i = {\mathbb{P}}(X = a_k$, $k \in \{1,2,\ldots, d\}$. Let $\hat{p}_n = (\hat{p}_{n,1}, \ldots, \hat{p}_{n,d}) \in {\mathbb{R}}^d$ be the random vector denoting the empirical distribution of a sequence of iid random variables $\{X_i\}_{i=1}^n$, where each $X_i$ has the same distribution as $X$. That is, $$\hat{p}_{n} = \left(\frac{1}{n}\sum_{i=1}^n 1\{X_i = a_1\}, \ldots, \frac{1}{n}\sum_{i=1}^n 1\{X_i = a_d\}\right).$$
Concentration of Bregman Divergence
-----------------------------------
We first establish that the Bregman divergence $D_\phi$ between the empirical distribution and the true distribution concentrates around the mean, where the rate can be expressed in terms of the gradient of the convex function $\phi$.
\[nonconvexConcentration\] Consider the random variable $Z = D_\phi(\hat{p}_n, p)$, the Bregman divergence between $\hat{p}_n$ and $p$, $$Z = D_\phi(\hat{p}_{n}, p) = \phi(\hat{p}_{n}) - \phi(p) - \langle \nabla\phi(p), \hat{p}_{n} - p \rangle,$$ where $\phi: [0,1]^d \to {\mathbb{R}}$ is a strictly convex function. Then the following concentration inequality holds for all $\epsilon > 0$: $${\mathbb{P}}\{Z - {\mathbb{E}}[Z] \ge \epsilon \} \le \exp(\frac{-n^2\epsilon^2}{4dM_\phi}\}),$$ where $M_\phi = \max_{t \in \Delta^{d-1}} \|\nabla\phi(t)\|_2$, and $\Delta^{d-1}$ is the standard $(d-1)$-simplex, which is the set $\{(t_1,t_2, \ldots, t_d) \in {\mathbb{R}}^d| \sum_{i=1}^d t_i = 1, t_i \ge 0, \forall i\}$.
Let $(X_1, \ldots, X_{i-1}, X_i, X_{i+1}, \ldots, X_n)$ be iid random variables from distribution $p$. Define another sequence of random variables $(X_1, \ldots, X_{i-1}, X_i', X_{i+1}, \ldots, X_n)$, in which only the $i$-th element in the sequence is different. Let the corresponding empirical distribution be $\hat{p}_n'$. Then $$Z' = D_\phi(\hat{p}_{n}', p) = \phi(\hat{p}_{n}') - \phi(p) - \langle \nabla\phi(p), \hat{p}_{n}' - p \rangle.$$ The difference of $Z$ and $Z'$ is $$\begin{aligned}
Z' - Z = \phi(\hat{p}_{n}') - \phi(\hat{p}_{n}) + \langle \nabla\phi(p), \hat{p}_{n} - \hat{p}_n' \rangle.\end{aligned}$$ Notice that by construction, $\hat{p}_{n} - \hat{p}_n'$ is a vector with an element being $1/n$, an element being $-1/n$, and all other elements being zeros. Therefore by the Cauchy-Schwarz inequality, $$\begin{aligned}
|\langle \nabla\phi(p), \hat{p}_{n} - \hat{p}_n' \rangle| \le \|\nabla \phi(p)\|_2 \|\hat{p}_{n} - \hat{p}_n'\|_2 \\
= \frac{\sqrt{2}}{n} \|\nabla \phi(p)\|_2 \le \frac{\sqrt{2}}{n} M_\phi.\end{aligned}$$ Also by the Taylor’s expansion, $$\begin{aligned}
\phi(\hat{p}_{n}') - \phi(\hat{p}_{n}) &= \|\nabla\phi(\xi)\|_2 \|\hat{p}_{n} - \hat{p}_n'\|_2 \\
&\le M_\phi \|\hat{p}_{n} - \hat{p}_n'\|_2 \\
&= \frac{\sqrt{2}}{n} M_\phi,\end{aligned}$$ where $\xi$ is a random vector which is a convex combination of $\hat{p}_{n}$ and $\hat{p}_{n}'$. Therefore by the triangle inequality, $$\begin{aligned}
|Z - Z'| &\le |\langle \nabla\phi(p), \hat{p}_{n} - \hat{p}_n' \rangle| + |\phi(\hat{p}_{n}') - \phi(\hat{p}_{n})| \\
&\le\frac{2\sqrt{2}}{n} M_\phi.\end{aligned}$$ Hence by the bounded difference inequality [@Talagrand772:1995], $$\begin{aligned}
{\mathbb{P}}\{Z - {\mathbb{E}}[Z] \ge \epsilon \} &\le \exp(\frac{-n^2\epsilon^2}{4dM_\phi}).\end{aligned}$$
Notice that Bregman divergence is only convex with respect to its first argument, which in the previous case is $\hat{p}_n$. To construct a convex ambiguity region, we need to reverse the order of $\hat{p}$ and $p$ to make the unknown true distribution the first argument. Hence we also prove the following concentration inequality:
\[convexConcentration\] Consider the random variable $Y = D_\phi(p,\hat{p}_n)$, the Bregman divergence between $p$ and $\hat{p}_n$: $$Y = D_\phi(p, \hat{p}_{n}) = \phi(p) - \phi(\hat{p}_{n}) - \langle \nabla\phi(\hat{p}_{n}), p - \hat{p}_{n}\rangle,$$ where $\phi: [0,1]^d \to {\mathbb{R}}$ is a strictly convex function. Then we have the following concentration inequality for all $\epsilon > 0$: $${\mathbb{P}}(Y - {\mathbb{E}}Y \ge \epsilon) \le \exp(-\frac{n^2\epsilon^2}{4d(M_\phi + L_\phi)^2}),$$ where $L_\phi$ is the Lipschitz constant of $\nabla\phi$, and $M_\phi = \max_{t \in \Delta^{d-1}} \|\nabla\phi(t)\|$. $\Delta^{d-1}$ is the standard $(d-1)$-simplex, which is the set $\{(t_1,t_2, \ldots, t_d) \in {\mathbb{R}}^d| \sum_{i=1}^d t_i = 1, t_i \ge 0, \forall i\}$.
Let $(X_1, \ldots, X_{i-1}, X_i, X_{i+1}, \ldots, X_n)$ be iid random variables from distribution $p$. Define another sequence of random variables $(X_1, \ldots, X_{i-1}, X_i', X_{i+1}, \ldots, X_n)$, in which only the $i$-th element in the sequence is different. Let the corresponding emprical distirbution be $\hat{p}_n'$. Then $$Y' = D_\phi(p, \hat{p}_{n}') = \phi(p) - \phi(\hat{p}_{n}') - \langle \nabla\phi(\hat{p}_{n}'), p - \hat{p}_{n}'\rangle.$$ The difference of $Y$ and $Y'$ is $$\begin{aligned}
Y' - Y &= \phi(\hat{p}_{n}) - \phi(\hat{p}_{n}') \\
&\qquad+ \langle \nabla\phi(\hat{p}_n), p - \hat{p}_{n} \rangle - \langle \nabla\phi(\hat{p}_n'), p - \hat{p}_{n}' \rangle.\end{aligned}$$ By the proof of Theorem \[nonconvexConcentration\], $$\phi(\hat{p}_{n}) - \phi(\hat{p}_{n}') \le \frac{\sqrt{2}}{n}M_\phi.$$ Meanwhile $$\begin{aligned}
\langle \nabla\phi(\hat{p}_n), p - &\hat{p}_{n} \rangle - \langle \nabla\phi(\hat{p}_n'), p - \hat{p}_{n}' \rangle \\
&= \langle \nabla\phi(\hat{p}_n) - \nabla \phi(\hat{p}_n'), p \rangle \\
&\qquad- \langle \nabla \phi(\hat{p}_n), \hat{p}_n \rangle + \langle \nabla \phi(\hat{p}_n'), \hat{p}_n' \rangle.\end{aligned}$$ Since $\nabla \phi$ is defined on the compact region $[0,1]^d$, we can assume without loss of generality that it has Lipschitz constant $L_\phi$. Then by the Cauchy-Schwarz inequality, $$\begin{aligned}
|\langle \nabla\phi(\hat{p}_n) - \nabla \phi(\hat{p}_n'), p \rangle| &\le \|p\|_2 \|\nabla\phi(\hat{p}_n) - \nabla\phi(\hat{p}_n')\|_2 \\
&\le L_\phi \|\hat{p}_n - \hat{p}_n'\|_2 = \frac{\sqrt{2}}{n}L_\phi,\end{aligned}$$ and similarly $$\begin{aligned}
&\qquad|- \langle \nabla \phi(\hat{p}_n), \hat{p}_n \rangle + \langle \nabla \phi(\hat{p}_n'), \hat{p}_n' \rangle| \\
&= |\langle \nabla \phi(\hat{p}_n), \hat{p}_n' - \hat{p}_n \rangle + \langle \nabla \phi(\hat{p}_n') - \nabla \phi(\hat{p}_n), \hat{p}_n' \rangle| \\
&\le |\langle \nabla \phi(\hat{p}_n), \hat{p}_n' - \hat{p}_n \rangle| + |\langle\nabla \phi(\hat{p}_n') - \nabla \phi(\hat{p}_n), \hat{p}_n' \rangle| \\
&\le \frac{\sqrt{2}}{n} M_{\phi} + \frac{\sqrt{2}}{n}L_\phi.\end{aligned}$$ Therefore $$|Y' - Y| \le 2\left(\frac{\sqrt{2}}{n} M_{\phi} + \frac{\sqrt{2}}{n}L_\phi\right).$$ By the bounded difference inequality, $${\mathbb{P}}(Y - {\mathbb{E}}Y \ge \epsilon) \le \exp(-\frac{n^2\epsilon^2}{4d(M_\phi + L_\phi)^2}).$$
Weak Convergence of Bregman Divergence
--------------------------------------
In this section, we will show that in the asymptotic case, Bregman divergence between the true parameters of a distribution and the maximum likelihood estimator of the parameters will converge in distribution to a finite weighted sum of independent $\chi^2$ distributed random variables. This result allows us to construct asymptotic ambiguity sets according to the quantiles of the asymptotic distribution.
Suppose there exists a family of probability distributions ${\mathbb{P}}_\theta$ parametrized by $\theta\in\Theta\subset{\mathbb{R}}^d$. Suppose we have iid data $\{X_i\}_{i=1}^n$, and $\hat{\theta}_n$ is the maximum likelihood estimator of $\theta$. Then $$\lim_{n \to \infty} n D_{\phi}(\theta, \hat{\theta}_n) \convd \frac{1}{2} \sum_{i = 1}^r \beta_i Z_i^2,$$ where $Z_i$’s are independent standard Gaussian random variables, $D_\phi$ denotes the Bregman divergence characterized by $\phi$, $\beta_i$’s are the non-zero eigenvalues of the matrix $H\Sigma$ and $r = rank(\Sigma^T H \Sigma)$, with $H$ the Hessian of $\phi$ at $\theta$ and $\Sigma$ the inverse Fisher information matrix.
First, write the Taylor expansion of $\phi$ around $\hat{\theta}_n$, $$\begin{aligned}
\phi(\theta) &= \phi(\hat{\theta}_n) + \langle \theta - \hat{\theta}_n, \nabla \phi(\hat{\theta}_n) \rangle \\
&\qquad+ \frac{1}{2}(\theta - \hat{\theta}_n)^T H(\hat{\theta}_n) (\theta - \hat{\theta}_n) + o(\|\theta - \hat{\theta}_n\|_2^2),\end{aligned}$$ where $H(\hat{\theta})$ is the Hessian of $\phi(x)$ at $x = \hat{\theta}$. Notice that by the properties of maximum likelihood estimators, as $n \to \infty$, $$\sqrt{n}(\theta - \hat{\theta}_n) \convd N(0,\mathcal{I}^{-1}) \ed N(0,\Sigma),$$ where $$(\mathcal{I})_{ij}= -{\mathbb{E}}\frac{\partial ^2 \log L}{\partial \theta_i \partial \theta_j}$$ is the Fisher information matrix of the underlying true distribution, with $L$ being the likelihood function. Also, $$H(\hat{\theta}_n) \to H(\theta)$$ in probability, and $$n \cdot o(\|\theta - \hat{\theta}_n\|_2^2) \to 0$$ in probability. Therefore by the Slutsky’s theorem, $$\begin{aligned}
nD_{\phi}(\theta, \hat{\theta}_n) &= n(\phi(\theta) - \phi(\hat{\theta}_n) - \langle \theta - \hat{\theta}_n, \nabla \phi(\hat{\theta}_n) \rangle) \\
&= \frac{1}{2}\sqrt{n}(\theta - \hat{\theta}_n)^T H \sqrt{n}(\theta - \hat{\theta}_n) \\
&\qquad+ n\cdot o(\|\theta - \hat{\theta}_n\|_2^2) \\
&\convd \frac{1}{2} X^T H X,\end{aligned}$$ where $X \ed N(0,\Sigma)$. Let $S \in {\mathbb{R}}^{d\times s}$ be a square root of $\Sigma$. Since $\Sigma$ and $H$ are positive semidefinite, by spectral theorem, we can write $S^THS = R^T\Lambda R$, where $\Lambda = diag(\beta_1, \ldots, \beta_r)$, which is the diagonal matrix of non-zero eigenvalues of $S^THS$, hence is also the diagonal matrix of non-zero eigenvalues of $H\Sigma$, $r = rank(\Sigma H \Sigma)$, and $R$ is the matrix of corresponding orthonormal eigenvectors. Then $$\begin{aligned}
X^T H X &\ed (SY)^T H SY \ed Y^T R^T\Lambda R Y \\
&\ed Z^T \Lambda Z = \sum_{i=1}^r \beta_i Z_i^2,\end{aligned}$$ where $Z_i$ are independent standard Gaussian random variables. Therefore, we have the quadratic form of Gaussian variables $$\sqrt{n}(\theta - \hat{\theta}_n)^T H \sqrt{n}(\theta - \hat{\theta}_n) \ed \sum_{i = 1}^r \beta_i Z_i^2.$$ This completes the proof.
*Remark*: Even though Bregman divergence is asymmetric, $n D_{\phi}(\hat{p}_n, p)$ has the same asymptotic distribution as $n D_{\phi}(p,\hat{p}_n)$ by a similar proof.
Noting that $\hat{p}_n$ is the maximum likelihood estimator of $p$, we immediately arrive at the following corollary.
\[asym\] For a discrete distribution $p = (p_1, \ldots, p_d)$ and the empirical distribution $\hat{p}_n = (\hat{p}_{n,1}, \ldots, \hat{p}_{n,d})$ generated from $n$ iid samples, we have $$\lim_{n \to \infty} n D_{\phi}(p,\hat{p}_n) \convd \frac{1}{2} \sum_{i = 1}^r \beta_i Z_i^2,$$ where $Z_i$ are independent standard Gaussian random variables, $r = rank(\Sigma^T H \Sigma)$, $H$ is the Hessian of $\phi$, $\Sigma$ is the inverse Fisher information matrix, and $\beta_1, \ldots, \beta_r$ are the nonzero eigenvalues of $H\Sigma$.
Wasserstein-Bregman Divergence
------------------------------
In this section, we first define the new Wasserstein-Bregman divergence between probability distributions. We then show that under some mild conditions, the divergence function is differentiable with respect to the parameters almost everywhere. This result allows the gradient descent algorithm to minimize the divergence between a target distribution and a parametric distribution. Therefore, Wasserstein-Bregman divergence can be used as an objective function in distribution learning.
Let $\phi: {\mathbb{R}}^d \to {\mathbb{R}}$ be a strictly convex function and $D_\phi: {\mathbb{R}}^d \times {\mathbb{R}}^d \to {\mathbb{R}}$ be the associated Bregman divergence with $D_\phi(x,y) = \phi(x) - \phi(y) - \langle \nabla \phi(y), x-y\rangle$. *Wasserstein-Bregman divergence* $W_{D_\phi} ({\mathbb{P}}, {\mathbb{Q}})$ is defined as $$\begin{aligned}
W_{D_\phi} ({\mathbb{P}}, {\mathbb{Q}}) &= \inf_{\gamma \in \Pi({\mathbb{P}},{\mathbb{Q}})} \int D_\phi(x,y) d\gamma(x,y) \\
&= \inf_{\gamma \in \Pi({\mathbb{P}},{\mathbb{Q}})} {\mathbb{E}}_{X,Y \sim \gamma} [D_\phi(X,Y)].\end{aligned}$$
As an example, if $\phi(x) = \|x\|^2$, $W_{D_\phi} ({\mathbb{P}}, {\mathbb{Q}})$ reduces to $W_2 ({\mathbb{P}}, {\mathbb{Q}})^2$.
By the nonnegativity of Bregman divergence, it is easy to verify that $W_{D_\phi} ({\mathbb{P}}, {\mathbb{Q}})$ is always nonnegative, and $W_{D_\phi} ({\mathbb{P}}, {\mathbb{Q}})= 0$ if and only if ${\mathbb{P}}= {\mathbb{Q}}$.
Let $\mathcal{X} \subset {\mathbb{R}}^d $ be a compact metric set, ${\mathbb{Q}}$ is a fixed distribution defined on $\mathcal{X}$, $g_\theta(Z)$ is a function of $Z$, with parameter $\theta \in {\mathbb{R}}^d$, and $Z$ being a random variable over another space $\mathcal{Z}$. Let ${\mathbb{P}}_\theta$ denote the distribution of $g_\theta(Z)$. Then
1. If $g$ is continuous in $\theta$, then $W_{D_\phi}({\mathbb{Q}}, {\mathbb{P}}_\theta)$ is also continuous in $\theta$.
2. If $g$ is locally Lipschitz with local Lipschitz constants $L(\theta,z)$ such that ${\mathbb{E}}_{Z\sim {\mathbb{P}}_\theta} L(\theta, Z)^2 < \infty$, then $W_{D_\phi}({\mathbb{Q}}, {\mathbb{P}}_\theta)$ is differentiable almost everywhere.
Remark: in this theorem, $g_\theta(Z)\sim {\mathbb{P}}_\theta$ is the parametric distribution that attempts to replicate the distribution ${\mathbb{Q}}$.
Because $\phi$ is strictly convex, its gradient $\nabla \phi$ has positive definite Jacobian matrix, which is also the Hessian of $\phi$. Then by the inverse function theorem, $\nabla \phi$ is invertible. Denote its inverse with $(\nabla \phi)^{-1}$ and the composition of ${\mathbb{P}}_\theta$ and $(\nabla \phi)^{-1}$ as ${\mathbb{P}}_\theta \circ (\nabla \phi)^{-1}$, then ${\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta \circ (\nabla \phi)^{-1}} [||Y||_2^2] = {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\|\nabla \phi(Y)\|_2^2]$.
Expand $W_{D_\phi} ({\mathbb{Q}}, {\mathbb{P}}_\theta)$ by the linearity of inner products, $$\begin{aligned}
W_{D_\phi} &({\mathbb{Q}}, {\mathbb{P}}_\theta) \\
&= \inf_{\gamma \in \Pi({\mathbb{Q}},{\mathbb{P}}_\theta)} \int [\phi(x) - \phi(y) - \langle \nabla \phi(y), x-y\rangle] d\gamma(x,y) \\
&= {\mathbb{E}}_{X \sim {\mathbb{Q}}} [\phi(X)] - {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\phi(Y)] + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\langle\phi(Y), Y \rangle] \\
&\qquad+ \inf_{\gamma \in \Pi({\mathbb{Q}},{\mathbb{P}}_\theta)} {\mathbb{E}}_{X,Y \sim \gamma} [\langle -\nabla \phi(Y), X \rangle] \\
&= \frac{1}{2}[ {\mathbb{E}}_{X \sim {\mathbb{Q}}} [\|X\|_2^2] + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\|\nabla \phi(Y)\|_2^2] \\
&\qquad + \inf_{\gamma \in \Pi({\mathbb{Q}},{\mathbb{P}}_\theta)} {\mathbb{E}}_{X,Y \sim \gamma} [\langle -2\nabla \phi(Y), X \rangle]] \\
&\qquad + {\mathbb{E}}_{X \sim {\mathbb{Q}}} [\phi(X)] - {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\phi(Y)] \\
&\qquad + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\langle\nabla\phi(Y), Y \rangle] \\
&\qquad- \frac{1}{2}\left[{\mathbb{E}}_{X \sim {\mathbb{Q}}} [\|X\|_2^2] + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\|\nabla\phi(Y)\|_2^2]\right] \\
&= \frac{1}{2} W_2({\mathbb{Q}}, {\mathbb{P}}_\theta \circ (\nabla \phi)^{-1})^2 + {\mathbb{E}}_{X \sim {\mathbb{Q}}} [\phi(X)] \\
&\qquad- {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\phi(Y)] + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\langle\nabla\phi(Y), Y \rangle] \\
&\qquad- \frac{1}{2}\left[{\mathbb{E}}_{X \sim {\mathbb{Q}}} [\|X\|_2^2] + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta \circ (\nabla \phi)^{-1}} [||Y||_2^2]\right].\end{aligned}$$ Therefore we can express the new $W_{D_\phi}({\mathbb{Q}}, {\mathbb{P}}_\theta)$ as the distorted squared Wasserstein distance $\frac{1}{2} W_2({\mathbb{Q}}, {\mathbb{P}}_\theta \circ (\nabla \phi)^{-1})^2$ plus some error correction terms, which do not depend on the choice of coupling $\gamma$.
Now it suffices to show that $W_{L^2}({\mathbb{P}}_r,{\mathbb{P}}_\theta)$ is almost everywhere differentiable. First, observe that for two vectors $\theta, \theta' \in {\mathbb{R}}^d$, let $\pi$ be the joint distribution of $(g_\theta(Z), g_{\theta'}(Z))$ where $Z \sim p(z)$, then $$\begin{aligned}
W_{L^2}({\mathbb{P}}_\theta,{\mathbb{P}}_\theta') &\le \left[{\mathbb{E}}_{(X,Y) \sim \pi} [\|X-Y\|_2^2] \right]^{\frac{1}{2}} = \left[ {\mathbb{E}}\|g_\theta(Z) - g_{\theta'}(Z)\|_2^2 \right]^{\frac{1}{2}}.
\end{aligned}$$ The continuity of $g_\theta$ ensures that $\|g_\theta(Z) - g_{\theta'}(Z)\|^2 \to 0 $ point-wise as $\theta \to \theta'$. Since ${\mathcal{X}}$ is compact, $\|g_\theta(Z) - g_{\theta'}(Z)\|^2$ is uniformly bounded. Therefore by the bounded convergence theorem, $$W_{L^2}({\mathbb{P}}_\theta,{\mathbb{P}}_\theta') \le \left [{\mathbb{E}}\|g_\theta(Z) - g_{\theta'}(Z)\|_2^2 \right]^{\frac{1}{2}} \to 0, \text{ as } \theta \to \theta'.$$ Hence by the triangle inquality, as $\theta \to \theta'$, $
|W_{L^2}({\mathbb{P}}_r, {\mathbb{P}}_\theta) - W_{L^2}({\mathbb{P}}_r, {\mathbb{P}}_{\theta'})| \le W_{L^2}({\mathbb{P}}_\theta, {\mathbb{P}}_\theta') \to 0.
$ This proves the continuity.
Now assume $g_\theta$ is locally Lipschitz, i.e., for each pair $(\theta,z)$, there exists a constant $L(\theta,z)$ and an open neighborhood $N(\theta,z)$ around $(\theta,z)$ such that $\forall (\theta',z') \in N(\theta,z)$, $$\|g_\theta(z) - g_{\theta'}(z')\|_2 \le L(\theta,z)(\|\theta - \theta'\|_2 + \|z - z'\|_2).$$ By fixing $z' = z$ and taking expectation of squares of both sides, we get $${\mathbb{E}}\|g_\theta(Z) - g_{\theta'}(Z)\|_2^2 \le \|\theta - \theta'\|_2^2 {\mathbb{E}}[L(\theta,Z)^2],$$ for all $\theta'$ in an open neighborhood of $\theta$. Therefore, $$|W_{L^2}({\mathbb{P}}_r, {\mathbb{P}}_\theta) - W_{L^2}({\mathbb{P}}_r, {\mathbb{P}}_{\theta'})| \le W_{L^2}({\mathbb{P}}_\theta, {\mathbb{P}}_\theta') \le \left[{\mathbb{E}}\|g_\theta(Z) - g_{\theta'}(Z)\|_2\right]^{\frac{1}{2}} \le \|\theta - \theta'\|_2 {\mathbb{E}}[L(\theta, Z)^2]^{\frac{1}{2}},$$ i.e., $W_{L^2}({\mathbb{P}}_r, {\mathbb{P}}_\theta)$ is locally Lipschitz and by Rademacher’s theorem, is differentiable almost everywhere.
Discussion
==========
DRSO and Ambiguity Set
----------------------
Suppose one chooses the divergence between probability distributions to be $d({\mathbb{P}},{\mathbb{Q}})$, where ${\mathbb{P}}$ and ${\mathbb{Q}}$ are probability measures defined on the set $\mathcal{X} \subset {\mathbb{R}}^n$. Let $\mathcal{M}_+$ denotes the set of all probability distributions defined over the set $\mathcal{X}$. Then the ambiguity set $\mathcal{P}$ can be defined as a ball centered at the nominal distribution ${\mathbb{Q}}$: $$\mathcal{P} = \{{\mathbb{P}}\in \mathcal{M}_+: d({\mathbb{P}},{\mathbb{Q}}) \le \delta\}.$$
The nominal distribution ${\mathbb{Q}}$ may come from prior knowledge of the model, or directly from data. In the data-driven setting where we are given iid samples $\{X_i\}_{i=1}^n$, the nominal distribution ${\mathbb{Q}}$ is chosen to be the empirical distribution $\hat{{\mathbb{P}}}_n$.
- When the sample size $n$ is large (relative to $d$), one can appeal to the asymptotic distribution of $D(p, \hat{p}_n)$ to construct an ambiguity set using Theorem \[asym\]. More specifically, an ambiguity set can be constructed as follows: $$\mathcal{P} = \{p : D_\phi(p, \hat{p}_n) \leq \frac{1}{2n}F^{-1}(\alpha)\},$$ where $F^{-1}(\alpha)$ is the quantile function of $\sum_{i = 1}^r \beta_i Z_i^2$, which is a weighted sum of independent $\chi^2$ random variables with one degree of freedom. This quantile can be approximated via a Monte Carlo approximation. For a large $K$ (say $K = 10000$), one can simulate $rK$ independent standard normal random variables $Z_{1, 1}, ..., Z_{1, r}, Z_{2, 1}, ..., Z_{2, r}, ..., Z_{K, 1}, ..., Z_{K, r},$ and compute $R_j = \sum_{i = 1}^r \beta_i Z_{i, j}^2$ for each $j = 1, ... K$. Then one can use take the $\alpha$-th empirical quantile of $(R_1, ..., R_K)$ as an approximation to $F^{-1}(\alpha)$. Note that $\mathcal{P}$ is convex since Bregman divergence is convex with respect to the first argument.
- When the sample size $n$ is of moderate size or small, one must appeal to concentration results to obtain a valid ambiguity set. In order to apply Theorem \[nonconvexConcentration\] or Theorem \[convexConcentration\] for the construction of the ambiguity set, we have to first derive an upper bound for ${\mathbb{E}}D_\phi(p, \hat{p}_{n})$ or ${\mathbb{E}}D_\phi(\hat{p}_{n}, p)$, respectively.
For ${\mathbb{E}}D_\phi(p, \hat{p}_{n})$, clearly $$\begin{aligned}
{\mathbb{E}}&D_\phi(p, \hat{p}_{n}) = {\mathbb{E}}[\phi(p) - \phi(\hat{p}_{n}) - \langle \nabla\phi(\hat{p}_{n}), p - \hat{p}_{n}\rangle] \\
&= {\mathbb{E}}[\phi(p) - \phi(\hat{p}_{n})] - {\mathbb{E}}[\langle \nabla\phi(p) - \nabla\phi(\hat{p}_{n}), p - \hat{p}_{n}\rangle] \\
&\le M_{\phi} \sqrt{\sum_{i=1}^d \frac{p_i(1-p_i)}{n}} + L_\phi {\mathbb{E}}\|p - \hat{p}_n\|_2^2 \\
&= M_{\phi} \sqrt{\sum_{i=1}^d \frac{p_i(1-p_i)}{n}} + L_\phi \sum_{i=1}^d \frac{p_i(1-p_i)}{n}\\
&= M_{\phi} \sqrt{\frac{d}{4n}} + L_\phi \frac{d}{4n},\end{aligned}$$ where the inequality is by the Cauchy-Schwarz inequality and the Taylor’s theorem.
Similarly, for ${\mathbb{E}}D_\phi(\hat{p}_{n}, p)$, $$\begin{aligned}
{\mathbb{E}}\left[D_\phi(\hat{p}_{n}, p)\right]
&= {\mathbb{E}}[ \phi(\hat{p}_{n}) - \phi(p)] \\
&= {\mathbb{E}}[\langle\nabla \phi(\xi), \hat{p} - p\rangle] \\
&\le {\mathbb{E}}[\|\nabla \phi(\xi)\|\|\hat{p} - p\|_2] \\
&\le M_{\phi} {\mathbb{E}}[\|\hat{p} - p\|_2] \\
&\le M_{\phi} \sqrt{{\mathbb{E}}[\|\hat{p} - p\|_2^2]} \\
&= M_{\phi} \sqrt{\sum_{i=1}^d \frac{p_i(1-p_i)}{n}}\\
&\leq M_{\phi} \sqrt{\frac{d}{4n}},\end{aligned}$$ where $\xi$ is between $\hat{p}$ and $p$, the first inequality is by Cauchy-Schwarz, and the third inequality is by the Jensen’s inequality.
- As described immediately after the proof of Theorem \[nonconvexConcentration\], the resulting ambiguity set might be intractable to be computed because of its potential nonconvex nature. On the other hand, Theorem \[convexConcentration\] results in a convex ambiguity set, which is the Bregman ball of $\hat{p_n}$ with radius $M_\phi \sqrt{\frac{d}{4n}} + L_\phi \left(\frac{d}{4n}\right) + \epsilon$: $$\left\{p: D_\phi(p, \hat{p_n}) \leq M_{\phi} \sqrt{\frac{d}{4n}} + L_\phi \left(\frac{d}{4n}\right) + \epsilon \right\}.$$
Distribution Learning
---------------------
For distribution learning, the Wasserstein-Bregman divergence $W_{D_\phi}$ can be served as the objective function in the optimization problem. As shown in the proof of Theorem 3.4, the Wasserstein-Bregman divergence $W_{D_\phi}$ has an interesting decomposition in terms of squared divergence plus a penalty term: $$W_{D_\phi}({\mathbb{Q}}, {\mathbb{P}}_\theta) = D + P,$$ where $$\begin{aligned}
D &=\frac{1}{2} W_2({\mathbb{Q}}, {\mathbb{P}}_\theta \circ (\nabla \phi)^{-1})^2 \\
P &= {\mathbb{E}}_{X \sim {\mathbb{Q}}} [\phi(X)] - {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\phi(Y)] + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta} [\langle\nabla\phi(Y), Y \rangle] \\
&\qquad- \frac{1}{2}\left[{\mathbb{E}}_{X \sim {\mathbb{Q}}} [\|X\|_2^2] + {\mathbb{E}}_{Y \sim {\mathbb{P}}_\theta \circ (\nabla \phi)^{-1}} [||Y||_2^2]\right].\end{aligned}$$
$$\begin{tikzcd}
& {\mathbb{P}}_\theta \circ (\nabla \phi)^{-1} \arrow[dr, dashrightarrow]{d}{W_{2}^2({\mathbb{Q}}, {\mathbb{P}}_\theta \circ \nabla \phi)} \\
{\mathbb{P}}_\theta \arrow{ur}{\nabla \phi} \arrow[rr, dashrightarrow]{d}{W_{D_\phi}({\mathbb{Q}}, {\mathbb{P}}_\theta)} && {\mathbb{Q}}\end{tikzcd}$$
From a high-level perspective, optimizing the Wasserstein-Bregman divergence between two distributions is basically optimizing the $L_2$-Wasserstein distance between one of the distribution and a $\nabla \phi$-transformed distribution (the $D$ term), with a penalty term accounting for the influence of $\phi$ in the divergence measure (the $P$ term). See Figure \[fig:schematic\].
By considering Wasserstein-Bregman divergence, we retain the choice of choosing a symmetric measure (say, $D_\phi(x, y) = ||x - y||_2^2$) or choosing an asymmetric measure (say, $D_\phi(x , y) = \sum_{i = 1}^d x_i \log\left(\frac{x_i}{y_i}\right) - \sum_{i = 1}^d (x_i - y_i)$). In particular, in the special case when the Bregman divergence is chosen to be the $L_2$ distance, we get $L_2$-Wasserstein distance, as in [@Arjovsky:2017vh]. In contrast, all Wasserstein distances are always symmetric, since a metric is used within the definition.
Future Work
-----------
It remains an open problem on how to select the underlying convex function $\phi$ in Bregman divergence for a given problem. Proposition \[FisherInfo\] provides some insight on how $\phi$ is selected. Based on the proposition, the amount of information containing in $p$ depends on the curvature of $\phi$. Choosing $\phi(x) = ||x||_2^2$ can be somewhat conservative in the sense that the amount of information is independent of the value of $p$.
On the other hand, it remains to be further investigated as to the definite advantage of replacing the metric $d(x, y)$ in the Wasserstein distance by the Bregman divergence $D_{\phi}(x, y)$. Nevertheless, Wasserstein-Bregman divergence can serve as a viable candidate for measuring distributional divergence when asymmetry is desirable.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The INTEGRAL observatory has been performing a deep survey of the Galactic central radian since 2003, with the goal of both extracting a catalog of sources and gaining insight into the Galactic diffuse emission. This paper concentrates on the estimation of the total point sources emission contribution. It is now clear that unresolved point sources contribute to the observed diffuse emission; the increasing sensitivity of instruments with time has lead to a steady decrease in estimates of this “diffuse emission”.
We have analysed the first year data obtained with the spectrometer and imager SPI on board INTEGRAL. First, a catalog of 63 hard X-ray sources detected, time-averaged, during our 2003 Galactic plane survey, is derived. Second, after extracting the spectra of the sources detected by SPI, their combined contribution is compared to the total (resolved and unresolved) emission from the Galactic ridge. The data analysis is complex: it requires us to split the total emission into several components, as discrete sources and diffuse emission are superimposed in SPI data. The main result is that point source emission dominates in the hard X-ray/soft $\gamma$-ray domain, and contributes around 90 % of the total emission around 100 keV, while above 250 keV, diffuse electron-positron annihilation, through its three-photon positronium continuum with a positronium fraction $\sim$ 0.97 and the 511 keV electron-positron line, dominates over the sources.
author:
- 'L. Bouchet, J.P.Roques, P. Mandrou'
- 'A. Strong, R. Diehl'
- 'F. Lebrun'
- 'R. Terrier'
- '*Received 2005 May 14 ; accepted 2005, August 19*'
title: 'SPI/INTEGRAL observation of the Galactic central radian: contribution of discrete sources and implication for the diffuse emission [^1] '
---
Introduction
============
Observations carried out for more than three decades indicate that the total hard X-ray/soft $\gamma$-ray emission from the Galaxy results from the superposition of multiple physical processes whose relative contribution depends on the energy.
The spectrum of this emission is reasonably well measured and understood above $\sim$ 1 MeV from OSSE and COMPTEL (Kinzer et al., 1999, Strong et al., 1994, 2004a). At energies above $\sim$ 100 MeV, the dominant emission process is the decay of $\pi^0$ mesons produced in the interaction of cosmic-ray nucleons with the interstellar matter (Bertsch et al., 1993). Between 1 and 70 MeV, electron bremsstrahlung and inverse Compton scattering are expected to dominate over discrete source emission (Sacher & Schönfelder, 1984, Skibo et al., 1993, Strong et al. 2004a ).
In the low-energy domain, the spectrum has been measured with RXTE in the 10-35 keV energy band (Valinia & Marshall, 1998) and with Ginga in the 3-16 keV band (Yamasaki et al., 1997). More recently ASCA (Sugizaki et al., 2001) observations in the 0.7-10 keV band, and Chandra observations (Ebisawa et al., 2001) in the 2-10 keV band, show that there is a genuine diffuse soft X-ray component, and this result is confirmed by XMM-Newton measurements (Hands et al., 2004) in the 2-10 keV band.
However, in the hard X-ray/soft $\gamma$-ray band (50-500 keV), the situation is more complicated as multiple components are believed to contribute to the total emission. These include discrete sources, the positron annihilation line, three-photon positronium continuum radiation and a soft $\gamma$-ray component corresponding to the diffuse continuum emission induced by cosmic-ray interactions (CR diffuse).
Unfortunately, measurements of the Galactic diffuse emission in the hard X-ray/soft $\gamma$-ray band and its interpretation are inherently difficult because of the presence of numerous hard X-ray discrete sources in the Galactic plane. Moreover, generally, hard X-ray/soft $\gamma$-ray instruments either have large fields of view and no imaging capabilities or have imaging capability but no sensitivity to extended emission. As a result, discrimination between diffuse emission and point sources remains a difficult task. For this reason, simultaneous multiple-instrument observations had been performed, with coordinated observations of the Galactic Center region with OSSE/CGRO (Kurfess et al., 1991) and the imaging instrument SIGMA/GRANAT (Paul et al, 1991). The main results were that, in the hard X-ray/soft $\gamma$-ray regime, the spatial distribution of “diffuse emission” is broad and relatively flat in longitude, with an extension in latitude of $\sim$ $5.5^\circ$ width, while a few discrete sources contribute at least 50$\%$ of the total emission (Purcell et al., 1996). However SIGMA had a sensitivity of about 25 mcrab (2$\sigma$) for a typical 24 hours observation. As a result, weak sources escaped detection in its survey, and the unresolved emission was still contaminated significantly by discrete sources. The extension of the spectral shape of the Galactic diffuse emission from the hard X-ray to $\gamma$-ray regime, and how much of the emission was due to discrete sources remained to be precisely determined. Further attempts to derive the diffuse emission characteristics followed (Kinzer et al. 1999, 2001; Boggs et al., 1999; Valinia et al., 2000).
Theoretical studies were unable to explain even 50% of the Galactic emission as originating in the interstellar medium (ISM). Two main processes can lead to an interstellar soft $\gamma$-ray emission. The first one is inverse Compton scattering of high energy (GeV) cosmic ray electrons on the ambient photons field; but these electrons would also produce radio synchrotron emission in the Galactic magnetic field at a level much higher than one actually observed. The second process is bremsstrahlung of a population of electrons of a few hundred keV, radiating through interactions with interstellar gas. Because these electrons will lose their energy through ionization and Coulomb collisions, the total power required to compensate for these energy losses is of the order of $10^{41} - 10^{43}$ erg$~$s$^{-1}$. This power, comparable or higher than that of cosmic ray protons, would affect interstellar-medium ionization equilibrium and give rise to excessive dissociation of interstellar molecules. A possible interstellar process has been proposed by Dogiel et al. (2002). Alternatively, a dozen point sources with intensities around 25 mCrab (the SIGMA 2$\sigma$ sensitivity) could account for most of the remaining diffuse emission. This demonstrates the need for highly sensitive imaging instruments for a precise determination of the diffuse Galactic emission.\
Actually, a new vision of the hard X ray sky is provided by INTEGRAL with the detection of a significant amount of new sources and the discovery of a new class of objects: the highly absorbed sources (Dean et al., 2005) may represent 20% of the total number of sources, but their contribution could not be evaluated by X-ray survey, with an obvious implication for the diffuse emission.
A recent result based on ISGRI/IBIS/INTEGRAL data (Hereafter ISGRI) shows that known binary sources account for the main part of the total Galactic emission (from 86% to 74%) from 20 keV to 220 keV (Lebrun et al., 2004, Terrier et al. 2004). However the limited sensitivity of ISGRI, above 200 keV, does not allow an extension of this study at higher energies. Early study of the diffuse continuum emission using SPI/INTEGRAL (hereafter SPI) shows that diffuse emission dominates above $\sim$ 200 keV (Strong et al. 2003) and that the ratio of diffuse emission to total emission vary from $10\%$ to $100\%$ along SPI energy domain (Strong et al. 2004b).
SPI (see next section) has a moderate angular resolution over a large field, thus is sensitive to extended sources as well as to point sources imaging. SPI can thus be used in a self consistent way to measure both the Galactic diffuse and discrete source emission, avoiding the complication due to different regions observed, different instrument responses, different models assumed and diverse ways in which results are presented (per radian, per Field-Of-View etc.) when comparisons with other instruments are made.
We present the catalog of the sources detected with SPI as well as their global spectral emission. An early SPI sources catalog can be found in Bouchet et al. (2004). Moreover a broadband spectral analysis including point sources and diffuse emission components (annihilation, positronium and CR diffuse continuum) has been performed in a self consistent way, i.e using the same instrument for both sources detection and diffuse emission estimate. We draw attention to an independent, complementary study on this topic by Strong et al. (2005).
Instrument
==========
The spectrometer SPI (Vedrenne et al., 2003) is one of the two main instruments onboard ESA’s INTEGRAL (INTErnational Gamma-Ray Astrophysics Laboratory) observatory launched from Baikonour, Kazakhstan, on 2002 October 17.
It consists of an array of 19 actively cooled high resolution Germanium (Ge) detectors with an area of 508 cm$^{2}$ and a thickness of 7 cm. It is surrounded by a 5-cm thick BGO shield. The detectors cover the 20 keV - 8 MeV energy range with an energy resolution ranging from 2 to 8 keV as a function of energy. In addition to its spectroscopic capability, SPI can image the sky with a spatial resolution of $2.6^\circ$ over a field of view of $30^\circ$, thanks to its coded mask. Despite such a modest angular resolution, it is possible to locate intense sources with an accuracy of few arc minutes (Dubath et al., 2005).\
A complete INTEGRAL orbit lasts ${\sim}$3 days, but scientific data cannot be accumulated when the instrument is crossing the radiation belts, reducing the useful observing time ${\sim}$ 2.5 days. The instrument in-flight performance is given in Roques et al. (2003).
Observations
============
Each INTEGRAL orbital revolution consists of several exposures (typically, 30-40 minutes pointings dithering around the target). Due to the small number of detectors, imaging which relies on these observations in dithering mode (Jensen et al., 2003), the pointing direction varying by steps of $2^\circ$ within a $5 \times 5$ square or a 7-point hexagonal pattern.
The data used to perform this analysis were recorded from 2003, March 3 to 2003, October 19, including in total 42 revolutions (table 1). They cover the region $-50^\circ$ $\leq$ l $\leq$ $50^\circ$, $-25^\circ$ $\leq$ b $\leq$ $25^\circ.$ Most of these data have been obtained through the GCDE (Galactic Center Deep Exposure) part of the Integral core program (Winkler et al., 2003).
Data polluted by solar flares and radiation belt entries are excluded. After image analysis and cleaning, there is 2552 pointings, which represent 5.75$\times$10$^6$ seconds of effective observing time.
Data analysis
=============
The data analysis is done in 2 steps. The first one concerns the determination of the positions of the emitting sources. We have used SPIROS, a software delivered in the INTEGRAL OSA package to produce images. In a second step, we have developed a dedicated software to extract fluxes from the positions obtained with SPIROS.
Image generation
----------------
SPIROS V6 (SPI Iterative Removal of Sources) algorithm has been used to derive source positions, with the MCM background option (mode 5). [^2] SPIROS (Skinner & Connell, 2003), a part of the Integral Science Data Centre (ISDC) package produces synthetic and simplified sky images : they consist of a limited number of sky pixels, those which contain excesses above a given threshold. An important limitation is that sources are considered as constant during the image reconstruction process. As a consequence, at low energy ( $<$ 50 keV), some exposures exhibit an unacceptable $\chi^{2}$ fit of raw data to the reconstructed sky image convolved by the response matrix along with biased residuals distribution. We found this effect to be due to intensity variations of the most intense sources, which are not taken into account. To suppress the effect of the variability of the strongest sources, a special iterative scheme described in §4.2.2 has been used.
SPIROS works both with known and unknown sources. If an input catalog is given, SPIROS will first build a sky model with the proposed sources then looks for a number a new ones required by the data. One must take care of the influence of the source position errors on the image generation. If the source position introduced in SPIROS is different from the true position (even slightly), the iterative removal of sources algorithm leads to subsequent artefacts in the image. This problem may become important for strong sources or when a large number of sources has inaccurate positions. Thus, whenever possible, the exact source positions should be introduced in SPIROS as [*a-priori*]{} knowledge (input catalog) in order to avoid these cumulative errors. But, [*a contrario*]{}, to use an input catalog with too many (non-emitting) sources also leads to unstable solution. Our analysis philosophy (see §5) takes into account both limitations.
Time dependent sources and background fluxes determination algorithm (time-model-fit)
-------------------------------------------------------------------------------------
A complete model fitting procedure (called “time-model-fit”) based on the likelihood statistics has been developed (See Annexe A-1). The input sky model can include both point sources and diffuse components. The main features of the algorithm are:\
- A self determination of the background distribution on the detector plane.\
- A time dependent background normalization determination\
- A time dependent flux determination for each point source with its own timescale (meaningless for diffuse components).
### Background determination
The SPI imaging system, although using a combination of a coded mask with a position sensitive detector, needs, due to the small number of detector pixels, a dithering scheme to increase the number of sky pixels that can be reconstructed. This results in a time modulation of the sky signal. In order not to confuse such modulation with background variations, the background versus time profile of the 19 detectors has to be evaluated. The construction of a time-dependend background model constitutes a key point of SPI data analysis.
Analyses of “empty-field” observations have shown that, for any energy band, the relative count rates (uniformity map) of the 19 Ge detectors are constant while the global amplitude (normalization factor) varies with time. We thus have to determined the relative count-rates of the 19 Ge detectors (background counts-ratio pattern depending on the energy band) leaving normalization as the only free parameter for the background intensity. This is included inside the model fitting procedure as described in annexe A-2 with the background amplitude able to vary on the pointing ($\sim$2500 s) timescale.
### Taking into account source variability
The SPI image reconstruction relies on the dithering. As a result, variability of sources has to be explicitly included in the system of equations to be solved. This is included in our model fitting procedure by means of additionnal equations (Annexe A-3). For each source, the allowed variation time scale can be chosen. In practice, it has to be carefully used as the number of parameters (unknown fluxes) necessary to describe the sky will increase, correspondingly the significances decrease. There is also a mathematical limitation: since there are 19 independent data per pointing, it means that, whatever the number of pointings, there can not be more than 19 variable sources on the time scale of a pointing.
However, most of the sources are weak enough to neglect the influence of their variability on the image reconstruction, and thus can be considered as constant. Even flaring sources are too weak and/or shortliving (e.g. SGR 1806-20) to have any effect on long period.\
In this work, the “time-model-fit” tool has been used for a dedicated variability treatment applied only to the most intense sources, namely, 4U 1700-377, SCO X-1 and OAO 1657-4154, below 50 keV. We build for them the light curves in a one pointing time scale. The corresponding contributions in the count space are then substracted from the original data to derive a “corrected” data set to use as SPIROS input. This cleaning is done from the original data set for each iteration, as the strong sources contributions can be affected by new sources introduction.\
Applying this scheme to the whole data set in the 25-50 keV energy band improves the global reduced $\chi^{2}$ from 3.24 (standard procedure) to 1.33 (Table 3).
Catalogue generation
====================
Images are built using all our data in the following energy bands : 300 - 600 keV, 150-300 keV, 50-150 keV and 25-50 keV. In order to minimise bias (see §4.1), the first step of the analysis was performed without any [*a-priori* ]{} information about known sources. The [*a priori*]{} knowledge is introduced progressively. For each image, SPIROS is parametrized in order to search for up to 30 new excesses above 2$\sigma$ in addition to the input catalog built from the previous iteration. This process (i.e. to fix found positions) suppresses the instabilities of such an IROS algorithm. Moreover, such a procedure leads to a minimal sky model able to represent the data.
The catalog generation process is the following:
- The first image is obtained from SPIROS with an empty input catalog and gives a list of source candidates.
- A catalog is built with: (i) All identified sources above 5$\sigma$ with their celestial positions. (ii) The unidentified excesses above 5$\sigma$ with the positions found. The source identification process is described §5.1. This current catalog is completely regenerated at each iteration as significances can evolve.
- We run SPIROS again with this catalog as input, obtain a new list of sources/excesses, and restart with step 2
- This iterative analysis continues until no other potential excess $\geq$ 2$\sigma$ can be found.
A flow chart illustrates this process in Figure 1.\
The final position catalog for a given energy band is built from the last iteration, where sources and excesses with a significance threshold fixed to 4 $\sigma$ are accepted. The use of a dedicated catalog for each energy band allows us to restrict the number of free parameters to the minimum needed, thus decreasing the flux detection limit at high energy. Figure 2 and 3 present the images obtained by SPIROS in the 25-50 keV and 50-150 keV energy bands using our corresponding catalogs in input.
Source identification
---------------------
To obtain source identification X-ray source catalogs are used. Primarly, the ISGRI catalogs (Bird et al., 2004, Revnivtsev et al., 2004) are used and, if nothing is found, the ISDC catalog is used (Ebisawa et al., 2003 [^3]).
Each excess is considered associated with the nearest known emitting X/$\gamma$ ray sources whose distance is less than $1^\circ$. This rather high value, 3 times the theoretical value for a $5\sigma$ detection (Dubath et al., 2005), is used because the source localisation precision is degraded when more than one source is present, especially for crowded regions. Note however that this value is much less than the geometrical resolving power (2.6$^\circ$), as the dithering pattern amounts to improving the imaging system.\
So, two known sources closer than $1^\circ$ can be retrieved depending on the statistics (Dubath et al., 2005). The Galactic Center region is an exception as too many sources are present in a small area. So we represent the $1^\circ$ region around the Galactic Center by only one source, namely 1E 1740.7-2942.
A more sophisticated scheme is used for the 25-50 keV band: the dataset used is divided in two parts, one constructed with odd pointing numbers the other with even pointing numbers; then two sub-images are built. A non-identified excess is considered as source candidate if it is detected in both images (cross validation method). This method drastically reduces the number of false source detections introduced by systematics, mainly due to variable sources.\
The number of excesses as function of the threshold level is shown in table 2. Above 150 keV and below 5$\sigma$, the number of unidentified excesses increase rapidely. From this, empirical thresholds of 5$\sigma$ and 4 $\sigma$ respectively below and above 150 keV have been fixed.
Final flux extraction
---------------------
Once the position catalogs completed, we build a complete sky model including the detected point sources plus diffuse emission morphologies (described is §6.1).\
Fluxes and significances for each of this sky model components are thus computed in a set of energy bands, using “time-fit -model” (§4.2). The fluxes of point sources and diffuse emission are assumed constant (except for the 20-50 keV band, where 4U 1700-377, SCO X-1 and OAO 1657-4154 vary on the time scale of one pointing). The background normalization is adjusted on a time scale of one pointing.\
Excesses above 4.5$\sigma$ are kept to form the final catalogs. Their fluxes and significances are computed again to be consistent; thus the final catalogs contain some sources below 4.5$\sigma$ that were previously detected above this threshold. This process is described in Figure 4.
The final $\chi^{2}$ obtained with this method between the proposed sky model and the data are shown in table 3.
The catalog
-----------
The use of a dedicated input catalog for each energy band allows us to optimize the signal-to-noise ratio per energy band.\
These catalogs are far from exhaustive: In the computed 1-year-averaged emission, weak or short-transient sources can be completely washed out. Moreover, as we looked for a minimal sky model, and given the modest SPI angular resolution, only one source is necessary to represent the complex regions (1E 1740.7-2942 for instance).
This work, based on the whole dataset, has been complemented by applying the same procedure in 3 subsets, sorting data according to the average Galactic longitude of each observation (table 1): positive (l $>5^\circ$), negative (l $< -5^\circ$) and central ($-5^\circ$ $\leq$ l $\leq$ $5^\circ$) longitudes, that have more or less equal exposures. These new datasets allow us to add four sources in the 25-50 keV band catalog. These sources are labelled with an asterisk.
As the final result, in the 25-50 keV domain the catalog contains 63 excesses above 4$\sigma$; 59 have been identified with known hard X-ray objects in the IBIS catalog, 4 are tentatively associated with X-ray sources (labelled with \*\*). Among the tentatively identified sources, the source at (l,b)=(1.94$^\circ$,-2.02$^\circ$) has a quite high significance of 11.2$\sigma$, while its flux is only 6.3 mCrab. It could represent several weak (eventually extended) sources below ISGRI detection threshold that are integrated within the SPI angular resolution. It could also results from variable sources not taken into account or from errors accumulation due, for example, to a wrong center-of-gravity of the many sources that compose our so-called ‘1E 1740.7-2942’ source actually fixed at the 1E 1740.7-2942 position. Intensive simulations are needed to understand the behaviour of the instrument in such a crowded region, and is beyond the scope of this paper.
Mean fluxes and significances are given in 3 energy bands: 25-50 keV, 50-150 keV,150-300 keV (table 4). In the 50-150 keV band, the catalog contains 20 excesses above 4$\sigma$ (table 5); all but one correspond to confirmed sources, the last identification is only tentative. Finally, 4 known objects emit significantly above 150 keV (table 6).
Spectral components of the Galactic ridge emission
==================================================
We will now concentrate on the relative contribution of point sources to the Galatic ridge emission as a function of the energy range.
In the soft $\gamma$ ray regime, multiple components contribute to the total emission. These include discrete sources, positron annihilation line and three-photon positronium annihilation continuum, and a diffuse continuum resulting probably from cosmic ray interactions with the interstellar medium.
SPI is sensitive to both discrete sources and diffuse emission. As a consequence the discrete sources should always be extracted simultaneously with diffuse emission. However it is difficult to separate/distinguish extended and discrete sources emission if too many free parameters are to be determined: the error bars then become very large, and it may be impossible to derive meaningful information from the analysis. Therefore as much [*a priori*]{} information as possible is included in the analysis:\
- precise external source locations is used in catalogs.\
- spatial morphologies derived from previous works for each diffuse component are used and fixed [*a priori*]{}.\
Sky model and components of soft $\gamma$-ray Galactic emission
---------------------------------------------------------------
### Point sources model versus energy
In order to estimate the total point source emission, it is obvious that all the discrete sources (also weak, transient and possibly undetected) should be included in the analysis. On the other hand, increasing the number of sources decreases the significance of the measurement. Thus only the significant sources are introduced in the analysis, their number depending on the considered energy band. Three spectral regions are considered:\
- In the 20-150 keV energy range, the 63 sources of the 25-50 keV catalog have been introduced.
- In the 150-300 keV energy range, the 20 sources of our 50-150 keV catalog is used.
- Above 300 keV, the 4 sources detected above 150 keV have been considered.
### CR diffuse continuum
The diffuse emission morphology in the 50-400 keV estimated by measurements from OSSE/GRO (Purcell et al., 1996, Kinzer et al., 1999) and confirmed with subsequent simultaneous RXTE and OSSE observations (Valinia et al., 2000), is broadly distributed in longitude with a $5^\circ - 6^\circ$ FWHM in latitude and a $\sim \pm35^\circ$ extent in longitude. However since most of this emission at least in the low energy range is due to point sources (Lebrun et al., 2004), it cannot be used to represent the CR diffuse component. A better model may be the CO map (Dame et al., 2001) since it is a tracer of the interstellar matter. In the absence of a better model and for simplicity, the CR diffuse continuum spatial morphology is modelled by the CO map. Nevertheless, this point constitutes a weakness, as finally the morphology is poorly determined.
### Positron annihilation line and three-photon positronium continuum
The annihilation line detected by SPI has been used to study the 511 keV spatial morphology (Knödlseder et al., 2003). This component is equally well described by models that represent the stellar bulge, by halo populations or by an azimuthally symmetric Gaussian with FWHM of $\sim$ $8^\circ$ (Knödlseder et al., 2005). We thus modelled it with this last simpler hypothesis.\
A recent study of the positronium annihilation emission with SPI shows that its all-sky distribution is consistent with that of the 511 keV electron-positron annihilation (Weidenspointmer et al., 2005). In the absence of any other information, the three-photon positronium continuum is assumed to have the same spatial distribution as the annihilation line.
Extraction of sky model component fluxes
----------------------------------------
Component intensities of the sky model are extracted channel by channel with the time-model-fit algorithm described in §4.2. Then, count rates are converted into photon spectra for each component. The total point source contribution to the Galactic emission is obtained by summing all the point source spectra in the central radian, excluding those from Sco X-1 and Cen A which are located at high galactic latitude (b $\ge$ 20 $^\circ$).\
Figure 5 shows the composite spectrum of the three major components.
The summed point source emission (upper one) can be represented by a power law of index 2.64$^{+0.12}_{-0.13}$ with an exponential cutoff of 504$^{+1100}_{-218}$ and a 100 keV flux of 4.36$^{+0.96}_{-0.69}$ $\times$ 10$^{-4}$ photons cm$^{-2}$ s$^{-1}$ keV$^{-1}$, for a reduced $\chi^{2}$ of 1.74 for 13 degrees of freedom (dof).
The CR diffuse emission spectrum is more complex. Below 50 keV it clearly exhibits a soft component. In the 300-500 keV domain, there is clearly some cross-talk between the positron annihilation and the CR diffuse continuum components. This can be easily understood as we try to fit simultaneously two sky model distributions that have a common part in the inner Galaxy where the exposure time is the highest.\
The CR diffuse continuum fit above 50 keV, excluding 300-500 keV points (figure 5), gives a photon power law index of $\sim$ -1.8 $^{+0.5}_{-0.4}$ and a 100 keV flux of (6 $\pm$2) $\times$ 10$^{-5}$ photons cm$^{-2}$ s$^{-1}$ keV$^{-1}$ for a reduced $\chi^{2}$ of 0.3(8 dof).\
In addition, a fit using only data whose pointing direction is outside the central radian, and supposed to contain insignificant electron-positron annihilation emission, gives a power index of 1.7$^{+0.5}_{-0.4}$ above 50 keV.\
These two indexes are fully compatible with the hard component expected from a ‘cosmic-ray interaction’ model (Skibo et al., 1993), which can be approximated by a power law with index of $\sim$ -1.65 over the 50 keV-10 MeV range. As we poorly constrain this parameter, we decided to fix the power law index to 1.65.\
Finally, in the 20-1800 keV energy range, and excluding the 300-500 keV band,the CR diffuse continuum spectrum is modelled by two components to give a reduced $\chi^{2}$ of 0.5 (8 dof): The first one is a power law with a photon index 1.7 $^{+0.5}_{-0.4}$ and a cutoff energy of 20 keV, having a 50 keV flux of (3.1 $\pm$ 0.7) $\times$ 10$^{-4}$ photons cm$^{-2}$ s$^{-1}$ keV$^{-1}$. The second one is a power law with a photon index fixed to 1.65 and a 100 keV flux of (4.8 $\pm$ 2.2) $\times$ 10$^{-5}$ photons cm$^{-2}$ s$^{-1}$ keV$^{-1}$ that represents the CR diffuse above 50 keV up to several MeV (including OSSE and COMPTEL results).
From this study, the contribution of point sources to the total Galactic emission can be derived: 85% in the 25-50 keV energy range, 90% in the 50-110 keV and 85% in the 110-140 keV range. This fraction then decreases as the positronium emission begins to be significant.\
Second, the 511 keV flux extracted in a 10 keV width channel is $\sim$ 0.9$\times$ 10$^{-3}$ photons cm$^{-2}$ s$^{-1}$. This value is compatible with that obtained in a completely different way by Knödlseder et al. (2005).\
From these data, we also extract a positronium fraction (defined as in Brown & Leventhal, 1987) of $\sim$ 0.9 as a rough estimate. In fact, the CR diffuse component in the 300-500 keV energy range is clearly overestimated, leading to an underestimation of the positronium flux. A more complex and adequate method has thus been proposed for this spectral region (see next section).\
Complete photon model fitting (above 300 keV)
---------------------------------------------
The photon spectra extraction method applied above (§6.2) presents the advantages to be spectral model independent and relatively easy to implement. Its “simplicity” lies on approximations well-suited only for energies below $\sim$ 300 keV. Indeed, the SPI response is averaged over all the exposures for a given component, and the energy redistribution matrix is not properly taken into account. The latter can lead to significant errors particularly in the case of photon spectra with positive slope, such as the positronium continuum (i.e. above 300 keV).
For these reasons a complete model fitting software has been developed, in which the emission photon spectral models of each sky components are convolved with the full SPI response (Sturner et al., 2003) for each pointing. This algorithm is detailed in Annex B. In the other hand, this kind of algorithm allows to fix the shape of the spectra as a function of the energy, via an analytique emission model. As a consequence, the problem of “cross-talk” previously encountered is drastically reduced.\
Each photon spectral component follows an appropriate emission model whose parameters can be adjusted simultaneously to the whole set of data along with the background intensity. Given the size of the dataset such an algorithm can lead to prohibitive computing time and memory space if the number of free parameters is too high. Fortunately, the benefit of this method is more significant at high energy, where the source number is low.
Moreover, we limit the number of free parameters by fixing some of them. First, as the used energy band is limited to 300 keV- 1 MeV, the slope of the power law that modelled the CR diffuse emission is fixed to 1.65 while the flux is to be fit. Second, the point source emission consists of 4 discrete sources (table 4) whose spectra are represented a by power law with an exponential cut-off. Because the source statistics are low, the indices and normalizations of these power laws have been previously determined below 150 keV by the first method (§6.2), while the cutoff energy is to be fitted. Results may slightly depend on the sources spectral shape assumed above 300 keV.\
The annihilation spectrum (positronium continuum plus line) is then determined by the complete model fitting.
A line flux of (0.93$\pm$0.15) 10$^{-3}$ photons cm$^{-2}$ s$^{-1}$ and a positronium fraction of 0.97$^{+0.09}_{-0.07}$ are found. The 100 keV flux of the diffuse emission is 5$\pm$3 $\times$ 10$^{-5}$ photons cm$^{-2}$ s$^{-1}$ keV$^{-1}$, in perfect agreement with previous determinations. The reduced $\chi^{2}$ of the fit is 0.99 (630337 dof), the reduced $\chi^{2}$ for each energy band are distributed between 0.95 and 1.2.\
The result is illustrated in figure 6.
Discussion and conclusions
==========================
Despite the SPI instrument’s moderate angular resolution, and benefitting from the large number of pointings, a first catalog of hard X-ray sources detected by SPI/INTEGRAL in the Galactic plane between $-50^\circ$ $\leq$ l $\leq$ $50^\circ$, $-25^\circ$ $\leq$ b $\leq$ $25^\circ$ has been derived.\
This catalog contains 63 sources and gives their fluxes up to 300 keV. It has been built by means of model-fitting procedure simultaneously applied to all pointings of the whole dataset of 5.7$\times$10$^6$ seconds, and represents the minimal averaged sky model needed to decribe the data. Thus transient sources could escape detection. Moreover, due to the modest SPI angular resolution, our sky model in crowded regions is simplified, with one point source representing the combined emission of all sources near 1E 1740.7-2942. On the other hand, this catalog is perfectly suited to estimate the point source contribution to the Galactic ridge emission.\
The diffuse emission components (continuum, 511 keV and positronium) have been studied by fixing their spatial distributions, as a first order model, and taking into account our catalog.\
The main conclusions are :
- Point sources contribute at least for 85-90% of the total Galactic emission in the energy range 25 to 140 keV. This independantly confirms the ISGRI result that point sources largely dominate the total Galactic emission (Lebrun et al., 2004) at least up to 250 keV. The interstellar emission is around 15 % of the total emission up to 150 keV. The low energy part ($<$ 250 keV) of the total spectrum is source dominated while the high energy part ($>$ 250 keV) is diffuse dominated.
- The CR diffuse emission (50 - 250 keV) can be fitted with a power law joining smoothly the high energy continuum as measured by OSSE (Kinzer et al., 1999) and COMPTEL (Strong et al., 1994) above 1 MeV.
- No additional spectral component is needed above the power law spectrum to model the diffuse continuum above 50 keV. The soft diffuse component reported by Kinzer et al. (1999) in this region is mostly due to point sources not taken into account by this instrument.
- Below 50 keV, a real diffuse soft component can still exist, but may also be related to very steep/weak X-ray sources not included in the present analysis.
- There is no evidence of 511 keV point source emission, in agreement with the conclusion drawn by Knödlseder et al. (2005).
- A positronium fraction 0.97$^{+0.09}_{-0.07}$ and a 511 keV flux of (0.93 $\pm$0.15) 10$^{-3}$ photons cm$^{-2}$ s$^{-1}$. This positronium fraction is compatible with the OSSE determination: 0.93 $\pm$ 0.04 (Kinzer et al., 2001) and SPI/INTEGRAL determinations of 0.94 $\pm$ 0.06 (Churazov et al., 2005).
These results constitute a first step as they are based on SPI data taken during the first year of operation and will be obviously refined in the future. An alternative approach is proposed by Strong et al. (2005).
The catalogs are available electronically as ApJ ‘supplementary electronic materials’.
Acknowledgments {#acknowledgments .unnumbered}
===============
The SPI/INTEGRAL project has been completed under the responsibility and leadership of CNES. We are grateful to ASI, CEA, CNES, DLR, ESA, INTA, NASA and OSTC for support.
Bird A., Barlow E.J., Bassani L., et al, 2004, ApJL, 607, L33
Bertsch D.L., Dame T.M., Fichtel C.E., et al., 1993, ApJ, 416, 587
Boggs S.E., Lin R.P., Slassi-Sennou S., et al., 1999, ApJ, 544, 320
Bouchet L., Jourdain E., Roques J.P., et al., Proc. 5th INTEGRAL Workshop, ESA SP-552, p147
Brown B. & Leventhal M., 1987, ApJ, 319, 637
Churazov E., Sunyaev R., Sazonov S., et al, 2005,MNRAS 357, 1377
Dame T.M., Hartmann D., & Thadeus P., 2001, ApJ, 547, 792
Dean A.J., Hill, A. B., Stephen J. B., et al, 2005, A&A, submitted
Dogiel V.A., Schönfelder, V., & Strong, A.W. 2002a, A&A, 382, 730
Dubath P., Knodlseder J., Skinner G.K., et al., 2005, MNRAS, 357, 420
Ebisawa K., Meada Y., Kaneda H. & Yamauchi S., 2001, Science, 293, 1633
Ebisawa K., Bourban A., Bodaghee A., et al, 2003, A&A, 411, L59
Hands A.D.P., Warwick R.S., Watson M.G. et Helfand D.J., 2004, MNRAS, 361, 31
Jensen P.L., Clausen K., Cassi C.,, et al., 2003, A&A, 411, L7
Kinzer R.L., Purcell W.R. & Kurfess J.D., 1999, ApJ, 515, 215
Kinzer R.L., Milnes, P.A.., Kurfess J.D., et al., 2001, ApJ, 559, 282
Knödlseder J., Lonjou V., Jean P., et al, 2003, A&A, 411, L457
Knödlseder J., Pierre J., Lonjou V., et al., 2005, A&A in press
Kurfess J.D., Johnson W.N., Kinzer R.L., et al., 1991, Adv Space Res., 11(8), 323
Lebrun F., Terrier R., Bazzano A., et al., 2004, Nature, 428, 293
Paul J., Manbdrou P., Ballet J., et al., 1991, Adv Space Res., 11(8), 289
Purcell W.R., Bouchet L., Johnson W.N., et al., 1996, A&AS, 120, 389
Revnivtsev M.G., Sunyaev R.A., Varshalovich D.A., et al, 2004, ASTL, 30, 382
Roques J.P., Schanne S., Von Kienlin A., et al, 2003, A&A, 411, L91
Sacher W., & Schoenfelder V., 1984, ApJ, 279, 817
Skibo J.G, Johnson, W.N., Kinzer, R.L., et al, 1993, A&AS, 25, 1293
Skinner G. & Connell P., 2003, A&A, 411, L123
Strong A.W., et al., 1994, A&A, 292, 82
Strong A.W., et al, 2003, A&A, 411, L447
Strong A.W., Moskalenko I.V & Reimer O, 2004a, ApJ, 613, 956
Strong, A. W., et al. 2004b, Proc. 5th INTEGRAL Workshop, ESA SP-552, p507; astro-ph/0405023
Strong A.W. et al., 2005, A&A in press, astro-ph/0509290
Sturner, S.J., Schrader, C.R., Weidenspointner, G., et al., 2003, A&A, 411, L81
Sugizaki, M., Kazuhisa K., Kaneda H., et al., 2001, ApJS, 134,77
Terrier, R., et al., 2004, Proc. 5th INTEGRAL Workshop, ESA SP-552, p513; astro-ph/0405207
Valinia A. & Marshall F.E., 1998, ApJ, 505, 134
Valinia A., Kinzer R.L. & Marshall F.E., 2000, ApJ, 534, 277
Vedrenne G., Roques J.P., Schonfelder V., et al, 2003, A&A, 411, L63
Weidenspointner G., Knödlseder J., Pierre J., et al, 2005, to be submitted to A%A
Winkler C., Gehrels N., Schonfelder V., et al., 2003, A&A, 411, L349
Yamasaki N.Y., Ohashi T., Takahara F., et al, 1997, ApJ, 481, 821
[cccccc]{}
47 & 2003/03/03 - 03:32 & 2003/03/05 - 18:02 & 136.0 & -27.4 & 0.3\
49 & 2003/03/09 - 03:00 & 2003/03/11 - 14:43 & 46.8 & -20.7 & 1.1\
50 & 2003/03/12 - 02:49 & 2003/03/14 - 16:44 & 92.1 & -27.9 & 0.8\
51 & 2003/03/15 - 02:34 & 2003/03/17 - 16:30 & 81.0 & -12.1 & 0.8\
52 & 2003/03/18 - 02:19 & 2003/03/20 - 16:36 & 91.6 & -14.0 & -0.1\
53 & 2003/03/21 - 02:04 & 2003/03/23 - 15:43 & 143.2 & -3.1 & 0.2\
54 & 2003/03/24 - 01:50 & 2003/03/26 - 15:56 & 116.7 & -15.8 & -1.7\
55 & 2003/03/27 - 01:35 & 2003/03/29 - 15:50 & 26.5 & -10.3 & 10.7\
56 & 2003/03/30 - 01:22 & 2003/04/01 - 15:27 & 49.6 & 4.1 & 4.0\
57 & 2003/04/02 - 01:10 & 2003/04/04 - 15:10 & 0.7 & -0.1 & -1.1\
58 & 2003/04/05 - 00:59 & 2003/04/07 - 14:20 & 18.1 & -5.2 & -2.9\
59 & 2003/04/08 - 00:49 & 2003/04/10 - 14:44 & 48.7 & 12.7 & 1.1\
60 & 2003/04/11 - 00:34 & 2003/04/13 - 14:59 & 43.6 & -5.7 & -0.4\
61 & 2003/04/14 - 00:20 & 2003/04/16 - 14:17 & 51.2 & -0.6 & 0.7\
62 & 2003/04/17 - 00:06 & 2003/04/19 - 13:35 & 22.8 & 6.1 & 4.4\
64 & 2003/04/22 - 23:37 & 2003/04/25 - 13:41 & 189.9 & 18.1 & -0.1\
65 & 2003/04/25 - 23:20 & 2003/04/28 - 13:11 & 201.7 & 24.2 & 0.1\
66 & 2003/04/28 - 23:06 & 2003/05/01 - 13:25 & 145.0 & 26.2 & -2.5\
97 & 2003/07/30 - 16:59 & 2003/08/02 - 05:57 & 180.0 & -1.4 & 23.7\
100 & 2003/08/08 - 15:18 & 2003/08/11 - 05:07 & 137.3 & -18.7 & 0.6\
101 & 2003/08/11 - 15:57 & 2003/08/14 - 05:19 & 181.7 & -2.1 & 23.5\
103 & 2003/08/17 - 14:39 & 2003/08/20 - 04:32 & 119.2 & -7.8 & 0.6\
104 & 2003/08/20 - 14:25 & 2003/08/23 - 04:30 & 156.2 & -42.1 & 31.1\
105 & 2003/08/23 - 15:06 & 2003/08/26 - 03:41 & 177.0 & -0.1 & -0.6\
106 & 2003/08/26 - 14:51 & 2003/08/29 - 03:38 & 166.7 & -0.7 & 0.6\
107 & 2003/08/29 - 14:35 & 2003/09/01 - 03:17 & 197.6 & 0.8 & 0.8\
108 & 2003/09/01 - 14:22 & 2003/09/04 - 02:54 & 201.5 & 1.5 & -0.0\
109 & 2003/09/04 - 14:07 & 2003/09/07 - 03:05 & 193.5 & 10.1 & -0.8\
110 & 2003/09/07 - 13:52 & 2003/09/10 - 02:46 & 139.2 & 0.1 & -0.6\
111 & 2003/09/10 - 13:39 & 2003/09/13 - 02:33 & 199.3 & 0.5 & -0.3\
112 & 2003/09/13 - 13:26 & 2003/09/16 - 02:21 & 195.8 & 0.3 & -0.3\
113 & 2003/09/16 - 13:14 & 2003/09/19 - 02:09 & 171.7 & -0.1 & -0.3\
114 & 2003/09/19 - 13:07 & 2003/09/22 - 01:58 & 177.3 & -0.2 & -0.3\
115 & 2003/09/22 - 12:55 & 2003/09/25 - 02:01 & 193.3 & 0.0 & -0.2\
116 & 2003/09/25 - 11:47 & 2003/09/28 - 01:41 & 148.5 & -14.8 & -2.0\
117 & 2003/09/28 - 11:35 & 2003/10/01 - 01:50 & 188.5 & 22.0 & 0.9\
118 & 2003/10/01 - 11:21 & 2003/10/04 - 01:45 & 180.8 & -19.2 & -1.2\
119 & 2003/10/04 - 11:08 & 2003/10/07 - 01:12 & 178.9 & -1.9 & -1.3\
120 & 2003/10/07 - 10:55 & 2003/10/10 - 01:02 & 190.6 & 2.6 & 0.1\
121 & 2003/10/10 - 10:45 & 2003/10/13 - 00:46 & 198.9 & 14.3 & 0.8\
122 & 2003/10/13 - 10:35 & 2003/10/16 - 00:32 & 199.0 & 13.6 & 0.5\
123 & 2003/10/16 - 10:25 & 2003/10/19 - 00:22 & 176.3 & 26.9 & 0.3\
[lcccc]{}
7$\sigma$ & 50(47) & 16(15) & 2(2) & 1(1)\
6$\sigma$ & 56(53) & 20(19) & 4(4) & 3(3)\
5$\sigma$ & 62(58) & 24(19) & 4(4) & 4(3)\
4$\sigma$ & 69(64) & 30(21) & 7(4) & 11(5)\
[lcc]{}
300 - 600 & 1.027 & 45867\
150 - 300 & 1.013 & 45869\
50 - 150 & 1.125 & 45855\
25 - 50 & 3.242 & 45809\
25 - 50\* & 1.328 & 45806\
[rlccccc]{}
1 & Cen A & -50.48 & 19.38 & 54.7 $\pm$ 2.9 ( 18.9) & 72.7 $\pm$ 6.3 ( 11.5) & 85.4 $\pm$ 21.8 ( 3.9)\
2 & 4U 1516-569 & -37.88 & 0.04 & 18.0 $\pm$ 3.0 ( 6.0) & 15.0 $\pm$ 6.6 ( 2.3) & $<$ 46.1\
3 & 4U 1538-522 & -32.58 & 2.18 & 9.6 $\pm$ 2.0 ( 4.9) & $<$ 9.2 & $<$ 33.5\
4 & 4U 1608-522 & -29.06 & -0.86 & 17.8 $\pm$ 1.6 ( 11.0) & 17.8 $\pm$ 3.9 ( 4.6) & $<$ 28.5\
5 & 4U 1636-536 & -27.08 & -4.81 & 19.2 $\pm$ 1.5 ( 12.8) & 12.5 $\pm$ 3.6 ( 3.5) & $<$ 26.1\
6 & IGRJ16167-4957\* & -26.92 & 0.51 & 5.8 $\pm$ 1.4 ( 4.3) & $<$ 7.2 & $<$ 26.7\
7 & IGR J16318-4848 & -24.37 & -0.44 & 20.9 $\pm$ 1.5 ( 13.7) & $<$ 7.4 & $<$ 27.5\
8 & 4U 1630-47 & -23.08 & 0.24 & 76.1 $\pm$ 1.5 ( 50.1) & 49.8 $\pm$ 3.7 ( 13.6) & $<$ 26.9\
9 & GX 339-4 & -21.06 & -4.32 & 8.6 $\pm$ 1.2 ( 7.0) & 11.8 $\pm$ 2.9 ( 4.1) & $<$ 20.5\
10 & IGR J16418-4532 & -20.81 & 0.52 & 14.1 $\pm$ 1.2 ( 11.4) & 9.6 $\pm$ 2.9 ( 3.3) & $<$ 21.2\
11 & 4U 1705-440 & -16.67 & -2.33 & 28.3 $\pm$ 1.2 ( 22.9) & 18.1 $\pm$ 2.9 ( 6.3) & $<$ 21.1\
12 & 4U 1702-429 & -16.11 & -1.31 & 8.4 $\pm$ 1.3 ( 6.5) & 6.9 $\pm$ 3.1 ( 2.2) & $<$ 22.5\
13 & OAO 1657-4154 & -15.63 & 0.32 & 83.0 $\pm$ 1.7 (50.3) & 34.1 $\pm$ 2.3 ( 14.7) & $<$ 17.1\
14 & 4U 1735-444 & -13.94 & -6.99 & 6.8 $\pm$ 0.9 ( 7.3) & $<$ 4.2 & $<$ 15.1\
15 & IGR J17195-4100 & -13.03 & -2.12 & 11.1 $\pm$ 0.8 ( 13.9) & 7.1 $\pm$ 1.9 ( 3.7) & $<$ 14.1\
16 & 4U 1700-377 & -12.25 & 2.18 & 164.1 $\pm$ 1.4 (119.7) & 75.8 $\pm$ 2.0 ( 38.3) & 16.6 $\pm$ 7.4 ( 2.2)\
17 & GX 349+2 & -10.88 & 2.76 & 14.2 $\pm$ 1.2 ( 11.4) & $<$ 6.1 & $<$ 22.9\
18 & IGR J17252-3616\* & -8.49 & -0.36 & 11.8 $\pm$ 1.4 ( 8.5) & $<$ 3.2 & $<$ 11.8\
19 & 4U 1746-37 & -6.48 & -4.97 & 7.7 $\pm$ 0.6 ( 13.0) & 6.7 $\pm$ 1.4 ( 4.7) & $<$ 10.6\
20 & GX 354-0 & -5.69 & -0.15 & 23.0 $\pm$ 0.6 ( 37.2) & 7.5 $\pm$ 1.5 ( 4.9) & $<$ 11.3\
21 & GRS 1724-308 & -3.67 & 2.31 & 20.3 $\pm$ 0.5 ( 38.7) & 12.0 $\pm$ 1.4 ( 8.9) & $<$ 10.1\
22 & 4U 1822-371 & -3.14 & -11.29 & 29.6 $\pm$ 0.8 ( 38.1) & 9.6 $\pm$ 1.8 ( 5.2) & $<$ 13.4\
23 & IGR J17488-3253 & -3.03 & -2.64 & 15.4 $\pm$ 0.9 ( 16.9) & 9.1 $\pm$ 2.3 ( 4.0) & $<$ 16.9\
24 & IGR J17464-3213 & -2.74 & -1.83 & 6.2 $\pm$ 1.1 ( 5.7) & 9.3 $\pm$ 2.7 ( 3.5) & $<$ 19.6\
25 & GRS 1734-292 & -1.11 & 1.41 & 7.8 $\pm$ 0.6 ( 12.3) & $<$ 3.4 & 15.6 $\pm$ 6.2 ( 2.5)\
26 & Sco X-1 & -0.90 & 23.77 & 216.8 $\pm$ 1.6 (136.6) & 20.9 $\pm$ 2.3 ( 9.2) & 18.9 $\pm$ 8.5 ( 2.2)\
27 & 1E 1740.7-2942 & -0.87 & -0.10 & 80.4 $\pm$ 1.1 ( 71.8) & 78.9 $\pm$ 2.7 ( 28.8) & 45.6 $\pm$ 10.0 ( 4.6)\
28 & SLX 1744-299 & -0.72 & -0.89 & 25.6 $\pm$ 1.2 ( 20.6) & $<$ 6.3 & 24.7 $\pm$ 11.5 ( 2.1)\
29 & V2400Oph & -0.15 & 8.73 & 14.2 $\pm$ 0.9 ( 16.4) & 4.8 $\pm$ 2.1 ( 2.3) & $<$ 15.2\
30 & EXMSB 1709-232\* & 0.59 & 9.27 & 16.8 $\pm$ 3.4 ( 4.9) & $<$ 4.2 & $<$ 15.2\
31 & SLX 1735-269 & 0.81 & 2.41 & 10.1 $\pm$ 0.5 ( 18.7) & 3.8 $\pm$ 1.4 ( 2.8) & $<$ 10.0\
32 & RX J1832-33 & 1.54 & -11.36 & 4.1 $\pm$ 0.7 ( 5.9) & 3.5 $\pm$ 1.7 ( 2.1) & $<$ 12.1\
33 & XTE J1807-294 & 1.93 & -4.27 & 5.0 $\pm$ 0.6 ( 9.1) & $<$ 2.8 & $<$ 10.4\
34 & AX J1758.0-2818\*\* & 1.94 & -2.02 & 6.3 $\pm$ 0.6 ( 11.2) & $<$ 3.0 & 14.9 $\pm$ 5.7 ( 2.6)\
35 & GX 1+4 & 1.95 & 4.80 & 12.0 $\pm$ 0.5 ( 22.2) & 7.3 $\pm$ 1.3 ( 5.5) & $<$ 10.0\
36 & GX 3+1 & 2.30 & 0.80 & 10.5 $\pm$ 0.6 ( 17.6) & $<$ 3.1 & $<$ 11.5\
37 & 4U 1820-303 & 2.79 & -7.92 & 14.8 $\pm$ 0.6 ( 25.1) & $<$ 2.8 & 11.8 $\pm$ 5.2 ( 2.3)\
38 & GRS 1758-258 & 4.51 & -1.36 & 61.2 $\pm$ 1.0 ( 59.9) & 76.5 $\pm$ 2.5 ( 30.7) & 59.6 $\pm$ 9.2 ( 6.5)\
39 & 4U 1849-31 & 4.97 & -14.35 & 4.3 $\pm$ 1.0 ( 4.5) & $<$ 4.3 & $<$ 14.9\
40 & GX 5-1 & 5.09 & -1.01 & 20.6 $\pm$ 1.1 ( 19.4) & 6.7 $\pm$ 2.6 ( 2.6) & 19.2 $\pm$ 9.5 ( 2.0)\
41 & SAX J1810.8-2609\*\* & 5.20 & -3.43 & 2.5 $\pm$ 0.5 ( 4.8) & $<$ 2.6 & $<$ 9.7\
42 & IGR J17597-2201 & 7.58 & 0.77 & 9.5 $\pm$ 0.6 ( 15.4) & $<$ 3.0 & $<$ 11.2\
43 & 4U 1745-203 & 7.72 & 3.81 & 4.3 $\pm$ 0.6 ( 7.7) & $<$ 2.7 & $<$ 10.1\
44 & GX 9+9 & 8.53 & 9.04 & 4.5 $\pm$ 0.7 ( 6.8) & 3.3 $\pm$ 1.6 ( 2.1) & $<$ 11.3\
45 & GS 1826-24 & 9.28 & -6.08 & 80.5 $\pm$ 0.6 ( 125.9) & 54.0 $\pm$ 1.5 ( 35.8) & 30.7 $\pm$ 5.6 ( 5.5)\
46 & SGR 1806-20 & 10.01 & -0.21 & 5.8 $\pm$ 0.7 ( 8.8) & $<$ 3.2 & $<$ 11.9\
47 & AX J1812.2-1842/SNR 011.2-00.3\*\* & 11.90 & -0.16 & 3.7 $\pm$ 0.7 (5.3) & 6.4 $\pm$ 1.7 ( 3.8) & $<$ 12.7\
48 & PKS 1830-211 & 12.15 & -5.72 & 4.8 $\pm$ 0.7 ( 7.0) & 7.6 $\pm$ 1.6 ( 4.6) & $<$ 12.3\
49 & GX 13+1 & 13.52 & 0.12 & 7.9 $\pm$ 0.8 ( 10.2) & 4.7 $\pm$ 1.9 ( 2.5) & $<$ 14.0\
50 & GX 17+2 & 16.44 & 1.28 & 17.1 $\pm$ 0.8 ( 21.1) & $<$ 3.9 & $<$ 14.2\
51 & M 1812-12 & 18.03 & 2.40 & 20.0 $\pm$ 0.9 ( 22.7) & 16.1 $\pm$ 2.1 ( 7.7) & $<$ 15.3\
52 & AX J1825.1-1253\*\* & 18.49 & -0.14 & 3.5 $\pm$ 0.9 ( 4.0) & $<$ 4.3 & $<$ 15.9\
53 & SNR 021.5-00.9\* & 21.51 & -0.88 & 5.4 $\pm$ 0.9 ( 6.0) & $<$ 4.8 & $<$ 17.5\
54 & IGR J18325-0756\* & 23.71 & 0.57 & 6.2 $\pm$ 0.9 ( 6.8) & 6.3 $\pm$ 2.5 ( 2.5) & $<$ 18.9\
55 & 4U 1850-087 & 25.36 & -4.32 & 5.7 $\pm$ 1.1 ( 5.2) & $<$ 5.2 & $<$ 19.0\
56 & Kes 73 & 27.39 & 0.01 & 5.1 $\pm$ 1.1 ( 4.5) & 8.1 $\pm$ 2.8 ( 2.9) & $<$ 20.5\
57 & XTE J1855-026 & 31.08 & -2.09 & 17.9 $\pm$ 1.4 ( 13.0) & 10.6 $\pm$ 3.3 ( 3.3) & $<$ 23.8\
58 & 1916-053 & 31.35 & -8.49 & 6.4 $\pm$ 1.6 ( 4.0) & $<$ 7.3 & 31.9 $\pm$ 13.3 ( 2.4)\
59 & AX J1852.6+0038 & 33.65 & 0.02 & 6.2 $\pm$ 1.6 ( 3.9) & $<$ 7.6 & $<$ 28.0\
60 & 4U 1901+03 & 37.19 & -1.25 & 22.9 $\pm$ 2.2 ( 10.6) & $<$ 10.4 & $<$ 37.9\
61 & SS 433 & 39.70 & -2.24 & 13.6 $\pm$ 2.8 ( 4.8) & $<$ 13.2 & $<$ 47.6\
62 & 4U 1909+07\* & 41.90 & -0.81 & 13.5 $\pm$ 3.3 ( 4.2) & $<$ 17.7 & $<$ 59.4\
63 & GRS 1915+105 & 45.37 & -0.21 & 91.6 $\pm$ 15.4 ( 6.0) & $<$ 36.9 & $<$ 97.4\
[lccc]{}
Cen A & -50.48 & 19.38 & 76.8 $\pm$ 6.7 ( 11.5)\
4U 1608-522 & -29.06 & -0.86 & 19.5 $\pm$ 3.9 ( 5.0)\
4U 1630-47 & -23.08 & 0.24 & 55.5 $\pm$ 3.2 ( 17.6)\
OAO 1657-4154 & -15.63 & 0.32 & 37.0 $\pm$ 2.0 ( 18.2)\
4U 1705-440 & -16.67 & -2.33 & 22.5 $\pm$ 2.2 ( 10.2)\
4U 1700-377 & -12.25 & 2.18 & 80.2 $\pm$ 1.6 ( 48.8)\
4U 1746-37 & -6.48 & -4.97 & 9.4 $\pm$ 1.3 ( 7.2)\
GX 354-0 & -5.69 & -0.15 & 5.3 $\pm$ 1.3 ( 4.2)\
GRS 1724-308 & -3.67 & 2.31 & 11.1 $\pm$ 1.2 ( 9.0)\
4U 1822-371 & -3.14 & -11.29 & 9.3 $\pm$ 1.7 ( 5.4)\
IGR J17488-3253 & -3.03 & -2.64 & 10.2 $\pm$ 1.9 ( 5.3)\
IGR J17464-3213 & -2.74 & -1.83 & 8.2 $\pm$ 1.9 ( 4.4)\
Sco X-1 & -0.90 & 23.77 & 22.4 $\pm$ 2.4 ( 9.3)\
1E 1740.7-2942 & -0.87 & -0.10 & 87.3 $\pm$ 1.3 ( 67.1)\
GX 1+4 & 1.95 & 4.80 & 6.1 $\pm$ 1.2 ( 5.0)\
GRS 1758-258 & 4.51 & -1.36 & 82.9 $\pm$ 1.2 ( 67.8)\
GS 1826-24 & 9.28 & -6.08 & 57.0 $\pm$ 1.5 ( 38.4)\
AX J1812.2-1842/SNR 011.2-00.3\*\* & 11.90 & -0.16 & 5.7 $\pm$ 1.5 ( 3.9)\
PKS 1830-211 & 12.15 & -5.72 & 8.0 $\pm$ 1.5 ( 5.2)\
M 1812-12 & 18.03 & 2.40 & 16.9 $\pm$ 2.0 ( 8.6)\
[lccc]{}
-2.74 & -1.83 & IGR J17464-3213 & 21.2 $\pm$ 4.0 (5.3)\
-0.87 & -0.10 & 1E 1740.7-2942 & 61.4 $\pm$ 4.1 (14.8)\
4.51 & -1.36 & GRS 1758-258 & 75.8 $\pm$ 4.0 (18.9)\
9.28 & -6.08 & GS 1826-24 & 28.6 $\pm$ 4.3 (6.6)\
Algorithms
==========
Maximum likelihood algorithm
----------------------------
The relation between the expected data, instrument aperture response, background and the $N_s$ sources which light the detectors is
$$E(d,p,e)=\sum_{j=1}^{N_s} M(d,p,e,\theta_j,\phi_j) t(d,p) I(\theta_j,\phi_j,e)+B(d,p,e)$$
where E(d,p,e) and B(d,p,e) are respectively the expected data and background in the energy bin e, pointing p and detector d. $I(\theta_j,\phi_j,e)$ is the intensity of the source number j, or sky image pixel j, in the direction $(\theta_j,\phi_j)$ in the energy bin e. $M(d,p,e,\theta_j,\phi_j)$ the instrument response of the data element (d,p,e) to the sky image pixel j in the direction $(\theta_j,\phi_j)$. t(d,p) is the duration exposure for the detector d and pointing p.
The background term is rewritten as
$$B(d,p,e)=t(d,p)U(d,e)$$
U(d,e) is the uniformity map that fixed the relative counts ratio between the 19 Ge detectors (§4.1). If the background amplitude varies on the pointing timescale, then an amplitude coefficient A(p) is introduced to describe the variable background $B_v$,
$$B_v(d,p,e)=t(d,p)U(d,e)A(p)$$
For a given energy bin e and to simplify the equation representation, the sky image pixel direction $(\theta_j,\phi_j)$ is represent by the sky pixel number j. The equation reduces to
$$E(d,p)=\sum_{j=1}^{N_s} M(d,p,j) t(d,p) I(j)+ U(d) t(d,p) A(p)$$
For the maximum likelihood algorithm, the cash statistics (Cash, 1979) is used and the following equation is to be maximized with respect to the parameters I and A (which are constrained to be positive):
$$L(I_1,...,I_N;A_1,..A_{N_p})=E(d,p)-N(d,p)log(E(d,p))$$
where N(d,p) are the measured data (for energy bin e) and $N_P$ the number of pointing.
Background uniformity map estimation
------------------------------------
The background uniformity estimation is obtained in two steps. First an uniformity map $\widetilde{U}(d)$ is assumed and the likelihood is optimized with respect to I and A to give the best estimation $\widetilde{I}$ and $\widetilde{A}$. Second, knowing the current best estimate of the sources intensity and current background. The background map is obtained through the minimization for each detector of
$$\sum_{p=1}^{N_P} \frac{\big( N(d,p)-S(d,p)-\beta (d) B(d,p) \big) ^2}{\sigma^2(d,p)}$$
where S(d,p) is the total source counts contribution, B the current background and $\beta(d)$ a multiplicative factor of order 1 of the current background map for detector d.
$$\begin{array}{l}
S(d,p)=\sum_{j=1}^{N_s} M(d,p,j) t(d,p) \widetilde{I}(j) \\
B(d,p)=U(d) t(d,p) \widetilde{A}(p) \\
\sigma(d,p)^2=E(d,p) \\
\end{array}$$
The minimization with respect to $\beta(d)$ gives
$$\beta(d)= \frac {\sum_{p=1}^{N_P} \big(N(d,p)-S(d,p) \big) B(d,p) / \sigma^2(d,p)}
{\sum_{p=1}^{N_P} B^2(d,p) / \sigma^2(d,p)}$$
The best uniformity map is then $$U(d)=\widetilde{U}(d) \beta(d)$$
This process is repeated several times until the equivalent $\chi^{2}$ stops to decrease , the equivalent $\chi^{2}$ being defined as
$$\chi_{eqv.}^2=\sum_{d,p} \frac{\Big( E(d,p)-D(d,p)) \Big) ^2}{E(d,p)}$$
Variable source
---------------
The expected counts E from a given source k counts is
$$E_{k}(d,p)=M(d,p,k) t(d,p) I(k)$$
For a variable source this equation is to be expanded in several equations, corresponding to intervals whic the source is considered as constant
$$E_{k}(d,p) = \left\{ \begin{array}{ll}
M(d,p,k) t(d,p) I_1(k) & \textrm{if~} p=1,\cdots,p_1 \\
M(d,p,k) t(d,p) I_2(k) & \textrm{if~} p=p_1+1,\cdots, p_2\\
\cdots & \cdots \\
M(d,p,k) t(d,p) I_L(k) & \textrm{if~} p=p_{L-1}+1,\cdots,P \\
\end{array} \right.$$
Finally, instead of a single intensity for source k, there is L intensities $I_1(k),\cdots,I_L(k)$ to be determined.
Photon model-fitting
====================
The function to be minimized is $$\chi^2 = \sum_{e=1}^{N_E} \sum_{p=1}^{N_P} \sum_{d=1}^{N_D} \frac{\big( N(d,p,e)-E(d,p,e) \big) ^2}{\sigma^2(d,p,e)}$$
where $N_E$ is the number of detector energy bin (counts space), $N_P$ the number of pointing and $N_D$ the number of detector. N(d,p,e) and E(d,p,e) are respectively measured and expected counts for detector d, pointing p and detector energy bin e. The following equation relies the response matrix, sources photon spectra, background and expected counts
$$\sum_{j=1}^{N_s} \sum_{e_{ph}=1}^{N_{E_{ph}}} R(d,p,j,E,e_{ph})t (d,p) F(j,e_{ph}) + B(d,p,e) = N(d,p,e)$$
where $R(d,p,j,E,E_{ph})$ is the response (aperture response and detector redistribution matrix) of the data element (d,p,e) to 1 incident photon in the sky direction $(\theta_j,\phi_j)$ for the incident in the energy bin energy $e_{ph}$.$N_s$ and $N_{E_{ph}}$ are respectively the number of sources and energy bin in the photon space. $F(j,e_{ph})$ the incident photon spectrum in the energy bin $E_{ph}$ (photon space) for the source j.
For each one of the $N_s$ sources, the incident photon spectrum is described by few parameters described by the vector $\vec \theta_j$
$$F_{j,e_{ph}}(\vec \theta_j) \equiv F(j,e_{ph}) \qquad j=1,\cdots,N_{s}; e_{ph}=1, \cdots, N_{E_{ph}}$$
The background will depends on the assumed photon spectra (§B.1) and thus on the parameters set of parameters $\theta $
$$B_{d,p,e}( \theta ) \equiv B(d,p,e) \qquad \textrm{with} \qquad
\theta \equiv ( \vec {\theta_1};\vec {\theta_2}, \cdots, \vec {\theta_{N_s}} )$$
The set of parameters $\theta $ are adjusted to minimized the function $f(\theta) $ ( $ \equiv \chi^2 $ ) through a non-linear minimization method.
The background
--------------
Given the parameters which describes the photon spectra and thus the photon spectrum of each sources $\widetilde{F}(j,e_{ph})$, the background is determined in each energy bin e. Given the best current background pattern $\widetilde{U}(d)$, the background is modeled as
$$B(d,p)=t(d,p) \widetilde{U}(d) A(p)$$
The current best background amplitude $\widetilde{A}(p)$ is determined exposure by exposure through the minimization with respect to A(p) of
$$\sum_{d=1}^{N_D} \frac{\big( N(d,p)-S(d,p)-A(p) t(d,p) \widetilde{U}(d) A(p) \big) ^2}{\sigma^2(d,p)}$$
where $N_D$ is the number of detector and $$\begin{array}{l}
S(d,p)=\sum_{j=1}^{N_s} \sum_{e_{ph}=1}^{N_{E_{ph}}} R(d,p,j,E,e_{ph}) (d,p) \widetilde{F}(j,e_{ph}) \\
\sigma^2(d,p) = N(d,p,e)
\end{array}$$
background uniformity pattern
-----------------------------
Having the sources parameters best estimate, the background uniformity pattern U(d) can be optimized as in §[A.2]{}
[^1]: Based on observations with INTEGRAL, an ESA project with instruments and science data centre funded by ESA member states (especially the PI countries: Denmark, France, Germany, Italy, Spain, and Switzerland), Czech Republic and Poland with participation of Russia and USA.
[^2]: Software and manual are available on the ISDC site (http://isdc.unige.ch/).
[^3]: available at http://isdc.unige.ch
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The fact, that $^4$He atoms on different concentric circular paths around the axis of a quantum vortex move with identically equal angular momentum, which represents an important aspect of superfluidity of He-II, has been used to discover a model which can explain the [*typical nature*]{} of experimentally observed $N$ (number of $^4He$ atoms) dependence of the rotational constant ($B$) of the rotor part of a cluster M:He$_N$. It reveals how exactly superfluidity is related to the said dependence of $B$ on $N$. We believe that this model, when used with simulation techniques, would render results that would agree closely with experiments.'
---
On the free rotation of a molecule embedded in helium-4 clusters
**Y. S. Jain$^1$ and S. Dey$^{1,2}$**
[**$^1$Department of Physics, North-Eastern Hill University, Shillong - 793022, India**]{}
[**$^2$Department of Basic Sciences, Assam Don-Bosco University, Guwahati - 781017, India**]{}
Key words : microscopic-superfluidity, He-clusters, He-nano-droplets.
PACS : 67.25. dw; 36.40.Mr; 36.40.-c
© by authors.
Ever since the spectroscopic study of SF$_6$ molecule embedded in superfluid helium-4 (He-II) was performed by Goyal [*et.al.*]{} [@goyal], rotational and vibrational dynamics of several molecules (say, OCS, CO$_2$, CO, N$_2$O, HCCCN, [*etc.*]{}, represented by M) embedded in bulk He-II and its nano-droplets and clusters of the form M:He$_N$ (where $N = 1, 2, 3, ...$ is the number of He atoms) have been extensively investigated. While, in a break through work, Grebenov [*et.al.*]{} [@grebe] found that OCS molecule embedded in $^4$He droplets, isolated in the normal phase of liquid $^3$He, rotates almost like a free rotor if the droplet has about 60 or more $^4$He atoms, systematic experimental study of M:He$_N$ clusters have demonstrated non-trivial dependence of their rotational constant $B$ (or moment of inertia, $I$) and vibrational frequency shift $\Delta\nu$ (of select modes of vibrations) with $N$ which concludes that superfluidity of $^4$He atoms has observable impact on $B$ of such small clusters as well [@mcke2; @mcke3; @mcke4; @surin; @topic]. A number of theoretical models, such as, (i) super-molecule model [@hartm], (ii) two fluid model [@grebe; @kwon], (iii) quantum hydrodynamic model [@calth], [*etc.*]{}, have been used to explain the initial observations. Simulation techniques have also been used, recently, to explain the phenomenon but with limited success at quantitative scale [@moron; @paesa; @paoli; @paes2; @blino; @zilli; @miur2; @miura]. It is not surprising because all these models associate the phenomenon with the superfluidity of $^4$He atoms which by itself is not clearly understood [@ander; @tom].
The conventional microscopic theory (CMT) [@ander; @tom] of a [*system of interacting bosons*]{} (SIB), such as liquid $^4$He, uses single particle basis (SPB) for its description. It considers that particles occupy different quantum states of a single particle placed in a box of volume V of the system and these states are described by plane waves ($u_k(r)$ = $A\exp{(i{\bf k}.{\bf r})}$ where symbols have their usual meaning); in other words a single particle represents the basic unit with an assumption that its momentum remains a good quantum number even in the superfluid state of the system. The theory concludes that: (1) the state of liquid $^4$He at a temperature ($T$) can be identified by the momentum distribution $N(p)$ of its particles where different number of particles $N_p(T)$ have different $p$, and (2) with the onset of superfluid transition, $N(p)$ does not change significantly except for the existence of a fraction of particles, $n_{p=0}(T) = N_{p=0}/N$ having $p=0$ in LT phase. Accordingly, even the ground state (G-state) of liquid $^4$He has different number of atoms in the states of different momenta, [*viz.*]{}, $N_{p=0}$ in $p$(=$\hbar k) = 0$ state and $N_{p\not=0}$ in several states of non-zero momenta, [**k**]{}$_1$, [**k**]{}$_2$, [**k**]{}$_3$, ... ... [*etc.*]{} (expressed in wave number). Based on different estimates by a large number of theoretical and experimental studies, $N_{p=0}$ ($p=0$ condensate) is believed to fall around 10$\%$ [@ander; @enss] leaving $N_{p\not=0}$ (non-condensate) to about 90$\%$. In what follows CMT identifies that $p=0$ condensate as the origin of superfluidity and related properties of He-II and the same view point is attributed to the superfluidity exhibited by microscopic systems of $^4$He atoms ([*viz.*]{}, droplets and clustes) too. The advances in CMT made over the last several years [@tom] suggest that superfluid phase of liquid $^4$He also has (in addition to $p=0$ condensate) pair condensate (similar to Copper pair condensation in superconductors) or a composite condensate ($p=0$ condensate, pair condensate, 3 particle condensate, [*etc.*]{}) as the origin of superfluidity.
We note that SPB used in CMT not only complicates the process of finding different aspects of a SIB, [*e.g.*]{}, the expectation value of inter-particle interaction which becomes infinitely repulsive for short distances, but also ignores the reality that the states of wave superposition assumed by the particles at low $T$ can not be described by $u_k(r)$. It is well known that: (i) two particles (say P1 and P2) in the state of their wave superposition are described by $\Psi{(1,2)} = [u_k(r)u_k(r) \pm
u_k(r)u_k(r)$)\] which basically represents a pair of particles moving with equal and opposite momenta ([**q**]{}, -[**q**]{}) with respect to their CM which moves with momentum [**K**]{} in the laboratory frame and (ii) positions ([**r**]{}$_1$, and [**r**]{}$_2$), momenta ([**k**]{}$_1$ and [**k**]{}$_2$) and energies (${\epsilon}_1$ and ${\epsilon}_2$) of two particles as separate entities lose their meaning in this state.
While superfluidity undoubtedly has a relation with the experimental observation of the rotation of a molecule embedded in a $^4He$ droplet and the non-trivial $N-$dependence of $B$ of M:$^4$He$_N$ cluster, it certainly has no relation with $p=0$ condensate because, as estblished unequivocally by one of us (Jain [@jain2]), $p=0$ condensate does not exist in the superfluid phase of a SIB. Starting with an assumption that the G-state of liquid $^4$He has non-zero values of both, $N_{p=0}$ and $N_{p\not=0}$, as concluded by CMT, Jain [@jain2] finds that: (i) such an $N(p)$ does not represent a state of minimum possible energy as expected for the G-state of every physical system, (ii) all particles in the true G-state of a SIB have identically equal enegy ($\varepsilon_o = h^2/8md^2$ with $h$ being the Planck constant, $m$ the mass of a particle and $d =
({\rm V}/N)^{1/3}$) and corresponding non-zero momentum ($q_o = \pi/d$), and (iii) the real form BEC that exists in a SIB is the macroscopic condensation of bosons as ([**q**]{}, -[**q**]{}) pairs in their G-state characterized by $q=q_o=\pi/d$ and $K=0$.
Motivated by all such observations, one of us [@jain1] used more realistic pair of particle basis [PPB]{} to conclude his non-conventional microscopic theory (NCMT) which emphasizes a pair as the basic unit of the system. The theory not only explains different properties of He-II at quantitative scale [@jain1; @simanta; @jain4] but also reveals that: (i) particles (G-state) of a SIB have identically equal energy ($\varepsilon_o$) and corresponding non-zero $q=q_o=\pi/d$ which agrees with a recent study by Jain [@jain1], (ii) they constitute a kind of close packed arrangement of their representative wave packets (CPA-WP) of identically equal size $\lambda/2 = h/2p = d$, (iii) they are allowed to move only coherently in order of their locations, obviously, with no relative motion and mutual collision, and (iv) they occupy phase ($\phi$) positions which differ by $\Delta\phi = 2n\pi$ (with $n=$ 1,2,3, ...). In addition, the theory finds that all the three characteristics of the G-state are retained by the superfluid phase over the entire range of temperature from $T = 0$ to $T_{\lambda}$ and the entire system assumes a kind of collective binding for which it behaves like a macroscopic molecule.
Since $^4$He atoms in M:He$_N$ clusters are confined to a space of few Å size, it is evident that each $^4$He atom has non-zero energy and corresponding non-zero momentum for the confinement. This undoubtedly proves the absence of $p=0$ condensate in these systems and we use these aspects of Jain’s NCMT [@jain1] to frame a model which provides a better account for the [*typical nature*]{} of non-trivial dependence of $B$ on $N$ revealed from experimental observations on selected M:He$_N$ clusters. In this context it may be mentioned that our intial efforts [@sam] tried to use some simple thoughts to explain the effect by presuming that: (i) each added atom which takes the cluster from M:He$_N$ to M:He$_{N+1}$ can significantly change the positions of other $^4$He atoms from the axis of rotation, and (ii) with $N$ increasing beyond its certain value (depending on several physico-chemical aspects of M), $^4$He atoms start occupying the second position from M ([*e.g.*]{} in M(zero)-$^4$He(first)-$^4$He(second)) and these atoms interact so weekly with the rotor-part of cluster (M and few $^4$He atoms, -at first position, which interact directly with M) that they do not follow the rotation of the rotor. In a sense the net potential seen by the rotor part of the cluster appears to remain constant with a change in the angular position of the rotor. In other words the rotor (when rotating about its axis) seems to role over a equi-potential surface; else if there are hills and valleys in the surface, the height of hills is much lower than the energy of rotational excitation of the rotor. Although these efforts rendered a satisfying account of the phenomenon, they could not find a clear relation to superfluidity of $^4$He atoms and this motivated us in concluding this model which not only associates an important aspect of superfluid He-II with the non-trivial dependence of $I$ on $N$ but also clarifies how only few $^4$He atoms attached directly to M take part in the rotation and rests do not follow the rotation effectively.
In what follows from the experimentally observed $N$ dependence of $B$ (represented typically by a curve depicted in Fig.(1)), we observe the following:
\(A) $B$ decreases when $N$ is increased from $N=N^i$ (the lowest $N$ for which experimental data are available) to certain $N = N^*$ that may depend on several factors like the size and symmetry of the structure of M, the strength of M-He interaction, [*etc*]{}.
\(B) It remains nearly constant when $N$ is increased beyond $N^*$ only by 1 or 2 or so but increases with further increase in $N$ up to another value, say $N_1$.
\(C) When $N$ is increased beyond $N_1$, $B$ is observed to decrease and increase alternately over different ranges of $N$, $N_1-N_2$, $N_2-N_3$, $N_3-N_4$, and $N_4-N_5$ and so on.
\(D) $\Delta\nu$ is observed first to increase linearly for first few $^4$He atoms (with $N < N^*$) but beyond this point it decreases with nearly a linear dependence on $N$; however, the slope of this decrease has different values over the ranges, $N^*-N_1$, $N_1-N_2$, $N_2-N_3$, $N_3-N_4$, and so on.
It is well known that $\Delta\nu$ is a simple consequence of a change in potential $V(Q_1, Q_2, ... ... Q_S)$ (governing all the $S$ possible vibrational modes of M) with changing $N$. It could be explained in terms of a small change in the related potential constant appearing in the harmonic component in the expansion of $V(Q_1, Q_2, ... ... Q_S)$. Although, it is difficult to argue in favour of increase or decrease in the value of potential constant of the chosen vibration, yet, however, a simple logic indicates that $\nu$ should increase for first few $^4$He atoms which occupy position in the closest vicinity of M since He-atoms are saturated with the electron charge density for which they would give away a small fraction of their own electron density to M which should strengthen the forces that govern its different modes of vibration and this is corroborated by experimental observation. Although, with $^4$He atoms occupying second or third, ... positions counting from M (at zero-th position), it is difficult to argue whether $\nu$ would increase or decrease with $N$, however, it is clear the effect on $\nu$ should decrease with each added atom and this expectation agrees with decrease in slope of $\Delta\nu$ [*vs*]{} $N$ observed experimentally. In what follows from these points, the change in $\Delta\nu$ with $N$ has nothing to do with microscopic superfluidity of the $^4$He atoms in the cluster. Hence, in this paper, we simply concentrate to find the origin of (A), (B) and (C) and conclude a general model of the phenomenon.
[**1**]{}. To a good approximation, the experimental observation of decrease in $B$ (or increase in $I$) for $N^i$ to $N^*$ can be explained by using rigid rotor picture for the cluster since this falls in line with the fact that a $^4$He atom interacts more strongly with M than with another $^4$He atom, the structure of the cluster M:He$_N$ for $N=N^i$ to $N=N^*$ can, therefore, be presumed to have a rigid rotor structure for the first few rotational excitations of each cluster. We note that $N^*$ can be different for different M (depending on its physico-chemical nature), while $N^i$ can, in principle, be as small as 1.
[**2**]{}. However, the non-trivial dependence of $B$ on $N$ represented by the observations that B remains nearly unchanged when $N$ changes by one or two $^4$He-atoms beyond $N^*$ and there after it follows cycles of its increase from $N^*+\Delta N$ to $N_1$, $N_2$ to $N_3$, ... and decrease from $N_1$ to $N_2$, $N_3$ to $N_4$, ... . This unexpected observation, naturally indicates its relation with superfluidity of $^4$He-atoms. Consequently, we try to explain it interms of an important aspect of superfluid $^4$He exhibited by it ubder the influence of its rotation.
The fact, that different atoms on different concentric circles around the axis of a quantum vortex in He-II move coherently in order of their locations in a manner that they have no difference in their angular momentum [@wilks] implies that their angular velocity ($\omega$) changes as $r^{-2}$ where $r$ is the distance of the atom from the axis of the vortex. This differs from atoms of a rigid body rotor where all atoms move around the axis of rotation with identically equal $\omega$. In the following we consider the example of a set of $^4$He atoms moving on two concentric circles around the axis of rotation (as shown in Fig.2A) under the condition of: (i) constant angular velocity and (ii) constant angular momentum. To this effect we evaluate the kinetic energy of the set by using, $$E = \sum_i^6\frac{n_i}{2}mr_1^2\omega_1^2 +
\sum_j^{12}\frac{n_j}{2}mr_2^2\omega_2^2, \eqno(1)$$
where we have $n_i = 1, 2, 3, ...6$ and $n_j = 1, 2, 3, ...12$ with indices $i$ and $j$ to identify different atoms on orbits 1 and 2, respectively. Presuming that the radii of orbits 1 and 2 satisfy $$r_2 = 2r_1 = 2r_o \eqno(2)$$
and all atoms move as a single rigid body with $$\omega_1 = \omega_2 = \omega_o, \eqno(3)$$
we have
$$E = \frac{1}{2}\left[6mr_o^2\omega_o^2 +
48mr_o^2\omega_o^2\right] = \frac{1}{2}\left[54mr_o^2\omega_o^2\right]
= \frac{1}{2}I\omega_o^2 \quad \quad {\rm or} \quad I = 54mr_o^2. \eqno(4)$$
Since the condition of constant angular momentum, applied to two atoms moving on different orbits of radius $r_1$ and $r_2$ ([*cf.*]{}, Fig.2B) renders $$mr_1^2\omega_1 = mr_2^2\omega_2 = C \quad ({\rm constant})
\quad \quad {\rm or} \quad
\omega \propto r^{-2}, \eqno(5)$$
which implies that corresponding linear velocity $v = r\omega$ changes as $r^{-1}$. As expected, this agrees with the well known dependence of $v$ on $r$ in a quantum vortex observed in superfluid $^4$He [@wilks]. Using Eqn.(5) in Eqn.(1), we get
$$E = \frac{1}{2}\left[6mr_1^2\omega_1^2 +
12mr_2^2\omega_1^2\left(\frac{r_1}{r_2}\right)^4\right] \eqno(6)$$
which for the orbits satisfying Eqn.(2) and $\omega_1 = \omega_o$ renders
$$E = \frac{1}{2}\left[6mr_o^2\omega_o^2 +
12mr_o^2\omega_o^2\frac{1}{4}\right] \quad =
\frac{1}{2}\left[9mr_o^2\omega_o^2\right]; \quad \quad
{\rm with} \quad I = 9mr_o^2. \eqno(7)$$
Eqn.(6) clearly reveals that the contribution to $I$ from an atom at a distance $r_2$, under the condition of constant angular momentum (Eqn. 5), gets reduced by a significant factor of $(r_1/r_2)^4$ (since $r_1 < r_2$) in comparison to that found under the condition of constant angular velocity (Eqn.(4)). As an example, the contribution to $I$ by an atom added to an orbit satisfying $r_2 = 2r_1$, Eqn.(2) gets reduced to a value as low as 1/16 ([*i.e.*]{} $\approx$6.3$\%$) and for the orbit satisfying $r_2 = 3r_1$, the said contribution is as low as 1/81 (or $\approx$1.3$\%$). This speaks of the smallness of the contribution of an added atom to the $I$ of the cluster when it goes to an orbit of higher $r$ and evinces that the said atom has a $+ve$ contribution indicating that $I_{N+1} > I_N$.
In what follows the above stated inferences, we can expect a small increase or almost no change in $I$ for each added atom to the cluster with $N=N^*$. However, it gives no clue for the experimental observations of $I_{N+1} < I_N$ for $N > N^*$. To this effect our critical thinking reveals that the phenomenon is possible only if the added atom reduces the distance of $^4$He atoms (all the $N$ atoms or a few of them) in M:$^4$He$_N$ cluster from the axis of rotation to an extent that contribution to $I$ by the added single atom is over compensated by the decrease in $I_N$. It is also possible if the added atom transforms the structure of M:$^4$He$_N$ cluster in a manner that one atom from first orbit of radius $r_1$ moves to the orbit of radius $r_2$ of the added atom (as shown in Fig.2C); in this case the net change $\Delta I = I_{N+1} - I_N$ becomes $$\Delta I = 2mr_2^2\frac{r_1^4}{r_2^4} - mr_1^2 =
mr_1^2\left[2\frac{r_1^2}{r_2^2} -1\right] \eqno(8)$$
which assumes a $-ve$ value for $r_2 > \sqrt{2}r_1$ indicating that $I_{N+1} < I_N$ when $r_2 > \sqrt{2}r_1$. Using this possibility for $N$ increasing beyond $N^*$ by 1 atom, we have
$$\Delta I = -0.5mr_o^2\eqno(9)$$
by using Eqns.(2) and (8). Presuming further that another atom moves similarly from orbit-1 to orbit-2 when an added atom to the cluster occupies orbit-2 as shown in Fig.2D, we have
$$\Delta I = 4mr_2^2\frac{r_1^4}{r_2^4} - 2mr_1^2 =
mr_1^2\left[4\frac{r_1^2}{r_2^2} -2\right] = - 1.0mr_o^2 \eqno(10)$$
for orbits satisfying Eqn.(2). Generalizing Eqns.(8) and Eqn.(10), we have
$$\Delta I = mr_1^2\left[2n_c\frac{r_1^2}{r_2^2} - n_c\right]
= mr_o^2\left[\frac{2n_c}{\alpha^2} -n_c\right] \eqno(11)$$
which represents the change in $I$ when $n_c$ atoms (above $N^*$) added to orbit-2 (making total $N = N^* + n_c$) induce $n_c$ atoms from orbit-1 to jump to orbit-2. Eqn.(11) reveals that $I$ has no change if $\alpha (=r_2/r_1) = \sqrt{2}$, it decreases by $\Delta I = -0.5n_cmr_o^2$ for $r_2 = 2r_1 = 2r_o$ ([*i.e.*]{}, $\alpha = 2$) and by a maximum of $\Delta I = -n_cmr_o^2$ for $r_2 > > r_1 (= r_o)$. Such changes in $I$ for $n_c = 1, 2, 3, ...$ for different $\alpha = r_2/r_1$ are depicted in Fig.3 for their better perception.
For a possible situation where no atom jumps from orbit-1 to orbit-2 when $(n_c+1)$-th atom is added to orbit-2, it is evident that the added atom increases $I$ by $mr_o^2/\alpha^2$. We have $$\Delta I = mr_o^2\left[\frac{2n_c}{\alpha^2} -n_c\right]
+ mr_o^2\frac{1}{\alpha^2}. \eqno(12)$$
which again means $I_{N+1} > I_N$ provided the added atom makes no change in $\alpha$. However, if it does and changes in $r_2$ and $r_1$ are such that $\alpha$ increases to $\alpha^* = \alpha +
\Delta\alpha$, then by using Eqn.(12), we find $$\delta{(\Delta{I})} = \Delta I(\alpha^*) - \Delta I(\alpha)
= -mr_o^2\frac{2(2n_c+1)\Delta\alpha}{\alpha^3}. \eqno(13)$$
We note that this $-ve$ change in $I$ can overcompensate the increase in $I$ by $mr_o^2/\alpha^2$ if
$$\Delta\alpha \ge \frac{\alpha}{2(2n_c+1)} \quad \left(=
\frac{1}{2(2n_c+1)}\frac{r_2}{r_1}\right)\eqno(14)$$
which is obtained by equating RHS of Eqn.(13) to $ mr_o^2/{\alpha^2}$. This indicates that $I$ of the cluster can have continuous decrease with increase in $N$ (possibly from $N^*$ to $N_1$) if $\alpha = r_2/r_1$ increases by an appropriate value of $\Delta\alpha$ with each added atom. In other words an agreement between theory and experiment can be seen by using $n_c$, $\alpha$, and $\Delta\alpha$ as adjustable parameters. Note that increase in $\alpha$ is possible both by decrease in $r_1$ and increase in $r_2$ when an atom is added to the cluster. However, it appears that desired increase in $r_2$ is more probable than decrease in $r_1$. As revealed by Eqn.(14), decrease in $I$ is possible if $\Delta\alpha$/$\alpha$ increases by more than 16.6$\%$, 10$\%$ and 7.1$\%$, respectively, in case of $n_c =$ 1. 2, and 3. These aspects are depicted in Fig.4 for their better understanding.
Summing up the possible explanation for the phenomenon in the light of our preceding analysis, we may mention that :
\(1) With increasing $N$ from $N^i$ to $N^*$, $^4$He atoms in M:$^4$He$_N$ cluster (for $N \le N^*$) seem to have reasonably strong binding with M for which the cluster as whole represents a rigid rotor, to a good approximation, and its $I$ increases (or corresponding $B$ decreases) with $N$ in agreement with experiments ([*cf.*]{} Fig.1, for $N \le N^*$).
\(2) For nearly no change in $I_{N}$ from $I_{N^*}$, when $N$ is set to have a value $N^* + \Delta N$ (where $\Delta N$ has only small value such as 1 or 2 or so), it appears that each of the $\Delta N$ atoms go to orbit 2 for which $\Delta I (= I_{N^*+1}- I_{N^*}$) is $\approx$ 6.3$\%$ of the contribution of an atom in orbit 1. Such a small increase can be easily compensated if the contribution to $I_{N^*+1}$ due to each of the $N^*$ atoms in orbit 1 gets reduced by 6.3/$N^*\%$ presumably due to small decrease, $\Delta r$, in the distance of atoms from the axis of rotation of M:$^4$He$_{N^*}$ and this does not demand necessarily a decrease in M-$^4$He bond length; a decrease by $\Delta r$ in the projection of the bond on the plane $\perp$ to the axis of rotation would suffice and this can be estimated to fall around 6.3/2$N^*\%$ of the said projection which equals to $\approx 0.8\%$ if $N^*= 4$, or 0.5$\%$ if $N^* = 6$ or $\approx 0.4\%$ if $N^* = 8$. Such a small change can easily be expected as a possible effect of an atom added to M:$^4$He$_{N^*}$ or another added to M:$^4$He$_{N^*+1}$; this naturally explains the said observation.
\(3) For the remaining part of $I_N$ [*vs.*]{} $N$ curve ([*i.e.*]{}, for $N > N^*+\Delta N$) where $I_N$ is observed to have significant decrease with increase in $N$, changes in $\alpha = r_2/r_1$ along with the jump of an atom from orbit 1 to orbit 2 seem to take place when an atom is added to orbit 2. Depending on the physico-chemical nature and size of M, the decrease in $I_N$ for $N$ changing from $N=N^*+\Delta N$ to $N=N_1$ can be explained by choosing $n_c =1$ and 2 ... (in different steps) clubbed with appropriate values of $\alpha$ and $\Delta\alpha$. This speaks of the sensitivity of the changes in $I_N$ on $\alpha$ and $\Delta\alpha$ as well as $\Delta r$ (change in the said projection of M-$^4$He bond length). This naturally simplifies the basis our understanding of the observed increase and decrease in $I$ (Fig.1) with $N$. (4) The decrease in $B$ (increase in $I$) for $N$ increasing from $N_1$ to $N_2$ is as per normal expectation. However, each atom added to the cluster in this range contributes only very small fraction of the contribution to $I$ by an atom in orbit 1. Guided by this fact it appears that value of $\alpha$, $\Delta\alpha$ and $\Delta r$ should explain the $N-$ dependence of $I_N$ not only in this range but for all values of $N > N_1$.
Identifying the ranges, $0- N_1$, $N_1-N_2$, $N_2-N_3$, [*etc.*]{} (Fig. 1) as cycles of inccrease and decrease in $I_N$, it appears that M:$^4$He$_N$ cluster has different shells of $^4He$ atoms around M and each cycle represents the completion of one shell. While $^4$He atoms in the first shell have direct bond with M, those in second, third, ... shells are separated from M, respectively, by 1, 2, ... $^4$He atoms in between. The maximum number of $^4He$ atoms, in a particular shell increases in proportion of $R^2$ (where $R$ is the radius of the shell which changes from one shell to next shell in units of the diameter of the sphere which represents the shape and size of $^4$He atom; however it also depends on the shape and size of M. Assuming that M has a shape and size of a $^4$He atom, a rough estimate reveals that the number of $^4$He atoms in first, second and third shells, should be [*around*]{} 6, 18 and 40, respectively. However, it may be emphasized that these numbers agree approximately with experimental values because M may have linear or a complex structure. Further, it may also be mentioned that it is not the length of M-$^4$He bond which changes much with atoms added to the cluster but the projection of this bond on the plane $\perp$ to the axis of the rotor which should be considered to explain the changes in $I_N$ with added $^4$He atoms that we observe through experiments.
The experimental observations seem to indicate that the part of M:$^4$He$_N$ cluster which rotates, to a good approximation as a rigid rotor, has fewer than $N^*$ $^4$He atoms in clusters of $N >> N^*$. This agrees with our suggestion that $n_c$ (= 1 or 2, or so) atoms move from orbit 1 to orbit 2 with increasing $N$ beyond $N^*$. However, the physics of this possibility is not yet very clear.
In the light of the fact that $^4$He atoms in superfluid state make a close packed arrangement of their wave packets (CPA-WP) supported by a number of experimental observations such as the observation of Stark effect of roton transition seen through microwave absorption [@jains] and the unequivocal conclusions of a number of theoretical studies [@jain2; @jain1; @simanta], the rotor part of the cluster in CPA-WP type arrangement of $^4$He atoms may in certain cases experience a kind of low energy potential barrier with an axial symmetry of the order $n$ (as shown in Fig.5). A theoretical analysis for such a case [@jain3] reveals that the effective $I$ of the rotor has a lower value that depends on the height and symmetry of $V_n$ ([*cf.*]{}, Fig. 5). This renders an additional reason for a small decrease in $I$.
As concluded by Jain’s NCMT of a system of interacting bosons such as liquid $^4$He [@jain1], atoms in He-II not only move coherently in order of their locations but also have identically equal angular momentum when they move on different concentric paths around the axis of a quantum vortex; in fact this theory for the first time answers a question raised by Wilks [@wilks] in relation to Feynman’s account for the origin of quantized circulation [@feyn1]. Wilks has rightly argued that Feynman’s basis for quantum vortices in He-II is equally valid for He-I but the latter does not show any quantum vortex. Using a basic aspect of quantum vortices observed in He-II, we discover a model which has enough potential to explain qualitatively the typical nature of experimentally observed $N$ dependence of the rotational constant $B$ of the rotor part of the cluster M:He$_N$. Naturally, the question, how exactly superfluidity is related to the said dependence of $B$ is answered with utmost clarity. We hope that this model, when used with simulation techniques on individual cluster, would render results that would agree closely with experiments. This would not only help in improving the model but also for having a clear understanding of the phenomenon. However, we could not take up this task for the want of facilities of computer simulations at our end.
[35]{} S. Goyal, D. L. Schutt and G. Scoles, Phys. Rev. Lett. 69, 933 (1992). S. Grebenev, J. P. Toennies and A. F. Vilesov, Science 279, 2083 (1998). A. R. W. Mckellar, Y. Xu and W. Jager, Phys. Rev. Lett. 97, 183401 (2006). A. R. W. Mckellar, J. Chem. Phys. 127, 044315 (2007).
A. R. W. McKellar. J. Chem. Phys., 128, 044308 (2008). L. A. Surin, A. V. Potapov, B. S. Dumesh, S. Schlemmer, Y. Xu, P. L. Raston and W. Jager. Phys. Rev. Lett., 101, 233401 (2008). W. Topic, W. Jager, N. Blinov, P. N. Roy, M. Botti and S. Moroni. J. Chem. Phys., 125, 144310 (2006). M. Hartmann, R. E. Miller, J. P. Toennies and A. F. Vilesov, Phys. Rev. Lett. 75, 1566 (1995).
\(a) Y. Kwon, P. Huang, M. V. Patel, D. Blume and K. B. Whaley. J. Chem. Phys., 113, 6469 (2000).
\(b) Y Kwon and K.B. Whaley, J. Chem. Phys.119, 1986 (2003) 10p
C. Callegari, Molecular Rotation and Dynamics in Superfluid $^4$He Nanodroplets, Ph.D thesis, Princeton University, 2000.
S. Moroni, A. Sarsa, S. Fantoni, K. E. Schmidth and S. Baroni. Phys. Rev. Lett., 90, 143401 (2003).
F. Paesani, A. Viel, F. A. Gianturco and K. B. Whaley. Phys. Rev. Lett., 90, 073401 (2003).
S. Paolini, S. Fantoni, S. Moroni and S. Baroni. J. Chem. Phys., 123, 114306 (2005). F. Paesani and K. B. Whaley. J. Chem. Phys., 121, 4180 (2004). N. Blinov, X. Song and P. N. Roy. J. Chem. Phys., 120, 5916 (2004). R. E. Zillich, F. Paesani, Y. Kwon and K. B. Whaley. J. Chem. Phys., 123, 14301 (2005). S. Miura. J. Phys., 17, S3259 (2005). S. Miura. J. Chem. Phys., 126, 114309 (2007), and references therein .
J.O. Anderson, Rev. Mod. Phys. 76, 599(2004).
M.D. Tomchenko, Asian J. Phys. 18, 245-254 (2009); Uneven horizon or several words about the superfluid $^4$He theory, arXiv:0904.4434.
C. Enss and S. Hunklinger, Low Temperature Physics, Springer Verlag, Berlin, (2005).
Y.S. Jain, [*The p=0 condensate is a myth*]{}, arxiv:cond-mat/1008.240v2
\(a) Y.S. Jain, Ind. J. Phys. 79, 1009-14 (2005); more detailed theory is available in (b).
(b)Y.S. Jain, Macro-orbitals and microscopic theory of a system of interacting bosons, arXiv:cond-mat/0606571; a most recently edited version of this paper which discusses several other aspects of the system (submitted for publication) is available as pdf file with the author.
S. Chutia, Ph.D. Thesis, Department of Physics, NEHU, Shillong-793022, India
Y.S. Jain, [*A study of thermal excitations of liquid helium-4 using macro-orbital microscopic theory*]{}, arxiv:cond-mat/0609418.
S Dey, J P Gewali, A K Jha, L Chhaigte and Y S Jain, Ind. J. Phys, Ind. J. Phys. Vol. 85, 1309-1330 (2011).
J. Wilks, The Properties of Liquid and Solid Helium, Clarendon Press, Oxford, 1967. Y.S. Jain, L. Chhangte, S. Chutia and S. Dey, Current Science 101, 769-775 (2011).
Y. S. Jain, B. Singh and B. N. Khanna. Pramana,-the J. Phys. 18, 511 (1982). R.P. Feynman, Applications of Quantum Mechanics to Liquid Helium, Progress in Low Temperature Physics (C.J. Gorter, editor). 1 17-53 (1954), North-Holland, Amsterdam Chapter II
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate here a supermatrix model with a mass term and a cubic interaction. It is based on the Lie superalgebra $\mathfrak{osp}(1|32,\mathbb{R})$, which could play a rôle in the construction of the eleven-dimensional M-theory. This model contains a massive version of the IIB matrix model, where some fields have a tachyonic mass term. Therefore, the trivial vacuum of this theory is unstable. However, this model possesses several classical solutions where these fields build noncommutative curved spaces and these solutions are shown to be energetically more favorable than the trivial vacuum. In particular, we describe in details two cases, the $SO(3) \times SO(3) \times SO(3)$ (three fuzzy 2-spheres) and the $SO(9)$ (fuzzy 8-sphere) classical backgrounds.'
---
$\mbox{ }$
----------------
KUNS-1796
NEIP-02-006
hep-th/0209057
Sep 2002
----------------
\
\
\
Curved-space classical solutions of a massive supermatrix model
Takehiro A[zuma]{}$^{\dag}$[^1] and Maxime B[agnoud]{}$^{\ddag}$\
$^{\dag}$[*Department of Physics, Kyoto University, Kyoto 606-8502, Japan*]{}\
$^{\ddag}$ [*Institut de Physique, Université de Neuchâtel, CH-2000 Neuchâtel, Switzerland*]{}
Introduction
============
Despite the fact that perturbative superstring theory provides us with a consistent unified theory of fundamental interactions, we still lack a completely satisfactory nonperturbative formulation of superstring theory. As a consequence, although this theory has a plethora of possible vacua (the dynamics in some of these vacua has already been studied in details), there is no way to select the true vacuum of the theory and compare the physical implications of superstring theory with known phenomenological data. It is thus instrumental to find a constructive definition of superstring theory in order to predict the real world or/and falsify the theory.
One of the successful proposals for a constructive definition of superstring theory[@9610043; @9612115; @9703030; @9708123] is a formulation through a large $N$ reduced model. A candidate model of this kind is the so-called IIB matrix model[@9612115; @9705128], which is defined by the following action:[^2] $$\begin{aligned}
S = \frac{1}{g^{2}} Tr \left( \frac{1}{4} [A_{\mu},
A_{\nu}][A^{\mu}, A^{\nu}] + \frac{1}{2} {\bar \psi} \Gamma^{\mu}
[A_{\mu}, \psi] \right), \label{AZ1IKKT}
\end{aligned}$$ where the indices $\mu, \nu, \cdots$ run over the 10-dimensional Minkowskian spacetime. It is a large $N$ reduced model[@EguchiKawai; @TEK; @TEK2] of 10-dimensional ${\cal N} =1$ super Yang-Mills theory with $U(N)$ gauge symmetry. Here, $\psi$ is a 10-dimensional Majorana-Weyl spinor field, and $A_{\mu}$ and $\psi$ are $N \times N$ Hermitian matrices. The IIB matrix model has ${\cal N}=2$ supersymmetry, which exhibits a particular structure that allows us to interpret the eigenvalues of the large $N$ matrices describing the bosonic fields as space-time coordinates[@9802085; @9903217] (IIB matrix model is extensively reviewed in [@9908038]).
Another intriguing attempt for a constructive definition of superstring theory is a background-independent matrix model based on the Lie superalgebra $\mathfrak{osp}(1|32,\mathbb{R})
$[@0002009; @0006137; @0102168; @0103003; @0201183]. It is a natural generalization of the IIB matrix model, in which both bosons and fermions are unified into a single supermultiplet. $\mathfrak{osp}(1|32,\mathbb{R})$ has been known as the unique maximal simple Lie superalgebra with 32 fermionic generators[@0003261]. In a 10-dimensional representation, the smallest irreducible spinors are the 16-component chiral spinors, so that the 32 fermionic generators can be decomposed in two chiral spinors of equal or opposite chiralities. The former and the latter respectively correspond to the type IIA and IIB superstring theories. In this sense, we speculate that the IIB matrix model would be extracted from this supermatrix model by integrating out some degrees of freedom. In the papers [@0102168; @0103003], it has been attempted to clarify the relation between a purely cubic $\mathfrak{osp}(1|32,\mathbb{R})$ supermatrix model and the IIB matrix model by paying particular attention to the structure of the supersymmetry algebra. In the paper [@0201183], such a cubic supermatrix model supplemented by a mass term has been investigated to elucidate its relation with the BFSS matrix model[@9610043].
Since large $N$ reduced models are expected to be eligible frameworks to describe gravitational interactions, it is essential to have the possibility of describing curved spacetimes manifestly in their framework. As the IIB matrix model only possesses flat noncommutative spacetime as a classical solution (the same holds true of the non-gauged $\mathfrak{osp}(1|32,\mathbb{R})$ supermatrix model[@0102168]), it is impossible to study perturbations around curved backgrounds.
Some generalization is thus necessary in order to overcome this difficulty. A possible approach is to identify large $N$ matrices with differential operators[@0102168; @0201129; @0202138; @0204078]. Large $N$ matrices have both aspects of differential operators and spacetime coordinates. The former appears clearly in the twisted Eguchi-Kawai model[@TEK; @TEK2] while the latter is the essential feature of the IIB matrix model. These two aspects can be related by expanding the IIB matrix model around its flat noncommutative background[@9908141]. This bilateral character is interpreted as the T-duality of string theory. The advantage of identifying matrices with differential operators lies in the fact that differential operators act on fields on a curved spacetime in a natural way.
Another approach is to consider a matrix model which has some curved space as a classical solution, so that it becomes possible to perform perturbations around this curved background. To achieve this, some modification of the IIB matrix model[@0101102; @0103192; @0204256; @0207115] is needed. In [@0101102], a Chern-Simons term has been added to the IIB matrix action to construct a noncommutative gauge theory on the $SO(3)$ fuzzy sphere[@Madore]. Another possible alteration is the addition of a tachyonic mass term to the bosonic part of the IIB matrix model[@0103192]: $$\begin{aligned}
\label{MIIBMM}
S = \frac{1}{g^{2}} Tr \left( \frac{1}{4} [A_{\alpha}, A_{\beta}]
[A_{\alpha}, A_{\beta}] + \lambda^{2} A_{\alpha} A_{\alpha} \right),
\end{aligned}$$ where the indices $\alpha, \beta$ run over $\alpha, \beta = 1, 2, 3,
4$. The indices are contracted with respect to the four-dimensional Euclidean space metric, and the model has $SO(4)$ global symmetry. Its equations of motion $$\begin{aligned}
[A_{\beta}, [A_{\alpha}, A_{\beta}]] + 2 \lambda^{2} A_{\alpha} = 0
\end{aligned}$$ have classical solutions given by a set of fields satisfying some Lie algebra. Thus, such a massive IIB matrix model can be expanded around various curved spaces. In [@0103192], expansions around the two-dimensional fuzzy sphere and the two-dimensional fuzzy torus have been studied.
In this paper, we take this latter approach in order to describe a curved background spacetime by considering an $\mathfrak{osp}(1|32,\mathbb{R})$ supermatrix model with a mass term. We analyze how the massive supermatrix model incorporates the non-commutative curved-space classical solutions.
This paper is organized as follows: In Section 2, we give a brief review of the Lie superalgebra $\mathfrak{osp}(1|32,\mathbb{R})$ and the associated supermatrix model. In Section 3, we suggest an ansatz that allows us to solve the equations of motion of the massive supermatrix model and leads to solutions of the fuzzy-sphere type. We describe in detail two of these solutions, one exhibiting $SO(3) \times SO(3) \times SO(3)$ symmetry and the other exhibiting $SO(9)$ symmetry and compare their stability properties. This leads us to a more general discussion of a possible brane nucleation process in such totally reduced matrix models. Then, we make a few remarks on the structure of the supersymmetry transformations in our model. Finally, we summarize the results presented in this work in section 4 and indicate there a few directions for future research on this topic.
$\mathfrak{osp}(1|32,\mathbb{R})$ supermatrix model with a mass term
====================================================================
L. Smolin proposed a cubic matrix model[@0002009; @0006137] based on the Lie superalgebra $\mathfrak{osp}(1|32,\mathbb{R})$. The action is constructed from a matrix $M$ belonging to $\mathfrak{osp}(1|32,\mathbb{R})$, whose entries are promoted to large $N$ Hermitian matrices. In this paper, we follow the notation of [@0103003], in which details about $\mathfrak{osp}(1|32,\mathbb{R})$ are given. Its relation with the 11-dimensional super-Poincaré algebra is described in [@0003261; @0201183].
The (even part) of the Lie superalgebra $\mathfrak{osp}(1|32,\mathbb{R})$ is constituted from matrices of the form: $$\begin{aligned}
M = \left( \begin{array}{cc} m & \psi \\ i {\bar \psi} & 0
\end{array} \right),
\end{aligned}$$ where $\psi$ is a 32-component Majorana spinor and $m$ belongs to the Lie algebra $\mathfrak{sp}(32,\mathbb{R})$. We take the metric of the 10-dimensional Minkowskian spacetime as $$\begin{aligned}
\eta^{\mu \nu} = \textrm{diag}(-1, +1, \cdots, +1).
\end{aligned}$$
Action
------
Here, we consider an $\mathfrak{osp}(1|32,\mathbb{R})$ supermatrix model with a mass term included, expecting similarities with the massive IIB matrix model studied in [@0103192]. We consider the following action, with a mass term added to the pure cubic $\mathfrak{osp}(1|32,\mathbb{R})$ supermatrix model: $$\begin{aligned}
S &=& Tr_{\mathfrak{u}(N)} \left[ str_{\mathfrak{osp}(1|32)} \left(
- 3 \mu M^{2} + i M [M, M]
\right) \right] \nonumber \\
&=& Tr_{\mathfrak{u}(N)} \left[ 3 \mu (- {m_{p}}^{q}{m_{q}}^{p}
+ 2i {\bar \psi} \psi) + i \left( {m_{p}}^{q}
[ {m_{q}}^{r}, {m_{r}}^{p} ] - 3 i {\bar \psi}^{p}
[ {m_{p}}^{q}, \psi^{q}] \right) \right]. \label{AZaction}
\end{aligned}$$ where $p, q, r, \cdots = 1,\ldots, 32$. In this model, each element of the $\mathfrak{osp}(1|32,\mathbb{R})$ supermatrices is promoted to an $N \times N$ Hermitian matrix. This action is invariant under $U(N)$ gauge transformations and $OSp(1|32,\mathbb{R})$ orthosymplectic transformations, and these two symmetries are decoupled, since they do not act on the same indices. Since we want to consider this model in a 10-dimensional spacetime context, we decompose the bosonic part $m$ as follows: $$\begin{aligned}
m = W \Gamma^{\sharp} + A_{\mu} \Gamma^{\mu} + B_{\mu} \Gamma^{\mu \sharp}
+ \frac{1}{2!} C_{\mu_{1} \mu_{2}} \Gamma^{\mu_{1} \mu_{2}}
+ \frac{1}{4!} H_{\mu_{1} \cdots \mu_{4}} \Gamma^{\mu_{1} \cdots
\mu_{4} \sharp}
+ \frac{1}{5!} Z_{\mu_{1} \cdots \mu_{5}} \Gamma^{\mu_{1} \cdots \mu_{5}},
\end{aligned}$$ where $\mu_i=0,\ldots ,9$ and $\Gamma^{\sharp}$ is the chirality operator. Then, the relevant part of the action (\[AZaction\]) is expressed as (writing simply $Tr$ instead of $Tr_{\mathfrak{u}(N)}$ from now on): $$\begin{aligned}
S &=& 96 \mu Tr \left( - W^{2} - A_{\mu} A^{\mu} + B_{\mu} B^{\mu}
+ \frac{1}{2} C_{\mu_{1} \mu_{2}} C^{\mu_{1} \mu_{2}}
- \frac{1}{4!} H_{\mu_{1} \cdots \mu_{4}} H^{\mu_{1} \cdots
\mu_{4}} - \right. \nonumber \\
&-& \left. \frac{1}{5!} Z_{\mu_{1} \cdots \mu_{5}} Z^{\mu_{1} \cdots
\mu_{5}} + \frac{i}{16} {\bar \psi} \psi \right)
+ 32 i Tr \Bigg( 3 C_{\mu_{1} \mu_{2}} [B^{\mu_{1}}, B^{\mu_{2}}]
+ C_{\mu_{1} \mu_{2}} [ {C^{\mu_{2}}}_{\mu_{3}}, C^{\mu_{3} \mu_{1}}]
\Bigg) + \nonumber \\
&+& \textrm{ cubic interactions involving } (W,A,H,Z,\psi), \label{AZaction3}
\end{aligned}$$ while the full result can be found in the first appendix and the detailed computation in [@0103003]. In the purely cubic supermatrix model (without mass term, which has been studied in [@0002009; @0006137; @0102168; @0103003]), the rank-2 field $C_{\mu_{1} \mu_{2}}$ possesses a cubic interaction term but has no quadratic term. This has been a severe obstacle to the appearance of a Yang-Mills-like structure in the supermatrix model, because it has been impossible to identify $C_{\mu_{1} \mu_{2}}$ with the commutators of the rank-1 fields $[B_{\mu_{1}}, B_{\mu_{2}}]$ (or $[A_{\mu_{1}},
A_{\mu_{2}}]$). In the 11-dimensional case, this difficulty has been overcome in [@0201183] through the addition of a mass term, and we thus expect this model to contain the massive IIB matrix model, the bosonic part of which has been studied in [@0103192] to investigate perturbation theory around noncommutative curved-space backgrounds.
Resolution of the equations of motion
=====================================
We proceed to search for possible curved-space classical configurations solving the equations of motion that follow from the action (\[AZaction3\]). To get a clearer picture of the problem, we now set the fermions and the positive squared-mass bosonic fields to zero: $$\begin{aligned}
\psi = W = A_{\mu} = H_{\mu_{1} \cdots \mu_{4}} = Z_{\mu_{1} \cdots
\mu_{5}} = 0. \label{AZ2trivial}
\end{aligned}$$ Since their masses are positive (at least in the spatial directions, while the time-like direction of quantum fields is generally unphysical), (\[AZ2trivial\]) is a stable classical solution. Furthermore, we choose to identify the tachyonic 10-dimensional vector field $B_{\mu}$, rather than the well-defined $A_{\mu}$ with the bosonic fields of the massive IIB matrix model, in order to obtain a possibly stable curved-space classical solution. The classical equations of motion for the remaining tachyonic fields $B_{\mu}$ and $C_{\mu \nu}$ following from (\[AZaction3\]) are $$\begin{aligned}
B_{\mu} &=& - i \mu^{-1} [B^{\nu}, C_{\mu \nu}], \label{AZeqB+} \\
C_{\mu \nu} &=& - i \mu^{-1} ( [B_{\mu}, B_{\nu}]
+ [ {C_{\mu}}^{\rho}, C_{\nu \rho}] ). \label{AZeqC+}
\end{aligned}$$ Although it is difficult to solve these equations in full generality, the equation of motion for $C_{\mu \nu}$ suggests to take $C_{\mu \nu} \propto [B_{\mu}, B_{\nu}]$ for $B_{\mu}$’s satisfying a fairly simple commutator algebra. If we look for objects having a clear geometrical interpretation, it is tempting to look for solutions building fuzzy spheres.
$SO(3) \times SO(3) \times SO(3)$ classical solution
----------------------------------------------------
The simplest tentative solution is the product of three fuzzy 2-spheres with the symmetry $SO(3) \times SO(3) \times SO(3)$. Such a system is described by $N \times N$ hermitian matrices building a representation of the $\mathfrak{so}(3)$ Lie algebra in the following way: $$[B_{i}, B_{j}] = i \mu r
\epsilon_{ijk} B_{k}, \hspace{12mm} B_{1}^{2} + B_{2}^{2} + B_{3}^{2} =
\mu^{2} r^{2} \frac{N^{2} - 1}{4} {\bf 1}_{N \times N}
\textrm{ for } (i,j,k=1,2,3)\label{AZso3so3so3}$$ with similar relations for $i,j,k=4,5,6$ and $i,j,k=7,8,9$, trivial commutators for indices that do not belong to the same group of 3, and $B_{0} = 0$ ($\epsilon_{ijk}$ is defined as usually).
This set of fields (\[AZso3so3so3\]) describes a space formed by the Cartesian product of three fuzzy spheres located in the directions $(x_{1}, x_{2}, x_{3})$, $(x_{4}, x_{5}, x_{6})$ and $(x_{7}, x_{8},
x_{9})$, whose radii are all $\mu r \sqrt{N^{2}-1}/2$. $(N^{2}-1)/4$ is the quadratic Casimir operator of the $\mathfrak{so}(3)$ Lie algebra. Note that any positive-integer value of $N$ is possible here, since $N$ indexes the dimensions of irreducible representations. For $SO(3)$, the irreps have dimensions $N=2j+1$, for all integer values of the spin $j$. However, we can also use spinorial representations with half-integer spins in this case. We have to consider this classical solution instead of the single $SO(3)$ fuzzy sphere $$\begin{aligned}
[B_{i}, B_{j}] = i \mu r \epsilon_{ijk} B_{k} \textrm{ (for
$i,j,k =1,2,3$)}, \hspace{3mm} B_{\mu} = 0 \textrm{ (for $\mu=0, 4,
5, \cdots, 9$)},
\end{aligned}$$ because the solution $B_{4} = \cdots = B_{9} = 0$ is unstable in the directions 4 to 9 due to the negative squared mass[^3] of the rank-1 fields $B_{\mu}$. Without restricting the generality, we can focus on the first sphere located in the direction $(x_{1}, x_{2}, x_{3})$, since the three fuzzy spheres all share the same equations of motion.
In the framework of fuzzy 2-spheres, we can solve the equations of motion (\[AZeqB+\]) and (\[AZeqC+\]) with the following ansatz for the rank-2 field $C_{ij}$: $$\begin{aligned}
C_{ij} = f(r) \epsilon_{ijk} B_{k}, \label{AZeqC+ansatz}
\end{aligned}$$ where $f(r)$ is a function depending on the radius parameter $r$. Indeed, the equation of motion (\[AZeqC+\]) reduces then to: $$\begin{aligned}
\epsilon_{ijk} B_{k} (-f(r) + r + r f^{2}(r) ) =
0. \label{AZeqC+1}
\end{aligned}$$ (\[AZeqC+1\]) has two solutions: $f_{\pm} (r) = \frac{1 \pm
\sqrt{1 - 4r^{2}}}{2r}$. When we plug this result in the equation of motion for $B$ (\[AZeqB+\]), this leads to $$\begin{aligned}
B_{i} ( 1- 2 r f_{\pm}(r) ) = 0.
\end{aligned}$$ This gives the same condition on the radius parameter $r$ for both $f_{+}(r)$ and $f_{-}(r)$, namely: $$\begin{aligned}
\sqrt{1 - 4r^{2} } = 0. \label{AZeqBC+so3}
\end{aligned}$$ Therefore, when we assume the ansatz (\[AZeqC+ansatz\]), we obtain the classical solution (\[AZso3so3so3\]) with the radius parameter set to $r = \frac{1}{2}$, which is fortunately real. Indeed, $r^2 \leq 0$ would indicate that the fuzzy sphere solution is unstable. For example, in the IIB massive matrix model described by (\[MIIBMM\]), the sign of the squared radius of the fuzzy 2-sphere is linked to the sign of the mass term in the action and it would become negative for a correct-sign mass term, which is to be expected, since in that case, the trivial commutative solution becomes the stable vacuum of the theory.
We next want to discuss the stability of the $SO(3) \times SO(3) \times SO(3)$ classical solution in more qualitative terms[@0101102; @0206075]. To this end, we compare the energy of the trivial commutative solution $B_{\mu} = 0$ with that of the fuzzy-sphere solution. The classical energy for $B_{\mu} = 0$ is obviously $E_{\tiny{ B_{\mu}=0}} = - S_{\tiny{ B_{\mu}=0}} = 0$. [^4]
In the $SO(3) \times SO(3) \times SO(3)$ fuzzy-sphere background, the 2-form field $C_{ij}$ is $$\begin{aligned}
C_{ij} = \epsilon_{ijk} B_{k}.
\end{aligned}$$ Therefore, the total energy is $$\begin{aligned}
E_{\tiny SO(3)^3}
&=& - S_{\tiny SO(3)^3} = - 64 \mu \sum_{\mu = 1}^{9} Tr(B_{\mu}
B^{\mu} )
= - 3 \times 64 \mu \sum_{i=1}^{3} Tr(B_{i}
B_{i}) \nonumber \\
&=& - 12 \mu^{3} N (N-1)(N+1). \label{AZ2so3energy}
\end{aligned}$$ This result shows that the $SO(3) \times SO(3) \times SO(3)$ fuzzy-sphere classical solution has a lower energy compared to the trivial commutative solution and hence a higher probability.
Other curved-space solutions and the fuzzy 8-sphere
---------------------------------------------------
So far, we have analyzed the simplest curved-space solution, the $SO(3) \times SO(3) \times SO(3)$ triple fuzzy spheres. Here, we consider other curved-space classical solutions. The fuzzy $2k$-spheres[@9712105; @0105006; @0111278; @0207111], which exhibit a $SO(2k+1)$ symmetry, are constructed by the following $n$-fold symmetric tensor product of $(2k+1)$-dimensional gamma matrices: $$\begin{aligned}
B^{SO(2k+1)}_{p} = \frac{\mu r}{2} [ (\Gamma^{(2k)}_{p} \otimes {\bf 1}
\otimes \cdots \otimes
{\bf 1}) + \cdots + ({\bf 1} \otimes \cdots \otimes {\bf 1} \otimes
\Gamma^{(2k)}_{p} )]_{\textrm{sym}},
\end{aligned}$$ where $p$ runs over $1,2, \cdots, 2k+1$. $\Gamma^{(2k)}_{p}$ are $2^{k}
\times 2^{k}$ gamma matrices, and build a representation of the $SO(2k+1)$ Clifford algebra; i.e. $\{ \Gamma^{(2k)}_{p},
\Gamma^{(2k)}_{q} \} = 2
\delta_{pq} {\bf 1}_{2^{k} \times 2^{k}}$. These matrices satisfy the following algebraic relations: $$\begin{aligned}
& & B^{SO(2k+1)}_{p} B^{SO(2k+1)}_{p} = \frac{\mu^{2} r^{2}}{4} n(n+2k)
{\bf 1}_{N_k \times N_k},
\label{AZ2radius} \\
& & B^{SO(2k+1)}_{pq} B^{SO(2k+1)}_{pq} =
- (\frac{\mu r}{2})^{4} 8 k n (n + 2k) {\bf 1}_{N_k \times N_k},
\label{AZ2radius2} \\
& & [B^{SO(2k+1)}_{pq}, B^{SO(2k+1)}_{s}] = \mu^{2} r^{2} ( -
\delta_{ps} B^{SO(2k+1)}_{q} +
\delta_{qs} B^{SO(2k+1)}_{p}), \label{AZ2pm} \\
& & [B^{SO(2k+1)}_{pq}, B^{SO(2k+1)}_{st}] = \mu^{2} r^{2}
(\delta_{qs} B^{SO(2k+1)}_{pt} +
\delta_{pt} B^{SO(2k+1)}_{qs} - \delta_{ps} B^{SO(2k+1)}_{qt} -
\delta_{qt} B^{SO(2k+1)}_{ps}), \nonumber \\
\label{AZ2mm}
\end{aligned}$$ where $B^{SO(2k+1)}_{pq} = [B^{SO(2k+1)}_{p},
B^{SO(2k+1)}_{q}]$ furnishes (up to a normalization factor) a representation of the $\mathfrak{so}(2k+1)$ Lie algebra and $N_k$ is the dimension of the fully symmetrized irreducible representation for the $SO(2k+1)$ fuzzy sphere. The commutation relations (\[AZ2pm\]) and (\[AZ2mm\]) are inherited from those of the gamma matrices. Thanks to these relations, we expect that the equations of motion can be solved for all even-dimensional fuzzy spheres in a similar fashion to the fuzzy 2-spheres. In other words, this means that the $SO(2k+1)$ fuzzy spheres will provide us with a whole set of curved classical solutions for some precise values of the parameter $r$. In addition, the $B^{SO(2k+1)}_{p}$’s satisfy the following self-duality relation: $$\begin{aligned}
\epsilon_{p_{1} \cdots p_{2k+1}} B^{SO(2k+1)}_{p_{1}}
B^{SO(2k+1)}_{p_{2}} \cdots
B^{SO(2k+1)}_{p_{2k}} = \left( \frac{\mu r}{2} \right)^{2k-1} m_{k}
B^{SO(2k+1)}_{p_{2k+1}}, \label{AZ2duality}
\end{aligned}$$ where $$\begin{aligned}
m_{1} = 2i, \hspace{3mm}
m_{2} = 8(n+2), \hspace{3mm}
m_{3} = -48 i (n+2)(n+4), \hspace{3mm}
m_{4} = - 384 (n+2)(n+4)(n+6), \nonumber \\
\label{AZ2soksphere}
\end{aligned}$$ which is a generalization of the $\mathfrak{so}(3)$ Lie algebra[^5] and a consequence of the duality relation for odd-dimensional Gamma matrices. We give the computation of these coefficients in the second appendix.
Another possible classical solution of our massive supermatrix model is the single $SO(9)$ fuzzy sphere. The analysis goes in the same way as in the $SO(3) \times SO(3) \times SO(3)$ fuzzy spheres. Here, the indices $p, q, \cdots$ run over $1, 2,
\cdots, 9$. For the $SO(9)$ fuzzy-sphere classical solution, we likewise assume the following ansatz for the rank-2 fields $C^{SO(9)}_{pq}$: $$\begin{aligned}
C^{SO(9)}_{pq} = - i \mu^{-1} g(r) B^{SO(9)}_{pq}.
\end{aligned}$$ Then, the equation of motion (\[AZeqC+\]) implies $$\begin{aligned}
\frac{-i}{\mu} B^{SO(9)}_{pq} ( -g(r) + 1 + 7 r^{2} g^{2}(r) ) = 0.
\end{aligned}$$ We again have two choices for the function $f(r)$: $$\begin{aligned}
g_{\pm}(r) = \frac{1 \pm \sqrt{1 - 28 r^2}}{14 r^{2}}.
\end{aligned}$$ The equation of motion (\[AZeqB+\]) for the rank-1 field $B^{SO(9)}_{p}$ gives $$\begin{aligned}
B^{SO(9)}_{p} ( 1 - 8r^{2} g_{\pm} (r) ) = 0.
\end{aligned}$$ Now, unlike the case of the $SO(3) \times SO(3) \times SO(3)$ fuzzy spheres, $1 - 8 r^{2} g_{-}(r) = 0$ does not have any real positive solution for $r$. However, there is exactly one such solution for $1 - 8r^{2}
g_{+}(r) = 0$, which is $r = \frac{1}{8}$.
More generally, for an $SO(2k+1)$ fuzzy sphere, the same ansatz would give $$\begin{aligned}
g_{\pm}(r) &=& \frac{1 \pm \sqrt{1 - 4(2k-1)r^2}}{2(2k-1)r^{2}} \; ,
\nonumber\\
1-2kr^{2} g_{\pm} (r)&=& 0 \; , \textrm{ solvable only for } g_{+} (r)
\textrm{ at} \\
r&=&\frac{1}{2k}.\nonumber
\end{aligned}$$
We discuss the stability of the $SO(9)$ fuzzy-sphere classical solution by computing its classical energy. At the classical level, we obtain $$\begin{aligned}
E_{\tiny SO(9)} = - \frac{5}{8} \mu^{3} n(n+8)
N_4. \label{AZ2so9energy}
\end{aligned}$$ $N_4$ is given in [@0105006][@0111278] by[^6] $$\begin{aligned}
N_4 = \frac{(n+1)(n+2)(n+3)^{2} (n+4)^{2} (n+5)^{2}
(n+6)(n+7)}{302400}.
\end{aligned}$$ In contrast with the $SO(3) \times SO(3) \times SO(3)$ case, $N$ can take here only certain precise values. For example, the smallest non-trivial representation ($n=1$) has dimension 16, the following one ($n=2$) 126, then 672, etc... The classical energy for the $SO(3) \times SO(3) \times SO(3)$ fuzzy-sphere solution is of the order ${\cal O}( - \mu^3 n^3 ) = {\cal
O}(- \mu^3 N^3)$ while that of the $SO(9)$ fuzzy-sphere solution is of the order ${\cal O}(- \mu^3 n^{12}) = {\cal
O}( - \mu^3 N^{\frac{6}{5}})$. Therefore, at large $N$, the $SO(3) \times SO(3) \times SO(3)$ triple fuzzy-sphere solution is energetically favored compared to the $SO(9)$ solution at the equal size $N$ of the matrices. The presence of a spherical solution for all $N$ in the $SO(3) \times SO(3) \times SO(3)$ case may indeed be a stabilizing factor. On the other hand, at equal value of $n$, whose physical meaning is less clear, the fuzzy 8-sphere solution has lower energy.
The single $SO(q)$ fuzzy spheres for $q \leq 8$ do not constitute a stable classical solution of our model. When the $SO(q)$ sphere occupies the direction $x_{1}, x_{2}, \cdots, x_{q}$, the solution $B^{SO(q)}_{q+1} = B^{SO(q)}_{q+2} = \cdots = B^{SO(q)}_{9} =0$ is trivially unstable because of the negative mass squared. Whereas, the Cartesian product of several fuzzy spheres, such as $SO(3)
\times SO(6)$, is a possible candidate for a stable classical solution.
Nucleation process of spherical branes
--------------------------------------
Starting from a vacuous spacetime, it is interesting to try to guess how spherical brane configurations could be successively produced through a sequence of decays into energetically more favorable meta-stable brane systems. The reader may have noticed that we have so far limited ourselves to the study of curved branes building irreducible representations of their symmetry groups. This could seem at first to be an unjustified prejudice, but it turns out that such configurations are energetically favored at equal values of $N$. For example, for $SO(3)$, an irreducible representation ${\cal R}_{N}$ of dimension $N$ contributes as $$E_{{\cal R}_{N}} = - 4 \mu^{3} (N^3-N) \nonumber$$ per fuzzy 2-sphere, while a reducible representation ${\cal R}_{N_1} \oplus
\ldots \oplus {\cal R}_{N_m}$ of equal dimension $N_1+\ldots+N_m=N$ would contribute as $$E_{{\cal R}_{N_1} \oplus \ldots \oplus {\cal R}_{N_m}} = - 4 \mu^{3}
\sum_{i=1}^{m} (N_i^3-N_i).\nonumber$$ This is obviously a less negative number, especially for big values of $N$. A similar conclusion was reached in [@0101102] for the case of a Euclidean 3-dimensional IIB matrix model with a Chern-Simons term and it seems to be a fairly general feature of matrix models admitting non-trivial classical solutions. This property is particularly clear for low-dimensional branes, since the classical energy is of order ${\cal O}(- \mu^3 N^3)$ for $SO(3)$, but it remains true for any $SO(2k+1)$ fuzzy-sphere solution, whose energy is of order ${\cal O}( - \mu^3 N^{1+4/(k(k+1))})$, which also shows that low-dimensional configurations are favored. As hinted for in the preceding subsection, this latter fact can be physically understood by remarking that there are more irreps available for low-dimensional fuzzy spheres, which makes it easier for them to grow in radius through energetically favorable configurations. A third obvious fact is that configurations described by representations of high dimensionality are preferred.
Put together, these comparisons give us a possible picture for the branes nucleation process in this and similar matrix models. As they appear, configurations of all spacetime dimensions described by small representations will be progressively absorbed by bigger representations to form irreducible ones, that will slowly grow in this way to bigger values of $N$. Parallel to that, branes of higher dimensionalities will tend to decay into a bunch of branes of smaller dimensionalities, finally leaving only 2-spheres and noncommutative tori of growing radii. If the size of the Hermitian matrices is left open, as is usually the case in completely reduced models, where the path integration contains a sum on that size, no configuration will be truly stable, since the size of the irreps will grow continuously.
Of course, this is a relatively qualitative study, which could only be proven correct by a full quantum statistical study of the model. However, it seems to be an interesting proposal for the possible physics of such theories.
Supersymmetry
-------------
We next comment on the structure of the supersymmetry. The biggest difference with the purely cubic supermatrix model, due to the addition of the mass term, is that this model is not invariant under the inhomogeneous supersymmetry $$\begin{aligned}
\delta_{\tiny \textrm{inhomogeneous}} m=0, \hspace{3mm}
\delta_{\tiny \textrm{inhomogeneous}} \psi = \xi, \label{AZinhomo}
\end{aligned}$$ which is a translation of the fermionic field. However, this model has 2 homogeneous supersymmetries in 10 dimensions, which are part of the $\mathfrak{osp}(1|32,\mathbb{R})$ symmetry: $$\begin{aligned}
\delta_{\epsilon} M = \left[ \left( \begin{array}{cc} 0 & \epsilon \\ i
{\bar \epsilon} & 0 \end{array} \right),
\left( \begin{array}{cc} m & \psi \\ i
{\bar \psi} & 0 \end{array} \right) \right]
= \left( \begin{array}{cc} i (\epsilon {\bar \psi} - \psi {\bar
\epsilon} ) & - m \epsilon \\ i {\bar \epsilon} m & 0
\end{array} \right), \label{AZhomo}
\end{aligned}$$ which transforms the bosonic and fermionic fields as $$\begin{aligned}
\delta_{\epsilon} m = i (\epsilon {\bar \psi} - \psi {\bar
\epsilon} ), \hspace{3mm}
\delta_{\epsilon} \psi = - m \epsilon. \label{AZ3SUSY}
\end{aligned}$$
In the IIB matrix model, the supersymmetry has to balance between a quartic term $Tr([A_{\mu},A_{\nu}])^2$ and a trilinear contribution $Tr {\bar \psi} \Gamma^{\mu} [A_{\mu}, \psi]$ in the action (\[AZ1IKKT\]), which implies that the SUSY transformation of the fermionic field has to be bilinear in the bosonic field. On the other hand, the homogeneous supersymmetries are all linear in the fields in the purely cubic supermatrix model[@0102168; @0103003]. By incorporating the mass term, we are allowed to integrate out the rank-2 field $C_{\mu_{1} \mu_{2}}$ by solving the classical equation of motion iteratively as in [@0201183][^7].
Thanks to this procedure, the homogeneous SUSY transformation for the fermionic field becomes $$\begin{aligned}
\delta_{\epsilon} \psi = \frac{i}{2} [B_{\mu_{1}}, B_{\mu_{2}}]
\Gamma^{\mu_{1} \mu_{2}} \epsilon + \cdots, \label{AZ2SUSYferm}
\end{aligned}$$ while the transformation of the field $B_{\mu}$ is $$\begin{aligned}
\delta_{\epsilon} B_{\mu} = - \frac{1}{32} tr_{32 \times 32}
( i (\epsilon {\bar \psi} - \psi {\bar \epsilon}) \Gamma_{\mu
\sharp} ) = - \frac{i}{16} {\bar \epsilon} \Gamma_{\mu \sharp}
\psi. \label{AZ2SUSYvec}
\end{aligned}$$
In that sense, the mass term is essential to realize the Yang-Mills-like structure for the homogeneous supersymmetries. On the other hand, if we want to preserve the homogeneous supersymmetries, we cannot just put a mass term for $C_{\mu_{1} \mu_{2}}$ by hand, the $\mathfrak{osp}(1|32,\mathbb{R})$ symmetry forces all fields to share the same mass, since they all lie in the same multiplet. In particular, we are forced to introduce a mass term for the fermions as well, which breaks the inhomogeneous supersymmetries. In other words, it seems difficult to have Super Yang-Mills-type structure for both homogeneous and inhomogeneous supersymmetries in the context of supermatrix models.
Indeed, in contrast with the purely cubic supermatrix model[@0102168; @0103003], which has twice as many SUSY parameters, the massive supermatrix model has only ${\cal N}=2$ SUSY in 10 dimensions, because it lacks the inhomogeneous supersymmetries. In consequence, we cannot realize the translation of the vector field $A_{\mu}$ as a commutator of two linear combinations of the homogeneous and inhomogeneous supersymmetries (\[AZ3SUSY\]) as in the IIB matrix model, where it leads to the interpretation of the eigenvalues of $A_{\mu}$ as spacetime coordinates. On the contrary, $$\begin{aligned}
[\delta_{\epsilon}, \delta_{\chi}] m = i [(\epsilon {\bar \chi} -
\chi {\bar \epsilon}), m], \hspace{3mm}
[\delta_{\epsilon}, \delta_{\chi}] \psi = i (\epsilon {\bar \chi} -
\chi {\bar \epsilon} ) \psi
\end{aligned}$$ vanishes up to an $\mathfrak{sp}(32,\mathbb{R})$ rotation. This problem is a serious obstacle for the identification of the supersymmetry of this model with that of the IIB matrix model. More analysis will be reported elsewhere.
Conclusion and outlook
======================
In this paper, we have investigated a supermatrix model based on $\mathfrak{osp}(1|32,\mathbb{R})$ with a mass term and a cubic interaction. To be able to describe the gravitational interaction in terms of large $N$ reduced models, we must understand how the reduced models can describe physics in curved spacetimes. Although the IIB matrix model only possesses flat noncommutative spacetime as a classical solution, by adding a tachyonic mass term as in [@0103192], we can obtain new classical solutions building curved space backgrounds. Following this idea, we have expected that massive supermatrix models could also exhibit similar properties leading to non-trivial classical solutions. In particular, we have investigated fuzzy-sphere solutions with symmetries $SO(3) \times SO(3) \times SO(3)$ and $SO(9)$, and calculated the parameter determining the quantization step separating the radii of configurations described by different representations of the Lie algebra. We have then discussed their respective likelihood by comparing their energy at the classical level, which gave us a way to understand a possible dynamical evolution of the solutions through successively more favorable brane configurations.
It is an intriguing issue to search for other stable curved-space classical solutions. For example, an $SO(3) \times SO(6)$ fuzzy sphere could be a promising candidate. Indeed, the expansion around this classical solution may be related in some way to the BMN matrix model[@0202021; @0205185; @0206239; @0207190], which appears as the discrete light-cone quantization of D0-brane in the M-theory pp-wave background[@0202021]. However, to study this case explicitly, we first have to analyze how the equations of motion (\[AZeqB+\]) and (\[AZeqC+\]) can be treated in the case of odd fuzzy spheres. Indeed, as is outlined in [@0207111], the commutator $[B_i,B_j]$ does not correspond to a representation of $SO(2k)$ for odd fuzzy spheres, which makes the analysis much more involved. However, the construction of an $SO(3) \times SO(6)$ classical solution would show how transverse 5-branes can appear in this model. The $SO(4) \times SO(5)$ case should proceed along similar lines. Another case that can be investigated is a solution of the type $SO(2) \times SO(2) \times SO(5)$, in which the two first circles build a noncommutative torus as in [@0103192].
Another difficult problem that could be tackled in the future is a detailed analysis of the stability under perturbations of such noncommutative spaces in this and other similar models, to see if these solutions are really local minima of the potential energy.
We next mention the relation between our massive $\mathfrak{osp}(1|32,\mathbb{R})$ supermatrix model and the IIB matrix model. In the quantum field theory, some different models which possess the same symmetry are equivalent in the continuum limit. This property is known as universality. We expect that some similar mechanism may hold true of the large $N$ reduced models, and hence that various matrix models may have the same large $N$ limit. If we believe in the matrix-model version of the universality conjecture, it is possible that our massive supermatrix model could be equivalent to the IIB matrix model. In this sense, it is interesting to investigate further whether our model shares the maximal 10-dimensional ${\cal N}=2$ SUSY with the IIB matrix model and how.
#### Acknowledgment
\
The authors would like to express their gratitude to Luca Carlevaro, Jean-Pierre Derendinger, Frank Ferrari, Hikaru Kawai, Yusuke Kimura and Christian Römelsberger for valuable discussion. The works of T.A. were supported in part by Grant-in-Aid for Scientific Research from Ministry of Education, Culture, Sports, Science and Technology of Japan ($\#$01282). M.B. acknowledges the financial support provided through the European Community’s Human Potential Programme under contract HPRN-CT-2000-00131 Quantum Spacetime and the Swiss Office for Education and Science as well as the Swiss National Science Foundation.
Massive supermatrix model action
================================
$$\begin{aligned}
S &=& 96 \mu Tr \left( - W^{2} - A_{\mu} A^{\mu} + B_{\mu} B^{\mu}
+ \frac{1}{2} C_{\mu_{1} \mu_{2}} C^{\mu_{1} \mu_{2}}
- \frac{1}{4!} H_{\mu_{1} \cdots \mu_{4}} H^{\mu_{1} \cdots
\mu_{4}} \right. \nonumber \\
& & \hspace{15mm} \left.
- \frac{1}{5!} Z_{\mu_{1} \cdots \mu_{5}} Z^{\mu_{1} \cdots
\mu_{5}}
+ \frac{i}{16} {\bar \psi} \psi \right) \nonumber \\
&+& 32 i Tr \Bigg( - 3 C_{\mu_{1} \mu_{2}} [A^{\mu_{1}},
A^{\mu_{2}}] + 3 C_{\mu_{1} \mu_{2}} [B^{\mu_{1}}, B^{\mu_{2}}]
+ 6 W [A_{\mu}, B^{\mu}]
+ C_{\mu_{1} \mu_{2}} [ {C^{\mu_{2}}}_{\mu_{3}}, C^{\mu_{3}
\mu_{1}}] \nonumber \\
& & \hspace{12mm} + \frac{1}{4} B_{\mu_{1}} [H_{\mu_{2} \cdots
\mu_{5}}, Z^{\mu_{1} \cdots \mu_{5}}]
- \frac{1}{8} C_{\mu_{1} \mu_{2}}
( 4 [ {H^{\mu_{1}}}_{\rho_{1} \rho_{2} \rho_{3}}, H^{\mu_{2}
\rho_{1} \rho_{2} \rho_{3}}]
+ [ {Z^{\mu_{1}}}_{\rho_{1} \cdots \rho_{4}}, Z^{\mu_{2}
\rho_{1} \cdots \rho_{4}}] ) \nonumber \\
& & \hspace{12mm} + \frac{3}{(5!)^{2}} \epsilon^{\mu_{1} \cdots
\mu_{10} \sharp} \left( - W [Z_{\mu_{1} \cdots \mu_{5}},
Z_{\mu_{6} \cdots \mu_{10}}] + 10 A_{\mu_{1}} [H_{\mu_{2} \cdots
\mu_{5}}, Z_{\mu_{6} \cdots \mu_{10}}] \right) \nonumber \\
& & \hspace{12mm} + \frac{200}{(5!)^{3}}
\epsilon^{\mu_{1} \cdots \mu_{10} \sharp}
\left( 5 H_{\mu_{1} \cdots \mu_{4}} [{Z_{\mu_{5} \mu_{6} \mu_{7}}}^{\rho
\chi}, Z_{\mu_{8} \mu_{9} \mu_{10} \rho \chi}]
+10 H_{\mu_{1} \cdots \mu_{4}} [ {H_{\mu_{5} \mu_{6}
\mu_{7}}}^{\rho}, H_{\mu_{8} \mu_{9}
\mu_{10} \rho}] \right. \nonumber \\
& & \hspace{35mm} \left. + 6 {H^{\rho \chi}}_{\mu_{1} \mu_{2}}
[ Z_{\mu_{3} \mu_{4} \mu_{5} \rho \chi}, Z_{\mu_{6} \cdots
\mu_{10}} ] \right) \Bigg)\nonumber \\
&+& 3 Tr \left( {\bar \psi} \Gamma^{\sharp} [W, \psi]
+ {\bar \psi} \Gamma^{\mu} [A_{\mu}, \psi]
+ {\bar \psi} \Gamma^{\mu \sharp} [B_{\mu}, \psi]
+ \frac{1}{2!} {\bar \psi} \Gamma^{\mu_{1} \mu_{2}}
[C_{\mu_{1} \mu_{2}}, \psi] \right. \nonumber \\
& & \hspace{10mm} + \left. \frac{1}{4!} {\bar \psi} \Gamma^{\mu_{1} \cdots
\mu_{4} \sharp} [H_{\mu_{1} \cdots \mu_{4}}, \psi]
+ \frac{1}{5!} {\bar \psi} \Gamma^{\mu_{1} \cdots \mu_{5}}
[Z_{\mu_{1} \cdots \mu_{5}}, \psi] \right)
\end{aligned}$$
Notations and useful formulae
=============================
In this appendix, we give hints for the derivations of the coefficients $m_{k}$ in the self-duality relation (\[AZ2duality\]) for the $SO(2k+1)$ fuzzy sphere. We first define the $2^{k} \times 2^{k}$ gamma matrices in the $2k$-dimensional Euclidean space $\Gamma^{(2k)}_{p}$ by the following recursive relation: $$\begin{aligned}
& & \Gamma^{(2k+2)}_{p} = \Gamma^{(2k)}_{p} \otimes \sigma_{2} =
\left( \begin{array}{cc} 0 & -i \Gamma^{(2k)}_{p} \\ i
\Gamma^{(2k)}_{p} & 0 \end{array} \right),
\hspace{3mm}
\Gamma^{(2k+2)}_{2k+2} = {\bf 1}_{2^{k} \times 2^{k}} \otimes
\sigma_{1} = \left( \begin{array}{cc} 0 & {\bf 1}_{2^{k} \times
2^{k}} \\ {\bf 1}_{2^{k} \times 2^{k}} & 0 \end{array} \right),
\nonumber \\
& & \Gamma^{(2k+2)}_{2k+3} = {\bf 1}_{2^{k} \times 2^{k}} \otimes
\sigma_{3} = \left( \begin{array}{cc} {\bf 1}_{2^{k}
\times 2^{k}} & 0 \\ 0 & - {\bf 1}_{2^{k} \times 2^{k}}
\end{array} \right), \label{AZCrecur}
\end{aligned}$$ where the index $p$ runs over $p=1,2, \cdots, 2k+1$. The 2-dimensional gamma matrices are identical to the Pauli matrices: $\Gamma^{(2)}_{i} = \sigma_{i}$. Under this notation, we obtain $$\begin{aligned}
\sigma_{1} \sigma_{2} = i \sigma_{3}, \hspace{3mm}
\Gamma^{(4)}_{1} \Gamma^{(4)}_{2} \Gamma^{(4)}_{3} \Gamma^{(4)}_{4}
= \Gamma^{(4)}_{5}, \hspace{3mm}
\Gamma^{(6)}_{1} \Gamma^{(6)}_{2} \cdots \Gamma^{(6)}_{6} = - i
\Gamma^{(6)}_{7}, \hspace{3mm}
\Gamma^{(8)}_{1} \Gamma^{(8)}_{2} \cdots \Gamma^{(8)}_{8} = -
\Gamma^{(8)}_{9}. \nonumber \\ \label{AZCchirality}
\end{aligned}$$ It is trivial that $m_{1} = 2i$ as explained in the footnote \[ftso3\], while the computation of the coefficient $m_{2}$ can be found in [@9712105]. In this appendix, we give formulae that are useful to derive $m_{3}$ and $m_{4}$. We set $\frac{\mu r}{2} = 1$ and omit “$_{\textrm{sym}}$”, with the understanding that these formulae are only valid in the fully symmetrized representations.
In general, we have $$\begin{aligned}
& &\sum_{l=1}^{2k+1} (\Gamma^{(2k)}_{l} \otimes \Gamma^{(2k)}_{l}) =
({\bf 1}_{2^k \times 2^k} \otimes {\bf 1}_{2^k \times 2^k}), \nonumber \\
& &\sum_{l=1}^{2k+1} (\Gamma^{(2k)}_{l_{1} l_{2}} \otimes
\Gamma^{(2k)}_{l_{1} l_{2}}) = -2k ({\bf 1}_{2^k \times 2^k} \otimes
{\bf 1}_{2^k \times 2^k})
\end{aligned}$$ More specifically, to compute $m_4$, we also need $$\begin{aligned}
& & (\Gamma^{(6)}_{l_{1} l_{2} l_{3}} \otimes
\Gamma^{(6)}_{l_{1} l_{2} l_{3}} ) =
- 18 ({\bf 1}_{8 \times 8} \otimes {\bf 1}_{8 \times
8}), \label{AZCfreq43} \\
& & (\Gamma^{(6)}_{l_{1} l_{2}}
\otimes \Gamma^{(6)}_{l_{3}} \otimes \Gamma^{(6)}_{l_{1}
l_{2} l_{3}}) = -6 ({\bf 1}_{8 \times 8} \otimes {\bf 1}_{8 \times
8} \otimes {\bf 1}_{8 \times 8}), \label{AZCfreq44} \\
& & (\Gamma^{(6)}_{l_{1} l_{2}}
\otimes \Gamma^{(6)}_{l_{3} l_{4}} \otimes \Gamma^{(6)}_{l_{1}
\cdots l_{4}}) = 24 ({\bf 1}_{8 \times 8} \otimes {\bf 1}_{8 \times
8} \otimes {\bf 1}_{8 \times 8}), \label{AZCfreq45} \\
& & (\Gamma^{(6)}_{l_{1} l_{2}} \otimes \Gamma^{(6)}_{l_{3} l_{4}}
\otimes \Gamma^{(6)}_{l_{5} l_{6}} \otimes \Gamma^{(6)}_{l_{1}
\cdots l_{6}} )
= - 48 ({\bf 1}_{8 \times 8} \otimes {\bf 1}_{8 \times 8} \otimes {\bf
1}_{8 \times 8} \otimes {\bf 1}_{8 \times 8}). \label{AZCfreq46}
\end{aligned}$$
[99]{} T. Banks, W. Fischler, S. H. Shenker and L. Susskind, “M theory as a matrix model: A conjecture,” Phys. Rev. D [**55**]{}, 5112 (1997) \[[hep-th/9610043]{}\]. N. Ishibashi, H. Kawai, Y. Kitazawa and A. Tsuchiya, “A large-N reduced model as superstring,” Nucl. Phys. B [**498**]{}, 467 (1997) \[[hep-th/9612115]{}\]. R. Dijkgraaf, E. Verlinde and H. Verlinde, “Matrix string theory,” Nucl. Phys. B [**500**]{}, 43 (1997) \[[hep-th/9703030]{}\]. M. Fukuma, H. Kawai, Y. Kitazawa and A. Tsuchiya, “String field theory from IIB matrix model,” Nucl. Phys. B [**510**]{}, 158 (1998) \[[hep-th/9705128]{}\]. H. Itoyama and A. Tokura, “USp(2k) matrix model: F theory connection,” Prog. Theor. Phys. [**99**]{}, 129 (1998) \[[hep-th/9708123]{}\]. J. Castelino, S. Lee and W. Taylor, “Longitudinal 5-branes as 4-spheres in matrix theory,” Nucl. Phys. B [**526**]{}, 334 (1998) . H. Aoki, S. Iso, H. Kawai, Y. Kitazawa and T. Tada, “Space-time structures from IIB matrix model,” Prog. Theor. Phys. [**99**]{}, 713 (1998) . S. Iso and H. Kawai, “Space-time and matter in IIB matrix model: Gauge symmetry and diffeomorphism,” Int. J. Mod. Phys. A [**15**]{}, 651 (2000) . H. Aoki, S. Iso, H. Kawai, Y. Kitazawa, A. Tsuchiya and T. Tada, “IIB matrix model,” Prog. Theor. Phys. Suppl. [**134**]{}, 47 (1999) \[[hep-th/9908038]{}\]. H. Aoki, N. Ishibashi, S. Iso, H. Kawai, Y. Kitazawa and T. Tada, “Noncommutative Yang-Mills in IIB matrix model,” Nucl. Phys. B [**565**]{}, 176 (2000) \[[hep-th/9908141]{}\]. L. Smolin, “M theory as a matrix extension of Chern-Simons theory,” Nucl. Phys. B [**591**]{}, 227 (2000) \[[hep-th/0002009]{}\]. E. Bergshoeff and A. Van Proeyen, “The many faces of OSp(1$|$32),” Class. Quant. Grav. [**17**]{}, 3277 (2000) \[[hep-th/0003261]{}\]. L. Smolin, “The cubic matrix model and a duality between strings and loops,” [hep-th/0006137]{}. S. Iso, Y. Kimura, K. Tanaka and K. Wakatsuki, “Noncommutative gauge theory on fuzzy sphere from matrix model,” Nucl. Phys. B [**604**]{}, 121 (2001) \[[hep-th/0101102]{}\]. T. Azuma, S. Iso, H. Kawai and Y. Ohwashi, “Supermatrix models,” Nucl. Phys. B [**610**]{}, 251 (2001) \[[hep-th/0102168]{}\]. T. Azuma, “Investigation of matrix theory via super Lie algebra,” [hep-th/0103003]{}. Y. Kimura, “Noncommutative gauge theories on fuzzy sphere and fuzzy torus from matrix model,” Prog. Theor. Phys. [**106**]{}, 445 (2001) \[[hep-th/0103192]{}\]. S. Ramgoolam, “On spherical harmonics for fuzzy spheres in diverse dimensions,” Nucl. Phys. B [**610**]{}, 461 (2001) . P. M. Ho and S. Ramgoolam, “Higher dimensional geometries from Matrix brane constructions,” Nucl. Phys. B [**627**]{}, 266 (2002) . S. Chaudhuri, “Nonperturbative type I-I’ string theory,” [hep-th/0201129]{}. M. Bagnoud, L. Carlevaro and A. Bilal, “Supermatrix models for M-theory based on osp(1$|$32,R),” Nucl. Phys. B [**641**]{}, 61 (2002) . D. Berenstein, J. M. Maldacena and H. Nastase, “Strings in flat space and pp waves from N = 4 super Yang Mills,” JHEP [**0204**]{}, 013 (2002) . S. Chaudhuri, “Spontaneous breaking of diffeomorphism invariance in Matrix theory,” [hep-th/0202138]{}. T. Azuma and H. Kawai, “Matrix model with manifest general coordinate invariance,” Phys. Lett. B [**538**]{}, 393 (2002) . Y. Kimura, “Noncommutative gauge theory on fuzzy four-sphere and matrix model,” Nucl. Phys. B [**637**]{}, 177 (2002) . K. Dasgupta, M. M. Sheikh-Jabbari and M. Van Raamsdonk, “Matrix perturbation theory for M-theory on a PP-wave,” JHEP [**0205**]{}, 056 (2002) . P. Valtancoli, “Stability of the fuzzy sphere solution from matrix model,” [hep-th/0206075]{}. C. Sochichiu, “Continuum limit(s) of BMN matrix model: Where is the (nonabelian) gauge group?,” [hep-th/0206239]{}. S. Ramgoolam, “Higher dimensional geometries related to fuzzy odd-dimensional spheres,” [hep-th/0207111]{}. Y. Kitazawa, “Matrix Models in Homogeneous Spaces,” Nucl. Phys. B [**642**]{}, 210 (2002) . K. Sugiyama and K. Yoshida, “Giant graviton and quantum stability in matrix model on PP-wave background,” [hep-th/0207190]{}.
T. Eguchi and H. Kawai, “Reduction Of Dynamical Degrees Of Freedom In The Large N Gauge Theory,” Phys. Rev. Lett. [**48**]{}, 1063 (1982). A. Gonzalez-Arroyo and M. Okawa, “The Twisted Eguchi-Kawai Model: A Reduced Model For Large N Lattice Gauge Theory,” Phys. Rev. D [**27**]{}, 2397 (1983). A. Gonzalez-Arroyo and C. P. Korthals Altes, “Reduced Model For Large N Continuum Field Theories,” Phys. Lett. B [**131**]{}, 396 (1983). J. Madore, “The Fuzzy sphere,” Class. Quant. Grav. [**9**]{}, 69 (1992).
[^1]: e-mail address : [email protected]
[^2]: We draw the readers’ attention to our choice of sign for the bosonic term of the IIB matrix model. In this paper, we regard the action as [*minus*]{} the potential, which is a choice of sign opposite to the usual definition. We define the action (\[AZ1IKKT\]) in the 10-dimensional Minkowskian space, and the path integral is in our case defined as $$\begin{aligned}
Z = \int dA d \psi e^{+S_{E}}, \nonumber
\end{aligned}$$ where $S_{E}$ is defined in the 10-dimensional Euclidean space; i.e. $S_{E}$ is defined by Wick-rotating $A_{0}$ as $A_{0} \to i A_{0}$ and replacing the gamma matrices for the $SO(9,1)$ Clifford algebra with those for $SO(10)$ in the action $S$.
[^3]: The classical solution with $B_{0} = 0$ has no problem, because it has a positive mass unlike the other directions of the field $B$.
[^4]: Recall that in our notation, the action is minus the potential. Since we now consider a classical solution with $B_{0} = 0$ (thus no need of Wick rotation), the energy is simply minus the classical action in which we substitute the solution.
[^5]: It generalizes the fuzzy 2-sphere case where $B_i^{SO(3)} B_i^{SO(3)}$ is proportional to the identity on the totally symmetric space of dimension $n+1$. For $N = n+1$, the radius of the $SO(3)$ fuzzy sphere is indeed $$\begin{aligned}
\frac{\mu^{2} r^{2}}{4} n(n+2) = (\mu r)^{2} \frac{N^{2} -1}{4}. \nonumber
\end{aligned}$$ The relation (\[AZ2radius\]) actually corresponds to the Casimir of the $\mathfrak{so}(3)$ Lie algebra. And (\[AZ2duality\]) is trivially equivalent to the commutation relation $[B^{SO(3)}_{i}, B^{SO(3)}_{j}] =
i \mu r \epsilon_{ijk} B^{SO(3)}_{k}$. \[ftso3\]
[^6]: Generally, $N_k$ is known to be of the order ${\cal O}(n^{\frac{k(k+1)}{2}})$, and more explicitly, $$\begin{aligned}
N_1 = (n+1), \hspace{2mm}
N_2 = \frac{(n+1)(n+2)(n+3)}{6}, \hspace{2mm}
N_3 = \frac{(n+1)(n+2)(n+3)^{2}(n+4)(n+5)}{360}. \nonumber
\end{aligned}$$
[^7]: The Yang-Mills-like structure of the homogeneous SUSY transformation on the fermion comes from $-
\frac{1}{2} C_{\mu \nu} \Gamma^{\mu \nu} \epsilon$, which is a part of $\delta_{\epsilon} \psi = - m \epsilon$. The explicit form of the iterative solution of the equations of motion (\[AZeqB+\]) and (\[AZeqC+\]) is $$\begin{aligned}
C_{\mu \nu} &=& - i \mu^{-1} [B_{\mu}, B_{\nu}]
+ i \mu^{-3} [[B_{\mu}, B_{\rho}], [B_{\nu}, B^{\rho}]] \nonumber \\
& & - i \mu^{-5} [[B_{\mu}, B_{\rho}], [[B_{\nu}, B_{\chi}],[B^{\rho},
B^{\chi}]] ]
+ i \mu^{-5} [[B_{\nu}, B_{\rho}], [[B_{\mu}, B_{\chi}],[B^{\rho},
B^{\chi}]] ]
+ {\cal O}(\mu^{-7}). \nonumber
\end{aligned}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A hyperaccreting stellar-mass black hole has been proposed as the candidate central engine of gamma-ray bursts (GRBs). The rich observations of GRBs by *Fermi* and *Swift* make it possible to constrain the central engine model by comparing the model predications against data. This paper is dedicated to studying the temporal evolution of central engine parameters for both prompt emission and afterglow phases. We consider two jet launching mechanisms, i.e., $\nu\bar{\nu}$ annihilations and the Blandford-Znajek (BZ) processe, and obtain analytical solutions to these two models. We then investigate the black hole central engine parameters, such as the jet power, the dimensionless entropy $\eta$, and the central engine parameter $\mu_0=\eta (1+\sigma_0)$ (where $\sigma_0$ is the initial magnetization of the engine) at the base of the jet. The black hole may be spun up by accretion, or spun down by the BZ process, leaving imprints in GRB lightcurves. Usually, a BZ jet is more powerful and is likely responsible for the late time central engine activities. However, an initially non-spinning black hole central engine may first launch a thermal “fireball” via neutrino annihilations, and then launch a Poynting-flux-dominated jet via the BZ process. Multiple flares, giant bumps, and plateaus in GRB afterglows can be well produced as the result of late time accretion onto the black hole.'
author:
- 'Wei-Hua Lei$^{1,2}$, Bing Zhang$^{2}$, Xue-Feng Wu$^{3}$, and En-Wei Liang$^{4}$'
title: 'Hyperaccreting Black Hole as Gamma-Ray Burst Central Engine. II. Temporal evolution of central engine parameters during Prompt and Afterglow Phases'
---
Introduction
============
The nature of the central engine of gamma-ray bursts (GRBs) remains a mystery. It is generally believed that long GRBs are connected with core-collapse supernovae (Woosley 1993; Paczy$\acute{n}$ski 1998;MacFadyen & Woosley 1999), and short GRBs are likely related to mergers of two neutron stars or a neutron star and a black hole (Eichler et al. 1989; Paczy$\acute{n}$ski 1991; Fryer et al. 1999). These scenarios lead to the formation of a stellar mass black hole (BH) or a millisecond magnetar.
Two types of GRB central engine models have been discussed in the literature, i.e., the BH model and magnetar model. One popular model invokes a stellar-mass BH surrounded by a neutrino-cooling-dominated accretion flow (NDAF). Two mechanisms are considered to power the relativistic jet in a GRB for a BH central engine: the neutrino-antineutrino annihilation mechanism, which liberates the gravitational energy from the accretion disk (Popham et al. 1999, hereafter PWF99; Di Matteo et al. 2002, hereafter DPN02; Gu et al. 2006; Chen & Beloborodov 2007; Janiuk et al. 2007; Lei et al. 2009; Liu et al. 2015); and the Blandford-Znajek (Blandford & Znajek 1977, hereafter BZ) mechanism, which extracts the spin energy from the Kerr BH (Lee et al. 2000; Li 2000; Lei et al. 2013).
Thanks to *Swift* and *Fermi*, the observations have collected rich information on GRBs, which put further constraints on the GRB central engine models. For example, since a good fraction of GRBs are followed by X-ray flares (some have giant bumps and plateaus), the GRB central engine must be long-lived. In some GRBs (e.g., GRB 080916C), the broadband spectra show no evidence of quasi-thermal emission from a fireball photoshpere (Abdo et al. 2009), suggesting that at least for some GRBs, the central engine has to be strongly magnetized (Zhang & Pe’er 2009). These observational constraints motivate us to systematically investigate the GRB BH central engine models. We planned to present our results in two papers. In Paper I (Lei et al. 2013, hereafter Paper I), we addressed the fundamental problem of baryon loading in GRB jets. We found that a magnetically dominated jet can be much cleaner and is more consistent with the requirement of large Lorentz factors in GRBs (Paper I). With the estimated Lorentz factor from the baryon loading rate, Yi et al. (2017) and Xie et al. (2017) found that some empirical correlations, such as jet power vs. Lorentz factor $\Gamma_0$ (Liang et al. 2010; Liang et al. 2015; Ghirlanda et al. 2012; Lü et al. 2012) and minimum variability timescale (MTS, Wu et al. 2016) vs. the Lorentz factor $\Gamma_0$, favor the scenario in which the jet is driven by the BZ mechanism. A direct comparison between NDAF and BZ processes have been discussed, mostly considering the energy output only, in a number of works (PWF99; Kawanaka et al. 2013; Liu et al. 2015). However, a dedicated study on the evolution of central engine parameters, especially the baryon-loading-related dimensionless “entropy” $\eta$ (for the neutrino model), the magnetization parameter $\sigma_0$, and the central engine parameter $\mu_0=\eta (1+\sigma_0)$ (for BZ model), are still lacking. In the observational front, the temporal behavior of GRBs in the prompt emission and early afterglow phases may provide meaningful clues to the central engine models. It is therefore interesting to compare the predictions from the BH central engine models with the temporal behaviour of GRBs. This is the purpose of this Paper II. We continue to investigate the evolution of the BH central engine based on Paper I.
This paper is organised as follows. In Section 2, we will study the two jet launching mechanisms within the context of Kerr metric in detail. We then apply our results to the prompt emission phase in Section 3 and the late central engine activity in Section 4. Finally, we summarize our results and discuss some related issues in Section 5.
Black Hole Central Engine Model: Neutrino annihilation and Magnetic Powers
==========================================================================
For a spinning BH with hyper-accretion disk, energy can be extracted to power GRB by neutrino annihilations from the NDAF or by the BZ mechanism from the rotating BH. In this section, we will study these two mechanisms in detail.
Neutrino Model
--------------
The neutrino model as the central engine of GRBs has been widely discussed (PWF99; Narayan, Piran & Kumar 2001, hereafter NPK01; Kohri & Mineshige 2002; DPN02; Chen & Beloborodov 2007; Janiuk et al. 2004, 2007; Gu et al. 2006; Liu et al. 2007, 2015; Lei et al. 2009; Xie et al. 2016; for a review see Liu et al. 2017). The typical mass accretion rate in such a model is high ($0.01$ to $10 M_{\sun} s^{-1}$). Under such a condition, the gas photon opacity is also very high and radiation becomes trapped (Katz 1977; Begelman 1978; Abramowicz et al. 1988). However, neutrinos can still escape and tap the thermal energy of the disk produced by viscous dissipation before being advected into the BH. In this model, GRBs are powered by the energy liberated via the $\nu\bar{\nu}\rightarrow e^+e^-$ process in regions of low baryon density.
DPN02 showed that the neutrino emission will be greatly suppressed by neutrino trapping for an accretion rate $\dot{M} \geq 1 M_{\sun} s^{-1}$. However, their results are based on a Newtonian disk model. Gu et al. (2006), Chen & Beloborodov (2007) and Lei et al. (2010) argued that the general relativistic effects are also important. In this paper, we adopt a model of a steady-state disk around a Kerr BH, in which neutrino loss and transfer are taken into account.
The accretion rate likely varies at the central engine of a GRB. As a first step, we assume a constant mass accretion rate to get the general properties of an NDAF, and leave the study of the evolution of the disk in Sections 3 and 4.
Because the gas cools efficiently, we are entitled to discuss the NDAF model within the context of a thin disk (Sharkura & Sunyaev 1973). The accuracy of the thin-disk approximation is not perfect at large radii, where the disk is thick. On the other hand, the details of the outer region have little effect on the solution for the neutrino-cooled disk (Chen & Beloborodov 2007).
The basic equations of NDAF (equations for continuity, state, conservation of angular momentum and energy balance) in the Kerr metric are given as follows (PWF99; DNP02; Reynoso, Romero & Sampayo 2006; Lei et al. 2009), $$\dot {M} = - 4\pi r v_r \rho H,
\label{eq:GRNDAF1}$$
$$\dot{M} \sqrt {G M_\bullet r } \frac{D}{A} = 4\pi r^2H\alpha P\sqrt {\frac{A}{BC}},
\label{eq:GRNDAF2}$$
$$\begin{aligned}
P = & & \frac{11}{12} aT^4 + \frac{\rho kT}{m_{\rm p} } \left(\frac{1 + 3X_{nuc} }{4}\right) +
\frac{2\pi hc}{3} \left(\frac{3}{8\pi m_{\rm p} } \right)^{4 / 3} \nonumber \\
& & \left(\frac{\rho }{\mu _{\rm e} } \right)^{4 /3} + \frac{u_\nu }{3},
\label{eq:GRNDAF3}\end{aligned}$$
$$Q^ + = Q^ - ,
\label{eq:GRNDAF4}$$
where $H=\sqrt{Pr^3 B/(\rho G M_\bullet C)}$ is the half thickness of the disk, $v_r$ is the radial velocity of the gas, $\alpha$ is the viscosity parameter, $a$ is the radiation constant, $k$ is the gas Boltzmann constant, and $m_{\rm p}$ is the proton rest mass. $A, B, C, D$ and $f$ are the general relativistic correction factors for a thin accretion disk around a Kerr BH (Riffert [&]{} Herold 1995). $$\label{eqA}
A = 1 - \frac{2G M_\bullet}{c^2r} + (\frac{G M_\bullet a_\bullet }{c^2r})^2,$$
$$\label{eqB}
B = 1 - \frac{3G M_\bullet}{c^2r} + 2a_\bullet (\frac{G M_\bullet}{c^2r})^{3 / 2},$$
$$\label{eqC}
C = 1 - 4a_\bullet (\frac{G M_\bullet}{c^2r})^{3 / 2} + 3(\frac{G M_\bullet a_\bullet
}{c^2r})^2,$$
$$\label{eqD}
D = B f,$$
where the BH spin parameter $a_\bullet=J_\bullet c/G M_\bullet^2$, and $M_{\bullet}$ and $J_\bullet$ are the BH mass and angular momentum, respectively. The expression for $f$ is given by Page & Thorne (1974) (in their Equation (15n)) as, $$\begin{aligned}
f & = & \frac{\chi^2}{(\chi^3-3\chi+2a_\bullet)} [\chi-\chi_{\rm ms} -\frac{3}{2}a_\bullet \ln(\frac{\chi}{\chi_{\rm ms}})- \nonumber \\
& & \frac{3(\chi_1-a_\bullet)^2}{\chi_1(\chi_1-\chi_2)(\chi_1-\chi_3)} \ln(\frac{\chi-\chi_1}{\chi_{\rm ms}-\chi_1}) - \nonumber \\
& & \frac{3(\chi_2-a_\bullet)^2}{\chi_2(\chi_2-\chi_1)(\chi_2-\chi_3)} \ln(\frac{\chi-\chi_2}{\chi_{\rm ms}- \chi_2})- \nonumber \\
& & \frac{3(\chi_3-a_\bullet)^2}{\chi_3(\chi_3- \chi_1)(\chi_3-\chi_2)} \ln(\frac{\chi-\chi_3}{\chi_{\rm ms}-\chi_3}) ],\end{aligned}$$ where $\chi=(r/r_{\rm g})^{1/2}$, $\chi_{\rm ms}=(r_{\rm ms}/r_{\rm g})^{1/2}$, and $r_{\rm g}=G M_\bullet/c^2$. The radius of the marginally stable orbit is (Bardeen et al. 1972) $$r_{\rm ms}=r_{\rm g} [3+Z_2 - sgn(a_\bullet)[(3-Z_1)(3+Z_1+2Z_2)]^{1/2}],$$ for $0\leq a_{\bullet} \leq 1$, where $Z_1 \equiv 1+(1-a_\bullet^2)^{1/3} [(1+a_\bullet)^{1/3}+(1-a_\bullet)^{1/3}], \ \ Z_2\equiv (3a_\bullet^2+Z_1^2)^{1/2}$, and $\chi_1= 2 \cos(\frac{1}{3} \cos^{-1}a_\bullet - \pi/3)$, $\chi_2= 2 \cos(\frac{1}{3} \cos^{-1}a_\bullet + \pi/3)$, $\chi_3=- 2 \cos(\frac{1}{3} \cos^{-1}a_\bullet )$ are the three roots of $\chi^3 - 3\chi+ 2 a_\bullet = 0$. It is easy to check that $f(r=r_{\rm ms})=0$ and $f(r \gg r_{\rm ms})\sim 1-\sqrt{r_{\rm ms}/r}$.
In Equation (\[eq:GRNDAF3\]), the total pressure consists of four terms, radiation pressure, gas pressure, degeneracy pressure, and neutrino pressure. The factor $11/12$ in the term of radiation pressure includes the contribution of relativistic electron-positron pairs. In the degeneracy pressure term, $\mu_e$ is the mass per electron, which is taken as 2 in agreement with NPK and PWF. $u_\nu $ is the neutrino energy density defined as (Popham [&]{} Narayan 1995) $$u_\nu = (7 / 8)aT^4\sum {\frac{\tau _{\nu _i } / 2 + 1 / \sqrt 3 }{\tau
_{\nu _i } / 2 + 1 / \sqrt 3 + 1 / (3\tau _{a,\nu _i } )}},
\label{eq:uv}$$ where $\tau _{\nu _i } = \tau _{a,\nu _i } + \tau _{s,\nu _i } $ is the sum of the absorptive and scattering optical depths calculated for each neutrino flavor $(\nu _e ,\nu _\mu ,\nu _\tau )$. The absorptive optical depths for the three neutrino flavors are (Kohri et al. 2005)
$$\tau _{a,\nu _e } \simeq 2.5\times 10^{ - 7}T_{11}^5 H + 4.5\times 10^{ -
7}T_{11}^2 X_{\rm nuc} \rho _{10} H,
\label{eq:tauae}$$
$$\tau _{a,\nu _\mu } = \tau _{a,\nu _\tau } \simeq 2.5\times 10^{ -
7}T_{11}^5 H,
\label{eq:tauvv}$$
where $T_{11}=T/10^{11} \rm K$, $\rho_{10} = \rho/10^{10} \rm g \ cm^{-3}$. $X_{\rm nuc} \simeq 34.8\rho _{10}^{ - 3 / 4} T_{11}^{9 / 8} \exp ( - 0.61 /T_{11} )$ is the mass fraction of free nucleons (PWF99; DPN02).
The total scattering optical depth is given by (DPN02) $$\tau _{s,\nu _i } \simeq 2.7\times 10^{ - 7}T_{11}^2 \rho _{10} H.
\label{eq:taus}$$
In Equation (\[eq:GRNDAF4\]), $Q^ + = Q_{\rm vis} $ represents viscous dissipation, and $Q^ - = Q_\nu + Q_{\rm photo} + Q_{\rm adv} $ is the total cooling rate due to neutrino losses $Q_\nu $, photodisintegration $Q_{\rm photo} $ and advection $Q_{\rm adv} $. We employ a bridging formula for calculating $Q_\nu $, which is valid in both the optically thin and thick cases. The expressions for $Q_\nu $, $Q_{\rm photo} $ and $Q_{\rm adv} $ are (DPN02)
$$Q_\nu = \sum {\frac{(7 / 8\sigma T^4)}{(3 / 4)(\tau _{\nu _i } / 2 + 1 /
\sqrt 3 + 1 / (3\tau _{a,\nu _i } ))}} ,
\label{eq:Qv}$$
$$Q_{\rm photo} = 10^{29}\rho _{10} v_r \frac{dX_{\rm nuc} }{dr}H\mbox{ } {\rm erg \cdot
cm^{-2}s^{-1} },
\label{eq:Qphoto}$$
$$Q_{\rm adv} \simeq v_r \frac{H}{r}(\frac{11}{3}aT^4 + \frac{3}{2}\frac{\rho
kT}{m_{\rm p} }\frac{1 + X_{\rm nuc} }{4} + \frac{4u_\nu }{3}),
\label{eq:Qadv}$$
The heating rate $Q_{\rm vis} $ is expressed as $$Q_{\rm vis} = \frac{3G M_\bullet \dot{M} }{8\pi r^3} f.
\label{eq:Qvis}$$ We solve numerically Equations (\[eq:GRNDAF1\])-(\[eq:Qvis\]) to find the disk temperature $T$ and density $\rho $ versus the disk radius given $a_\bullet $, $m$ and $\dot {m}$ (where $m_\bullet=M_\bullet/M_{\sun}$, and $\dot {m}=\dot{M}/M_{\sun} s^{-1}$). We take $X_{\rm nuc} = 1$ for fully photodisintegrated nuclei, which is appropriate in the inner disk. Furthermore, $\alpha = 0.1$ is adopted.
In the calculations, we ignore the cooling rate arising from photodisintegration, because it is much less than the neutrino cooling rate in the inner disk (Janiuk et al. 2004). We also approximately take the free nucleon fraction $X_{\rm nuc} \simeq 1$. For the disks formed by the collapses of massive stars, the photodisintegration process that breaks down $\alpha$-particles into neutrons and protons is important in the disk region at very large radii. However, the effect of photodisintegration becomes less significant for the region at small radii, which contains fewer $\alpha$-particles. See Kohri et al. (2005), Chen & Beloborodov (2007) and Liu et al. (2007) for details, which showed that photodisintegration is not important for $r \leq 10^2 r_g$. On the other hand, for disks formed by the mergers of compact star binaries, we reasonably take all the nucleons to be free ($X_{\rm nuc} \simeq 1$) and neglect the photodisintegration process, since we mainly focus on the inner region of the disk.
The neutrino power from the accretion flow is given by, $$\label{eq26}
\dot{E}_\nu = 4\pi \int_{r_{\rm ms} }^{r_{\rm out} } {Q_\nu rdr} .$$
We are interested primarily in the properties of the inner accretion flow, where neutrino processes are important. As argued by PWF99, NPK01 and DPN02, for $r > 100r_g $, the neutrino cooling is not important and photons are completely trapped. The flows are fully advection-dominated at that region. We therefore concentrate on discussing in the region from $r_{\rm ms} $ to $r_{\rm max } = 100 r_{\rm g} $.
In order to get the neutrino annihilation power, we model the disk as a grid of cells in the equatorial plane. A cell $k$ has its neutrino mean energy $\varepsilon _{\nu _i }^k $ and luminosity $l_{\nu _i }^k $, and the height above (or below) the disk is $d_k $. The angle at which neutrinos from cell $k$ encounter anti-neutrinos from another cell $k'$ at that point is denoted as $\theta _{k{k}'} $. Then the neutrino annihilation power at that point is given by the summation over all pairs of cells (PWF99; Rosswog et al. 2003),
$$\begin{aligned}
\dot{E}_{\nu \bar{\nu}} = & & A_1 \sum_k \frac{l^k_{\nu_i}}{d_k^2} \sum_{k'}
\frac{l^k_{\nu_i}}{d_{k'}^2}
(\epsilon^k_{\nu_i}+\epsilon^{k'}_{\bar{\nu}_i})(1-cos\theta_{kk'})^2
+ \nonumber \\
& & A_2 \sum_k \frac{l^k_{\nu_i}}{d_k^2} \sum_{k'} \frac{l^k_{\nu_i}}{d_{k'}^2} \frac{\epsilon^k_{\nu_i}+\epsilon^{k'}_{\bar{\nu}_i}}{\epsilon^k_{\nu_i} \epsilon^{k'}_{\bar{\nu}_i}}(1-cos\theta_{kk'})\end{aligned}$$
where $A_1 \approx 1.7\times 10^{ - 44} \ {\rm cm \cdot erg^{-2} \cdot s^{-1} }$ and $A_2 \approx 1.6\times 10^{ - 56} \ {\rm cm \cdot erg^{ - 2}s^{-1} }$.
The total neutrino annihilation luminosity is obtained by integrating over the whole space outside the BH and the disk. As a typical case, we show the results of neutrino power $\dot{E}_{\nu}$ (left panel) and neutrino annihilation power $\dot{E}_{\nu\bar{\nu}}$ (right panel) for a BH with mass $m_\bullet=3$ and with different accretion rate and BH spin in Figure \[fig\_dotEvv\] (points in the figure), in which $\alpha = 0.1 $ is adopted. For comparison, we also show the results from previous works, such as PWF99 (open symbols), and Xue et al. (2013, filled symbols). Inspecting Figure \[fig\_dotEvv\], one finds that the results by PWF99 overestimate the neutrino annihilation power in the high accretion rate region. A reasonable understanding for this disagreement is the lack of neutrino trapping in PWF99 solutions.
Generally, our resulting curves (thick dotted lines in Figure \[fig\_dotEvv\]) exhibit broken a power law shape with two breaks. The first break marks the transition of the inner disk from neutrino-dominated to advection-dominated. Following Chen & Beloborodov (2007), we take the accretion rate at this break as $\dot{m}_{\rm ign}$, i.e., the disc temperature is not high enough to ignite the neutrino emitting reactions if $\dot{m} <\dot{m}_{\rm ign}$. The second break is due to the neutrino trapping effects (see DPN02 for details), and the corresponding accretion rate is denoted by $\dot{m}_{\rm trap}$. If $\dot{m} > \dot{m}_{\rm trap}$, the emitted neutrinos become trapped in the disc and advected into the BH. Therefore, for convenience, we summarize our numerical results with smooth power law fits with two breaks (shown with solid lines in Figure \[fig\_dotEvv\]), i.e., $$\begin{aligned}
\dot{E}_{\nu} \simeq && \dot{E}_{\nu, \rm ign} \left[\left(\frac{\dot{m}}{\dot{m}_{\rm ign}} \right)^{-\alpha_\nu } + \left(\frac{\dot{m}}{\dot{m}_{\rm ign} }\right)^{-\beta_\nu } \right]^{-1} \nonumber \\
&& \times \left[1+\left(\frac{\dot{m}}{\dot{m}_{\rm trap} } \right)^{\beta_\nu - \gamma_\nu } \right]^{-1} ,
\label{eq_Ev}\end{aligned}$$
$$\begin{aligned}
\dot{E}_{\nu \bar{\nu}} \simeq && \dot{E}_{\nu \bar{\nu}, \rm ign} \left[ \left(\frac{\dot{m}}{\dot{m}_{\rm ign}} \right)^{-\alpha_{\nu \bar{\nu}} } + \left(\frac{\dot{m}}{\dot{m}_{\rm ign} } \right)^{-\beta_{\nu \bar{\nu}} } \right]^{-1} \nonumber \\
&& \times \left[1+(\frac{\dot{m}}{\dot{m}_{\rm trap} })^{\beta_{\nu \bar{\nu}} - \gamma_{\nu \bar{\nu}} }\right]^{-1} ,
\label{eq_Evv}\end{aligned}$$
where, $$\begin{aligned}
&&\left\lbrace
\begin{tabular}{l}
$\dot{E}_{\nu, \rm ign}=10^{(51.4-0.3 a_\bullet^2)} \left(\frac{m_\bullet}{3} \right)^{\log(\dot{m}/\dot{m}_{\rm ign}) -1.5} {\rm erg \ s^{-1}}, $ \\
$\alpha_\nu = 2.3, \ \beta_\nu = 1.12, \ \gamma_\nu = 0.4, $
\end{tabular}
\right.\nonumber \\
&&\left\lbrace
\begin{tabular}{l}
$\dot{E}_{\nu \bar{\nu}, \rm ign}=10^{(48.0+0.15 a_\bullet)} \left(\frac{m_\bullet}{3} \right)^{\log(\dot{m}/\dot{m}_{\rm ign}) -3.3} {\rm erg \ s^{-1}}, $ \\
$\alpha_{\nu \bar{\nu}} = 4.7, \ \beta_{\nu \bar{\nu}} = 2.23, \ \gamma_{\nu \bar{\nu}} =0.3, $
\end{tabular}
\right.\nonumber \\
&& \dot{m}_{\rm ign} = 0.07-0.063 a_\bullet, \ \dot{m}_{\rm trap} = 6.0-4.0 a_\bullet^3,\end{aligned}$$ where $\dot{m}_{\rm ign}$ and $\dot{m}_{\rm trap}$ are the igniting and trapping accretion rates, respectively. For $m_\bullet=3$ and $\alpha=0.1$, $\dot{m}_{\rm ign}=0.07$ and $\dot{m}_{\rm trap}=6.0$ for $a_\bullet=0$, and $\dot{m}_{\rm ign}=0.01$ and $\dot{m}_{\rm trap}=2.6$ for $a_\bullet=0.95$. Similar results are obtained by Kohri et al. (2005) and Chen & Beloborodov (2007).[^1]
{width="8cm"} {width="8cm"}
With the second terms in $\dot{E}_{\nu, \rm ign}$ and $\dot{E}_{\nu \bar{\nu}, \rm ign}$, our analytical solutions can also apply to an NDAF with the BH mass in the range from $m_\bullet=3$ to $10$. To illustrate the accuracy of these power law fits we compare our fits (solid lines) with the numerical solutions and with the analytic formula obtained by several other authors, such as Fan et al. (2005, thin dotted lines) [^2] and Zalamea & Beloborodov (2011, thin dashed lines). [^3] in Figure \[fig\_dotEvv\].
From Figure \[fig\_dotEvv\], we find that our analytical solution (Equation (\[eq\_Evv\])) agrees quite well with the analytical solution by Zalamea & Beloborodov (2011) for $\dot{m}>\dot{m}_{\rm ign}$, and the numerical solution by PWF99 (or the analytical one by Fan et al. 2005) for small BH spin and low accretion rates. Zalamea & Beloborodov (2011) did not treat the NDAF with $\dot{m}<\dot{m}_{\rm ign}$, and roughly setted $\dot{E}_{\nu\bar{\nu}}$ as constant for $\dot{m}>\dot{m}_{\rm trap}$. Fan et al. (2005) only fitted for $0.01<\dot{m}<0.1$. Our analytical solutions, however, cover all the three regions (the whole range of accretion rate) rather smoothly. Therefore, for convenience, we will adopt our analytical solutions (i.e., Equations (\[eq\_Ev\]) and (\[eq\_Evv\])) directly in the following calculations.
The baryon loading of jet is the fundamental problem in GRBs. In Paper I (Lei et al. 2013), we obtained the baryon loading rate for the jet driven by neutrino-annihilation, $$\begin{aligned}
\dot{M}_{\rm j, \nu\bar{\nu}} & \simeq & 7.0 \times 10^{-7} A^{0.85} B^{-1.35} C^{0.22} \theta_{\rm j,-1}^2 \alpha_{-1}^{0.57} \epsilon_{-1}^{1.7} \nonumber \\
& & \left( \frac{R_{\rm ms} }{2} \right) ^{0.32} \dot{m}_{-1}^{1.7} \left(\frac{m_\bullet}{3}\right)^{-0.9} \left(\frac{\xi}{2} \right)^{0.32} \ M_{\sun} {\rm s}^{-1}. \nonumber \\\end{aligned}$$ for $\dot{m}>\dot{m}_{\rm ign}$, where $\theta_{\rm j}$ is the jet half opening angle, $\xi \equiv r/r_{\rm ms}$ is the disk radius in uints of $r_{\rm ms}$, $\epsilon$ is the neutrino emission efficiency, i.e., $\epsilon = \dot{E}_{\nu} /\dot{M}c^2$. For $\dot{m} <\dot{m}_{\rm ign}$, the neutrino cooling becomes unimportant. The dependence of $\dot{M}_{\rm j, \nu\bar{\nu}} $ on the accretion rate $\dot{m}$ might be replaced with $\dot{M}_{\rm j, \nu\bar{\nu}} \propto \dot{m}^{3.8}$.
We can thus define an important quantity in GRB central engine, the dimensionless “entropy” parameter $\eta$ as $$\eta \equiv \frac{\dot{E}_{\rm m}}{ \dot{M}_{\rm j, \nu\bar{\nu}} c^2 }.
\label{eta}$$ where $\dot{E}_{\rm m} = \dot{E}_{\nu \bar{\nu}}+ \dot{M}_{\rm j, \nu\bar{\nu}} c^2$ is the total matter energy outflow luminosity.
This $\eta$ parameter describes the maximum available Lorentz factor in neutrino annihilation model (supposing that the neutrino annihilation energy is totally converted into kinetic energy of baryons after acceleration), i.e., $\Gamma_{\rm max} \simeq \eta$.
To evolve these central engine parameters (such as $\dot{E}_{\nu \bar{\nu}}$ and $\eta$) with time, we need to consider the evolution of BH, since most of these parameters have significant dependences on the BH spin. During the hyper-accreting process, the equations for BH evolution are, $$\frac{dM_\bullet c^2}{dt} = \dot{M} E_{\rm ms},
\label{dMvv}$$
$$\frac{dJ_\bullet}{dt} = \dot{M} L_{\rm ms},
\label{dJvv}$$
where $E_{\rm ms}$ and $L_{\rm ms}$ are the specific energy and the specific momentum corresponding to the inner most radius $r_{\rm ms}$ of the disk, which are defined in Novikov & Thorne (1973) as $E_{\rm ms} = (4\sqrt{ R_{\rm ms} }-3a_{\bullet}) /(\sqrt{3} R_{\rm ms})$, $L_{\rm ms} = (G M_\bullet/c) (2 (3 \sqrt{R_{\rm ms}} -2 a_\bullet) )/(\sqrt{3} \sqrt{R_{\rm ms}} )$, where $R_{\rm ms} = r_{\rm ms}/r_{\rm g} $.
As $a_\bullet = J_\bullet c/(G M_\bullet^2)$, by incorporating the above two equations, we find that the BH will be spun up by the accretion with a rate as $$\frac{da_\bullet}{dt} = \dot{M} L_{\rm ms} c/(G M_\bullet^2) -2 a_\bullet \dot{M} E_{\rm ms} /(M_\bullet c^2)$$
The duration of the burst, in such a model, is determined by the viscous timescale of the accreting gas. In most accretion flows, the viscous time is significantly longer than the dynamical time, so the accretion model naturally explains the large difference between the duration of bursts and their minimum variability timescales.
Another topic of NDAF is about its stability, since it will shape the GRB lightcurve. The stability properties of NDAFs were first discussed by NPK01. They found that their NDAF is unstable only if it is optically thin and radiation pressure dominated, which could conceivably play a role in determining the temporal behavior of some bursts. In other cases, their NDAF solution is viscously, thermally and gravitationally stable. After considering neutrino trapping, DPN02 found that NDAFs are viscously and thermally stable, but are only gravitationally unstable for an extremely large accretion rate like $\dot{m} \sim 10$ and for $r \geq 50$. By including microphysics and photodisintegration, Janiuk et al. (2007) suggested that for sufficiently large accretion rates ($\dot{m} \geq 10$), the inner regions of the disk become opaque and develop a viscous and thermal instability. However, these models did not consider the effect of magnetic fields. Lei et al. (2009) pointed out that an NDAF torqued by magnetic coupling is viscously and thermally unstable for $\dot{m} \geq 0.086$. Janiuk and Yuan (2010) extended their work by introducing the BH spin and magnetic field. It is shown that the instability can occur when $\dot{m} \geq 0.5$ for a fastly spinning BH. Recently, Xie et al. (2016) suggested that the inner-boundary torque should be taken into account for NDAFs, and obtained an unstable solution as a possible interpretation for the variability of GRB prompt emission and X-ray flares. Shibata et al. (2007), on the other hand, performed an axisymmetric general relativity magnetohydrodynamic (GRMHD) simulation for neutrino-cooled accretion tori around a rotating BH. Their results suggest that the angular momentum transport and the consequent shock heating caused by magnetic stress will induce a time-varying neutrino power, which is favorable for explaining the variability of GRB lightcurves.
Magnetic Model
--------------
Blandford & Znajek (1977) proposed that the rotating energy and the angular momentum of a BH can be extracted by a surrounding magnetic field, and this energy mechanism has been referred to as the BZ process. If the magnetic field of BH is strong enough ($\sim 10^{15} {\rm G} $), the rotational energy extracted by this process can power GRBs (Paczy$\acute{n}$ski 1998; Mészáros & Rees 1997; Paper I; Tchekhovskoy & Giannios 2015). On the other hand, researches showed that the magnetic fields can be magnified up to $10^{15} \sim 10^{16} {\rm G}$ by virtue of MRI or dynamo process (Pudritz & Fahlman 1982 and references therein) in hyperaccretion disks.
The rotational energy of a BH with angular momentum $J_\bullet$ is a fraction of the BH mass $M_\bullet$, $$E_{\rm rot}= 1.8 \times 10^{54} f_{\rm rot}(a_\bullet) \frac{M_\bullet}{M_\sun} {\rm erg},$$
$$f_{\rm rot}(a_\bullet)=1-\sqrt{(1+\sqrt{1-a_\bullet^2})/2 },$$
For a maximally rotating BH ($a_\bullet=1$), $f_{\rm rot}(1)=0.29$.
The BZ jet power from a BH with mass $M_{\bullet}$ and angular momentum $J_\bullet$ is (Lee et al. 2000; Li 2000; Wang et al. 2002; McKinney 2005; Lei et al. 2005; Lei & Zhang 2011; Lei et al. 2013) $$\begin{aligned}
\dot{E}_{\rm B}=1.7 \times 10^{50} a_{\bullet}^2 m_{\bullet}^2
B_{\bullet,15}^2 F(a_{\bullet}) \ {\rm erg \ s^{-1}} \\ \nonumber
\simeq 1.1 \times 10^{50} a_{\bullet}^2 m_{\bullet}^2
B_{\bullet,15}^2 \ {\rm erg \ s^{-1}},
\label{eq_Lmag}\end{aligned}$$ where $B_{\bullet,15}=B_{\bullet}/10^{15} {\rm G}$ and $F(a_\bullet)=[(1+q^2)/q^2][(q+1/q) \arctan q-1]$. Here $q= a_{\bullet} /(1+\sqrt{1-a^2_{\bullet}})$, and $2/3\leq F(a_{\bullet}) \leq \pi-2$ for $0\leq a_{\bullet} \leq 1$. It apparently depends on $M_{\bullet}$, $B_{\bullet}$, and $a_{\bullet}$. A strong magnetic field of the order $\sim 10^{15} \rm G$ is required to produce the high luminosity of a GRB. The accumulation of magnetic flux by an accretion flow may account for such a high magnetic field strength (e.g., Tchekhovskoy et al. 2011).
The dependence of $\dot{E}_{\rm B}$ on BH spin is shown in Figure \[fig:dotEBa\]. For comparison, we also plot the expressions given by BZ77 (derived in the limit $a_\bullet \ll 1$, but widely used, e.g., Thorne et al. 1986, PWF99)[^4] and by Tchekhovskoy et al. 2011) [^5]. It is found that the BZ power with the formula adopted here is quite close to that given by Tchekhovkoy (2011). However, the BZ77 expression can only apply to the case with low BH spin. Similar results were also obtained by recent GRMHD numerical simulations (Nagataki 2009, 2011).
![The magnetic power $\dot{E}_{\rm B} $ as a function of BH spin $a_\bullet$. Solid line is the result with Equation (\[eq\_Lmag\]). We also plot the result from Tchekhovskoy, Narayan & McKinney (2011) and BZ77 with dashed and dotted lines, respectively. In our calculations, we adopt the magnetic flux $\Phi_{\rm BH}=10^{27} \rm G \ cm^2$, BH mass $m_\bullet=3$, and $\kappa=0.053$. []{data-label="fig:dotEBa"}](fig_2.eps){width="8cm"}
The total magnetic torque applied on the BH is (Lee et al. 2000; Li 2000; Wang et al. 2002; McKinney 2005; Lei et al. 2005; Lei & Zhang 2011; Lei et al. 2013) $$\begin{aligned}
T_{\rm B} = \frac{\dot{E}_{\rm B}}{\Omega_{\rm F}}=3.4 \times 10^{45} a_\bullet^2 q^{-1} m_\bullet^3 B_{\bullet,15}^2 F(a_\bullet) {\rm \ g \ cm^2 \ s^{-2}}, \nonumber \\\end{aligned}$$ where $\Omega_{\rm F}=0.5\Omega_\bullet$ is usually taken to maximize the BZ power, and $$\Omega_\bullet =\frac{a_\bullet c}{2 r_\bullet}= \frac{c^3}{G M_\bullet} \frac{a_\bullet}{2 (1+\sqrt{1-a_\bullet^2})}$$ is the angular velocity of BH horizon.
The spin-down timescale by the BZ process can be estimated as (Lee et al. 2000; Lei et al. 2005) $$t_{\rm spindown} \simeq \frac{E_{\rm rot}}{\dot{E_{\rm B}} } \simeq 2.7 \times 10^3 {\rm s} \times B_{\rm \bullet, 15}^{-2} m_\bullet^{-1} .
\label{tspindown}$$ One can find that $t_{\rm spindown}$ is not sensitive to the initial BH spin, since both the rotational energy and spin power depend on it.
Considering a BH with an initially spin $a_{\bullet}(0)$ is slowing down by the BZ mechanism to a final spin $a_{\bullet,f}=0$. The final BH mass is then given by $$M_\bullet=M_{\bullet}(0) \exp {\int_{a_{\bullet}(0)}^0 \frac{-1}{2a_\bullet-4/q} da_\bullet}.$$ If $a_{\bullet}(0)=1$, the final BH mass will be $M_\bullet=(e^{1/4}/{\sqrt{2}})M_{\bullet}(0)=0.91 M_{\bullet}(0) $. We see that $9\%$ of the initial mass or $31\%$ of the rotational energy can be used to power GRB from the maximally rotating BH. The extracted energy is therefore less than a half of the initial rotational energy. Other energy increase the irreducible mass of the BH. For $a_{\bullet}(0)=0.5$, $M_\bullet=0.98 M_{\bullet}(0)$ or $2\%$ of the initial mass can be used to power a GRB[^6].
As the magnetic field on the BH is supported by the surrounding disk, there are some relations between $B_{\bullet}$ and $\dot{M}$. In a hyper-accreting flow in a GRB, it is possible that a magnetic flux is accumulated near the black hole horizon. Considering the balance between the magnetic pressure on the horizon and the ram pressure of the innermost part of the accretion flow (e.g. Moderski et al. 1997), one can estimate the magnetic field strength threading the BH horizon $B_{\bullet}^2 /(8\pi) = P_{\rm ram} \sim \rho c^2 \sim \dot{M} c/(4\pi r_{\bullet}^2)$, where $r_{\bullet}=(1+\sqrt{1-a_\bullet^2})r_{\rm g}$ is the radius of BH horizon. One thus has $$B_{\bullet} \simeq 7.4 \times 10^{16} \dot{m}^{1/2} m_\bullet^{-1} \left(1+\sqrt{1-a_\bullet^2} \right)^{-1} \rm{G}.
\label{Bmdot}$$ Inserting it into Equation (\[eq\_Lmag\]), we obtain the magnetic power and torque as a function of mass accretion rate and BH spin, i.e. $$\begin{aligned}
\dot{E}_{\rm B} & = & 9 \times 10^{53} a_\bullet^2 \dot{m} X(a_\bullet) \ {\rm erg \ s^{-1}} \nonumber \\
& \simeq & 1.5 \times 10^{53} a_\bullet^2 \dot{m} \ {\rm erg \ s^{-1}},\end{aligned}$$ $$\begin{aligned}
T_{\rm B} & = & 1.8 \times 10^{49} a_\bullet \dot{m} m_\bullet F(a_\bullet) \ {\rm \ g \ cm^2 \ s^{-2}} \nonumber \\
& \simeq & 1.2 \times 10^{49} a_\bullet \dot{m} m_\bullet \ {\rm \ g \ cm^2 \ s^{-2}},\end{aligned}$$ where $X(a_\bullet)=F(a_\bullet)/(1+\sqrt{1-a_\bullet^2} )^2$. It is found that $X(0)=1/6$, and $X(1)=\pi -2$.
Both neutrino annihilation and magnetic power depend on the disk mass accretion. In Figure \[fig\_Eb\_mdot\], we present the BZ power as function of accretion rate for different BH spin and compare it with the neutrino annihilation power. We find that: (1) the magnetic power is much greater than the neutrino annihilation power for a moderate to high spin BH; (2) The neutrino annihilation power dominates over the BZ power for BHs with a very small spin at high accretion rates; (3) compared with the magnetic power, $\dot{E}_{\nu \bar{\nu}}$ is much more sensitive to the mass accretion rate $\dot{m}$. Therefore, if the disk accretion rate is variable, the jet driven by the neutrino annihilation process should be highly variable. However, the MHD jet is usually subject to instabilities, such as kink instability (Wang et al. 2006) and magnetic reconnection (e.g. the Internal-Collision induced Magnetic Reconnection and Turbulence or ICMART in Zhang & Yan 2011). These MHD processes will add complexity to the GRB lightcurves.
![The magnetic power $\dot{E}_{\rm B} $ as a function of accretion rate for different BH spin $a_\bullet =0.01$ (thick red solid line) and 0.99 (thick blue solid lines). The dashed lines show neutrino annihilation power $\dot{E}_{\nu\bar{\nu}}$ calculated with Equation (\[eq\_Evv\]) for $a_\bullet =0.01$ (red dashed line) and 0.99 (blue dashed line). The thin blue line is produced with the analytic expression of Tchekhovskoy, Narayan & McKinney (2011) for $a_\bullet =0.99$, where the average magnetic flux $\langle\Phi_{\rm BH}^2/(\dot{M} r_{\rm g}^2)c\rangle^{1/2} \simeq 47$ and $\kappa=0.044$ are taken based on the numerical simulation Model A0.99f.[]{data-label="fig_Eb_mdot"}](fig_3.eps){width="8cm"}
In the magnetic model, baryons from the disk will be suppressed by the strong magnetic field lines. For $\dot{m}>\dot{m}_{\rm ign}$, the baryon loading rate for the BZ driven jet is (Paper I), $$\begin{aligned}
\dot{M}_{\rm j,BZ} & \simeq & 3.5 \times 10^{-7} A^{0.58} B^{-0.83} f_{\rm p,-1}^{-0.5} \theta_{\rm j,-1} \theta_{\rm B,-2}^{-1} \nonumber \\
& & \alpha_{-1}^{0.38} \epsilon_{-1}^{0.83} \dot{m}_{-1}^{0.83} \left(\frac{m_\bullet}{3}\right)^{-0.55} r_{z,11}^{0.5} \ M_{\sun} {\rm s}^{-1}.
\label{Eq_mdotj_BZ}\end{aligned}$$ For $\dot{m} <\dot{m}_{\rm ign}$, the dependence of $\dot{M}_{\rm j, \nu\bar{\nu}} $ on $\dot{m}$ will be $\dot{M}_{\rm j, BZ} \propto \dot{m}^{1.9}$. In Equation (\[Eq\_mdotj\_BZ\]), $f_{\rm p}$ is the fraction of protons, $r_{\rm z}$ is the distance from the BH in the jet direction, which is normalized to $10^{11}$cm. Because of the existence of a strong magnetic field, protons with an ejected direction larger than $\theta_{\rm B}$ with respect to the field lines would be blocked.
We can then define a parameter denoting the maximum available energy per baryon in the jet driven by the BZ process, $$\mu_0 \equiv \frac{\dot{E}}{\dot{M}_{\rm j,BZ} c^2} = \frac{\dot{E}_{\rm m}+\dot{E}_{\rm B}}{\dot{M}_{\rm j,BZ} c^2} = \eta (1+\sigma_0),
\label{eq:mu}$$ where $\dot{E}_{\rm m} = \dot{E}_{\nu \bar{\nu}}+ \dot{M}_{\rm j,BZ} c^2$, and $\sigma_0 = \dot E_{\rm B}/\dot E_{\rm m}$.
The acceleration behavior of the jet is subject to uncertainties. Generally, the jet will reach a terminating Lorentz factor $\Gamma$ that satisfies $$\Gamma_{\rm min} < \Gamma < \Gamma_{\rm max},
\label{eq_gamma_bz}$$ with the explicit value depending on the detailed dissipation process, such as kink instability (Wang et al. 2006), ICMART (Zhang & Yan 2011) and magnetic dissipate due to the shearing interaction between two component jets (e.g. Wang et al. 2014). In Equation (\[eq\_gamma\_bz\]), $\Gamma_{\rm min}=\max(\mu_{0}^{1/3},\eta)$ ($\eta=\dot{E}_{\nu \bar{\nu}} /(\dot{M}_{\rm j, BZ} c^2)$) and $\Gamma_{\rm max} = \mu_0$, which correspond to the beginning and the end of the slow acceleration phase in a hybrid outflow, respectively (see Gao & Zhang 2015 for a detailed discussion of the acceleration dynamics of an arbitrarily magnetized relativistic or hybrid jet).
As to the evolution of BH, we should consider both the accretion and BZ processes. The evolution equations are given by $$\frac{dM_\bullet c^2}{dt} = \dot{M} c^2 E_{\rm ms} - \dot{E}_{\rm B},
\label{dMbz}$$
$$\frac{dJ_\bullet}{dt} = \dot{M} L_{\rm ms} - T_{\rm B}
\label{dJbz}$$
the evolution equation for the BH spin is then $$\begin{aligned}
\frac{da_\bullet}{dt} = && (\dot{M} L_{\rm ms} - T_{\rm B})c/(G M_\bullet^2) - \nonumber \\
&& 2 a_\bullet (\dot{M} c^2 E_{\rm ms} - \dot{E}_{\rm B}) /(M_\bullet c^2)\end{aligned}$$
As a BH may be spun up by accretion or spun down by the BZ mechanism, the BH spin will reach an equilibrium value when $da_\bullet/dt =0$. If the magnetic field is related to the mass accretion rate as Equation (\[Bmdot\]), the final BH spin will be $a_\bullet^{\rm eq} \sim 0.87$.
The evolution of BH spin combining with the accretion profile will give rise to a reasonable GRB lightcurve. In addition, possible jet pression (Lei et al. 2007), episodic jet (Yuan & Zhang 2012) and episodic accretion (by magnetic barrier, see Proga & Zhang 2006; or by magnetically arrested disk (MAD), see Lloyd-Ronning et al. 2016) would enrich the structure of lightcurve.
Prompt Emission Phase
=====================
Now we apply the above theory to GRBs. Firstly, we study the prompt emission phase. During the this stage, the BH accretes the main part of the disk with a high accretion rate. We begin with a BH of mass $M_\bullet(0)=3 M_{\sun}$, spin $a_{\bullet}(0)$, accretion rate $\dot{M}(0)$ and with a disk of mass $M_{d}(0)$. Other parameters are taken their typical values ($r_{\rm z}=10^{11}cm, f_{\rm p}=0.1, \theta_{\rm j}=0.1, \theta_{\rm B}=0.01$).
To obtain the accretion rate profile, we adopt a simple model described in Kumar et al. (2008a, 2008b) and Metzger et al. (2008). In this model, the disk are treated as a single annulus ring with effective disk radius $r_d$, which is defined as $$j(r_{\rm d}) = (G M_\bullet r_{\rm d})^{1/2} = \frac{J_{\rm d}}{M_{\rm d}}
\label{Eq:jd}$$ where $M_{\rm d}$ and $J_{\rm d}$ are the total mass and angular momentum of the disk at time $t$. The accretion rate depends on the mass and accretion time-scale as $$\dot{M} = M_{\rm d}/t_{\rm acc}
\label{Eq:dMacc}$$ where $t_{\rm acc}=r_{\rm d}^2 / \nu \sim 2/(\alpha \Omega_K)$, and $\alpha$ is the dimensionless viscosity parameter(Shakura & Sunyaev 1973).
The mass and angular momentum of the disk change with time as $$\dot{M}_{\rm d} = - \dot{M}
\label{Eq:dMd}$$ $$\dot{J}_{\rm d} = - L_{\rm ms} \dot{M}
\label{Eq:dJd}$$
The evolutions of the BH are given by Equations (\[dMvv\])-(\[dJvv\]) for the neutrino annihilation model, and by Equations (\[dMbz\])-(\[dJbz\]) for the magnetic model.
Combing the evolution equations of the disk and the BH, one can get the values of $\dot{m}$, $m$ and $a_\bullet$ at each time step. With the formula obtained in Section 2, we can perform the evolution of the central engine parameters, such as $\dot{E}_{\nu \bar{\nu}}$, $\dot{E}_{\rm B}$, $\eta$ (for the neutrino model) and $\mu_0$ (for the magnetic model). The results are presented in Figures \[fig\_evl1\]-\[fig\_evl01\] for different sets of initial parameters.
{width="7cm"} {width="7cm"}\
{width="7cm"} {width="7cm"}
{width="7cm"} {width="7cm"}\
{width="7cm"} {width="7cm"}
{width="7cm"} {width="7cm"}\
{width="7cm"} {width="7cm"}
Figure \[fig\_evl1\] shows the case with an initial accretion rate of $\dot{m}(0)=1$ and initial disk mass of $m_{\rm d}(0)=0.1$. The parameters of the neutrino model and the magnetic model are plotted with dashed lines and solid lines, respectively. Different colors indicate distinct initial BH spin parameters, i.e., $a_{\bullet}(0)=0$ (red lines), 0.5 (green lines), and 0.95 (blue lines).
Top left panel exhibits the evolution of the accretion rate $\dot{m}$, which is insensitive to the BH parameters. So for the three examples exhibited in this figure, they share the same evolution curve for $\dot{m}$. The mass accretion rate decreases during the prompt phase due to angular momentum transfer. The vertical lines mark the igniting time $t_{\rm ign}$ when $\dot{m}$ becomes lower than the igniting accretion rate $\dot{m}_{\rm ign}(a)$, after which the neutrino cooling becomes unimportant.
For the neutrino model, the BH spin is always increasing until reaching the maximum spin $\sim 0.998$ if possible (see the dashed lines in the top right panel). For the magnetic model (solid lines in the top right panel), the evolution tracks have been divided into two branches by the equilibrium spin $a_\bullet^{\rm eq}$, i.e., the increasing branch for $a_\bullet(0)<a_\bullet^{\rm eq}$ (e.g., red and blue solid lines) and the decreasing branch for $a_\bullet(0)>a_\bullet^{\rm eq}$ (e.g., the blue solid line).
The jet power (lower left) at each time step depends on the values of accretion rate, BH spin and BH mass (the dependence on mass is weak). We find that the evolution of $\dot{E}$ generally tracts the accretion profile at late times since the evolution of the BH spin can be ignored when the majority of the disk mass is accreted. The evolution of $a_\bullet$ still has imprints on the $\dot{E}$ curve at earlier times, especially for $\dot{E}_{\rm B}$ with $a_\bullet(0)=0$ (red solid line in lower left panel). This case with lower $a_\bullet(0)=0$ is also an outlier in the three examples. Usually, we have $\dot{E}_{\rm B} > \dot{E}_{\nu \bar{\nu}}$ for all times. Only this one (the red lines) shows $\dot{E}_{\rm B} < \dot{E}_{\nu \bar{\nu}}$ at early times ($t<0.03$s). Our model, therefore, predicts that the jet composition can evolve from a thermally dominated jet to a magnetically dominated jet. Recently, the spectral study of GRB 160625B suggested a clear transition from fireball to Poynting flux dominated jet (Zhang et al. 2017), which might be an example of such a case.
For the parameters $\eta$ and $\mu_0$ (lower right panel), the evolution path in principal follows that of the jet power $\dot{E}$ before the igniting time $t_{\rm ign}$. Actually, such tracing properties are believed to be the physics behind the empirical relation $L_{\gamma} - \Gamma_0$ (Lü et al. 2012; Paper I; Yi et al. 2017). After $t_{\rm ign}$, the parameter $\mu_0$ begins to increase with time since the baryon loading rate drops very quickly in the BZ driven jet. For the case with $a_\bullet(0)=0.95$, we find a dip in evolution of $\mu_0$. It is worth mentioning that Gao & Zhang (2015) found a similar feature in the temporal profile of magnetic parameter $\sigma_0$ when analysing the data of GRB 110721A.
To illustrate the effects of disk mass, we present the results of the disk with an initial accretion rate of $\dot{m}(0)=1$ but with a large initial disk mass of $m_{\rm d}(0)=10$, as shown in Figure \[fig\_evl10\]. We find that the typical duration becomes longer compared with the first example (Figure \[fig\_evl1\]) since there are more masses to be accreted by the BH. For the same reason, $t_{\rm ign}$ is also greater. The bumps in the evolution curve of $\dot{E}$ represents the competition between the effects of accretion and BH spin.
In Figure \[fig\_evl01\], we study an example with a lower accretion rate. The duration becomes shorter because the flux is too weak to be observed at the final stage of accretion.
The results obtained here are based on a simple analytical model. There are a number of simulations on the GRB central engine (e.g., MacFadyen & Woosley 1999; Rosswog et al. 2003; Zhang et al. 2008; Janiuk et al. 2013; Janiuk 2017), which usually show complex behaviour of disk accretion. Direct comparisons between our results and theirs are beyond the scope of this paper. Rosswog (2007) presented an analytical model of the fall-back accretion of the bound debris based on previous 3D simulation of NS-NS and NS-BH mergers. Here, we adopt his results of the merger of NS-BH binaries with the NS mass fixed to $1.4 M_{\sun}$ and the BH mass adopted as $6 M_{\sun}$, $14 M_{\sun}$ and $16 M_{\sun}$, respectively. We estimate the fall-back accretion rate $\dot{m}_{\rm fb}$ from the fall-back accretion luminosity $L_{\rm acc}=dE_{\rm fb}/dt$ (Rosswog 2007), where $E_{\rm fb}$ denotes the difference between the potential plus kinetic energy at the start radius $r_{\rm i}$ and the potential energy at the dissipation radius $r_{\rm dis}$. Usually, the dissipation radius is taken as $r_{\rm dis}\simeq 10r_{\rm g}$ (Rosswog 2007). For comparison, we also plot the fall-back accretion rate with gray lines for the cases of different NS to BH mass ratios: $1.4:6$ (gray solid line), $1.4:14$ (gray dashed line), $1.4:16$ (gray dotted line) in Figure \[fig\_evl1\]. One can see that the evolution characteristics of the central engine presented here are generally consistent with those numerical simulation results.
Late Central Engine Activities
==============================
Many GRBs exhibit flares (Burrows et al. 2005; Chincarini et al. 2007; Falcone et al. 2007; Zhang 2007), plateaus (e.g. GRB 070110; Troja et al. 2007; Lyons et al. 2010; Lü & Zhang 2014; Lü et al. 2015; Gao et al. 2016a; Li et al. 2016; Chen et al. 2017), or giant bumps (e.g. GRB 121027A and GRB 111209A; Wu et al. 2013; Gao et al. 2016b) in X-ray lightcurves. These observations suggest that the GRB central engine is long-lived. Various models are invoked to interpret these activities, such as continuous energy injection from the spindown power of a magnetar, and the re-start of accretion onto a BH.
Here, we consider the BH central engine with fall-back accretion. The evolution of the fall-back accretion rate are described with a broken-power-law function of time as (Chevalier 1989; MacFadyen et al. 2001; Zhang et al. 2008; Dai & Liu 2012)
$$\begin{aligned}
\dot{M}_{\rm fb} = \dot{M}_{\rm p} \left[ \frac{1}{2}\left(\frac{t-t_0}{t_{\rm p}-t_0} \right)^{-1/2} + \frac{1}{2}\left(\frac{t-t_0}{t_{\rm p}-t_0} \right)^{5/3} \right]^{-1},
\label{dotm}\end{aligned}$$
where $t_0$ is the beginning time of the fall-back accretion in the local frame.
As an example, we assume a fall-back accretion starting at $t_0=1000$s, reaching the peak at $t_{\rm p}=1500$s. The peak accretion is adopted as $\dot{M}_{\rm p}=10^{-4} M_\sun s^{-1}$. Since $\dot{M}_{\rm p}$ is far below the igniting accretion rate, the neutrino annihilation power cannot explain the late time X-ray activities observed in both short and long GRBs (Fan et al. 2005). We ignore the contribution from neutrino annihilations, and assume that the jet is powered by the BZ process in the calculations. The baryon loading in this stage is quite uncertain since neutrino cooling is shut off and a strong wind kicks in, one cannot make robust predictions. In this paper, we do not calculate baryon loading and the parameter $\mu_0$ during the late BH central engine activity phase, although the jet is expected dirtier due to the strong disk wind expected in an advection dominated accretion flow (ADAF).
{width="7cm"} {width="7cm"}\
{width="7cm"}
First, we present the results of a fall-back accretion disk with rapid accretion surrounding a fast spinning BH $a_\bullet(0)=0.9$ (model I, thick solid lines in Figure \[fig\_evl\_late\]). The BH accretion just follows the fall-back rate, i.e., $\dot{M}=\dot{M}_{\rm fb}$. As shown in Figure \[fig\_evl\_late\], there is a weak evolution in the BH spin for this case. We find that the evolution of jet power just tracks that of the fall-back accretion rate.
If the viscosity parameter $\alpha$ is too small, the disk will undergo very slow accretion. We introduce a large viscosity timescale $\tau_{\rm vis}$ to model the slow accretion. The accretion rate onto the BH can be estimated as $$\dot{M} = \frac{1}{\tau_{\rm vis}} e^{-t/\tau_{\rm vis}} \int_{t_0}^t e^{t^\prime /\tau_{\rm vis}} \dot{M}_{\rm fb} dt^\prime.$$ We, therefore, in the second case (model II) take $\tau_{\rm vis}=10000$s. The results are presented with dashed lines in Figure \[fig\_evl\_late\]. The accretion rate becomes flat until $\tau_{\rm vis}$ and then begins to decay. Interestingly, we find a plateau in the jet power evolution.
Since the main part of the disk is already accreted, the mass accretion rate in this afterglow stage is very small. The disk will be dominated by advection. The feature of an adevction dominated disk is that it has a strong wind which is driven by a positive Bernouilli constant (Narayan and Yi 1994). Recently, Mu et al. (2016) took into account the effects of outflow in the accretion disk when interpreting X-ray flares. Due to the existence of mass loss into the wind, the accretion rate is expected to decrease inward in a scaling form $$\dot{M} \simeq \dot{M}_{\rm fb,r_d} \left(\frac{r_{\rm ms}}{r_{\rm d}}\right)^s,
\label{eqMdotaccr}$$ where $0 \leq s \leq 1$, $r_{\rm d}$ is the outer edger of the disk. We therefore consider a disk with $r_d=100 r_g$, $a_\bullet(0)=0.9$, and $s=0.5$, as shown with dot-dashed lines (model III) in Figure \[fig\_evl\_late\]. The accretion rate and jet power have been significantly suppressed by the outflow.
To check the effects of BH spin, we change the parameter $a_\bullet(0)=0.1$, as shown with dotted lines (model IV) in Figure \[fig\_evl\_late\]. Other parameters are the same as the first case. We find a strong evolution in BH spin, and the peak of the jet power tends to be flatter than the case with a high BH spin.
Finally, the fall-back material may contain large specific angular momentum, which in turn will shape the profile of accretion rate. Supposing that the gas at fall-back radius $r$ has an angular velocity $\Omega$ equal to a fraction $f_\Omega$ of the local Keplerian angular velocity $\Omega_{\rm k} (r)$, the specific angular momentum of this gas can then be written as (Kumar et al. 2008b) $$j_{\rm fb} \simeq 3.8 \times 10^{18} m_{\bullet,1}^{1/2} r_{10}^{1/2} f_\Omega(r) {\rm cm^2 s^{-1} }.$$ where $r_{10}=r/10^{10}\rm cm$ and $m_{\bullet,1}=m_\bullet/10$. The gas at $r$ will fall to the disk at a time around the fall-back time $t\sim t_{\rm fb} \simeq 2 (r^3/G M_\bullet)^{1/2}$. One therefore finds that the specific angular momentum increases with time as $j_{\rm fb} \simeq 4.7 \times 10^{18} t_2^{1/3} m_{\bullet, 1}^{2/3} f_\Omega(r) \ {\rm cm^2 s^{-1} }$. The evolution of the disk can be described with a model adopted in Kumar et al. (2008a, 2008b), $$\begin{aligned}
\dot{M}_{\rm d} = \dot{M}_{\rm fb} - \dot{M}, \nonumber \\
\dot{J}_{\rm d} = j_{\rm fb} \dot{M}_{\rm fb} - L_{\rm ms} \dot{M},\end{aligned}$$ where the accretion rate $\dot{M}$ is estimated with Equation (\[Eq:dMacc\]). In Figure \[fig\_evl\_late\], we present the results of a BH-fall-back disk system with $a_\bullet(0)=0.9$ and $f_\Omega=0.4$ (model V, thin solid lines). Since the angular momentum determines the fall-back radius (see Equation (\[Eq:jd\])) and $t_{\rm acc}\sim 2/(\alpha \Omega_{\rm K}(r_{\rm d}))$, the large angular momentum of the fall-back material leads to a longer accretion time $t_{\rm acc}$ and thereby a shallower lightcurve.
Discussions
===========
The central engine of GRBs is likely a hyperaccreting BH. The neutrino annihilation and BZ processes are two candidate mechanisms for powering GRB jets. In this paper, we obtained the analytical solutions to the neutrino and magnetic models, and studied the time evolution of the central engine parameters for these two models.
The evolution of accretion rate and BH spin have strong effects on the evolution of central engine parameters such as, $\dot{E}$, $\eta$ and $\mu_0$. The neutrino annihilation power is generally weaker that the BZ power. It fails to produce the long term X-ray activities observed in many GRBs. The magnetic model remains the leading candidate mechanism to interpret the X-ray flares, giant bumps and plateaus. For a BH central engine with small initial spin $a_\bullet(0)$, the jet may be first dominated by the neutrino annihilation power and then by the BZ power, leading to a transition from a thermally-dominated fireball to a Poynting flux dominated flow as is observed in some GRBs, e.g. GRB 160625B.
There are several predictions in our model, such as the transition from a thermal to a magnetic dominated jet, the evolution of $\mu_0$, and the late time plateaus. Systematic comparisons of these predictions against a large GRB sample are needed to test the BH central engine models. Some examples (e.g., GRB 160625B and GRB 110721A) have been observed that are consistent with our model predictions.
This work focuses on the BH-accretion central engine models. Metzger et al. (2011) and Beniamini et al. (2017) performed detailed investigations on the magnetar central engine model for GRBs. The comparison between these two models is desirable. In principle, the BH central engine is more complex, which contains two energy mechanisms (the neutrino annihilation and BZ processes) and two systems (the BH and disk). To predict a lightcurve, one needs to consider the evolution of both the central BH and the surrounding disk. Due to these intrinsic differences, our results show unique predictions on temporal evolutions of $\dot{E}$ and $\mu_0$, especially for the case with a small $a_\bullet(0)$. we hope our results can be used to distinguish the BH model from the magnetar model with observational data.
In this paper, we ignore the baryon loading during the late time central engine activities, since there is no good knowledge on the thermally driven wind at low accretion rates when neutrino cooling totally shuts off. Our analytical solutions are based on the numerical results of a standard NDAF model. We did not include the effects, such as magnetic coupling (Lei et al. 2009), inner boundary torque (Xie et al. 2016) and vertical structure (Liu et al. 2014). These effects may be important, but usually depend on some uncertain parameters. GRMHD simulations will help to give a better understanding of these issues.
We thank H. Gao and Q. Yuan for helpful discussions. The Numerical calculations were performed by using a high performance computing cluster (Hyperion) of HUST. This work is supported by the National Basic Research Program (’973’ Program) of China (grants 2014CB845800), the National Natural Science Foundation of China (U1431124, 11773010, 11673068, 11433009, 11603006, 11533003 and U1731239). X.F. W also acknowledges the support by the Key Research Program of Frontier Sciences (QYZDB-SSW-SYS005), the Strategic Priority Research Program “Multi-waveband Gravitational Wave Universe” (Grant No. XDB23000000) of the Chinese Academy of Sciences. E.W. L also acknowledges support by the Guangxi Science Foundation (2016GXNSFCB380005) and special funding for Guangxi distinguished professors (Bagui Yingcai & Bagui Xuezhe).
[99]{} Abdo, A. A., Ackermann, M., Arimoto, M., et al. 2009, Sci, 323, 1688 Abramowicz, M., Czerny, B., Lasota, J. P., & Szuszkiewicz, E. 1988, ApJ,332, 646 Atteia, J.-L., Heussaff, V., Dezalay, J.-P., et al. 2017, ApJ, 837, 119 Bardeen, J. M., Press, W. H., & Teukolsky, S. A. 1972, ApJ, 178, 347 Begelman, M. C. 1978, MNRAS, 184, 53 Beniamini, P., Giannios, D., Metzger, B. D., arXiv: 1706.05014 Blandford, R. D., & Znajek, R. L. 1977, MNRAS, 179, 433 (BZ77) Burrows D. N., Romano P., Falcone A., et al. 2005, Science, 309, 1833 Chen, W., Xie, W., Lei, W. H., Zou, Y. C., Lü, H. J., Liang, E. W., Gao, He., & Wang, D. X. 2017, accepted for publication in ApJ, arXiv:1709.08285 Chen, W. X., & Beloborodov, A. M. 2007, ApJ, 657, 383 Chevalier, R. A. 1989, , 346, 847 Chincarini, G., Morettti, A., Romano, P., et al. 2007, ApJ, 671, 1903 Dai, Z. G., & Liu R.-Y. 2012, , 759, 58 Di Matteo, T., Perna, R., [&]{} Narayan, R. 2002, ApJ, 579, 706 (DPN02) Eichler, D., Livio, M., Piran, T., & Schramm, D. N. 1989, Nature, 340, 126 Falcone, A. D., et al. 2007, ApJ, 671, 1921 Fan, Y. Z., Zhang, B., & Proga, D. 2005, ApJ, 635, L129 Fryer, C. L., Woosley, S. E., Herant, M., & Davies, M. B. 1999, ApJ, 520,650 Gao, H., Zhang, B. 2015, ApJ, 801, 103 Gao, H., Lei, W. H., You, Z. Q., & Xie, W. 2016a, ApJ, 826, 141 Gao, H., Zhang, B., & Lü, H. J. 2016b, Phys. Rev. D 93, 044065 Ghirlanda, G., Nava, L., Ghisellini, G., et al. 2012, MNRAS, 420, 483 Gu, W. M., Liu, T., [&]{} Lu, J. F., 2006, ApJ, 643, L87 Janiuk, A., Perna, R., Di Matteo, T., & Czerny, B. 2004, MNRAS, 355, 950 Januik, A., Yuan, Y., Perna, R., [&]{} Di Matteo, T. 2007, ApJ, 664, 1011 Janiuk, A., & Yuan, Y. 2010, A&A, 509, 55 Januik, A., Mioduszewski, P., & Moscibrodzka, M. 2013, ApJ, 776, 105 Januik, A. 2017, ApJ, 837, 39 Katz, J. 1977, ApJ, 215, 265 Kawanaka, N., Piran, T., & Krolik, J. H. 2013, ApJ, 766, 31 Kohri, K., & Mineshige, S. 2002, ApJ, 577, 311 Kohri, K., Narayan, R., & Piran, T. 2005, ApJ, 629, 341 Kumar, P., Narayan, R., & Johnson,J. L. 2008a, MNRAS, 388, 1729 Kumar, P., Narayan, R., & Johnson,J. L. 2008b, Science, 321, 376 Lei, W. H., Wang, D. X., & Ma, R. Y. 2005, ApJ, 619, 420 Lei, W. H., Wang, D. X., Gong, B. P., & Huang, C. Y. 2007, A&A, 468, 563 Lei,W. H., Wang,D. X. , Zhang, L., Gan,Z. M., Zou, Y. C. & Xie, Y. 2009, ApJ, 700, 1970 Lei,W. H., Wang,D. X. , Zhang, L., Gan,Z. M. & Zou, Y. C. 2010, Sciences in China (G), 2010, 53(s1), 98 Lei, W. H., & Zhang, B. 2011, , 740, L27 Lei, W. H., Zhang, B. & Liang, E. W. 2013, ApJ, 756, 125 (Paper I) Lee, H. K., Wijers, R. A. M. J., & Brown, G.E. 2000, Physics Reports, 325, 83 Li, L. X. 2000, , 61, 084016 Li, A., Zhang, B., Zhang, N. B., Gao, H., Qi, B., Liu, T. 2016, Phys. Rev. D, 94, 083010 Liang, E.-W., Yi, S.-X., Zhang, J., et al. 2010, ApJ, 725, 2209 Liang, E. W., Lin, T. T., Lü, J., et al. 2015, ApJ, 813, 116 Liu, T., Gu, W. M., Xue, L., [&]{} Lu, J. F. 2007, ApJ, 661, 1025 Liu, T., Yu, X. F., Gu, W. M., & Lu, J. F. 2014, ApJL, 791, 69 Liu, T., Hou, S. J., Xue, L., & Gu, W. M. 2015, ApJS, 218, 12 Liu, T., Gu, W. M., [&]{} Zhang, B. 2017, New Astronomy Review, in press (arXiv:1705.05516) Lloyd-Ronning, N. M., Dolence, J. C., & Fryer, C. L. 2016, MNRAS, 461, 1045 Lü, H. J., & Zhang, B. 2014, ApJ, 785, 74 Lü, H. J., Zhang, B., Lei, W. H., Li, Y., Lasky, P. D., 2015, ApJ, 805, 89 Lü, J., Zou, Y. C., Lei, W. H., et al. 2012, ApJ, 751, 49 Lyons N., O’Brien P. T., Zhang B., et al. 2010, MNRAS, 402,705 MacFadyen, A. I., & Woosley, S. E. 1999, ApJ, 524, 262 MacFadyen, A. I., Woosley, S. E., & Heger, A. 2001, , 550, 410 McKinney, J. C. 2005, , 630, L5 Mészáros, P., & Rees, M. J. 1997, ApJ, 428, L29 Metzger, B. D., Piro, A. L. & Quataert, E. 2008, MNRAS, 390, 781 Metzger, B. D., Giannios,D.,Thompson,T. A.,Bucciantini, N. & Quataert, E. 2011, MNRAS, 413, 2031 Moderski R., Sikora M., Lasota J. P. 1997, in Ostrowski M., Sikora M., Madejski, G., Belgelman M., eds, Proc. International Conf., Relativistic Jets in AGNs. Krakow, p. 110 Mu, H. J., Gu, W. M., Hou, S. J. et al. 2016, ApJ, 832, 161 Narayan R., Yi I., 1994, ApJ, 428, L13 Narayan, R., Piran, T., [&]{} Kumar, P. 2001, ApJ, 557, 949 (NPK01) Novikov, I. D., & Thorne, K. S., 1973, in Black Holes, ed. C. DewittMorette & B. S. DeWitt (New York: Gordon & Breach), 345 Paczy$\acute{n}$ski, B. 1991, Acta Astron., 41, 157 Paczy$\acute{n}$ski, B. 1998, ApJ, 494, L45 Page, D. N., & Thorne, K. S. 1974, ApJ, 191, 499 Proga, D., & Zhang, B. 2006, MNRAS, 370, L61 Popham, R., & Narayan, R. 1995, ApJ, 442, 337 Popham, R., Woosley, S. E., [&]{} Fryer, C. 1999, ApJ, 518, 356 (PWF99) Pudritz, R. E., & Fahlman, G. G. 1982, MNRAS, 198, 689 Reynoso, M. M., Romero, G. E., [&]{} Sampayo, O. A. 2006, A[&]{}A, 454, 11 Riffert, H., [&]{} Herold, H. 1995, ApJ, 450, 508 Rosswog, S., Ramirez-Ruiz, E., & Davies, M. 2003, MNRAS, 345, 1077 Rosswog, S. 2007, MNRAS, 376, L48 Sharkura, N. I., & Sunyaev, R. A. 1973, A& A, 24, 337 Shibata, M., Sekiguchi, Y., & Takahashi, R. 2007, Prog.Theor.Phys., 118, 257 Tchekhovskoy, A., Narayan, R., & McKinney, J. C. 2010, ApJ, 711, 50 Tchekhovskoy, A., Narayan, R., & McKinney, J. C. 2011, MNRAS Lett., 418, L79 Tchekhovskoy, A., & McKinney, J. C. 2012, MNRAS Lett., 423, L55 Tchekhovskoy A., Giannios D., 2015, MNRAS, 447, 327 Thorne, K. S., Price, R. H., Macdonald D. A., 1986, Black Holes: The Membrane Paradigm. Yale Univ. Press, New Haven Troja, E., Cusumano, G., O’Brien, P. T., et al., 2007, ApJ, 665, 599 Wang, D. X., Xiao, K., & Lei, W. H. 2002, , 335, 655 Wang, D.X., Lei, W.H., & Ye, Y.C. 2006, ApJ, 643, 1047 Wang, J. Z., Lei, W. H., Wang, D. X., et al. 2014, ApJ, 788, 32 Woosley, S. E. 1993, ApJ, 405, 273 Wu, X. F., Hou, S. J., & Lei, W. H. 2013, , 767, L36 Wu, Q., Zhang, B., Lei, W. H., et al. 2016, MNRAS, 455, 1 Xie, W., Lei, W. H., & Wang, D. X. 2016, ApJ, 833, 129 Xie, W., Lei, W. H., & Wang, D. X. 2017, ApJ, 838, 143 Xue, L., Liu, T., Gu, W. M. & Lu, J. F. 2013, ApJs, 207, 23 Yi, S. X., Lei, W. H., Zhang, B., Dai, Z. G., Wu, X. F., & Liang, E. W. 2107, JHEAp, 13, 1 Yuan, F., & Zhang, B. 2012, ApJ, 757, 56 Zalamea, I., & Beloborodov A. M. 2011, MNRAS, 410, 2302 Zhang, B., & Pe’er, A. 2009, ApJ, 700, L65 Zhang, B., Fan, Y. Z., Dyks, J., et al. 2006, ApJ, 642, 354 Zhang, B. Chin. J. Astron. Astrophys. 2007, 7, 1 Zhang, B., & Yan, H. R. 2011, ApJ, 726, 90 Zhang, W., Woosley, S. E., & Heger, A. 2008, , 679, 639 Zhang, B. B., Zhang, B., Castro-Tirado, A. J., et al. 2017, Nature Astronomy, in press (arXiv:1612.03089)
[^1]: In Chen & Beloborodov (2007), the characteristic accretion rates $\dot{m}_{\rm ign}$ and $\dot{m}_{\rm trap}$ are well approximated by the following formulae, $\dot{m}_{\rm ign}=K_{\rm ign} \alpha_{-1}^{5/3}$, and $\dot{m}_{\rm trap}=K_{\rm trap} \alpha_{-1}^{1/3} $. For $a_\bullet=0$, one has $K_{\rm ign}=0.071$ and $K_{\rm trap}=9.3$, whereas for $a_\bullet=0.95$, one has $K_{\rm ign}=0.021$ and $K_{\rm trap}=1.8$.
[^2]: Fan et al.(2005) found that the $\nu \bar{\nu}$ power for $0.01<\dot{m}<0.1$ can be well fitted with (see the thin dotted lines in the right panel of Figure \[fig\_dotEvv\]), $ \dot{E}_{\nu \bar{\nu}} \simeq 10^{43.6+4.3 a_\bullet} (\frac{\dot{m}}{0.01})^{4.89} {\rm erg \ s^{-1} } $.
[^3]: Zalamea & Beloborodov (2011) also obtained a simple formula for $\dot{E}_{\nu\bar{\nu}}$ as: 0 for $\dot{m}< \dot{m}_{\rm ign}$; $ 1.1\times 10^{52} (\frac{r}{r_{\rm ms}})^{-4.8} (\frac{m_\bullet}{3})^{-3/2} \dot{m}^{9/4} $ for $ \dot{m}_{\rm ign}< \dot{m}<\dot{m}_{\rm trap}$; $ 1.1\times 10^{52} (\frac{r}{r_{\rm ms}})^{-4.8} (\frac{m_\bullet}{3})^{-3/2} \dot{m}_{\rm trap}^{9/4} $ for $ \dot{m}>\dot{m}_{\rm trap}$. For $\alpha=0.1$, one has $\dot{m}_{\rm ign} = 0.071$ and $\dot{m}_{\rm trap}=9.3$ for $a_\bullet= 0$, and $\dot{m}_{\rm ign} = 0.021$ and $\dot{m}_{\rm trap}=1.8$ for $a_\bullet= 0.95$.
[^4]: BZ77 showed that magnetic power of a force-free jet from a slowly spinning BH ($a_\bullet \ll 1$) is $\dot{E}_{\rm B} = \frac{\kappa c}{4\pi} \Phi_{\rm BH}^2 \frac{a_\bullet^2}{16 r_{\rm g}^2}$, where $\kappa$ weakly depends on the field geometry (it is 0.053 for a split monopole geometry and 0.044 for a parabolic geometry), $\Phi_{\rm BH}$ is an absolute magnetic flux through the BH.
[^5]: Tchekhovskoy, Narayan & McKinney (2010) extended the magnetic power in BZ77 to high-spin BHs, (see also Tchekhovskoy, Narayan & McKinney 2011 and Tchekhovskoy & McKinney 2012), and obtained $
\dot{E}_{\rm B} = \frac{\kappa}{4\pi c} \Phi_{\rm BH}^2 \Omega_\bullet^2 f(\Omega_\bullet),
$ where $f(\Omega_\bullet) \simeq 1 + 1.38 (\Omega_\bullet r_{\rm g}/c)^2 - 9.2 (\Omega_\bullet r_{\rm g}/c)^4$ is a high-spin correction to BZ77.
[^6]: Atteia et al. (2017) found a maximum isotropic energy of GRBs when they studied the GRB energy distribution within redshifts $z=1-5$. Jet break measurements are needed to derive the beaming-corrected energy, which can be compared with our model predictions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Square ice is a statistical mechanics model for two-dimensional ice, widely believed to have a conformally invariant scaling limit. We associate a Peano (space filling) curve to a square ice configuration, and more generally to a so-called $6$-vertex model configuration, and argue that its scaling limit is a space-filling version of the random fractal curve ${\operatorname{SLE}}_{\kappa}$, Schramm–Loewner evolution with parameter $\kappa$, where $4<\kappa\leq 12+8\sqrt{2}$. For square ice, $\kappa=12$. At the “free-fermion point” of the 6-vertex model, $\kappa=8+4\sqrt{3}$. These unusual values lie outside the classical interval $2\le \kappa\le 8$.'
author:
- Richard Kenyon
- Jason Miller
- Scott Sheffield
- 'David B. Wilson'
bibliography:
- '../bipolar.bib'
- '../ice.bib'
- '../../exploration/activity/activity.bib'
title: 'The six-vertex model and Schramm–Loewner evolution'
---
Square ice was introduced by Pauling [@pauling:ice] as a model of hydrogen bonding in ice crystals in two dimensions [@baxter]. A square-ice configuration is an orientation of each edge of the square lattice, subject to the constraint that each vertex has two incoming and two outgoing edges (see the diagram below and Fig. \[fig:ice-tours\]). Recently actual square ice crystals were produced between sheets of graphene [@square-ice].
The classical $6$-vertex model from statistical mechanics generalizes square ice by adding energies to each of the $6$ types of local configuration at a vertex:
{width=".5\textwidth"}
Square ice is the uniform measure on 6-vertex configurations. The 6-vertex model partition function was famously solved by Lieb in 1967 [@lieb:ice]. A number of beautiful combinatorial identities arising in this model have intrigued mathematicians and physicists for many years [@razumov-stroganov:ASM; @cantini-sportiello]. In particular it is widely believed that the $6$-vertex model has conformally invariant scaling limits, however a mathematical proof of this fact is lacking.
We show here how to associate a discrete Peano (space filling) curve to configurations of the square ice model with appropriate boundary conditions (Fig. \[fig:ice-tours\]). We present evidence that the scaling limit of this curve is a random fractal curve called a Schramm–Loewner evolution (SLE).
For each $\kappa \leq 0$, an ${\operatorname{SLE}}_\kappa$ in the upper half plane is a random non-self-crossing random curve that extends from the origin to $\infty$, with the parameter $\kappa$ indicating how “windy” the path is. In recent decades, SLE has been thoroughly studied and celebrated within both physics and mathematics, and has led to many new results about two-dimensional statistical physics and the Liouville theory of quantum gravity — some of which go far beyond the results previously established using conformal field theory and other techniques.
The precise definition of SLE is interesting and indirect. Fix $\kappa > 0$, let $B(t)$ be a one-dimensional Brownian motion, and for each $z$ in the complex upper half plane $\mathbb H$, let $g_t(z)$ solve the ODE $$\frac{\partial g_t(z)}{\partial t}= \frac{2}{g_t(z)-\sqrt{\kappa}\, B(t)}\quad\quad\quad\quad g_0(z) = z\,,$$ which is defined until $T_z = \inf \{t : g_t(z)-W_t = 0 \}$. Then ${\operatorname{SLE}}_\kappa$ is the curve $\eta: \mathbb R_+ \to \mathbb H$ defined so that $\{z : T_z \leq t\}$ is the set of points hit or cut off from $\infty$ by $\eta([0,t])$.
For $\kappa\leq 4$, ${\operatorname{SLE}}_\kappa$ is a simple curve; for $4<\kappa<8$, the curve hits itself without crossing itself, forming bubbles; for $\kappa\geq 8$, the curve is space-filling [@rohde-schramm]. For $4<\kappa<8$, there is also a space-filling version of ${\operatorname{SLE}}_\kappa$ in which the bubbles get filled in recursively as they are made [@miller-sheffield:ig4].
The ${\operatorname{SLE}}_\kappa$ curves are either known or believed to characterize the scaling limits of various two-dimensional critical statistical physics models: dilute polymers ($\kappa=8/3$) [@LSW:SAW], dense polymers ($\kappa=8$) [@LSW:tree], loop-erased random walk ($\kappa=2$) [@LSW:tree], percolation interfaces ($\kappa=6$) [@smirnov:percolation], Ising model spin clusters ($\kappa=3$) [@smirnov:ising; @CDCHKS:ising], dimer systems ($\kappa=4$), contours of the Gaussian free field ($\kappa=4$) [@schramm-sheffield:discrete-GFF; @schramm-sheffield:continuum-GFF], the Ashkin–Teller model ($\kappa=4$), the Fortuin–Kasteleyn random cluster model ($2\leq\kappa\leq 8$), active spanning trees ($4<\kappa\leq 12$) [@kassel-wilson:active], and others. The dimension $D_{\!f}$ of the fractal increases with the parameter $\kappa$ according to the formula $D_{\!f}=\min(2,1+\kappa/8)$ [@rohde-schramm; @Beffara]. See [@rohde-schramm; @Cardy; @Schramm-ICM] for further background.
SLE is connected with conformal field theory (CFT) [@Cardy], where the central charge $c$ is related to $\kappa$ by $$\label{eqn:c-kappa}
c = (8 - 3\kappa) (\kappa - 6)/(2 \kappa)\,.$$ In CFT usually $c\geq-2$, which corresponds to $\kappa \in [2,8]$, the values relevant to [*conformal loop ensembles*]{} [@sheffield:cle-trees]. Before this work and [@kassel-wilson:active] it was widely assumed that only $\kappa \in [2,8]$ would appear in natural discrete models [@Cardy].
For the 6-vertex model Peano curve defined here, $\kappa$ depends on the vertex energies and spans the range $(4, 12+8\sqrt{2}]$, which in particular includes values outside of $[2,8]$. For square ice, $\kappa=12$, which corresponds to $c=-7$. The square ice Peano curve joins a tiny pantheon of models (including the uniform spanning tree and the Ising model) that have *independently* solvable random lattice analogs; these analogs are described in [@KMSW1], along with connections to Liouville quantum gravity and string theory.
Six-vertex configurations have a height function which plays an important role in their analysis [@vanbeijeren:roughening]. The heights are defined on the faces; around even-parity vertices, the heights increase by $1$ in the counterclockwise direction across outgoing edges, and decrease by $1$ in the counterclockwise direction across incoming edges (see Fig. \[fig:ice-tours\]).
We produce a Peano curve (Figs. \[fig:ice-tours\] and \[fig:ice-tour\]) from a 6-vertex configuration as follows: From the even index vertices, we bend the outgoing arrows $45^\circ$ left so that they terminate at the face centers, and from the odd index vertices, we bend the outgoing arrows $45^\circ$ right. Note that each arrow gets bent into the same face regardless of which way it is oriented, and each face receives two arrows from opposite sides. Because each face and each vertex now has degree two, the curved arrows form a collection of loops and chains which terminate at the boundary. Observe that the six-vertex heights, when scaled by $\pi/2$, give the winding angle of the green curve measured in radians. Because the height function is single-valued, the green curve cannot close up on itself to form loops. The boundary conditions were chosen so that there is only one chain, so it must form a single space-filling curve.
The six-vertex model can also specialize to the ${\mathrm{O}}(n)$ loop model. To obtain the ${\mathrm{O}}(n)$ model, we set (with $\omega_i=e^{-\epsilon_i}$) $\omega_1=\omega_2=\omega_3=\omega_4=1$ and $\omega_5=\omega_6=C$. The parameter $\Delta$ is defined by $$\label{eqn:Delta}
\Delta = \frac{\omega_1\omega_2+\omega_3\omega_4-\omega_5\omega_6}{2\sqrt{\omega_1\omega_2\omega_3\omega_4}}=\frac{2-C^2}{2}.$$
There is a weight-preserving mapping between six-vertex configurations and ${\mathrm{O}}(n)$ model loop configurations, so that the partition functions are equal [@BKW]: One splits each vertex in half (maintaining planarity) so that each half has one out-going and one in-coming edge. For any vertex with adjacent out-going arrows, there is one way to do this split, but for $C$-type vertices, there are two ways to split it. A split vertex is given a weight of $r$ if the arrows turn right, and weight $1/r$ if the arrows turn left. For the non-$C$-type vertices, the total weight is $r\times r^{-1} = 1$. For the $C$-type vertices, the total weight is $r^2+ r^{-2} = C$. Each loop has weight $r^4 + r^{-4} = n$. Thus $$n = C^2 - 2 \label{eqn:n-C}$$ and hence $n=-2\Delta$.
The ${\mathrm{O}}(n)$ model loops are widely believed to be described by the conformal loop ensemble ${\operatorname{CLE}}_{\kappa^\circ}$ (the loop version of SLE), where $$n = -2\cos(4\pi/\kappa^\circ) \label{eqn:n-kappa}$$ [@sheffield:cle-trees]. (Here ${}^\circ$ is a mnemonic for ${\mathrm{O}}(n)$.) The ${\operatorname{SLE}}$-parameter for the Peano curve coming from the associated 6-vertex model we call $\kappa'$. Interestingly, $\kappa'\neq\kappa^\circ$.
The variance in the height function of the six-vertex model was computed by Nienhuis [@nienhuis:6v]: When the height function $h$ is measured in radians, for small $a$, $\langle\exp(ia(h(x)-h(0)))\rangle = \exp(-a^2/g \log|x|)$, where $g$ is the Coulomb gas coupling constant. So the height variance, given by the quadratic term (in $a^2$), is $(1/g)\log|x|$. From [@nienhuis:6v (3.29)] we have $$\label{eqn:g-C}
\sin\frac{\pi g}{8} = \frac{C}{2}.$$
The theory of imaginary geometry, as developed by Miller and Sheffield, associates to a Gaussian free field (GFF) a space-filling SLE [@miller-sheffield:ig1; @miller-sheffield:ig2; @miller-sheffield:ig3; @miller-sheffield:ig4]. Roughly speaking, the GFF height function $h$ is divided by a parameter $\chi$ to obtain a field of orientations (measured in radians), and the orientation of the SLE curve is $e^{i h/\chi}$. Thus the Coloumb gas coupling constant $g$ and the parameter $\chi$ are (heuristically) related by $g=\chi^2$.
The space-filling SLE parameter $\kappa'$ and $\chi$ are related by $
\chi = \frac{\sqrt{\kappa'}}{2} - \frac{2}{\sqrt{\kappa'}}
$ [@miller-sheffield:ig4], so $$\label{eqn:g-chi-kappa}
\frac{1}{g} = \frac{1}{\chi^2} = \frac{4\kappa'}{(\kappa'-4)^2}\,.$$
If we parametrize $n$ by $n=-2\cos\theta$ with $0\leq\theta\leq\pi$, then , , , , and can be expressed as $$\begin{aligned}
n&=-2\cos\theta\notag\\
\Delta&=-\cos\theta\notag\\
C^2 &=2-2\cos\theta\notag\\
\chi^2=g &= 4\,\theta/\pi\notag\\
\kappa'&=4+8\,\theta/\pi+8\sqrt{\theta/\pi+\theta^2/\pi^2}\label{eqn:kappa'}\\
c'&=1-24\,\theta/\pi\,,\notag\end{aligned}$$ where $c'$ is the central charge associated with ${\operatorname{SLE}}_{\kappa'}$.
The table below gives some special cases. The limiting case $C\to0$ is included, but with $C=0$ the discrete models do not converge to SLE. Square ice is the $C=1$ row. The special value $C=\sqrt{2}$ is the “free fermion” point, where there is a mapping between the 6-vertex model and square-lattice dimers; in this case $\kappa'=8+4\sqrt{3}$.
There is a useful, and related, bijection between six-vertex configurations and *bipolar orientations*. Let ${{\mathcal G}}$ be a finite subgraph of ${{\mathbb Z}}^2$, that is, the part of ${{\mathbb Z}}^2$ bounded by a rectilinear integer polygon. Let $N$ and $S$ be distinct vertices on the outer boundary of ${{\mathcal G}}$. A *bipolar orientation* is an orientation of the edges of ${{\mathcal G}}$ which is acyclic (has no oriented cycles), has only one source, at $N$, and has only one sink, at $S$.
We give a bijection between bipolar orientations of ${{\mathcal G}}$ and 6-vertex configurations on another graph $H$, the “double” of ${{\mathcal G}}$, whose vertices are the vertices and faces of ${{\mathcal G}}$, with edges of $H$ connecting vertices of ${{\mathcal G}}$ to their incident faces of ${{\mathcal G}}$. Edges of ${{\mathcal G}}$ correspond to faces of $H$. (See Fig. \[bpfig\]a.)
At each vertex $v$ of ${{\mathcal G}}$, the outgoing edges in the bipolar orientation form a contiguous interval in the circular order around $v$, that is, there are no vertices for which the orientation is in-out-in-out. Equivalently the incoming arrows form a contiguous interval around $v$. In the corresponding 6-vertex configuration, outgoing arrows from $v$ point to the two faces that separate these intervals. For each face $f$ of ${{\mathcal G}}$, the bipolar orientation restricted to that face has a unique source and unique sink; the 6-vertex arrows point from this face to these two extremal vertices. It is easy to check that each edge of $H$ is oriented by precisely one of these two rules, so it has out-degree 2 everywhere, that is, it is a 6-vertex configuration.
Given an edge in a bipolar-oriented graph ${{\mathcal G}}$, there is a canonical path to the sink, obtained by travelling along that edge in the direction of its orientation and, when arriving at a vertex, taking the maximally left outgoing edge from the new vertex. The union of these paths forms a tree, the “SE-tree”, drawn in blue in Fig. \[bpfig\]b. The analogous “NW-tree”, which is the SE-tree for the bipolar orientation obtained by reversing all the arrows, is drawn in red in Fig. \[bpfig\]c. The SE-tree and NW-tree do not cross each other, so there is a curve winding between them, which is shown in green in Fig. \[bpfig\]d. This map from bipolar orientations to Peano curves was first described for general planar graphs in [@KMSW1]. This Peano curve is the same curve defined by the 6-vertex height function.
![\[fig:ice-tour\]The Peano curve, colored according to the time parameter, for the square ice model ($C=1, \kappa'=12$).](6v-rainbow-tour-c2_1){width="2.5in"}
Fig. \[fig:ice-tour\] shows a random sample of the Peano curve associated to a large square ice configuration on the square grid. For planar graphs, perfect samples for the 6-vertex models with $C\ge 1$ can be obtained from single-site Glauber dynamics and coupling from the past [@propp-wilson:cftp].
We used Monte Carlo simulations to check that the 6-vertex model Peano curve is described by ${\operatorname{SLE}}_{\kappa'}$. We produced 6-vertex configurations on an $L\times L$ torus for various values of $L$, to eliminate boundary effects. We measured the winding angle variance of the Peano curve, and also the dimension of the outer boundary of the Peano curve.
SLE theory predicts that the Peano curve’s winding angle variance scales as $$\label{eqn:winding}
\frac{4\kappa'}{(\kappa'-4)^2}\, \ln L\,.$$ [@miller-sheffield:ig4]. Since the winding of the curve is given by the height function, we measured the height function variance.
The outer boundary corresponds to paths within the blue SE-tree in Fig. \[bpfig\]b. Since the simulations are done on a torus, the “SE-tree” is actually a cycle-rooted spanning forest (CRSF), and we measured both the winding angle variance and the length $\ell$ of the cycle in the cycle-rooted spanning tree containing the edge at the origin. The SLE prediction is that the outer boundary’s winding angle variance scales as $$\label{eqn:outer-winding}
\frac{4}{\kappa'}\, \ln L\,,$$ and that its length scales as $\ell\sim L^{D_{\!f}}$ where $$\label{eqn:outer-dimension}
D_{\!f} = 1+\kappa/8 = 1+2/\kappa'\,.$$ We estimated the winding angle variance coefficients and the outer boundary dimension using samples for $L=256$ and $L=512$, as shown in Fig. \[fig:monte-carlo-wind-dim\].
The estimates for the winding angle variance coefficient is an excellent fit to the predicted value. Since the formula relating $\kappa'$ to $C^2$ was derived from Nienhuis’ formula , the left panel of Fig. \[fig:monte-carlo-wind-dim\] is essentially an experimental verification of Nienhuis’ formula.
The outer boundary winding angle variance and dimension estimates (middle and right panels of Fig. \[fig:monte-carlo-wind-dim\]) are both independent tests of the curve’s convergence to SLE. The estimated values are a close match to the predicted value, though when $C^2\approx 3.5$, the measured dimension deviates from the prediction by as much as $0.015$. Further tests of the distribution of the loop length $\ell$ and its dependence on $L$ suggest that the convergence to the asymptotic behavior occurs for larger values of $L$ when $C^2\approx 4$ than when, for example, $C^2\approx 2$. Overall, the experiments are consistent with convergence to ${\operatorname{SLE}}$.
R.K. was supported by an NSF grant and a Simons Foundation grant. J.M. was supported by an NSF grant. S.S. was supported by a Simons Foundation grant, an NSF grant, and two EPSRC grants.
rst \#1 \#2other[\#1]{}
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Department of Mathematics\
Purdue University\
West Lafayette, IN 47907\
U.S.A.
author:
- 'Su-Jeong Kang'
title: Hodge structure of a complete intersection of quadrics in a projective space
---
Let $V$ be a smooth projective variety of dimension $n$. One may say that $V$ has motivic dimension less than $d+1$ if the cohomology of $V$ comes from varieties of dimensions less than $d+1$ in some geometric way. More precisely, we say that $V$ has [*motivic dimension*]{} less than $d+1$ if there exists a (nonconnected) smooth projective variety $W$ of dimension less than $d+1$ and an algebraic correspondence $\Gamma$ on $W \times X$ such that $\Gamma$ induces a surjection $H^*(W) \to H^*(V)$.
Suppose the Generalized Hodge conjecture(GHC) ([@groth],[@lewis]) holds for $V$ and suppose the level of $V$ is less than $l$ (i.e. $\text{level}(H^i(V)) < l$ for all $i$). This implies that $V$ has a motivic dimension less than $l$. Conversely, if $V$ has motivic dimension less than $d+1$, i.e. there is a smooth projective variety $W$ of dimension less than $d+1$ and a surjection $\phi: H^*(W) \to H^n(V)$ induced by a correspondence $\Gamma$ on $W \times X$, then the level of the Hodge structure $H^n(V)$ is less than $d+1$ because a morphism of Hodge structures preserves the level. The existence of $W$ does not imply that the GHC holds for $V$ of course, because the dimension of the variety $W$ is not small enough to conclude the GHC for $V$. However, this gives a way to reduce the GHC for $V$ to the GHC for $W$, the variety with the smaller dimension. And we think that this can be a practical intermediate step for checking GHC for a smooth projective variety.
For a smooth complete intersection of $k$ quadrics $V=Q_1 \cap \cdots \cap Q_k$ in ${{\mathbb P}}^{n+k}$, it can be checked that the level of the Hodge structure $H^n(V)$ is less than $k$ (for example by checking Hodge number of it). In this paper, we show that a smooth complete intersection of $k$ quadrics in ${{\mathbb P}}^{n+k}$ has motivic dimension less than $k $ (Theorem \[thm:main\]). As a corollary of this, we get the Hodge conjecture for $V$ holds if $k<4$.
A brief outline of the proof of main theorem is as follows. By using the constuction in [@ogrady], we form a family of quadrics parametrized by ${{\mathbb P}}^{k-1}$ with the base locus $V$. In case when $n+k$ is odd, this is a family of even dimensional quadrics. Then we use the fact that an even dimensional quadric contains two irreducible families of linear spaces of the expected dimension, and we choose $W$ to be a double covering of ${{\mathbb P}}^{n+k}$. In the case when $n+k$ is even, then quadrics in the original family are odd dimensional. So in order to find $W$, we pass to the family of singular fibers over the discriminant variety which can be understood as a family of quadrics of even dimension. Then we form a double covering of the discriminant variety to find $W$.
All varieties in this paper will be defined over ${{\mathbb C}}$ and the cohomology without coefficient would be the singular cohomology with rational coefficient.
I would like to express my thanks to Professor Donu Arapura for suggesting me to look at this problem and for his encouragement. The definition of motivic dimension is due to him.
Main Theorem
============
Let $V =Q_1 \cap \cdots \cap Q_k$ be a smooth complete intersection of $k$ quadrics in a projective space ${{\mathbb P}}^{n+k}$. Then $V$ is a smooth subvariety of ${{\mathbb P}}^{n+k}$ of $\dim V =n$. Consider the family of quadrics $Q_t$ ($t \in {{\mathbb P}}^{k-1}$) with the base locus $V$. We can give more precise description of this family. Let $Q_l = \{F_l=0 \}$ for $l=1,\cdots, k$, where $F_l(x_0, x_1,\cdots , x_{n+k}) = \sum^{n+k}_{i,j=0} c^l_{ij}x_i x_j$, $[c^l_{ij}]_{0 \leq i,j \leq n+k}$ is a symmetric $(n+k+1) \times (n+k+1)-$matrix. Then for a general $t=(t_0,..., t_{k-1}) \in {{\mathbb P}}^{k-1}$, the fiber $Q_t$ is given by the equation $\sum^{k-1}_{l=0}t_lF_{l+1}(x_0, ..., x_{n+k})=0$. Let $\Delta \subset {{\mathbb P}}^{k-1}$ be the subvariety of ${{\mathbb P}}^{k-1}$ parametrizing all singular fibers in the family. Then $\Delta$ is a hypersurface in ${{\mathbb P}}^{k-1}$ of degree $n+k+1$. We assume that $\Delta$ is smooth. Set $$X = \{(t,x) \in {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}~|~ x \in Q_t \subset {{\mathbb P}}^{n+k} \} \subset {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$$ and let $p_1 : X \to {{\mathbb P}}^{k-1}$ and $p_2 : X \to {{\mathbb P}}^{n+k}$ be projections. Note that $X$ is a smooth projective variety of dimension $n+2k-2$. $$\begin{aligned}
\label{diag:proj}
\xymatrix@R=11pt{
X \ar[rr]^{p_2 \quad} \ar[d]^{p_1}&& {{\mathbb P}}^{n+k} \supset V \\
{{\mathbb P}}^{k-1} &&
}\end{aligned}$$
Let $i_X : X \hookrightarrow {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$ and $i_V: V \hookrightarrow {{\mathbb P}}^{n+k}$ be inclusions. Then by Lefschetz theorem, the restriction maps $i^*_X : H^{n+2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \to H^{n+2k-2}(X)$ and $i^*_V:H^{n}({{\mathbb P}}^{n+k}) \to H^{n}(V)$ are injections. Set
$$H^{n+2k-2}_0 (X) = H^{n+2k-2}(X) / {\text{im}}~ i^*_X$$ and $$H^n_0 (V) = H^n (V) / {\text{im}}~i^*_V$$
In fact, $$H^{n+2k-2}_0 (X) \cong \begin{cases}H^{n+2k-1}_c(({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})-X) & \text{if $n$ is even}\\H^{n+2k-2}(X) & \text{if $n$ is odd}\end{cases}$$ and $$H^n_0(V) \cong \begin{cases}H^{n+1}_c({{\mathbb P}}^{n+k} -V) & \text{if $n$ is even}\\
H^n(V) & \text{if $n$ is odd}\end{cases}$$ as Hodge structures.
Now we can state our main theorem precisely:
\[thm:main\] Let $V$ be a smooth complete intersection of $k$ quadrics in ${{\mathbb P}}^{n+k}$. Then $V$ has motivic dimension less than $k$. More precisely, there is a smooth projective variety $W$ of dimension $k-1$ (resp. $k-2$) and surjection of rational Hodge structures $$\Theta : H^{k-1}(\tilde{W})(-q) \to H^n(V) \qquad \text{if $n$ is odd and $k$ is even}$$ $$(\text{resp. } \Theta : H^{k-2}(\tilde{W})(-q) \to H^n(V) \qquad \text{if $n$ is odd and $k$ is odd} )$$ where $q=\frac{n-k+1}{2}$(resp. $\frac{n-k+2}{2}$) and $\tilde{W}$ is a disjoint union of finitely many copies of $W$, $$\Theta : \bigoplus_r H^{2r}(W)(-q_r)^{\oplus l_r} \to H^n_0(V) \qquad \text{if $n$ is even}$$ where $q_r=\frac{n-2r}{2}$ and $l_r$ are positive integers given by explicit formula.
We will prove this theorem by considering two cases depending on the parity of $n+k$ in the last two sections. One immediate corollary of this theorem is
If $k \leq 4$, then Hodge conjecture holds for $V$.
Intermediate step {#sec:intermediate-step}
=================
Consider the projection $p_2: X \to {{\mathbb P}}^{n+k}$. Note that for any $q=(q_0,..., q_{n+k}) \in {{\mathbb P}}^{n+k}- V$, $$\begin{aligned}
p^{-1}_2 (q) &=& \{ (t,q) \in {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}~|~ q \in Q_t \}\\
&=& \left\{ t=(t_0,\cdots, t_{k-1})\in {{\mathbb P}}^{k-1}~|~\sum^{k-1}_{l=0}\left( \sum^{n+k}_{i,j=0} c^{l+1}_{ij} q_i q_j \right)t_l =0 \right\} \cong {{\mathbb P}}^{k-2} \end{aligned}$$ and for any $q \in V$, $$p^{-1}_2(q) = \{(t,q) \in {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}~|~ q \in Q_t \} = {{\mathbb P}}^{k-1} \times \{q\}$$ since $V$ is the base locus of the family. Hence we have the following diagram, which we will use throughout this section: $$\begin{aligned}
\label{diag:xv}
\xymatrix{
({{\mathbb P}}^{n+k} - V) \times {{\mathbb P}}^{k-2} \ar@{^(->}[r] \ar[d]^{p_2} & X \ar[d]^{p_2} & {{\mathbb P}}^{k-1} \times V \ar@{_(->}[l]_{i_E \quad} \ar[d]^{p_2} \\
{{\mathbb P}}^{n+k} -V \ar@{^(->}[r] & {{\mathbb P}}^{n+k} & \ar@{_(->}[l] V
}\end{aligned}$$
\[lemma:rank1\] $$\dim H^{n}_0(V) = \dim H^{n+2k-2}_0(X)$$
From the diagram (\[diag:xv\]), we get $$\begin{aligned}
\label{eq:chiX}
\chi(X) &=& (k-1)(n+k+1) + \chi(V)\end{aligned}$$ Since $X$ is a very ample divisor in ${{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$ and $V$ is a smooth complete intersection in ${{\mathbb P}}^{n+k}$, by Lefschetz Theorem we have $$H^i({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \cong H^i(X), \qquad H^j({{\mathbb P}}^{n+k}) \cong H^j(V)$$ for any $i< \dim X=n+2k-2$ and $j < \dim V=n$.
(Case 1) If $n$ is even : then we have $$\begin{aligned}
\chi(V)&=&2\sum^{n-1}_{i=0}(-1)^i b_i(V) + b_{n}(V)=n + b_{n}(V)\end{aligned}$$ $$\begin{aligned}
\chi(X)&=&2\sum^{n+2k-3}_{i=0} (-1)^i b_i(X) + b_{n+2k-2}(X)\\
&=& 2 \left( \frac{k(k+1)}{2} +k \cdot \frac{n-2}{2} \right) +b_{n+2k-2}(X)\\
&=&
nk+k^2-k + b_{n+2k-2}(X)\end{aligned}$$ Therefore by (\[eq:chiX\]) we get $$nk+k^2-k +b_{n+2k-2}(X)=(k-1)(n+k+1)+n + b_{n}(V)\\$$ i.e. $$b_{n+2k-2}(X) = b_{n}(V) + k-1$$ Note that $$\dim H^{n+2k-2}_0(X)=b_{n+2k-2}(X)-\dim H^{n+2k-2}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})=b_{n+2k-2}(X)-k$$ and $$\dim H^{n}_0(V)=b_{n}(V)-\dim H^{n}({{\mathbb P}}^{n+k})=b_n(V) -1$$ therefore we get $$\dim H^{n+2k-2}_0(X)=\dim H^{n}_0(V)$$ in this case.
(Case 2) If $n$ is odd : then by the similar calculation, we get $$\chi(X) = nk+k^2 -b_{n+2k-2}(X), \qquad \chi(V) = (n+1)-b_{n}(V)$$ Then, again by (\[eq:chiX\]) $$\begin{aligned}
nk+k^2 - b_{n+2k-2}(X) = (k-1)(n+k+1)+(n+1)- b_{n}(V) =nk+k^2-b_{n}(V)\end{aligned}$$ i.e. $$b_{n+2k-2}(X) = b_n(V)$$ Hence, we get $$\dim H^{n+2k-2}_0(X) = \dim H^{n+2k-2}(X) = \dim H^{n}(V)=\dim H^n_0(V)$$ in this case also, which finishes the proof of the Lemma.
Set $E = {{\mathbb P}}^{k-1} \times V \subset X$. Then we have a ${{\mathbb P}}^{k-1}-$bundle $p_2: E \to V$. Let $i_E:E \hookrightarrow X$ be an inclusion. Then we have the following of morphism of Hodge structures $$\phi : H^{n}(V) \stackrel{p^*_2}{\longrightarrow} H^{n}(E) \stackrel{{i_E}_*}{\longrightarrow} H^{n+2k-2}(X)$$ where ${i_E}_*: H^{n}(E) \to H^{n+2k-2}(X)$ is the Gysin map. (Note that ${\text{codim}}(E,X)=k-1$)
\[thm:mainpartI\] $\phi: H^n(V) \to H^{n+2k-2}(X)$ induces a morphism $$\bar{\phi} : H^{n}_0 (V)(-k+1) \to H^{n+2k-2}_0(X)$$ which is an isomorphism of rational Hodge structures.
First we prove that the induced morphism $$\bar{\phi}: H^{n}_0(V) \to H^{n+2k-2}_0(X)$$ is well-defined. In the case when $n$ is odd, then $\bar{\phi}=\phi$, hence the morphism is well-defined. In case when $n$ is even, it is enough to show that $\phi({\text{im}}~i^*_V) \subseteq {\text{im}}~i^*_X$. Let $[E] \in H^{2k-2}(X)$ be the fundamental class of $E$. Note that $i^*_X: H^{2k-2}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}) \to H^{2k-2}(X)$ is an isomorphism by Lefschetz theorem since $2k-2 < \dim X = n+2k-2$. Hence there is $\gamma \in H^{2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k})$ such that $i^*_X(\gamma) =[E]$. In fact, we can write $\gamma = [{{\mathbb P}}^{k-1} \times S]$ where $[S] \in H^{2k-2}({{\mathbb P}}^{n+k})$ such that $$\label{eq:gamma}
i^*_X(\gamma) = [X \cap ({{\mathbb P}}^{k-1} \times S)] = [{{\mathbb P}}^{k-1} \times V]$$ To show this, we consider the following commutative diagram: $$\xymatrix{
H^0(E) \ar[r]_{{i_*}_E} \ar@/^1pc/[rr]^{(i_X \circ i_E)_*} & H^{2k-2}(X) \ar[r]_{{i_X}_*} & H^{2k}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \\
& H^{2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \ar[u]^{i^*_X}_{\cong} \ar[ur]_{\cup [X]}^{\cong}
}$$ Since $\gamma \in H^{2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k})$, we may write $\gamma=\sum^{k-1}_{i=0}[a_{i}H^i \times b_{i}L^{k-1-i}]$, where $H^i$(resp. $L^j$) is a linear space in ${{\mathbb P}}^{k-1}$(resp. ${{\mathbb P}}^{n+k}$) of codimension $i$(resp. $j$) By commutativity of the diagram, we have $$\begin{aligned}
[{{\mathbb P}}^{k-1}\times V] ={i_X}_*[E]={i_X}_*{i^*}_X(\gamma) =\gamma \cup [X]=\sum^{k-1}_{i=0}[(a_{i} H^i \times b_{i}L^{k-1-i}) \cap X ]\end{aligned}$$ Hence $a_{i}=0$ for $i \neq 0$, since $H^i \subsetneqq {{\mathbb P}}^{k-1}$ for $i \neq 0$ and then this forces $b_i =0$ for $i\neq 0$ since $[b_iL^{k-1-i}] \notin H^{2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k})$ for $i>0$. Hence $$\gamma = [{{\mathbb P}}^{k-1} \times b_0 L^{k-1}]$$ We take $S$ to be $b_0 L^{k-1}$ and we will use this $S$ later.
Now consider the following diagram: $$\begin{aligned}
\label{diag:int}
\xymatrix@=10pt{
\ar@/^1pc/[rrrr]^{\phi} H^{n}(V)\ar@{}[ddrr]|{(I)} \ar[rr]^{p^*_2} && H^n(E) \ar@{}[drr]|{(III)} \ar[rr]^{{i_E}_*} && H^{n+2k-2}(X) \\
&&& H^n(X) \ar[ul]_{i^*_E} \ar[ur]^{\cup [E]} & \\
H^{n}({{\mathbb P}}^{n+k}) \ar[uu]^{i^*_V} \ar[rr]_{{pr}^*_2} && H^{n}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \ar@{}[uur]|{(II)\qquad } \ar[ur]_{i^*_X} \ar[uu]^{j^*_E} \ar[rr]_{\Phi} && H^{n+2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \ar[uu]_{i^*_X} \ar@{}[uul]|{\qquad (IV)}
}\end{aligned}$$ where
1. $j_E: E \to {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$, $i_E: E \to X$ and $i_X:X \to {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$ are inclusions,
2. $pr_2: {{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k} \to {{\mathbb P}}^{n+k}$ is the projection to the second factor,
3. $\Phi:H^n({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \to H^{n+2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k})$ is defined by $\Phi(\alpha) = \alpha \cup \gamma$.
Commutativity of $(II)$ and $(III)$ are clear and commutativity of $(I)$ follows from the following commutative diagram: $$\xymatrix{
E = V \times {{\mathbb P}}^{n+k} \ar@{^(->}[rr]^{\quad j_E} \ar[d]^{p_2} && {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k} \ar[d]^{pr_2} \\
V \ar@{^(->}[rr]^{i_V} && {{\mathbb P}}^{n+k}
}$$ We show the commutativity of $(IV)$. Let $\alpha \in H^{n}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k})$. Then $$i^*_X \circ \Phi (\alpha)= i^*_X(\alpha \cup \gamma)=i^*_X(\alpha) \cup i^*_X(\gamma) =i^*_X(\alpha) \cup [E]$$ by the definition of $\gamma$. Hence $(IV)$ commutes. Therefore we have $$\phi \circ i^*_V = i^*_X \circ \Phi \circ pr^*_2$$ and hence $\phi({\text{im}}~ i^*_V) \subseteq {\text{im}}~ i^*_X$ and $\phi$ induces a well-defined morphism $\bar{\phi}: H^n_0(V) \to H^{n+2k-2}_0(X)$ in this case also.
To show that $\bar{\phi}$ is an isomorphism, note that we have $$\begin{aligned}
\label{eq:E}
H^{n}(E)&\cong& ( H^{n}(V) \otimes H^0({{\mathbb P}}^{k-1}) ) \oplus \left(\bigoplus^{k-1}_{i=1} H^{n-2i} ({{\mathbb P}}^{n+k}) \otimes H^{2i}({{\mathbb P}}^{k-1}) \right) \end{aligned}$$ by the Künneth formula and Lefschetz theorem.
(Case 1) If $n$ is odd : in this case $\phi=\bar{\phi}$ and we show that $\phi={i_E}_* \circ p^*_2$ is surjective. Since $n$ is odd, $H^{n-2i}({{\mathbb P}}^{k-1}) =0$ for all $i $. Hence (\[eq:E\]) gives an isomorphism $$p^*_2 : H^{n}(V) \stackrel{\cong}{\longrightarrow} H^{n}(E)$$ Now for the morphism ${i_E}_*: H^{n}(E) \to H^{n+2k-2}(X)$, consider the following Gysin exact sequence $$\label{eq:gysin}
\cdots \to H^{n+2k-3}(X-E) \to H^{n}(E) \stackrel{{i_E}_*}{\to} H^{n+2k-2}(X) \to H^{n+2k-2}(X-E) \to\cdots$$ Since $X -E \cong {{\mathbb P}}^{k-2} \times ({{\mathbb P}}^{n+k} -V)$, $$\begin{aligned}
H^{n+2k-2}(X-E)&=&H^{n+2k-2} ({{\mathbb P}}^{k-2} \times ({{\mathbb P}}^{n+k} -V))\\
&=& \bigoplus^{k-2}_{i=0} H^{(n+2k-2)-2i}({{\mathbb P}}^{n+k}-V)(-i)=0\end{aligned}$$ Hence, ${i_E}_* : H^{n}(E) \to H^{n+2k-2}(X)$ is surjective and hence we get the surjection $\phi = {i_E}_* \circ p^*_2$. Now lemma \[lemma:rank1\] implies that $\bar{\phi}=\phi$ is an isomorphism in this case.
(Case 2) If $n=2l$ is even : then (\[eq:E\]) gives that $p^*_2: H^{n}(V) \to H^{n}(E)$ is an injection. Since $X -E \cong {{\mathbb P}}^{k-2} \times ({{\mathbb P}}^{n+k} -V)$, we get $$H^{n+2k-3}(E)=H^{n+2k-3} ({{\mathbb P}}^{k-2} \times ({{\mathbb P}}^{n+k} -V)) = \bigoplus^{k-2}_{i=0} H^{(n+2k-3)-2i}({{\mathbb P}}^{n+k}-V)(-i)=0$$ Then Gysin exact sequence (\[eq:gysin\]) implies that ${i_E}_* : H^{n}(E) \to H^{n+2k-2}(X)$ is injective. Since $\phi= {i_E}_* \circ p^*_2$, we get the injection $\phi = {i_E}_* \circ p^*_2 : H^{n}(V) \to H^{n+2k-2}(X)$. We show that the induced map $\bar{\phi}$ is also an injection in this case: Suppose $\bar{\phi}(\bar{\alpha})=0$. Then $$\phi(\alpha)\in {\text{im}}[i^*_X: H^{n+2k-2}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})\to H^{n+2k-2}(X)]$$ where $\alpha \in H^{n}(V)$ which maps to $\bar{\alpha} \in H^{n}_0(V)$. Let $\beta \in H^{n+2k-2}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})$ such that $i^*_X(\beta) = \phi(\alpha)$. We claim that $\beta$ is a cycle in $H^0({{\mathbb P}}^{k-1}) \otimes H^{n+2k-2}({{\mathbb P}}^{n+k})$. $$\xymatrix@=12pt{
&&& H^{n+2k-2}(E)\\
H^n(V) \ar[r]^{p^*_2}&H^n(E) \ar[r]^{{i_E}_*} \ar@/^1pc/[urr]^{\cup c_{k-1}(N_{E/X})} & H^{n+2k-2}(X) \ar[ur]^{i^*_E}& \\
H^n({{\mathbb P}}^{n+k}) \ar[u]^{i^*_V} \ar[r]_{pr^*_2}& \ar[u]H^{n}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})\ar[r]_{\Phi}& H^{n+2k-2}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}) \ar[u]^{i^*_X} \ar[uur]_{(i_X \circ i_E)^*}&
}$$ By applying $i^*_E$ to $\phi(\alpha) =i^*_X(\beta) \in H^{n+2k-2}(X)$, we get $$\begin{gathered}
\label{eq:phialpha}
i^*_E(\phi(\alpha))= i^*_E {i_E}_* p^*_2(\alpha) = p^*_2(\alpha) \cup c_{k-1}(N_{E/X})
= i^*_E(i^*_X(\beta) )) =(i_X \circ i_E)^*(\beta)\end{gathered}$$ where the second equality comes from the self-intersection formula ([@fulton p103]). From the inclusions $E \stackrel{i_E}{\hookrightarrow} X \stackrel{i_X}{\hookrightarrow}{{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$, we have $$0 \to i^*_E (N_{X/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}}) \to N_{E/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}} \to N_{E/X} \to 0$$ Hence $$c ( N_{E/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}}) = c( i^*_E (N_{X/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}})) \cdot c(N_{E/X})$$ In particular, $$\begin{gathered}
\label{eq:ck0}
c_{k} ( N_{E/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}}) \\= c_1( i^*_E (N_{X/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}})) \cdot c_{k-1}(N_{E/X})= i^*_E([X]|_X) \cup_E [E]|_E\end{gathered}$$ since $X$ is a divisor in ${{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$, where $\cup_E$ is the cup product on $E$. Since the inclusion $i_X \circ i_E: E \to {{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$ is actually $(\text{id}_{{{\mathbb P}}^{k-1}}, i_V) :{{\mathbb P}}^{k-1} \times V \hookrightarrow {{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}$, we have $N_{E/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}}=p^*_2(N_{V/{{\mathbb P}}^{n+k}})$ and hence $$\label{eq:ck}
c_{k} ( N_{E/{{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}}) =p^*_2(c_k(N_{V/{{\mathbb P}}^{n+k}})) \in H^0({{\mathbb P}}^{k-1}) \otimes H^{2k}(V) \subset H^{2k}(E)$$ Now since $X$ is an ample divisor in ${{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}$, $[X] = [a H + b L] \in H^2({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})$ where $H$(resp. $L$) is a hyperplane in ${{\mathbb P}}^{k-1}$(resp. ${{\mathbb P}}^{n+k}$) and $a,b \geq 0$ such that $ab\neq 0$. Now by using (\[eq:gamma\]), $$\begin{aligned}
i^*_E([X]|_X) \cup_E [E]|_E &=& i^*_E ([X]|_X \cup_X [E])=i^*_E ([X]|_X \cup_X i^*_X (\gamma))\\
&=&i^*_E i^*_X([X] \cup_{{{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}} \gamma)\\
&=& i^*_E i^*_X ([a H +b L ]\cup_{{{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}} [{{\mathbb P}}^{k-1} \times S])\\
&=& i^*_E i^*_X ([a (H \times S) + b ({{\mathbb P}}^{k-1} \times (S \cap L))])\\
&\in & (H^2({{\mathbb P}}^{k-1})\otimes H^{2k-2}(V)) \oplus (H^0({{\mathbb P}}^{k-1}) \otimes H^{2k}(V))\end{aligned}$$ Then by (\[eq:ck0\]), (\[eq:ck\]), we have $a=0$ and hence $[X] = [bL]\in H^2({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})$. Set $P$ be a hypersurface of degree $b$ in ${{\mathbb P}}^{n+k}$ such that $[X] =[{{\mathbb P}}^{k-1}\times P]$. Since $X$ contains $E={{\mathbb P}}^{k-1} \times V$, we may assume that $V \subset P$. Then $N_{E/X} = p^*_2 (N_{V/P})$, and hence $$c_{k-1}(N_{E/X}) = p^*_2(c_{k-1}(N_{V/P})) \in H^0({{\mathbb P}}^{k-1})\otimes H^{2k-2}(V)$$ Now, since $p^*_2(\alpha) \in H^0({{\mathbb P}}^{k-1}) \otimes H^{n}(V)$, we have $$i^*_E \phi (\alpha) = p^*_2(\alpha) \cup c_{k-1}(N_{E/X}) \in H^0({{\mathbb P}}^{k-1})\otimes H^{n+2k-2}(V)$$ Then by (\[eq:phialpha\]), we have $$i^*_E \phi (\alpha)= (i_E \circ i_X)^*(\beta) = (\text{id}_{{{\mathbb P}}^{k-1}}, i^*_V)(\beta) \in H^{0} ({{\mathbb P}}^{k-1})\otimes H^{n+2k-2}(V)$$ Hence $\beta \in H^0({{\mathbb P}}^{k-1}) \otimes H^{n+2k-2}({{\mathbb P}}^{n+k})$.
Next note that by cupping with $\gamma = [{{\mathbb P}}^{k-1} \times S]$, $\Phi: H^{n}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k}) \to H^{n+2k-2}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})$ maps the Künneth component $H^{i}({{\mathbb P}}^{k-1})\otimes H^{n-i}({{\mathbb P}}^{n+k})$ of $H^n({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})$ to the Künneth component $H^{i}({{\mathbb P}}^{k-1}) \otimes H^{n+2k-2-i}({{\mathbb P}}^{n+k})$ of $H^{n+2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k})$ isomorphically. Hence we have $\eta \in H^0({{\mathbb P}}^{k-1}) \otimes H^{n}({{\mathbb P}}^{n+k})$ such that $\eta \cup \gamma = \Phi(\eta)= \beta$. Now $\eta \in {\text{im}}[pr^*_2: H^n({{\mathbb P}}^{n+k}) \to H^{n}({{\mathbb P}}^{k-1}\times {{\mathbb P}}^{n+k})]$ and we may consider $\eta \in H^{n}({{\mathbb P}}^{n+k})$ since $pr^*_2$ is injective. Then by commutativity of diagram (\[diag:int\]), we have $$\phi \circ i^*_V(\eta)= i^*_X \circ \Phi \circ {pr^*_2}(\eta)= i^*_X(\eta \cup \gamma)=i^*_X(\beta) =\phi(\alpha)$$ Since $\phi$ is injective, we have $\alpha = i^*_V(\eta)$. Hence $\bar{\phi}$ is also an injection. Now by lemma \[lemma:rank1\], we can conclude that $\bar{\phi}: H^{n}_0(V) \to H^{n+2k-2}_0(X)$ is an isomorphism.
Proof of Theorem \[thm:main\] when $n+k$ is odd {#sec:part-ii-n}
===============================================
Throughout this section, we assume that $n+k=2m+1$.
In order to prove theorem \[thm:main\], we use the construction of O’Grady[@ogrady]. We give a brief outline of his construction here. For detailed construction, see [@ogrady].
Recall the diagram (\[diag:proj\]) and consider the projection $p_1 : X \to {{\mathbb P}}^{k-1}$ and recall that $\Delta$ is the discriminant variety, which is a smooth hypersurface in ${{\mathbb P}}^{k-1}$ by our assumption. In case when $n+k$ is odd, for a general $t \in {{\mathbb P}}^{k-1}$, the fiber $p^{-1}_1(t) = Q_t$ is a smooth quadric of dimension $n+k-1=2m$ in ${{\mathbb P}}^{n+k}$. Hence it contains two irreducible families of $m-$planes parametrized by $F^1_t$ and $F^2_t$. Note that $F^1_t \cong F^2_t$ and $\dim F^i_t = \frac{m(m+1)}{2}$ [@gh]. Let $F$ be the abstract variety to which $F^i_t$ is isomorphic for $i=1,2$ and for any $t$. Let $W$ be a double covering of ${{\mathbb P}}^{k-1}$ branched over $\Delta$ and let $\sigma: W \to {{\mathbb P}}^{k-1}$ be the covering map. Set $$P =\{ M \subset X ~|~ p_1(M) = t \in {{\mathbb P}}^{k-1} \text{ a point, } p_2(M) \cong {{\mathbb P}}^{m} \subset {{\mathbb P}}^{n+k}\}$$ Then there is a natural map $\psi : P \to {{\mathbb P}}^{k-1}$ defined by $\psi(M) = p_1(M) \in {{\mathbb P}}^{k-1}$. Then the Stein factorization of $\psi$ is factored through $W$ and get a composition $$\psi : P \stackrel{f}{\longrightarrow} W \stackrel{\sigma}{\longrightarrow} {{\mathbb P}}^{k-1}$$ and set $$\begin{aligned}
\Gamma &=& \{ (M,x) \in P \times X ~|~ x \in M \subset X\} \\
&=& \{(M,x) \in P \times X ~|~ p_1(x) = p_1(M) \in {{\mathbb P}}^{k-1},~p_2(x) \in p_2(M)\cong {{\mathbb P}}^{m} \} \end{aligned}$$ Let $pr_1 : \Gamma \to P$ and $pr_2 : \Gamma \to X$ be the projections. We summarize the construction in the following diagram: $$\xymatrix@R=10pt{
& \Gamma \ar[dl]_{pr_1} \ar[dr]^{pr_2}& \\
P \ar[d]^f \ar[drr]^{\psi} && X \ar[d]^{p_1} \\
W \ar[rr]_{\sigma} && {{\mathbb P}}^{k-1}
}$$ Note that for any $w \in W$, $$f^{-1}(w) = \{ M \in X~|~ p_1(M) = \sigma(w), p_2(M) \cong {{\mathbb P}}^{m} \in {{\mathbb P}}^{n+k}\}
\cong\{ {{\mathbb P}}^m \subset Q_{\sigma(w)} \} \cong F$$ and for any $M \in P$, $$\begin{aligned}
pr^{-1}_1(M) &=& \{(M,x) \in P \times X~|~ x \in M \subset X \}\\
&\cong& \{x \in X ~|~ p_1(x) =p_1(M),~ p_2(x) \in p_2(M) \cong {{\mathbb P}}^{m} \} \cong {{\mathbb P}}^{m}\end{aligned}$$ Hence $f:P \to W$ is $F-$bundle and $pr_1: \Gamma \to P$ is ${{\mathbb P}}^m-$bundle.
Now we give a proof of theorem \[thm:main\] in case of $n+k$ odd.
Consider the morphism of Hodge structures $${pr_2}_* : H_{n+2k-2}(\Gamma) \to H_{n+2k-2}(X)$$ We show that ${pr_2}_*$ is a surjection. First note that $pr_2:\Gamma \to X$ is a surjection. Hence we can take an iterated hyperplane section $\Gamma_1$ of $\Gamma$ such that $\dim \Gamma_1 = \dim X = n+2k-2$ and $\Gamma_1$ surjects onto $ X$. Let $g = pr_2|_{\Gamma_1}:\Gamma_1 \to X$. Then $g$ is a generically finite map. Let $j : \Gamma_1 \to \Gamma$ be an inclusion. Then we have the following commutative diagram:
$$\xymatrix{
\Gamma \ar[dr]^{pr_2}& &&H_{n+2k-2}(\Gamma) \ar[drr]^{{pr_2}_*} && \\
\Gamma_1 \ar@{^(->}[u]^j \ar[r]^{g} & X && H_{n+2k-2}(\Gamma_1) \ar[u]^{j_*} \ar[rr]^{g_*} \ar[d]^{P_{\Gamma_1}}_{\cong} && H_{n+2k-2}(X) \ar[d]^{P_{X}}_{\cong}\\
&&&H^{n+2k-2}(\Gamma_1) \ar[rr]^{g_*} && H^{n+2k-2}(X) \ar@/^1pc/[ll]^{g^*}
}$$ where $P_{\Gamma_1}$ and $P_X$ are isomorphisms from Poincaré duality. Since $g_* ={pr_2}_* \circ j_*$, it is enough to show that $g_*$ is surjective, but it is clear since $g_* \circ g^* = \deg g \cdot \text{id}$.
Now since $pr_1: \Gamma \to P$ is ${{\mathbb P}}^m-$bundle, by Künneth formula we have $$\begin{aligned}
H_{n+2k-2}(\Gamma) &=& \bigoplus^{m}_{j=0} ( H_{n+2k-2-2j}(P) \otimes H_{2j}({{\mathbb P}}^m))
\\
&=& \bigoplus^{m}_{j=0} H_{n+2k-2-2j}(P)(j)\end{aligned}$$ We claim that for each $j$, there is an isomorphism $$H_{n+2k-2-2j}(P) = \bigoplus_{r+2s=n+2k-2-2j}(H_{r}(W) \otimes H_{2s}(F))$$ To show this, first note that we can use cohomology instead of homology since all our varieties considered are smooth. Let $U_W \subset W$ be a Zariski open set in $W$ such that $f^{-1}(U_W) \cong U_W \times F$. Since $F$ has a celluar decomposition [@ehr], $H_*(F)$ are generated by algebraic cycles. Let $\alpha_1 , \cdots, \alpha_l$ be the algebraic cycles generating $H_*(F)$. i.e. there are algebraic subvarieties $Z_1, \cdots, Z_l$ of $F$ such that the fundamental classes of them are $\alpha_1,..., \alpha_l$. Let $ p: U_W \times F \to F$ be the projection to the second factor and consider the algebraic cycles $p^*(\alpha_i)$ in $H^*(U_W \times F)$ which are supported on $U_W \times Z_i$ for $i=1,2,...,l$. Let $\beta_i$ be the closures of $p^*(\alpha_i)$ in $P$ for each $i=1,..., l$. Then this gives the splitting of the restriction map $H^*(P) \to H^*(F)$. So we may apply the Leray-Hirsch theorem [@sp]. Therefore we have $$H^q(P) = \bigoplus_{r+2s=q}(H^r(W) \otimes H^{2s}(F))$$ for any $q$. In particular, $$\label{eq:p}
H^{2d-(n+2k-2-2j)}(P) = \bigoplus_{s}(H^{2d-(n+2k-2-2j)-2s}(W) \otimes H^{2s}(F))$$ where $d =\dim P$.
(Case 1) If $n$ is odd (and hence $k$ is even) : in this case, from (\[eq:p\]) we have $$H^{2d -(n+2k-2-2j)}(P ) \cong H^{k-1}(W)\otimes H^{2d-(n+3k-3-2j)}(F)$$ for each $j$, since $W$ is simply connected, $H^r(W)=0$ for all odd $r$ such that $r \neq \dim W =k-1$, or equivalently $$H_{n+2k-2-2j}(P) \cong H_{k-1}(W) \otimes H_{n+k-1-2j}(F) = H_{k-1}(W)(m-j)^{\oplus l_j}$$ where $l_j = \dim H_{n+k-1-2j}(F)$. Therefore we have $$\begin{aligned}
H_{n+2k-2}(\Gamma) &=& \bigoplus^{m}_{j=0} H_{n+2k-2-2j}(P)(j)\\
&\cong& \bigoplus^{m}_{j=0} \left(H_{k-1}(W)(m-j )^{\oplus l_j} \right)(j) = H_{k-1}(W)(m)^{\oplus N}\end{aligned}$$ where $N = \sum^m_{j=0} l_j$. Thus, we have a surjection $${pr_2}_* : H_{k-1}(W)(m)^{\oplus N} \to H_{n+2k-2}(X)$$ and by Poincaré duality, we get a surjection $$\xymatrix{
H^{k-1}(W)(-m)^{\oplus N} \ar@{>>}[r] & H^{n+2k-2}(X) }$$ By choosing $\tilde{W}$ to be the disjoint union of $N$ copies of $W$ and by composing with $\bar{\phi}^{-1}$ in theorem \[thm:mainpartI\], we get a surjection $$\xymatrix{
\Theta : H^{k-1}(\tilde{W})(-m+k-1) \ar@{>>}[r] & H^n (V)
}$$ as we claimed.
(Case 2) If $n$ is even (and hence $k$ is odd) : in this case, (\[eq:p\]) gives $$H^{2d -(n+2k-2-2j)}(P ) \cong \bigoplus_{r} (H^{2r}(W) \otimes H^{2d-(n+2k-2-2j)-2r}(F))$$ for each $j$, or equivalently $$H_{n+2k-2-2j}(P) \cong \bigoplus_s \left(H_{2s}(W) \otimes H_{n+2k-2-2j-2s}(F) \right)$$ Therefore we have $$\begin{aligned}
H_{n+2k-2}(\Gamma) &=& \bigoplus^{m}_{j=0} H_{n+2k-2-2j}(P)(j)\\
&\cong& \bigoplus^{m}_{j=0}\bigoplus_s \left(H_{2s}(W) \otimes H_{n+2k-2-2j-2s}(F) \right)(j)\\
&=& \bigoplus_s H_{2s}(W) \otimes \left( \bigoplus^m_{j=0} H_{n+2k-2-2j-2s}(F) \right)(j)\\
&=& \bigoplus_s H_{2s}(W)(q_s)^{\oplus l_s}\end{aligned}$$ where $q_s = \frac{n+2(k-1)-2s}{2}$ and $l_s = \sum^m_{j=0} \dim H_{n+2k-2-2j-2s}(F)$. Thus, we have a surjection $$\xymatrix{
{pr_2}_* : \bigoplus_s H_{2s}(W)(q_s)^{\oplus l_s} \ar@{>>}[r] & H_{n+2k-2}(X)
}$$ Recall that $H^{n+2k-2}_0(X) = H^{n+2k-2}(X)/ {\text{im}}~i^*_X$. Hence we get a composition of surjections $$\xymatrix{
\bigoplus_s H^{2s}(W)(-q_s)^{\oplus l_s} \ar@{>>}[r] & H^{n+2k-2}(X) \ar@{>>}[r] & H^{n+2k-2}_0(X)
}$$ Therefore, by composing with $\bar{\phi}^{-1}$ from theorem \[thm:mainpartI\] we get a surjection of Hodge structures $$\xymatrix{
\Theta : \bigoplus_s H^{2s}(W)(-q_s+k-1)^{\oplus l_s} \ar@{>>}[r] & H^{n}_0(V)
}$$ (Note that $q_s -k+1 = \frac{n-2s}{2}$.) This completes the proof of theorem \[thm:main\] in case when $n+k$ is odd.
Proof of Theorem \[thm:main\] when $n+k$ is even {#sec:case-when-n}
================================================
Throughout this section, we assume that $n+k=2m$ is even.
Again we start by considering the projection $p_1: X \to {{\mathbb P}}^{k-1}$ in diagram (\[diag:proj\]). Recall that the discriminant variety $\Delta$ is a smooth hypersurface of degree $n+k+1$ in ${{\mathbb P}}^{k-1}$ by our assumption. Set $U ={{\mathbb P}}^{k-1} -\Delta$ and let $X_{\Delta} = p^{-1}_1(\Delta)$ and $X_{U} =p^{-1}_1(U)$. $$\begin{aligned}
\label{diag}
\xymatrix{
X_{\Delta} \ar@{^(->}[r] \ar[d]^{p_1} & X \ar[d]^{p_1} & \ar@{_(->}[l] \ar[d]^{p_1} X_U \\
\Delta \ar@{^(->}[r] & {{\mathbb P}}^{k-1} &\ar@{_(->}[l] U ={{\mathbb P}}^{k-1} - \Delta
}\end{aligned}$$ Note that for any $t \in \Delta$, $p^{-1}_1(t)$ is a singular quadric in the family. Since we have assumed that $\Delta$ is smooth, a singular fiber is a cone through a $0-$plane (i.e. a point) over a quadric of rank $n+k$ in ${{\mathbb P}}^{n+k-1}$, i.e. all singular fibers are cones over a smooth quadric $\tilde{Q}_t$ of dimension $n+k-2$ in ${{\mathbb P}}^{n+k-1}$. For any $t\in \Delta$, we denote $p^{-1}_1(t) = C_t$ a cone with a vertex $0_t$. We can form a family of quadrics of dimension $n+k-2$ over $\Delta$ as follows: Let $s: \Delta \to X_{\Delta}$ be a section of $p_1$ defined by $s(t) = 0_t$ for any $t\in \Delta$. Let $Y_0$ be a general hyperplane section of $X_{\Delta} - s(\Delta)$ and let $p_1|_{Y_0}: Y_0 \to \Delta$ be the obvious map. Let $Y$ be a smooth compactification of $Y_0$. Then by using theorem by Hironaka, we may assume that the rational map $Y \to \Delta$ is actually a morphism. We denote this morphism by $\pi:Y \to \Delta$. Then for a general $t \in \Delta$, $\pi^{-1}(t) = \tilde{Q}_t$ a smooth quadric of dimension $n+k-2$ in ${{\mathbb P}}^{n+k-1}$. $$\xymatrix{
Y \ar[dr]_{\pi} & \ar@{_(->}[l]Y_0 \ar@{^(->}[r] \ar[d]^{p_1|_{Y_0}} & X_{\Delta} \ar[d]^{p_1} \ar@{^(->}[r] & X \ar[d]_{p_1}\\
&\Delta \ar@{=}[r]&\Delta \ar@{^(->}[r] & {{\mathbb P}}^{k-1}
}$$
For any $p$ $$H^{p}(X_{\Delta}) \cong H^{p-2}(Y)(-1) \quad \text{as Hodge structures}$$
Consider the Leray spectral sequence associated to the map $p_1: X_{\Delta} \to \Delta$ $$'E^{pq}_2 = H^p(\Delta, R^q {p_1}_* {{\mathbb Q}}) \quad \Rightarrow \quad H^{p+q}(X_{\Delta},{{\mathbb Q}})$$ and one associated to the map $\pi : Y \to \Delta$ $$''E^{pq}_2 = H^p (\Delta, R^q {\pi}_* {{\mathbb Q}}) \quad \Rightarrow \quad H^{p+q}(Y,{{\mathbb Q}})$$ Since for any $t \in \Delta$, $C_t$ is a cone through a point $0_t$ over a smooth quadric $\tilde{Q}_t$ in ${{\mathbb P}}^{n+k-1}$, we have $$\begin{aligned}
R^q {p_1}_* {{\mathbb Q}}&=& R^q {(p_1|_{Y_0})}_! {{\mathbb Q}}= (R^{2(n+k-1)-q}{({p_1|}_{Y_0})}_*{{\mathbb Q}})^*(-n-k+1) \\
&=& (R^{2(n+k-1)-q}{\pi}_*{{\mathbb Q}})^*(-n-k+1) = R^{q-2}{\pi}_*{{\mathbb Q}}(-1)\end{aligned}$$ Hence we have $$'E^{pq}_2 ~=~ ''E^{p,q-2}_2(-1)$$ and so $$H^{p+q}(X_{\Delta}) \cong H^{p+q-2}(Y)(-1)$$ as Hodge sturctures [@arapura].
In particular, $$\label{eq:9}
H^{n+2k-2}(X_{\Delta}) \cong H^{n+2k-4}(Y)(-1)$$
\[lemma:inj\] There is an injection of Hodge structures $$0 \to H^{n+2k-2}_0 (X) \to H^{n+2k-4}(Y)(-1) $$
From the top row of the diagram (\[diag\]), we have an exact sequence of mixed Hodge structures
$$\label{eq:mhs}
\cdots \to H^{n+2k-2}_c (X_U) \to H^{n+2k-2}(X) \to H^{n+2k-2}(X_{\Delta}) \to \cdots$$
By using (\[eq:9\]) and by taking the exact functor $Gr^W_{n+2k-2}$, we get
$$\label{eq:7}
0 \to Gr^W_{n+2k-2} H^{n+2k-2}_c (X_U) \to H^{n+2k-2}(X) \to H^{n+2k-4}(Y)(-1) \to \cdots$$
Consider the morphism $p_1: X_U \to U$ and the Leray spectral sequence associated to it $$E^{pq}_2 = H^p_c (U , R^q {p_1}_* {{\mathbb Q}}) \quad \Rightarrow \quad H^{p+q}_c(X_U,{{\mathbb Q}})$$ Note that $$E^{pq}_2 = \begin{cases} 0 & \text{if $q$ is odd} \\ H^p_c(U,{{\mathbb Q}}) & \text{if $q$ is even} \end{cases}$$ since $(R^q {p_1}_* {{\mathbb Q}})_t = H^q (Q_t,{{\mathbb Q}})$ and $Q_t$ is a smooth quadric of dimension $n+k-1$(odd). Also, we have $$0 \to H^0_c(U) \to H^0({{\mathbb P}}^{k-1}) \to H^0(\Delta) \to H^1_c(U) \to 0$$ $$0 \to H^{2i-1}(\Delta) \to H^{2i}_c(U) \to H^{2i}({{\mathbb P}}^{k-1}) \to H^{2i}(\Delta) \to H^{2i+1}_c(U) \to 0$$ for $1 \leq i \leq k-2$ and $$H^{2k-2}_c(U) \cong H^{2k-2}({{\mathbb P}}^{k-1})$$ Now, since $\Delta$ is a smooth hypersurface in ${{\mathbb P}}^{k-1}$, by Lefschetz theorem we get $$H^j({{\mathbb P}}^{k-1}) \stackrel{\cong}{\longrightarrow} H^j(\Delta), \qquad \text{for } j < \dim \Delta= k-2$$ Hence $H^{2i-1}(\Delta) = 0$ and hence $H^{2i}_c(U)=H^{2i+1}_c(U)=0$ for $i$ such that $2i < \dim \Delta = k-2$. By applying duality on $H^{2i-1}(\Delta)$, we get $H^j_c(U)=0$ unless $j = k-1$ or $j=2k-2$. Hence the Leray spectral sequence degenerates at $E_2$ and we have a short exact sequence $$0 \to E^{2k-2,n}_{\infty} \to H^{n+2k-2}_c(X_U) \to E^{k-1,n+k-1}_{\infty} \to 0$$ Note that $E^{k-1,n+k-1}_{\infty} = E^{k-1,n+k-1}_2 =0$ since $n+k$ is even. Hence we have $$\label{eq:short}
E^{2k-2,n}_{\infty} \cong H^{n+2k-2}_c(X_U)$$
(Case 1) If $n$ is odd : Then $ E^{2k-2,n}_{\infty}=0$ also and hence we have $H^{n+2k-2}_c(X_U)=0$. So (\[eq:7\]) gives an injection $$0 \to H^{n+2k-2}(X) \to H^{n+2k-4}(Y)(-1)$$
(Case 2) If $n$ is even : In this case, (\[eq:short\]) gives $$\begin{aligned}
H^{n+2k-2}_c (X_U) &\cong& E^{2k-2,n}_{\infty}=E^{2k-2,n}_2 = H^{2k-2}_c(U,R^{n}{p_1}_* {{\mathbb Q}}) \\
&\cong& H^{2k-2}_c(U) \otimes H^n(Q_t) \cong H^{2k-2}({{\mathbb P}}^{k-1}) \otimes H^n({{\mathbb P}}^{n+k})\end{aligned}$$ by Lefschetz theorem. Hence, we can rewrite the exact sequence (\[eq:7\]) as follows: $$0 \to H^{2k-2}({{\mathbb P}}^{k-1}) \otimes H^{n}({{\mathbb P}}^{n+k})\stackrel{h_*}{\to} H^{n+2k-2}(X) \to H^{n+2k-4}(Y) \to \cdots$$ Recall that $H^{n+2k-2}_0(X) \cong H^{n+2k-2}(X)/{\text{im}}~i^*_X$ and note that ${\text{im}}~h_* \cap H^{n+2k-2}_0(X) = \emptyset$: In fact, $h_*$ can be factorized as $$\xymatrix{
H^{2k-2}({{\mathbb P}}^{k-1}) \otimes H^{n}({{\mathbb P}}^{n-k}) \ar@{^(->}[r] \ar@/^1pc/[rr]^{h_*}&H^{n+2k-2}({{\mathbb P}}^{k-1} \times {{\mathbb P}}^{n+k}) \ar@{^(->}[r]_{\qquad i^*_X} & H^{n+2k-2}(X)
}$$ Hence ${\text{im}}~ h_* \subseteq {\text{im}}~i^*_X$, and we get an injection $$0 \to H^{n+2k-2}_0(X) \to H^{n+2k-4}(Y)(-1)$$
Now we have a family $\pi:Y \to \Delta$ of quadrics of dimension $n+k-2$ which is even, so we can form a double covering of $\Delta$ as in the construction of O’Grady([@ogrady]). For a general $t \in \Delta$, the fiber $\tilde{Q}_t$, which is a smooth quadric of dimension $n+k-2=2(m-1)$, contains two $\frac{m(m-1)}{2}-$dimensional irreducible families $F^1_t, F^2_t$ of $(m-1)-$planes[@gh]. As in the case of $n+k$ odd, we can form the following diagram $$\xymatrix@R=10pt{
&\ar[dl]_{pr_1} \Gamma \ar[dr]^{pr_2}& &\\
P \ar[d]_f \ar[drr]^{\psi} && Y \ar[d]^{\pi} \ar[r]^{\pi_2 \quad} & {{\mathbb P}}^{n+k-1}\\
W \ar[rr]_{\sigma} && \Delta&
}$$ where
1. $P = \{ M \subset Y ~|~ \pi_1(M) = t \in \Delta \text{ a point}, \pi_2(M)\cong {{\mathbb P}}^{m-1} \subset {{\mathbb P}}^{n+k-1} \}$
2. $\Gamma = \{ (M, y) \in P \times Y~|~ \pi(y) =\pi(M) \in \Delta, \pi(y) \in \pi_2(M) \cong {{\mathbb P}}^{m-1} \subset \tilde{Q}_{\pi(M)} \}$, $pr_1$ and $pr_2$ are projections.
3. $\psi : P \to \Delta$ is a natural map defined by $\psi(M) = \pi(M) =t \in \Delta$
4. $f: P \to W$ is $F-$bundle and $pr_1:\Gamma \to P$ is ${{\mathbb P}}^{m-1}-$bundle, where $F$ is the abstract variety such that $F^i_t$ is isomorphic to it for $i=1,2$ and $t \in \Delta$
5. $\sigma: W \to \Delta$ is a double covering branched over the discriminant variety $\Delta_1$ of the family $\pi : Y\to \Delta$.
\[lemma:surjoddk\] There are surjections of Hodge structures
1. $$H^{k-2}(\tilde{W})(-m+1) \to H^{n+2k-4}(Y) \qquad \text{if $n$ is odd}$$ where $\tilde{W}$ is a disjoint union of finitely many copies of $W$,
2. $$\bigoplus_r H^{2r}(W)(-q_r)^{\oplus l_r} \to H^{n+2k-4}(Y) \qquad \text{if $n$ is even}$$ where $q_r=\frac{n+2k-4-2r}{2}$ and $ l_r=\sum_j \dim H_{n+2k-4-2j-2r}(F)$.
First note that by using the same arguments in the proof of theorem \[thm:main\] in the case of $n+k$ odd, we can show that $$\label{eq:pr2}
{pr_2}_*: H_{n+2k-4}(\Gamma) \to H_{n+2k-4}(Y)$$ is surjective and $$\begin{aligned}
H_{n+2k-4}(\Gamma) &=& \bigoplus^{m-1}_{j=0} (H_{n+2k-4-2j} (P) \otimes H_{2j}({{\mathbb P}}^{m-1}) )\\
&=& \bigoplus^{m-1}_{j=0} H_{n+2k-4-2j} (P)(j)\\
&=& \bigoplus^{m-1}_{j=0} \bigoplus_{s}(H_{n+2k-4-2j-2s}(W) \otimes H_{2s}(F))(j)
\end{aligned}$$
(Case 1) If $n$ is odd : In this case, $n+2k-4-2j-2s$ is odd. But $H_r(W)$ is zero for an odd number $r$ unless $r=k-2 =\dim W$. Hence we have $$H_{n+2k-4-2j}(P) =H_{k-2}(W) \otimes H_{n+k-2-2j}(F) = H_{k-2}(W)(m-1-j)^{\oplus l_j}$$ where $l_j=\dim H_{n+k-2-2j}(F)$. Thus we have $$\begin{aligned}
H_{n+2k-4}(\Gamma) &=& \bigoplus^{m-1}_{j=0} H_{k-2}(W)(m-1-j)^{\oplus l_j}(j)\\
&=&\bigoplus^{m-1}_{j=0} H_{k-2}(W)(m-1)^{\oplus l_j} = H_{k-2}(W)(m-1)^{\oplus N}\end{aligned}$$ where $N = \sum^m_{j=0} l_j$. Hence by (\[eq:pr2\]) and using Poincaré duality, we get a surjection $$\xymatrix{
{pr_2}_* : H^{k-2}(W)(-m+1)^{\oplus N} \ar@{>>}[r] & H^{n+k-4}(Y)
}$$ By setting $\tilde{W}$ to be a disjoint union of $N$ copies of $W$, we get the desired surjection in this case.
(Case 2) If $n$ is even : In this case, $n+2k-4-2j$ is even. Hence we have $$\begin{aligned}
H_{n+2k-4-2j}(P)&=& \bigoplus_{s}(H_{2s}(W) \otimes H_{n+2k-4-2j-2s}(F))\end{aligned}$$ Thus we have $$\begin{aligned}
H_{n+2k-4}(\Gamma) &=& \bigoplus^{m-1}_{j=0} H_{n+2k-4-2j} (P)(j)\\
&=& \bigoplus^{m-1}_{j=0}\left(\bigoplus_{s}(H_{2s}(W) \otimes H_{n+2k-4-2j-2s}(F)) \right) (j)\\
&=& \bigoplus_{s} H_{2s}(W) \otimes \left(\bigoplus^{m-1}_{j=0}H_{n+2k-4-2j-2s}(F) \right)(j)\\
&=& \bigoplus_{s} H_{2s}(W)(q_s)^{\oplus l_s}\end{aligned}$$ where $q_s = \frac{n+2k-4-2s}{2}$ and $l_s = \sum^{m-1}_{j=0} \dim H_{n+2k-4-2j-2s}(F)$. Hence by (\[eq:pr2\]) and using Poincaré duality, we get a surjection $$\xymatrix{
{pr_2}_* : \bigoplus_{r}H^{2r}(W)(-q_r)^{\oplus l_r} \ar@{>>}[r] & H^{n+2k-4}(Y)
}$$ in this case.
We can summarize lemmas \[lemma:inj\] and \[lemma:surjoddk\] in the following diagram: $$\begin{aligned}
\label{diag:theta}
\xymatrix{
& H^* (W) \ar@{>>}[d]^{\theta} \ar[dr]^{\psi_2 \circ \theta}&\\
0 \to H^{n+2k-2}_0 (X) \ar[r]^{\psi_1} & H^{n+2k-4}(Y)(-1) \ar[r]^{\psi_2 \qquad \quad } & Gr^{W}_{n+2k-2}H^{n+2k-1}_c(X_U) \to 0 }\end{aligned}$$ where with the notation in theorem \[lemma:surjoddk\] $$H^*(W) = \begin{cases} H^{k-2}(\tilde{W})(-m) & \text{if $n$ is odd} \\
\bigoplus_r H^{2r}(W)(-q_r-1)^{\oplus l_r} & \text{if $n$ is even} \end{cases}$$ and $\theta: H^*(W) \to H^{n+2k-4}(Y)(-1)$ is the surjection in lemma \[lemma:surjoddk\].
To finish the proof of theorem \[thm:main\] in this case, we need to lift the surjection $\theta$ to a surjection onto $H^{n+2k-2}_0(X)$. In order to do that, we construct a section $s_U$ of $\psi_2$ in geometric way. Following two lemmas will lead us the desired section $s_U$.
There is a decomposition $$H^{n+2k-4}(Y) \cong \begin{cases} H^{k-2}(\Delta, R^{n+k-2}\pi_*{{\mathbb Q}}) & \text{if $n$ is odd}\\
H^{k-2}(\Delta, R^{n+k-2}\pi_*{{\mathbb Q}}) \oplus \bigoplus^{k-2}_{p=0, p\neq \frac{k-2}{2}} (\alpha_p \otimes \beta_p){{\mathbb Q}}& \text{if $n$ is even}
\end{cases}$$ where $\alpha_p$ (resp. $\beta_p$) is an algebraic cycle generating $H^{2p}(\Delta)$ (resp. $H^{n+2k-4-2p}(\tilde{Q}_t)$).
Consider the Leray spectral sequence associated to the map $\pi: Y \to \Delta$. $$E^{pq}_2 = H^p(\Delta, R^q {\pi}_* {{\mathbb Q}}) \quad \Rightarrow \quad H^{p+q}(Y, {{\mathbb Q}})$$ Note that $$\label{eq:ss}
E^{pq}_2 = \begin{cases} 0 & \text{if either $q$ is odd or $p$ is odd, $p=k-2$} \\
H^p(\Delta,{{\mathbb Q}}) & \text{if $p$ is even and $q$ is even, $q \neq n+k-2$}\\
H^p (\Delta,R^{n+k-2}{\pi}_*{{\mathbb Q}}) &\text{if $p$ is even and $q=n+k-2$} \end{cases}$$ where $(R^{n+k}{\pi}_*{{\mathbb Q}})_t={{\mathbb Q}}^2$ for $t\in \Delta$. Then the spectral sequence degenerates at $E_2$. Let $F^{{{\bullet}}}$ be the filtration on $H^{n+2k-4}(Y)$ obtained from the spectral sequence. Then for each $p=0,1,..., 2(k-2)$ we have a short exact sequence $$\label{eq:split}
0 \to F^{p+1} \to F^p \to \text{Gr}^p_F H^{n+2k-4}(Y) \to 0$$ where $$\text{Gr}^p_F H^{n+2k-4}(Y) = E^{p, n+2k-4-p}_{\infty} = H^{p}(\Delta , R^{n+2k-4-p}{\pi}_*{{\mathbb Q}})$$
(Case 1) If $n$ is odd (and hence $k$ is odd): then $n+2k-4$ is odd. Hence $E^{pq}_2=0$ unless $(p,q)=(k-2, n+k-2)$. Hence we have $$\label{eq:xdelta}
H^{n+2k-4}(Y) \cong H^{k-2}(\Delta, R^{n+k-2}{\pi}_*{{\mathbb Q}})$$
(Case 2) If $n$ is even (and hence $k$ is even): then $n+2k-4$ is also even. Note $ F^{2p-1} = F^{2p}$ from (\[eq:ss\]). Hence we may write above sequence as $$\label{seq:splitting}
0 \to F^{2p+2} \to F^{2p} \stackrel{t_{2p}}{\to} \text{Gr}^{2p}_F H^{n+2k-4}(Y) \to 0$$ First by using descending induction on $p$, we show that for $p$ such that $k-2 < 2p \leq 2(k-2)$ there is a natural splitting of exact sequences (\[seq:splitting\]) such that $$F^{2p} \cong F^{2p+2} \oplus \text{Gr}^{2p}_FH^{n+2k-4}(Y) \cong \bigoplus^{k-2}_{l=p+1} (\alpha_l \otimes \beta_l) {{\mathbb Q}}$$ where $\alpha_l$ is an algebraic cycle generating $H^{2l}(\Delta) \cong {{\mathbb Q}}$ and $\beta_l$ is an algebraic cycle generating $ H^{n+2k-4-2l}(\tilde{Q}_t)(-1) \cong {{\mathbb Q}}$.
If $2p=2(k-2)$, then $$\begin{aligned}
F^{2(k-2)} &\cong& \text{Gr}^{2(k-2)}_F H^{n+2k-4}(Y) = E^{2(k-2), n}_{\infty}\\
&=& H^{2(k-2)}(\Delta) \otimes H^{n}(\tilde{Q}_t) \cong (\alpha_{k-2} \otimes \beta_{k-2}){{\mathbb Q}}\end{aligned}$$ where $\alpha_{k-2}$(resp. $\beta_{k-2}$) is an algebraic cycle generating $H^{2(k-2)}(\Delta)={{\mathbb Q}}$ (resp. $ H^{n}(\tilde{Q}_t)={{\mathbb Q}}$). Now suppose $k-2<2p<2(k-2)$ and consider the short exact sequnce (\[seq:splitting\]). Since $2p>k-2$, $\dim H^{2p}(\Delta)=\dim H^{n+2k-4-2p}(\tilde{Q}_t)=1$, hence we can choose an algebraic cycle $\alpha_p$ (resp. $\beta_p$) which generates $H^{2p}(\Delta)$ (resp. $H^{n+2k-4-2p}(\tilde{Q}_t)$) and hence $$\begin{aligned}
\text{Gr}^{2p}_F H^{n+2k-4}(Y) &=&H^{2p}(\Delta, R^{n+2k-4-2p}{\pi}_*{{\mathbb Q}}) = H^{2p}(\Delta) \otimes H^{n+2k-4-2p}(\tilde{Q}_t) \\
&=& (\alpha_p \otimes \beta_p){{\mathbb Q}}\end{aligned}$$ and by induction hypothesis $F^{2p+2}$ has a decomposition $$F^{2p+2} \cong \bigoplus^{k-2}_{l=p+1}(\alpha_l \otimes \beta_l) {{\mathbb Q}}$$ with a basis $\mathcal{B}_{p+1}=\{\alpha_{p+1} \otimes \beta_{p+1},..., \alpha_{k-2} \otimes \beta_{k-2}\}$, where $a_l$ and $\beta_l$ are algebraic cycles. Let $\mathcal{B}'_{p}=\{v_{p}, \alpha_{p+1} \otimes \beta_{p+1},..., \alpha_{k-2} \otimes \beta_{k-2}\}$ be a basis of $F^{2p}$ extending the basis $\mathcal{B}_{p+1}$. Then $t_{2p} (v_p ) = q_p (\alpha_p \otimes \beta_p)$ for some nonzero $q_p \in {{\mathbb Q}}$. Define $$s_p : \text{Gr}^{2p}_F H^{n+2k-4}(Y) \to F^{2p}$$ by $s_p(\alpha_p \otimes \beta_p) = \frac{1}{q_p}v_p$. Then $s_p$ is a section of $t_p$ and hence we have a decomposition $$F^{2p}\cong F^{2p+2} \oplus \text{Gr}^{2p}_F H^{n+2k-4}(Y) =\bigoplus^{k-2}_{l=p+1}(\alpha_l \otimes \beta_l){{\mathbb Q}}\oplus s_p(\alpha_p \otimes \beta_p){{\mathbb Q}}$$ with a basis $\mathcal{B}_p=\{\alpha_p \otimes \beta_p, \alpha_{p+1} \otimes \beta_{p+1}, ..., \alpha_{k-2}\otimes \beta_{k-2} \}$ of $F^{2p}$ by identifying $\alpha_p \otimes \beta_p$ with its image under $s_p$.
Now for splitting for $2p < k-2$ in the exact sequence (\[seq:splitting\]), the argument is same as the case when $2p>k-2$, since $\text{Gr}^{2p}_F H^{n+2k-4}(Y) = H^{2p}(\Delta) \otimes H^{n+2k-4-2p}(\tilde{Q}_t) = {{\mathbb Q}}$. Hence we can choose a natural splitting of each exact sequence (\[seq:splitting\]) and get a decomposition $$\label{eq:decomp}
H^{n+2k-4}(Y) \cong H^{k-2}(\Delta,R^{n+k-2}{\pi}_*{{\mathbb Q}}) \oplus \bigoplus^{k-2}_{p=0, p\neq \frac{k-2}{2}} (\alpha_p \otimes \beta_p){{\mathbb Q}}$$
Let $L$ be the subspace of $H^{n+2k-4}(Y)$ generated by $\{\alpha_p \otimes \beta_p ~|~ p =0,...,\widehat{\frac{k-2}{2}},..., k-2\}$. Then, $$H^{k-2}(\Delta, R^{n+k-2}\pi_* {{\mathbb Q}}) \cong L^{\perp}$$ where $\perp$ is the orthogonal complement with respect to the cup product on $H^{n+2k-4}(Y)$.
First we show that $$(\alpha_i \otimes \beta_i) \cup (\alpha_j \otimes \beta_j) =
\begin{cases} 0 & \text{if $i+j \neq k-2$} \\ \deg~ \Delta & \text{if $i+j=k-2$}\end{cases}$$ Since $$\begin{gathered}
H^{2i}(\Delta, R^{n+2k-4-2i}\pi_*{{\mathbb Q}}) \otimes H^{2j}(\Delta, R^{n+2k-4-2j}\pi_*{{\mathbb Q}})\\
\stackrel{\cup}{\longrightarrow} \quad H^{2(i+j)}(\Delta, R^{2n+4k-8-2(i+j)}\pi_*{{\mathbb Q}})\end{gathered}$$ $(\alpha_i \otimes \beta_i) \cup (\alpha_j \otimes \beta_j) =0$ for $i+j \neq k-2$. For $i+j=k-2$, we observe $\alpha_i, \beta_i$ closely. Note that for any $i \neq \frac{k-2}{2}$, $\alpha_i = h^i_{\Delta}$ where $h^i_{\Delta} \in H^{2i}(\Delta)$ is a class corresponding to an iterated hyperplane section of $\Delta$ of codimension $i$. For $\beta_i$, let $\Pi \in H^{n+k-2}(\tilde{Q}_t)$ be a class corresponding to ${{\mathbb P}}^{\frac{n+k-2}{2}} \subset \tilde{Q}_t$ and $H^r \in H^{2r}(\tilde{Q}_t)$ a class corresponding to an iterated hyperplane section of $\tilde{Q}_t$ of codimension $r$. Note that $\Pi$ can be chosen in either families $F^1_t$ or $F^2_t$ of ${{\mathbb P}}^{\frac{n+k-2}{2}}$ contained in $\tilde{Q_t}$ ([@reid]). Then, $$\beta_i = \begin{cases} H^{\frac{n+2k-4-2i}{2}} & \text{if $i > \frac{k-2}{2}$} \\
\Pi \cup H^{\frac{k-2}{2}-i} & \text{if $i<\frac{k-2}{2}$}
\end{cases}$$ Then for $i+j=k+2$, $$(\alpha_i \otimes \beta_i ) \cup (\alpha_j \otimes \beta_j)=\deg ~\Delta$$
Now we show $L \cap L^{\perp}=0$. Let $\eta=\sum^{k-2}_{i=0,i\neq \frac{k-2}{2}} c_i (\alpha_i \otimes \beta_i) \in L \cap L^{\perp}$. Then, we have $c_i \cdot \text{deg} \Delta = 0 $ for any $i \neq \frac{k-2}{2}$. Hence $\eta =0$. Hence $$H^{n+2k-4}(Y) = L \oplus H^{k-2}(\Delta, R^{n+k-2}\pi_* {{\mathbb Q}}) = L \oplus L^{\perp}$$ Hence we get $$H^{k-2}(\Delta, R^{n+k-2}\pi_* {{\mathbb Q}})= L^{\perp}$$
From the above lemmas, we have an injection $$\label{eq:h}
h: H^{k-2} (\Delta, R^{n+k} {p_1}_*{{\mathbb Q}}) \hookrightarrow H^{n+2k-2}(X_{\Delta})$$
We refer the diagram (\[diag:theta\]) for the following lemma.
\[lemma:section\] There is a section defined in geometric way $$s_U: Gr^{W}_{n+2k-2}H^{n+2k-1}_c(X_U) \to H^{n+2k-4}(Y)(-1)$$ of $\psi_2$.
First recall $H^{n+2k-4}(Y)(-1) \cong H^{n+2k-2}(X_{\Delta})$ and $\psi_2$ is the connecting homomorphism in the exact sequence of mixed Hodge structures (\[eq:mhs\]) $$\cdots \to H^{n+2k-2}(X) \to H^{n+2k-2}(X_{\Delta}) \stackrel{\psi_2}{\to} Gr^W_{n+2k-2} H^{n+2k-1}_c(X_U) \to 0$$ As in the proof of lemma \[lemma:inj\], we can show that the the Leray spectral sequence associated to ${p_1} : X_U \to U$ gives rise to an exact sequence $$0 \to E^{2k-2,n+1}_{\infty} \to H^{n+2k-1}_c(X_U) \to E^{k-1,n+k}_{\infty} \to 0$$ By taking the exact functor $\text{Gr}^W_{n+2k-2}$, we get an isomorphism $$l:Gr^W_{n+2k-2}H^{n+2k-1}(X_U) \stackrel{\cong}{\longrightarrow} Gr^W_{k-2}H^{k-1}_c(U)$$ since $E^{2k-2,n+1}_{\infty}=H^{2k-2}_c(U)(-\frac{n+1}{2}) \cong H^{2k-2}({{\mathbb P}}^{k-1})(-\frac{n+1}{2})$ is a pure Hodge structure of weight $n+2k-2,$ if $n$ is odd and $0$ if $n$ is even. Hence we have a following commutative diagram: $$\begin{aligned}
\label{diag:section}
\xymatrix{
H^{n+2k-2}(X_{\Delta}) \ar[r]^{\psi_2 \qquad } & Gr^W_{n+2k-2} H^{n+2k-1}_c(X_U) \ar[d]^{\cong}_l\\
H^{k-2}(\Delta, R^{n+k}{p_1}_* {{\mathbb Q}}) \ar@{^(->}[u]^{h}
\ar[r]^{\delta \quad } & Gr^W_{k-2}H^{k-1}_c(U) \left(-\frac{n+k}{2} \right)
}\end{aligned}$$ where $h$ is the injection of (\[eq:h\]) and $\delta: H^{k-2}(\Delta, R^{n+k}{p_1}_* {{\mathbb Q}})\to Gr^W_{k-2}H^{k-1}_c(U) \left(-\frac{n+k}{2} \right)$ is the connecting homomorphism of long exact sequence induced by a short exact sequence of sheaves on ${{\mathbb P}}^{k-1}$ $$0 \to j_! (R^{n+k}p_1 {{\mathbb Q}})|_U \to R^{n+k}{p_1}_*{{\mathbb Q}}\to {i_{\Delta}}_*(R^{n+k}{p_1}_* {{\mathbb Q}})|_{\Delta} \to 0$$ where $j: U \to {{\mathbb P}}^{k-1}$ and $i_{\Delta} : \Delta \to {{\mathbb P}}^{k-1}$ are inclusions. Hence in order to choose a section of $\psi_2$, it is enough to construct a section of $\delta$.
Note that $(R^{n+k}{p_1}_*{{\mathbb Q}})_t ={{\mathbb Q}}^2$ for $t\in \Delta$ and $R^{n+k}{p_1}_*{{\mathbb Q}}|_U ={{\mathbb Q}}_U$ is the constant sheaf. Let $G$ be the monodromy group of $R^{n+k}{p_1}_*{{\mathbb Q}}|_{\Delta}$. Then $R^{n+k}{p_1}_*{{\mathbb Q}}|^G_{\Delta} ={{\mathbb Q}}$ is a subsheaf of $R^{n+k}{p_1}_*{{\mathbb Q}}|_{\Delta}$ and let $R^{n+k}{p_1}_*{{\mathbb Q}}|^G_{\Delta} \hookrightarrow R^{n+k}{p_1}_*{{\mathbb Q}}|_{\Delta}$ be an injection. Let $$s: {i_{\Delta}}_* (R^{n+k}{p_1}_* {{\mathbb Q}}^G_{\Delta}) \to {i_{\Delta}}_* (R^{n+k}{p_1}_* {{\mathbb Q}}_{\Delta})$$ be the induced map. Then there is a sheaf $s^*(R^{n+k}{p_1}_*{{\mathbb Q}})$ on ${{\mathbb P}}^{k-1}$ which fits into the following commutative diagram: $$\xymatrix{
0 \ar[r]& j_! (R^{n+k}p_1 {{\mathbb Q}}|_U) \ar[r] \ar@{=}[d] & s^*(R^{n+k}{p_1}_*{{\mathbb Q}}) \ar[d] \ar[r] & {i_{\Delta}}_*(R^{n+k}{p_1}_* {{\mathbb Q}}|_{\Delta}^{G}) \ar[d]^{s} \to 0 \\
0 \ar[r] & j_! (R^{n+k}p_1 {{\mathbb Q}}|_U) \ar[r]& R^{n+k}{p_1}_*{{\mathbb Q}}\ar[r]& {i_{\Delta}}_*(R^{n+k}{p_1}_* {{\mathbb Q}}|_{\Delta}) \to 0 \\
}$$ From this, we get a commutative diagram $$\xymatrix@C=8pt{
\cdots H^{k-2}({{\mathbb P}}^{k-1},s^*(R^{n+k}{p_1}_*{{\mathbb Q}})) \ar[d] \ar[r] &H^{k-2}(\Delta, {{\mathbb Q}}) \ar[r]^{\delta_1\qquad } \ar[d]^{s} & Gr^W_{k-2}H^{k-1}_c(U,{{\mathbb Q}}) \ar@{=}[d] \to 0 \\
\cdots H^{k-2}({{\mathbb P}}^{k-1},R^{n+k}{p_1}_*{{\mathbb Q}})\ar[r] &H^{k-2}(\Delta, R^{n+k}{(p_1|_{\Delta})}_* {{\mathbb Q}}) \ar[r]^{\quad \delta} & Gr^W_{k-2}H^{k-1}_c(U) \to 0
}$$ i.e. $$\label{eq:delta}
\delta_1 = \delta \circ s$$ Now from the exact sequence $$\cdots \to H^{k-2}({{\mathbb P}}^{k-2}) \stackrel{i^*_{\Delta}}{\to} H^{k-2}(\Delta) \stackrel{\delta_1}{\to} Gr^{W}_{k-2}H^{k-1}_c(U) \to 0$$ we have an isomorphism $$Gr^{W}_{k-2}H^{k-1}_c(U) \stackrel{\cong}{\longrightarrow} H^{k-2}(\Delta)/{\text{im}}~ i^*_{\Delta}$$ hence we can choose a natural section $s_1 : Gr^W_{k-2}H^{k-1}_c(U) \to H^{k-2}(\Delta)$ of $\delta_1 : H^{k-2}(\Delta) \to Gr^W_{k-2}H^{k-1}_c(U)$, i.e. $\delta_1 \circ s_1 =\text{id}$. Then from (\[eq:delta\]) we get a section $s \circ s_1$ of $\delta$. By combining all these, we get a commutative diagram:
$$\xymatrix@R=13pt{
H^{n+2k-2}(X_{\Delta}) \ar[r]^{\psi_2 \qquad } & Gr^W_{n+2k-2} H^{n+2k-1}_c(X_U) \ar[d]^{\cong}_l\\
H^{k-2}(\Delta, R^{n+k}{p_1}_* {{\mathbb Q}}) \ar@{^(->}[u]^{h}
\ar[r]^{\delta \quad } & Gr^W_{k-2}H^{k-1}_c(U) \left(-\frac{n+k}{2} \right) \\
H^{k-2}(\Delta) \ar[r]^{\delta_1 \qquad } \ar[u]^{s}& \ar@/^1pc/[l]^{s_1} Gr^W_{k-2}H^{k-1}_c(U) \left(-\frac{n+k}{2} \right) \ar@{=}[u]
}$$ Set $$s_U = h \circ s \circ s_1 \circ l$$ then $$\psi_2 \circ s_U = \psi_2 \circ( h \circ s \circ s_1 \circ l) = l^{-1} \circ (\delta \circ s \circ s_1) \circ l = l^{-1}\circ l =\text{id}$$ Hence $s_U$ is a section of $\psi_2$.
Now we finish the proof of the Theorem \[thm:main\] in case when $n+k$ is even. Recall the diagram (\[diag:theta\]) $$\xymatrix{
& H^* (W) \ar@{>>}[d]^{\theta} \ar[dr]^{\psi_2 \circ \theta}&\\
0 \to H^{n+2k-2}_0 (X) \ar[r]^{\psi_1} & H^{n+2k-4}(Y)(-1) \ar[r]^{\psi_2 \qquad \quad } & \ar@/^ 1pc/[l]^{s_U} Gr^{W}_{n+2k-2}H^{n+2k-1}_c(X_U) \to \cdots
}$$ where $s_U$ is the section of $\psi_2$ constructed in lemma \[lemma:section\].
Let $$\Theta = \theta - s_U \circ \psi_2 \circ \theta : H^{*}(W) \to H^{n+2k-4}(Y)(-1)$$ First note that $ \psi_2 \circ \Theta =0 $: In fact, $$\begin{aligned}
\psi_2 \circ \Theta &=& \psi_2 \circ (\theta - s_U \circ \psi_2 \circ \theta) \\
&=& \psi_2 \circ \theta - \psi_2 \circ s_U \circ \psi_2 \circ \theta =\psi_2 \circ \theta - \psi_2 \circ \theta=0\end{aligned}$$ Hence ${\text{im}}~\Theta \subseteq \ker ~\psi_2$ and we may consider that $\Theta$ is mapped into $H^{n+2k-2}_0(X)$ since $\psi_1$ is an injection. $$\xymatrix{
& \ar[dl]_{\Theta=\theta - s_U \circ \psi_2 \circ \theta} H^* (W) \ar@{>>}[d]^{\theta} \ar[dr]^{\psi_2 \circ \theta}&\\
0 \to H^{n+2k-2}_0 (X) \ar[r]^{\psi_1} & H^{n+2k-4}(Y)(-1) \ar[r]^{\psi_2 \qquad \quad } & \ar@/^ 1pc/[l]^{s_U} Gr^{W}_{n+2k-2}H^{n+2k-1}_c(X_U) \to \cdots
}$$ Now we show this map is in fact a surjection. Let $\alpha \in H^{n+2k-2}_0(X)$. Since $\psi_1$ is an injection we may identify $\alpha$ with $\psi_1(\alpha)$. Then there is $\beta \in H^*(W)$ such that $\theta(\beta) = \psi_1(\alpha)$. Then $$\begin{aligned}
\Theta (\beta) &=& (\theta-s_U \circ \psi_2 \circ \theta)(\beta) \\
&=& \theta(\beta) -s_U \circ \psi_2 \circ \theta(\beta)=\psi_1(\alpha)-s_U \circ \psi_2 \circ \psi_1(\alpha) = \psi_1(\alpha) =\alpha\end{aligned}$$ Hence we have a surjection $$\xymatrix{
\Theta : H^*(W) \ar@{>>}[r] & H^{n+2k-2}_0(X)
}$$ as we claimed. This finishes the proof of theorem \[thm:main\].
[99]{} D. Arapura, [*The Leray spectral sequence is motivic*]{}, Invent. math. 160. (2005) 567 – 589
C. Ehresmann, [*Sur la topologie de Certains Espaces Homogenes*]{}, Annals of Math. 2nd Ser., Vol 35, No 2. (1934) 396 – 443
W. Fulton, [*Intersection Theory*]{}, 2nd ed., Springer-Verlag, Berlin, 1998.
A. Grothendieck, [*Hodge’s general conjecture is false for trivial reasons,*]{} Topology 8, (1969)
P. Griffiths, J. Harris, [*Principles of Algebraic geometry*]{}, John Wiley & Sons, 1978
J. Lewis, [*A survey of the Hodge conjecture*]{} CRM Monograph Series, 10. American Mathematical Society, (1999).
K. O’Grady, [*The Hodge structure of the Intersection of Three quadrics in an odd dimensional projective space*]{}, Math. Ann. 273, (1986) 277 – 285
M. Reid, [*The complete intersection of two or more quadrics*]{}, Ph.D. thesis (1972)
Spanier, [*algebriac Topology*]{}, McGraw-Hill, 1966
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For each positive integer $n$, the Fibonacci-sum graph $G_n$ on vertices $1,2,\ldots,n$ is defined by two vertices forming an edge if and only if they sum to a Fibonacci number. It is known that each $G_n$ is bipartite, and all Hamiltonian paths in each $G_n$ have been classified. In this paper, it is shown that each $G_n$ has at most one non-trivial automorphism, which is given explicitly. Other properties of $G_n$ are also found, including the degree sequence, the treewidth, the nature of the bipartition, and that $G_n$ is outerplanar.'
address:
- 'Department of Mathematics University of Manitoba, Winnipeg, Manitoba, Canada'
- 'Department of Computer Science, University of Manitoba, Winnipeg, Manitoba, Canada'
author:
- Andrii Arman
- 'David S. Gunderson'
- Pak Ching Li
title: 'Properties of the Fibonacci-sum graph'
---
Fibonacci numbers ,sum graph
Introduction
============
The Fibonacci sequence $\{F_n\}_{n\geq 0}$ is defined by $F_0=0, F_1 = 1$ and for $n\geq 2$, $F_n = F_{n-1}+F_{n-2}$. Each $F_i$ is called a [*Fibonacci number*]{}. As defined in [@FKMOP:14], for each $n\geq 1$, the [*Fibonacci-sum graph*]{} $G_n=(V,E)$ is the graph defined on vertex set $V=[n]=\{1,2,\ldots,n\}$ with edge set $$E=\{\{i,j\}:i,j\in V, i\neq j,\ i+j\mbox{ is a Fibonacci number}\}.$$ By this definition, each $G_n$ is a simple graph (no loops or multiple edges). For example, $G_6$ is depicted in Figure \[fi:G6\].
=\[circle, draw=black, minimum size=12pt,inner sep=0pt\]
(one) [$1$]{}; (two) \[right = of one\] [$2$]{}; (three) \[right = of two\] [$3$]{}; (four) \[right = of three\] [$4$]{}; (five) \[right = of four\] [$5$]{}; (six) \[right = of five\] [$6$]{}; (one) to (two); (two) to (three); (three) to \[bend right = 30\] node\[anchor=south\](five); (two) to \[bend right = 30\] node\[anchor=south\](six); (one) to \[bend left = 30\] node\[anchor=south\](four);
Inspired by a question of Barwell [@Barw:06] about $G_{34}$, K. Fox, Kinnersley, McDonald, Orlow, and Puleo [@FKMOP:14] classified those Fibonacci-sum graphs that have Hamiltonian paths, gave a description of such paths, and examined generalizations to other related graphs. Some of these results are reviewed in Section \[ss:hampaths\], and some basic properties of Fibonacci numbers are given in Section \[ss:fibnofacts\].
The main results in this paper are about additional properties of Fibonacci-sum graphs. Fibonacci-sum graphs are connected; this simple result was also observed by Costain [@Cost:08 p. 26], and is repeated here in Corollary \[co:Gnconnected\]. In Section \[se:degrees\], degrees of vertices in Fibonacci-sum graphs are made explicit. In Section \[se:bipartite\], it is shown that every $G_n$ is bipartite (Theorem \[th:bipartite\], first observed in [@Silv:77] and in a more general setting, in [@AEH:78]), and the nature of such bipartitions is examined in more detail. The structure of cycles and chords is examined in Section \[se:cycles\].
In Section \[se:treewidth\], it is shown that for all $n\geq 7$, the Fibonacci-sum graph $G_n$ has treewidth 2. In Section \[se:planarity\], each $G_n$ is found to be outerplanar (Theorem \[th:outerplanar\]). In Section \[se:automorphisms\], for each $n\geq 9$, all automorphisms of $G_n$ are given, and in each case, there are at most two. (Section \[se:automorphisms\] is long since the proof breaks down into 11 cases, each case requiring some detailed checking.)
Background {#se:background}
==========
Hamiltonian paths {#ss:hampaths}
-----------------
In 2006, Barwell [@Barw:06] asked for an ordering of $\{1,2,\ldots, 34\}$ so that consecutive pairs sum to a Fibonacci number. Such an ordering corresponds to a Hamiltonian path in $G_{34}$.
In 2014, K. Fox, Kinnersley, McDonald, Orlow, and Puleo [@FKMOP:14] answered Barwell’s question (by giving a recipe for showing how to find the Hamiltonian path) and generalized the result to whenever $34$ (which is $F_9$) is replaced by any Fibonacci number or one less than a Fibonacci number. In doing so, they characterized those $n$ for which $G_n$ has a Hamiltonian path, and in those cases, they identified all such paths.
\[[@FKMOP:14]\]\[th:FKMOPmain\] The Fibonacci-sum graph $G_n$ has a Hamiltonian path if and only if $n$ is either 9, 11, a Fibonacci number, or one less than a Fibonacci number. The Hamiltonian path is unique except when $i\equiv 1\pmod{3}$, $n\in \{F_i,F_i-1\}$, in which case there are only two Hamiltonian paths, paths which agree except on the last three vertices (and so the last four vertices form a cycle).
In the proof of Theorem \[th:FKMOPmain\], the authors also showed that each $G_n$ had at least one vertex of degree 1, thereby preventing any Hamiltonian cycles (this fact is also proved below in Lemma \[le:pendants\]).
From Figure \[fi:G6\], it is seen that indeed $G_6$ has no Hamiltonian path. Since 34 is a Fibonacci number, Theorem \[th:FKMOPmain\] guarantees that $G_{34}$ has a unique Hamiltonian path, partially answering the question of Barwell. The same authors [@FKMOP:14] also gave ways to find a Hamiltonian path, provided one exists:
\[th:FKMOPpath\] For $k\geq 5$, the subgraph of $G_{F_k}$ formed by edges whose sum is in $\{F_{k-1}, F_k, F_{k+1}\}$ is a Hamiltonian path.
All Hamiltonian paths guaranteed by Theorem \[th:FKMOPpath\] were found:
\[th:endpoints\] Let $k\geq 5$. If $k\equiv 0\pmod{3}$, then $G_{F_k}$ contains a unique Hamiltonian path with end vertices $F_k$ and $\frac{F_k}{2}$. If $k\equiv 2\pmod{3}$, the unique Hamiltonian path in $G_{F_k}$ has end vertices $F_k$ and $\frac{F_{k+1}}{2}$. If $k\equiv 1\pmod{3}$, there are two Hamiltonian paths, one with endpoints $F_k$ and $\frac{F_{k-1}}{2}$, the other with endpoints $F_k$ and $F_k-\frac{F_{k-4}}{2}$.
For example, to find the (unique) Hamiltonian path in $G_{34}$ (asked for by Barwell), Theorem \[th:endpoints\] says to use 34 and 17 as endpoints. Starting with 17, and simply applying Theorem \[th:FKMOPpath\] (using sums only in $\{21,34,55\}$) gives the path: 17, 4, 30, 25, 9, 12, 22, 33, 1, 20, 14, 7, 27, 28, 6, 15, 19, 2, 32, 23, 11, 10, 24, 31, 3, 18, 16, 5, 29, 26, 8, 13, 21, 34. (At each point in the algorithm, the next vertex is uniquely determined. )
As an example where two Hamiltonian paths exist (with $k=7$ in the statement of Theorem \[th:FKMOPpath\]) in $G_{13}$, using only sums in $\{8,13,21\}$, the two paths are: 13, 8, 5, 3, 19, 11, 2, 6, 7, 1, 12, 9, 4 and 13, 8, 5, 3, 19, 11, 2, 6, 7, 1, 4, 9, 12.
Fibonacci number facts {#ss:fibnofacts}
----------------------
Here are some well-known facts about Fibonacci numbers that are used below.
\[fa:evenfibs\] The Fibonacci number $F_k$ is even iff $k\equiv 0\pmod{3}$.
\[fa:sumevenfibs\] For each $\ell\geq 0$, $$\sum_{i=0}^\ell F_{2i}=F_{2\ell+1}-1.$$
For each $k\geq 3$, $$\frac{F_{k-3}+F_k}{2}=F_{k-1}.$$
\[fa:halffib\] For $k\geq 0$, $F_k<\frac{F_{k+2}}{2}<F_{k+1}$.
\[fa:twoFssumtoF\] The only pairs of Fibonacci numbers whose sum is another Fibonacci number are the consecutive pairs. That is, if for some $i,j,k$, $F_i+F_j=F_k$, then $j=i+1$ and $k=i+2$ or $j=i-1$ and $k=i+1$.
Degrees {#se:degrees}
=======
In a graph, vertices of degree 1 are called [*pendant vertices*]{}, or simply, pendants. When $n$ is a Fibonacci number, the Fibonacci-sum graph on $n$ vertices has a pendant.
\[le:Fkonenbr\] For each $k\geq 1$, in $G_{F_k}$, the vertex $F_k$ has only one neighbour, namely $F_{k-1}$.
The basic idea used to prove Lemma \[le:Fkonenbr\] can be extended; the following is essentially contained in [@FKMOP:14], but formulated slightly differently here.
\[le:lastvertex\] Let $n\geq 2$, and let $k$ be so that $F_k\leq n<F_{k+1}$. In $G_n$, vertex $n$ is adjacent to only $$\begin{aligned}
\begin{cases}
F_{k+1}-n &\mbox{ if } n\leq \frac{F_{k+2}}{2};\\
F_{k+1}-n \mbox{ and }F_{k+2} -n&\mbox{ if } n>\frac{F_{k+2}}{2}.
\end{cases}\end{aligned}$$
Let $x\in [1,n]$ be adjacent to $n$; in other words, let $i$ be so that $x+n=F_i$. Since $$F_k<1+F_k\leq x+n$$ and $$x+n \leq 2n-1<2F_{k+1}<F_{k+1}+F_{k+2}=F_{k+3},$$ it follows that $i\in\{k+1, k+2\}$.
When $n\leq \frac{F_{k+2}}{2}$, $x+n<2n\leq F_{k+2}$, so $i\neq k+2$. Thus $x=n-F_{k+1}$ is the only possible solution. When $n>\frac{F_{k+2}}{2}$, the possible solution $x=F_{k+2}-n$ is also realizable (with $x>F_k$).
Lemma \[le:lastvertex\] and induction imply the following known result (see, e.g., [@Cost:08]):
\[co:Gnconnected\] For each $n\geq 1$, $G_n$ is connected.
The next lemma is also implicit in [@FKMOP:14 proof of Thm. 4], but is stated and proved here separately for later reference.
\[le:pendants\] Let $n \geq 2$, and let $k$ be so that $F_k\leq n<F_{k+1}$. Then in $G_n$, the vertex $F_k$ has only one neighbour, namely $F_{k-1}$.
If $n=F_k$, then by Lemma \[le:Fkonenbr\], in $G_n$ the vertex $n$ is adjacent to only $F_{k-1}$. (This also follows from the first case of Lemma \[le:lastvertex\] since by Fact \[fa:halffib\], $n< \frac{F_{k+2}}{2}$).
So assume that $F_k < n < F_{k+1}$. In $G_{F_{k}}$, the vertex $F_k$ is adjacent to only $F_{k-1}$, and so in $G_n$, the vertex $F_{k}$ has only one neighbour smaller than $F_k$. To see that in $G_n$, vertex $F_k$ has no larger neighbours, consider some vertex $v$ satisfying $F_k<v\leq n$. Then $$F_{k+1}<F_k+F_k <\ F_k + v \ \leq F_k + n < F_{k} + F_{k+1} = F_{k+2}$$ shows that $F_k+v$ is not a Fibonacci number, and so $v$ is not adjacent to $F_k$. Hence in $G_n$, $F_k$ has only one neighbour (namely $F_{k-1}$).
Calculating the degree of any vertex in a Fibonacci-sum graph is relatively straightforward. To see the idea, consider the adjacency matrix for $G_{18}$ given in Figure \[fi:adjG18\],
---- ------- ------- ------- --- ------- --- --- ------- --- ---- ---- ---- ------- ---- ---- ---- ---- ----
$F_2$ $F_3$ $F_4$ $F_5$ $F_6$ $F_7$
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
1 0 1 1 1 1
2 1 1 1 1
3 1 1 1 1
4 1 0 1 1
5 1 1 1
6 1 1 1
7 1 1 1
8 1 1
9 1 1
10 1 1
11 1 1
12 1 1
13 1
14 1
15 1
16 1 1
17 1 0
18 1 1
---- ------- ------- ------- --- ------- --- --- ------- --- ---- ---- ---- ------- ---- ---- ---- ---- ----
\[th:degreeformula\] Let $n\geq 1$ and let $x\in [1,n]$. Let $k\geq 2$ satisfy $F_k\leq x <F_{k+1}$ and $\ell\geq k$ satisfy $F_\ell \leq x+n < F_{\ell+1}$. Then the degree of $x$ in $G_n$ is $$\deg_{G_n}(x)=\begin{cases}
\ell -k &\mbox{if $2x$ is not a Fibonacci number};\\
\ell -k -1&\mbox{if $2x$ is a Fibonacci number}.
\end{cases}\label{eq:degreecount1}$$
For each $s\in [n]$, since $F_k< x+s\leq x+n<F_{\ell+1}$, if $x+s$ is a Fibonacci number, then $k<\ell$ and $x+s \in \{F_{k+1}, \ldots, F_{\ell}\}$. Thus, $$\label{eq:degcount}
\deg_{G_n}(x)=|\{s\in [n]: s\neq x, x+s\in \{F_{k+1}, \ldots, F_{\ell}\} \}|,$$ and so $\deg_{G_n}(x)\leq \ell-k$. However, when $s=x$, if $2x$ is a Fibonacci number, since $G_n$ has no loops, the sum $x+x$ fails to generate an edge and so (\[eq:degcount\]) says $\deg_{G_n}(x)= \ell-k-1$. Otherwise, if $2x$ is not a Fibonacci number, then (\[eq:degcount\]) shows $\deg_{G_n}(x)=\ell-k$.
The expression (\[eq:degreecount1\]) can be made more specific:
\[th:degreeformula2\] Let $n\geq 1$ and let $x\in [1,n]$. Let $k\geq 2$ satisfy $F_k\leq x <F_{k+1}$ and $\ell\geq k$ satisfy $F_\ell \leq x+n < F_{\ell+1}$. Then $$\deg_{G_n}(x)=\begin{cases}
\ell -k -1&\mbox{if $x=1$ or $k\geq 4$ and $x=\frac{1}{2}F_{k+2}$};\\
\ell -k &\mbox{otherwise}.
\end{cases}\label{eq:degreecount2}$$
Assume that $2x$ is a Fibonacci number. Then $2x<2(F_{k+1})<F_{k+3}$, and so either $2x=F_{k+1}$ or $2x=F_{k+2}$.
Case 1: $2x=F_{k+1}$: Then $2F_k\leq 2x=F_{k+1}=F_k+F_{k-1}$ implies $F_k\leq F_{k-1}$, and so $k=2$, $x=s=1$ and $2x=2=F_3=F_{k+1}$. One value of $s$ in (\[eq:degcount\]) is lost, and so $\deg_{G_n}(1)=\ell-k-1$.
Case 2: Assume that $k\geq 3$, and let $2x=F_{k+2}$. Then $F_{k+2}=x+x\leq x+n<F{\ell+1}$, and so $k+2\leq \ell$. Thus one value for $s$ in (\[eq:degcount\]) is lost and again $\deg_{G_n}(\frac{1}{2}F_{k+2})=k-\ell-1$. In fact, when $k=3$, $x=2$, and in this case, $2x$ is not a Fibonacci number, so one may assume that $k\geq 4$ (and $x\geq 4$).
In all other cases, $2x$ is not a Fibonacci number, so by (\[eq:degcount\]), $\deg_{G_n}(x)=\ell-k$. Hence (\[eq:degreecount2\]) is verified.
Examples of degree sequences in $G_n$ are found in the diagrams used in the proof of Theorem \[th:Automorphism\]. The following consequence of Theorem \[th:degreeformula2\] can be verified (for small $n$) in the adjacency matrix for $G_{18}$ given in Figure \[fi:adjG18\].
For any $n \geq 2$, vertex $2$ has maximum degree in $G_n$. If $n+1$ is a Fibonacci number, then $\deg_{G_n}(1)=\deg_{G_n}(2)-1$; otherwise, $\deg_{G_n}(1)=\deg_{G_n}(2)$.
Using the degree sequence given in either Theorem \[th:degreeformula\] or \[th:degreeformula2\], the number of edges in $G_n$ can be computed; the proof of the following is left to the reader:
Let $n\geq 1$ and $k\geq 2$ be integers satisfying $F_k\leq n <F_{k+1}$. Then $$\label{eq:numberofedges}
|E({G_n})|=\begin{cases}
n+\frac{F_k+1}{2}-\frac{\left\lfloor\frac{4(k+1)}{3}\right\rfloor}{2}
& \mbox{if } n\leq \frac{F_{k+2}}{2};\\
2n+\frac{F_k+1}{2}-\frac{\left\lfloor\frac{4(k+1)}{3}\right\rfloor}{2}
-\left\lceil\frac{F_{k+2}-1}{2}\right\rceil
& \mbox{if }n>\frac{F_{k+2}}{2}.
\end{cases}$$
For $n=1,2,\ldots, 21$, the values of $|E(G_n)|$ are 0, 1, 2, 3, 4, 5, 7, 8, 9, 10, 12, 14, 15, 16, 17, 18, 19, 21, 23, 25, and 26.
Lemma \[le:pendants\] says that every $G_n$ has at least one pendant vertex (and so $G_n$ does not have a Hamiltonian cycle). In general, how many pendants can $G_n$ have? The following observation can be proved directly or by applying Theorems \[th:degreeformula\] and \[th:degreeformula2\].
\[th:manypendants\] Let $k\geq 3$ and $n$ satisfy $F_k\leq n <F_{k+1}$. If $n<\frac{F_{k+2}}{2}$, then $F_k,F_k+1, \ldots, n$ are the vertices of degree 1. If $n\geq \frac{F_{k+2}}{2}$, then $F_k, F_k+1, \ldots, F_{k+2}-n-1$ are vertices of degree 1. The only other possible degree 1 vertices are $$\begin{cases}
\frac{F_k}{2} & \mbox{ if } k\equiv 0\pmod{3} \mbox{ and } n<F_{k+1}-\frac{F_k}{2};\\
\frac{F_{k+1}}{2} & \mbox{ if } k\equiv 1\pmod{3};\\
\frac{F_{k+2}}{2} & \mbox{ if } k\equiv 2\pmod{3} \mbox{ and } n\geq \frac{F_{k+2}}{2}.
\end{cases}$$
In classifying automorphisms (see Theorem \[th:Automorphism\] for diagrams), Theorem \[th:manypendants\] is useful. For example, in $G_{51}$, using $k=9$ (and $F_9=34$) in Theorem \[th:manypendants\], $F_{k+2}-n-1=89-51-1=37$, so that vertices 34, 35, 36, 37 and 17 are pendants.
For any $p\geq 1$, there exists an $n$ so that $G_n$ has exactly $p$ pendants.
For $p\geq 1$, let $n=F_{6p+1}+p-1$. Then, the only degree 1 vertices in $G_n$ are $F_{6p+1},F_{6p+1}+1, \ldots,F_{6p+1}+p-1$.
For example, to find $n$ so that $G_n$ has exactly $p=3$ pendants, let $n=F_{19}+(3-1)=4181+2=4183$.
$G_n$ is bipartite {#se:bipartite}
==================
As mentioned in the introduction, it is known (see [@Silv:77], [@AEH:78], or the survey [@Cost:08]) that $G_n$ is bipartite, but for completeness, the short proof is given here.
\[th:bipartite\] For $n \geq 1$, $G_n$ is bipartite.
The proof is by induction on $n$.
[Base step:]{} For $n\leq 6$, $G_n$ is a tree (see Figure \[fi:G6\]) and so is bipartite.
[Induction step:]{} Let $m\geq 6$ and assume that $G_m$ is bipartite. It remains to show that $G_{m+1}$ is bipartite. Let $\ell$ be the positive integer such that $F_{\ell} \leq m+1 < F_{\ell+1}$.
Case 1: If $m+1 \leq \frac{F_{\ell+2}}{2}$, by Lemma \[le:lastvertex\], in $G_{m+1}$, the vertex $m+1$ is adjacent to (only) the vertex $F_{\ell+1}-(m+1)$. Since $G_m$ is bipartite, the addition of the single edge to create $G_{m+1}$ is still bipartite, concluding the inductive step in this case.
Case 2: If If $m+1 \geq \frac{F_{\ell+2}}{2}$, by Lemma \[le:lastvertex\], in $G_{m+1}$, the vertex $m+1$ is adjacent to (only) the two vertices $v_1 = F_{\ell+1} - (m+1)$ and $v_2 = F_{\ell+2} - (m+1)$. By Fact \[fa:halffib\], $m+1>F_\ell$, and so the vertex $v=m+1-F_\ell$ is adjacent to both $v_1$ and $v_2$ in $G_m$. By the induction hypothesis, $G_m$ is bipartite, so $v_1$ and $v_2$ are in the same bipartite class in $G_m$, in which case the vertex $m+1$ can be added to the opposite class, showing that $G_{m+1}$ is bipartite. This concludes Case 2 of the inductive step, and hence the proof.
As an example, Figure \[fi:G20bipartite\] shows the bipartition of $G_{20}$.
=\[circle, draw=black, minimum size=15pt,inner sep=1pt\]
(one) at (1,6)[$1$]{}; (three) at (3,6) [$3$]{}; (six) at (6,6) [$6$]{}; (eight) at (8,6) [$8$]{}; (nine) at (9,6) [$9$]{}; (eleven) at (11,6) [$11$]{}; (fourteen) at (14,6)[$14$]{}; (sixteen) at (15.8,6) [$16$]{}; (seventeen) at (17,6) [$17$]{}; (nineteen) at (19,6)[$19$]{}; (two) at (1,0) [$2$]{}; (four) at (4,0) [$4$]{}; (five) at (5,0) [$5$]{}; (seven) at (7,0) [$7$]{}; (ten) at (10,0) [$10$]{}; (twelve) at (12,0) [$12$]{}; (thirteen) at (13.1,0) [$13$]{}; (fifteen) at (15,0) [$15$]{}; (eighteen) at (18,0) [$18$]{}; (twenty) at (20,0) [$20$]{};
(one) to (two); (one) to (four); (one) to (seven); (one) to (twelve); (one) to (twenty); (two) to (three); (two) to (six); (two) to (eleven); (two) to (nineteen); (three) to (five); (three) to (ten); (three) to (eighteen); (four) to (nine); (four) to (seventeen); (five) to (eight); (five) to (sixteen); (six) to (seven); (six) to (fifteen); (seven) to (fourteen); (eight) to (thirteen); (nine) to (twelve); (ten) to (eleven); (fourteen) to (twenty); (fifteen) to (nineteen); (sixteen) to (eighteen);
Looking at $G_{20}$ (see Figure \[fi:G20bipartite\]), one might expect that the bipartition of $G_n$ is always nearly balanced. However, there are values of $n$ so that the “imbalance” is as large as one wants; before proving this, some lemmas are given.
\[le:2colourFSgraph\] Let $n\geq 1$ and (by Theorem \[th:bipartite\]) let $c:V(G_n)\rightarrow\{0,1\}$ be a 2-colouring defining the bipartition of $G_n$, and suppose that $c(1)=c(F_1)=1$. Then for each $k\in [1, \frac{n-1}{2}]$, $$\begin{aligned}
c(F_{2k})&=1, \mbox{ \ }c(F_{2k+1})=0, \label{eq:colorfibns}\\
c\left(\frac{F_{6k-3}}{2}\right)&=1, \mbox{ and } c\left(\frac{F_{6k}}{2}\right)=0.\label{eq:colorfibns3mod0}\end{aligned}$$
Consecutive Fibonacci numbers are adjacent in $G_n$, and so a simple inductive argument shows that (\[eq:colorfibns\]) holds. To see (\[eq:colorfibns3mod0\]), by Fact \[fa:evenfibs\], starting with $F_3$, every third Fibonacci number is even and $c(\frac{F_3}{2})=
c(1)=1$. Since for any $k\geq 1$, $$\frac{F_{6k}}{2}+\frac{F_{6k-3}}{2}=\frac{1}{2}(F_{6k-1}+F_{6k-2}+F_{6k-3})
=\frac{2F_{6k-1}}{2}=F_{6k-1},$$ pairs of Fibonacci numbers of the form $\{\frac{F_{6k}}{2}, \frac{F_{6k-3}}{2}\}$ are adjacent in $G_n$ and so alternate in colour accordingly.
\[le:recursioncolour\] Let $k\geq 4$ and let $N$ satisfy $F_k<N<F_{k+1}$. Let $c$ be the 2-colouring of $G_N$ as in Lemma \[le:2colourFSgraph\]. If $N\neq \frac{F_{k+2}}{2}$, then $c(N)=c(N-F_k)$. If $N=\frac{F_{k+2}}{2}$, then $$c(N)=c\left(\frac{F_{k+2}}{2}\right)=\begin{cases}
1 & \mbox{if }k\equiv 1\pmod{6}\\
0 & \mbox{if }k\equiv 4\pmod{6}.
\end{cases}$$
If $N=\frac{F_{k+2}}{2}$, then the value for $c(\frac{F_{k+2}}{2})$ is given by Lemma \[le:2colourFSgraph\]. So suppose that $N \neq \frac{F_{k+2}}{2}$. Then $F_{k+1}-N \neq N - F_{k}$ and vertex $N$ is adjacent to $F_{k+1}-N$. So, $c(N)=1-c(F_{k+1}-N)$. The vertex $F_{k+1}-N$ is adjacent to $N-F_{k}$ and therefore $c(N)=1-c(F_{k+1}-N)=1-(1-c(N-F_k))=c(N-F_k)$.
For the next two lemmas, for $N\geq 1$ define $$\label{eq:S(N)}
S(N)=\sum_{i=1}^N c(i)-\frac{N}{2},$$ and put $S(0)=0$.
\[le:SFibonacci\] For $k\geq 3$, and $S$ as defined in (\[eq:S(N)\]), $$S(F_{k})=\begin{cases}
\hspace*{.1in} 0 & \mbox{if }k\equiv 0,3\pmod{6}\\
\hspace*{.1in} \frac{1}{2} & \mbox{if }k\equiv 2,4\pmod{6}\\
-\frac{1}{2} & \mbox{if }k\equiv 1,5\pmod{6}.
\end{cases}$$
[Case 1]{}: $k\equiv 0,3\pmod{6}$.
For $i \neq \frac{F_{k}}{2}$, vertex $i$ is adjacent to $F_{k+1}-i$, so $c(i)+c(F_{k+1}-i)=1$. $$\begin{aligned}
S(F_{k})&=\sum_{i=1}^{F_k} c(i)-\frac{F_k}{2}=\sum_{i=1}^{\frac{F_k}{2}-1}
\left[c(i)+c(F_{k+1}-i)\right]+c\left(\frac{F_k}{2}\right)+c(F_k)-\frac{F_k}{2}\\
&=\frac{F_{k}}{2}-1+c\left(\frac{F_k}{2}\right)+c(F_k)-\frac{F_k}{2}
=c\left(\frac{F_k}{2}\right)+c(F_k)-1.\end{aligned}$$ By Lemmas \[le:2colourFSgraph\] and \[le:recursioncolour\], $c\left(\frac{F_{6l}}{2}\right)+c(F_{6l})-1=0+1-1=0$ and $c\left(\frac{F_{6l+3}}{2}\right)+c(F_{6l+3})-1=1+0-1=0$.
[Case 2]{}: $k\equiv 2,4\pmod{6}$. $$\begin{aligned}
S(F_{k})&=\sum_{i=1}^{F_k} c(i)-\frac{F_k}{2}=\sum_{i=1}^{\frac{F_k-1}{2}}
\left[c(i)+c(F_{k+1}-i)\right]+c(F_k)-\frac{F_k}{2}\\
&=\frac{F_{k}-1}{2}+c(F_k)-\frac{F_k}{2}=-\frac{1}{2}+c(F_k)
=-\frac{1}{2}+1=\frac{1}{2}.\end{aligned}$$
[Case 3]{}: $k\equiv 1,5\pmod{6}$. $$\begin{aligned}
S(F_{k})&=\sum_{i=1}^{F_k} c(i)-\frac{F_k}{2}
=\sum_{i=1}^{\frac{F_k-1}{2}}\left[c(i)+c(F_{k+1}-i)\right]+c(F_k)-\frac{F_k}{2}\\
&=\frac{F_{k}-1}{2}+c(F_k)-\frac{F_k}{2}=-\frac{1}{2}+c(F_k)
=-\frac{1}{2}+0=-\frac{1}{2}.\end{aligned}$$
To show that the bipartite imbalance is unbounded, a lemma is used:
\[le:SN\] Let $N$ and $k$ be integers satisfying $F_{k}<N<F_{k+1}$. With $S(N)$ as defined in (\[eq:S(N)\]), if $k \equiv 0,2,3,5\pmod{6}$, or if $k\equiv 1,4\pmod{6}$ and $N<\frac{F_{k+2}}{2}$, then $S(N)=S(F_k)+S(N-F_k)$.
$$\begin{aligned}
S(N) &=\sum_{i=1}^N c(i)-\frac{N}{2} &&\mbox{(by def'n)}\\
&=\sum_{i=1}^{F_k} c(i)-\frac{F_k}{2}+\sum_{i=F_k+1}^N c(i)- \frac{N-F_k}{2}\\
&=S(F_k)+\sum_{i=1}^{N-F_k} c(i)-\frac{N-F_k}{2}&&\mbox{(by Lemma \ref{le:recursioncolour})}\\
&=S(F_k)+S(N-F_k), &&\mbox{(by def'n)}\end{aligned}$$
as desired.
\[th:imbalance\] For each positive integer $z$, there exists $n$ so that if $G_n$ has bipartition $V(G_n)=A\cup B$, then $\left| |A| -\frac{n}{2}\right| = z$.
Let $z\in \mathbb{Z}^{+}$. Define $n=\sum_{i=1}^{2z} F_{6i+1}$. Then $F_{12z+1}\leq n<F_{12z+2}$ and so $n<\frac{F_{12z+3}}{2}=\frac{F_{(12z+1)+2}}{2}$. Using $k=12z+1\equiv 1\pmod{6}$, by Lemmas \[le:SFibonacci\] and \[le:SN\], $$S(n)=S\left(\sum_{i=1}^{2z} F_{6i+1}\right)
=\sum_{i=1}^{2z} S(F_{6i+1})=2z \cdot \frac{-1}{2}=-z.$$ Assuming that (as in Lemma \[le:2colourFSgraph\]) $A$ is the set of vertices with colour 1, the imbalance of $A$ is $\left| |A| -\frac{n}{2}\right|=|S(n)+\frac{n}{2}-\frac{n}{2}|=z$.
For any $n$, it is possible to determine the sizes of parts in bipartition of $G_n$ using techniques of Theorem \[th:imbalance\] and Zeckendorf’s [@Zeck:72] representation of $n$. An example covered by Theorem \[th:imbalance\] with $z=2$ is when $n=F_{20}=7164$, in which case the bipartition is 3580 vs 3584.
Cycles in $G_n$ {#se:cycles}
===============
The girth of a graph $G$, denoted $\mbox{girth}(G)$, is the length of a shortest cycle in $G$. As seen in Figure \[fi:G6\], if $n\leq 6$ then $G_n$ is acyclic.
For $n \geq 7$, $\mbox{girth}(G_n) = 4$.
For $n \geq 7$, a 4-cycle in $G_n$ is $(1,2,6,7)$. By Theorem \[th:bipartite\], $G_n$ is bipartite and so contains no triangles.
For large $n$, $G_n$ contains many 4-cycles. When $F_k\leq n <F_{k+1}$ and $n>\frac{F_{k+2}}{2}$, a 4-cycle is $(n-F_k, F_{k+2}-n, n, F_{k+1}-n)$ as shown Figure \[fi:last4cycle\], where edges are labelled with sums.
=\[circle, fill=black, minimum size=6pt,inner sep=0pt\] =\[circle, draw=black, minimum size=5pt,inner sep=0pt\] at (-5,0) (x1); at (-3.5,0) (x2); at (-2,0) (x3); at (0,0) (x4); at (1.5,0) (x5); at (4.5,0) (x6); at (6,0) (x7); (x1) to \[bend left = 30\] node\[anchor=south\][$F_{k-1}$]{}(x3); (x1) to \[bend right= 30\] node\[anchor=north\][$F_{k+1}$]{}(x5); (x3) to \[bend left = 40\] node\[anchor=south\][$F_{k+1}$]{}(x6); (x5) to \[bend right = 30\] node\[anchor=north\][$F_{k+2}$]{}(x6);
Let $k \geq 2$ and $n=F_{2k+3}-1$. Then $G_n$ contains a cycle of length $2k$.
Define $c_1=1$, $c_2=4$, and for each $i=3,\ldots,2k$, recursively define $c_i=F_{i+4}-c_{i-1}$. For each $i\geq 2$, $c_{i-1}+c_i$ is a Fibonacci number and so $\{c_{i-1},c_i\}$ is an edge in $G_n$. To show that $(c_1,c_2,...,c_{2k})$ is a cycle of length $2k$ in $G_n$, it remains to show that $c_{2k}=F_{2k+3}-1$, in which case $1+c_{2k}$ is also a Fibonacci number. When $k=2$, $(c_1,c_2,c_3,c_4)=(1,4,9, 12)$ is a cycle, so assume that $k\geq 3$. Calculating, $$\begin{aligned}
c_{2k}&=F_{2k+4}-c_{2k-1}\\
&=F_{2k+4}-(F_{2k+3}-c_{2k-2})\\
&=F_{2k+2}+c_{2k-2}\\
&=F_{2k+2}+F_{2k+2}-c_{2k-3}\\
&=F_{2k+2}+F_{2k+2}-(F_{2k+1}-c_{2k-4})\\
&=F_{2k+2}+F_{2k}+c_{2k-4}\\
&\ \ \vdots \\
&=\left(\sum_{i=3}^{k+1}F_{2i}\right)+c_2\\
& =\left(\sum_{i=0}^{k+1}F_{2i}\right)-(F_0+F_2+F_4)+4\\
&=\sum_{i=0}^{k+1}F_{2i}\\
&=F_{2k+3}-1,&&\mbox{(by Fact \ref{fa:sumevenfibs})}\end{aligned}$$ as desired.
\[th:nocross\] Let $n\geq 7$. If $C = (a_1,a_2,...,a_m)$ is a cycle in the Fibonacci-sum graph $G_n$, then there do not exist edges $\{a_i,a_k\}$ and $\{a_j,a_\ell\}$ in $C$ with $i < j < k < \ell$; in other words, there are no crossing chords inside $C$.
The proof is by induction on $n$.
When $n=7$, there is only one cycle, which has no chords and hence no crossing chords. Let $N\geq 7$, and assume that for all $n\leq N$, the statement holds. It remains to show that $G_{N+1}$ has no crossing chords.
Let $k$ be so that $F_k\leq N+1 <F_{k+1}$ and let $C = (a_1,a_2,...,a_m)$ be a cycle in $G_{N+1}$. If $N+1=F_k$, then $\deg(N+1)=1$, and so $N+1$ is not on $C$ and so by induction hypothesis, $G_{N+1}$ has no crossing chords.
So assume that $F_k<N+1<F_{k+1}$. Suppose that $N+1$ is on $C$, since otherwise the statement follows from induction hypothesis. By Lemma \[le:lastvertex\], the only neighbours of $N+1$ are $F_{k+1}-(N+1)$ and $F_{k+2}-(N+1)$. No chord in $C$ contains $N+1$ (since otherwise, the degree of $N+1$ would be at least 3). Similarly, $F_{k+2}-(N+1)$ has only two neighbours in $G$ (namely $N+1-F_k$ and $N+1$), and so is not the endpoint of a chord. Thus, for some $b\in\{1,\ldots,m\}$, let $a_{b}=N+1-F_k$, $a_{b+1}=F_{k+2}-(N-1)$, $a_{b+2}=N+1$, and $a_{b+3}=F_{k+1}-(N+1)$.
Then $C=(a_1, a_2, ..., a_{b-1}, N+1-F_k, F_{k+2}-(N+1), N+1, F_{k+1}-(N+1), a_{b+4},\ldots, a_m)$. Consider $C^{\prime}=(a_1, a_2, ..., a_{b-1}, N+1-F_k, F_{k+1}-(N+1), a_{b+4}, \ldots, a_m)$. Since $C^\prime$ is contained in $G_N$, by the induction hypothesis, $C^{\prime}$ has no crossing chords. Therefore, $C$ has no crossing chords.
Although any cycle in $G_n$ does not contain crossing chords, $G_n$ satisfies a property similar to being chordal (where every cycle induces a chord). First note that any copy of a 4-cycle in $G_n$ has no chords (since a chord would then create a triangle, which is impossible because $G_n$ is bipartite).
\[th:longcyclechord\] For any $n \geq 1$, every cycle of $G_n$ of length at least 6 contains a chord that has exactly two vertices of the cycle on one side of the chord (so forming a 4-cycle).
Fix $n$ and let $C$ be a cycle of length at least 6 in $G_n$. Let $m$ be the vertex in $C$ with the largest value, and let $k$ be so that $F_k \leq m \leq F_k+1$. Then vertices $F_{k+1}-m$ and $F_{k+2}-m$ are the only neighbours of $m$ in $G_m$, and so are both adjacent to $m$ in $C$. However, $F_{k+2}-m$ is adjacent to $m-F_k$ in $G_m$, and therefore $\{F_{k+1}-m,m-F_k\}$ is a desired chord in $C$.
Treewidth {#se:treewidth}
=========
The “treewidth” of a graph $G$ is a measure of how close $G$ is to being a tree. The notions of tree decompositions and treewidth (under the name “dimension”) were introduced by Halin [@Hali:76] in 1976, and independently, later by Seymour and Thomas (see [@SeTh:86]) in in their study of graph minors.
A [*tree decomposition*]{} of a graph $G$ is a pair $(T,\mathcal{ V})$, where $T$ is a tree and $\mathcal{ V}=\{V_t\subseteq V(G):t\in V(T)\}$ is a family of subsets of $V(G)$ indexed by vertices in $T$ so that three conditions hold: (i) $\cup_{i\in V(T)}=V$; (ii) for every edge $\{x,y\}\in E(G)$, there exists $V_t\in \mathcal{ V}$ containing both $x$ and $y$; (iii) if $t_it_jt_k$ is a path in $T$, then $V_i\cap V_k\subseteq V_j$. The width of a tree decomposition is $(T,\mathcal{V})$ is $\max\{|V_t|-1:t\in V(T)\}$ and the [*treewidth*]{} of $G$, denoted tw$(T)$ is the minimum width of any tree-decomposition of $G$.
There are different characterizations of treewidth, some of which are described in [@Dies:10]. For example, Seymour and Thomas [@SeTh:93] characterized treewidth in terms of brambles. Another characterization of treewidth uses chordal graphs and clique numbers. For a graph $H$, let $\omega(H)$ denote the clique number of $H$, the order of the largest complete subgraph of $H$.
\[le:twchordal\] For any graph $G$, $$\mbox{tw}(G)=\min\{\omega(H)-1:G\subseteq H, H \mbox{ chordal}\}.$$
So the treewidth of a tree is 1, the treewidth of a cycle is 2, and the treewidth of the complete graph $K_n$ is $n-1$.
For $n=1,\ldots, 6$, the treewidth of $G_n$ is equal to 1 since these graphs are trees. However, for $n\geq 7$ each $G_n$ contains a cycle and so has treewidth at least 2.
\[th:treewidth\] For each $n\geq 7$, $G_n$ has treewidth 2.
The proof given here is by induction and uses Lemma \[le:twchordal\]. A second proof is given in Section \[se:planarity\] as a Corollary to Theorem \[th:outerplanar\] on outerplanarity.
For the base step, by inspection, $G_7$ has treewidth 2.
For the inductive step, let $m \geq 8$ and assume that $G_{m-1}$ has treewidth 2. Then by definition, there is a chordal graph $H_{m-1}$ on $m$ vertices that contains $G_{m-1}$ and has clique number 3. It remains to show that for $G_{m}$ there is a chordal graph $H_{m}$ on $m$ vertices that contains $G_{m}$ and has clique number 3. Assume that $F_{k}\leq m < F_{k+1}$. By Lemma \[le:lastvertex\], the degree of $m$ in $G_m$ is either 1 or 2.
If $\deg_{G_m}(m)=1$, then $H_{m}$ can be taken to be $H_{m-1}$ plus the edge $\{m, F_{k+1}-m\}$. So assume that $\deg_{G_m}(m)=2$. Then $m$ is adjacent to $F_{k+2}-m$ and $F_{k+1}-m$, both of which are adjacent to $m-F_{k}$. Construct $H_{m}$ by adding vertex $m$ to $H_{m-1}$ and the edges $\{m, F_{k+2}-m\}$, $\{m, F_{k+1}-m\}$ and $\{m, m-F_k\}$. Then $H_m$ contains $G_{m}$ as a subgraph and $\deg_{H_m}(m)=3$.
Any clique with 3 or more vertices in $H_m$ that contains vertex $m$ does not contain both of the vertices $F_{k+2}-m$ and $F_{k+1}-m$ since they are not adjacent. Therefore any clique in $H_m$ that contains $m$ has at most 3 vertices. By the induction hypothesis, any clique in $H_m$ that does not contain $m$ has at most 3 vertices. Therefore, any clique in $H_m$ has at most 3 vertices. This completes the proof of induction step and hence the theorem.
Planarity {#se:planarity}
=========
The following theorem is known (see, e.g., [@Cost:08 p.27] for a suggested proof outline); however, a proof is included here for completeness.
\[th:planar\] For $n \geq 1$, $G_n$ is planar.
The proof is by induction on $n$. For $n\leq 6$, $G_n$ is planar (see Figure \[fi:G6\]). Fix $m\geq 6$ and assume that $G_m$ is planar (with a fixed plane drawing). It remains to show that $G_{m+1}$ s planar. Let $\ell$ be the positive integer such that $F_{\ell} \leq m+1 < F_{\ell+1}$. Consider two possible scenarios.
Case 1: Suppose $m+1 \leq \frac{1}{2}F_{\ell+2}$. By Lemma \[le:lastvertex\], the vertex $v=m+1$ has degree 1 in $G_{m+1}$ , in which case it follows that $G_{m+1}$ is planar, since adding any pendant edge to a planar drawing of a graph can be done without crossing any edges.
Case 2: Suppose $m+1> F_{\ell+2}$. Then in $G_{m+1}$, the vertex $v=m+1$ is adjacent to vertices $v_1 = F_{\ell+1} - (m+1)$ and $v_2 = F_{\ell+2} - (m+1)$. The vertex $v_3=m+1-F_\ell$ is adjacent to both $v_1$ and $v_2$ in $G_{m+1}$ and in $G_m$. The vertex $v_2$ has degree 1 in $G_m$. Let $F$ be one of the faces with $v_1v_3$ as a bordering edge, and, if necessary, move $v_2$ into $F$. Then $v$ can also be placed in $F$ (so that the edges $vv_1$ and $vv_2$ do not cross any other edges; [*e.g.,*]{} put $v$ inside the triangular region formed by $v_1v_3v_2$), giving a planar drawing $G_{m+1}$. This concludes the inductive step, and hence the proof.
Below in Theorem \[th:outerplanar\], it is shown that each $G_n$ is outerplanar; the following lemma is used for the proof of this fact.
\[le:nobooksubgraph\] Let $H$ be the graph on eight vertices consisting of three 4-cycles sharing a single edge. For $n \geq 1$, $G_n$ does not contain a subgraph isomorphic to $H$.
The proof is by induction on $n$.
For $n \leq 7$, the statement of the theorem is trivially true because $H$ has eight vertices.
Fix $m \geq 8$ and assume that $G_{m-1}$ does not contain $H$. To complete the induction step, it remains to show that $G_{m}$ does not contain $H$. Let the integer $k$ be so that $F_{k}\leq m < F_{k+1}$.
In hopes of a contradiction, suppose that $G_{m}$ contains a copy of $H$. By the induction hypothesis, $m$ is a vertex of $H$. Since vertex $m$ belongs to $H$, the degree of $m$ is 2. Therefore, $m>\frac{F_{k+2}}{2}$. The only neighbours of $m$ in $G_{m}$ are $F_{k+1}-m$ and $F_{k+2}-m$. Vertex $F_{k+2}-m>F_{k+2}-F_{k+1}=F_k$ is also a vertex of degree 2. The other neighbour of $F_{k+2}-m$ is $m-F_{k}$. Therefore, vertices $F_{k+2}-m$, $F_{k+1}-m$, $m$ and $m-F_{k}$ belong to $H$.
Vertex $m-F_{k}$ has degree at least 4 in $G_{m}$, since $m-F_k$ belongs to $H$. That may only happen if $m-F_{k}<F_{k-2}$, because $m-F_{k}$ added to other vertex has to make $F_{k-2}$, $F_{k-1}$, $F_{k}$ and $F_{k+1}$. Therefore, $\frac{F_{k+2}}{2}<m<F_{k}+F_{k-2}$; in particular $F_{k}+F_{k-3}<m$. Vertex $m-F_{k}$ has degree precisely 4 in $G_m$ and is adjacent to $F_{k+2}-m$, $F_{k+1}-m$, $F_{k}+F_{k-2}-m$ and $2F_{k}-m$. Then, $$F_{k}>2F_{k}-(F_{k}+F{k-3})>2F_{k}-m>2F_k-(F_k+F_{k-2})=F_{k-1}.$$
Vertex $2F_{k}-m$ has degree 2 and is adjacent to $m-F_{k}$ and $m-F_{k-2}$. Vertex $m-F_{k-2}$ is not adjacent to vertex $F_{k+1}-m$, and therefore vertex $m$ is not in a copy of $H$ in $G_{m}$. By the inductive hypothesis, there is no copy of $H$ in $G_{m-1}$, and so $G_{m}$ does not contain $H$ as a subgraph. This ends the proof of induction step and a proof of the lemma.
\[th:outerplanar\] For $n \geq 1$, $G_n$ is outerplanar.
The proof is by induction on $n$. For values $n\leq 7$, the statement of the theorem is true by inspection.
Let $m \geq 8$ and assume that $G_{m-1}$ is outerplanar (with a fixed outerplanar drawing). It remains to show that $G_{m}$ is outerplanar. Let $k$ be so that $F_{k}\leq m < F_{k+1}$.
If $\deg_{G_m}(x)=1$, then drawing the edge $\{m, F_{k+1}-m\}$ in the outer face of the outerplanar drawing of $G_{m-1}$ produces an outerplanar drawing of $G_{m}$.
If $\deg_{G_m}(x)=2$, then $m$ is adjacent to vertices $F_{k+2}-m$ and $F_{k+1}-m$. Vertex $F_{k+2}-m$ has degree 2 in $G_{m}$ and is also adjacent to vertex $m-F_{k}$. Vertex $m-F_{k}$ is adjacent to $F_{k+1}-m$.
[Claim]{}: The edge $\{F_{k+1}-m, m-F_{k}\}$ borders the outer face of the outerplanar drawing of $G_{m-1}$.\
[Proof of claim]{}: Suppose that this is not the case and derive a contradiction. So assume $\{F_{k+1}-m, m-F_{k}\}$ borders two inner faces of the drawing of $G_{m-1}$. These two inner faces of the outerplanar drawing of $G_{m-1}$ do not have two edges in common. Then by Theorem \[th:longcyclechord\] there are two cycles $\{F_{k+1}-m, m-F_{k}, x , y\}$ and $\{F_{k+1}-m, m-F_{k}, z, t\}$ with all four vertices $x,y,z,t$ being different. Then vertices $\{m, F_{k+2}-m, F_{k+1}-m, m-F_{k}, x, y, z, t\}$ induce a copy of $H$ in $G_m$ which contradicts Lemma \[le:nobooksubgraph\], proving the claim.
To complete the proof, by the claim, put vertices $m$ and $F_{k+2}-m$ in the outer face next to $\{F_{k+1}-m, m-F_{k}\}$ in the drawing of $G_{m-1}$ to form an outerplanar drawing of $G_m$. This completes the proof of the induction step and hence the theorem.
As promised in Section \[se:treewidth\], Theorem \[th:treewidth\] is now also a corollary to Theorem \[th:outerplanar\] because of the well-known result (see, e.g., [@Bodl:88], [@Bodl:96], or [@CoWa:83]) that outerplanar graphs have treewidth at most 2.
Automorphisms of $G_n$ {#se:automorphisms}
======================
Any automorphism of $G_n$ takes endpoints of a Hamiltonian path to endpoints of another Hamiltonian path. By Theorem \[th:FKMOPmain\], Fibonacci-sum graphs have either no, one, or two Hamiltonian paths so one might expect that in most cases, there are few automorphisms.
For $n\leq 2278$, calculations show that the $G_n$s with a trivial automorphism group are those where $n$ is in the intervals $[7,10]$, $[17,21]$, $[30,50]$, $[72,92]$, $[127,215]$, $[305,393]$, $[538,914]$, $[1292,1668]$. The rest have order two.
All but the first interval for each type have lengths that are Fibonacci numbers. For those with just one automorphism, lengths are 4, 5, 21, 21, 89, 89, 377, 377, which are, respectively, 4, $F_5$, $F_8$, $F_8$, $F_{11}$, $F_{11}$, ...; similarly, for order 2 the intervals have lengths, 6, 8, 21, 34, 89, 123, all but the first being Fibonacci numbers.
Denote the automorphism group of a graph $G$ by $\mbox{Aut}(G)$. In the following theorem, “id” denotes an identity map. In the proof of the next theorem, standard functional notation is used. For any set $V$ and any function $f$ with domain $X$, write $f(X)=\{f(x):x\in X\}$. Say that $f:X\rightarrow X$ “fixes” $x\in X$ iff $f(x)=x$. For any subset $W\subseteq X$ say that “$f$ fixes $W$” iff $f(W)=W$. Note that a function can fix a set without necessarily fixing every vertex in that set.
\[th:Automorphism\] Let $n\geq 9$ be an integer that is divisible by 3.
If $N \in \left[F_n, \frac{3}{2} F_n \right)\cup \left[\frac{1}{2}F_{n+3}, F_{n+2}+\frac{1}{2}F_{n-3} \right)\cup \left[F_{n+3}-\frac{1}{2}F_n, F_{n+3} \right)$, then Aut$(G_N)=\{id\}$.
If $N \in \left[\frac{3}{2}F_{n}, \frac{1}{2}F_{n+3}\right)$, then Aut$(G_N)=\{id, \phi\},$ where $\phi$ interchanges $\frac{1}{2}F_n$ with $\frac{3}{2}F_n$ and fixes all other vertices.
If $N \in \left[F_{n+2}+\frac{1}{2}F_{n-3}, F_{n+3}-\frac{1}{2}F_{n}\right)$, then Aut$(G_N)=\{id, \phi\}$, where $\phi$ interchanges $\frac{1}{2}F_n$ with $\frac{3}{2}F_n$, interchanges $F_{n+2}+\frac{1}{2}F_{n-3}$ with $\frac{1}{2}F_{n+3}$, and fixes all other vertices.
The proof is by induction on $n$. For each integer $n\geq 9$ (divisible by 3), let $A(n)$ be the statement of the theorem.
[Base step]{}: For each $i\in [34,144)=[F_9, F_{12})$ the statement $A(i)$ is verified by direct calculation.
[Induction step]{}; Let $k\geq 12$ satisfy $3 \mid k$, and assume that for every $m\in [F_9, F_k)$, the statement $A(m)$ is true. It remains to show that for all $N\in \left[F_k ,F_{k+3}\right)$, the statement $A(N)$ is true. The proof consists of five cases, each case corresponding to one of the five half-open intervals given in the statement of the theorem. All but the last case have subcases. For each case, a diagram indicating degrees in certain intervals is given (not to scale) spanning two lines because of space limitations.
Throughout the proof, let $\phi:[N]\rightarrow [N]$ denote an automorphism of $G_N$.
[Case]{} 1: $N \in \left[F_k, \frac{1}{2}F_{k+2}\right)$. There are three subcases to consider.
[Case]{} 1a: $N \in \left[F_k, F_k+\frac{1}{2}F_{k-3}\right)$:
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{}; (-2.2,-2) circle (1pt) node\[anchor=north\] [1]{}; (-2.2,-2) circle (0pt) node\[anchor=south\] [$N$]{};
in [1,...,5]{} (-2.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,16]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,5]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
(5.4,1) circle (0pt) node\[anchor=south\] [[$F_{k+1}-N$]{}]{}; (5.4,1) – (5.4,0);
in [1,...,2]{} (5.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [2]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
Let $S=\{1,...,F_k-1\}$. Then $S$ is a set of vertices of degree at least 2 together with the vertex $\frac{1}{2}F_k$. The vertex $\frac{1}{2}F_k$ is the only degree 1 vertex adjacent to a degree 4 vertex (namely, $\frac{1}{2}F_{k-3}$), and so $\phi$ fixes $\frac{1}{2}F_k$. Since any automorphism sends vertices of degree at least 2 vertices of degree at least 2, $\phi$ sends all remaining vertices of $S$ into $S$ and so $\phi(S)=S$. The map $\psi=\phi |_S $ is an automorphism of $G_{F_k-1}$, and so by the induction hypothesis, $\psi$ is the identity on $S$.
Now consider any $y \geq F_k$. The degree of $y$ is 1 and $y$ is adjacent to $F_{k+1}-y$. Since $F_{k+1}-y\leq F_{k-1}$, $\phi(F_{k+1}-y)=F_{k+1}-y$. Vertex $y$ is adjacent to $F_{k+1}-y$, therefore $\phi(y)$ is adjacent to $F_{k+1}-y$. Since $y$ is the only degree 1 vertex adjacent to $F_{k+1}-y$, it follows that $\phi(y)=y$.
Therefore, the only automorphism of $G_N$ is the identity.
[Case]{} 1b: $N \in \left[ F_k+\frac{1}{2}F_{k-3}, \frac{1}{2}F_{k+2}\right)$: Note that $\frac{1}{2}F_{k+2}$ is not an integer.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(-1.8,-2) circle (1pt) node\[anchor=north\] [1]{}; (-1.8,-2) circle (0pt) node\[anchor=south\] [$N$]{};
in [1,...,7]{} (-1.8-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,16]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [2]{};
(4.2,1) circle (0pt) node\[anchor=south\] [[$F_{k+1}-N$]{}]{}; (4.2,1) – (4.2,0);
in [1,...,2]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
(4.2,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,5]{} (4.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
The proof is nearly the same as for Case 1a, with the exception that now the degree of $\frac{1}{2}F_k$ is 2. Let $S=\{1,...,F_k-1\}$. Since any automorphism maps vertices of degree at least 2 to vertices of degree at least 2, $\phi(S)=S$, and the remainder of the proof is identical.
[Case]{} 1c: $N \in \left(\frac{1}{2}F_{k+2},\frac{3}{2}F_k\right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(-1,-2) circle (1pt) node\[anchor=north\] [2]{}; (-1,-2) circle (0pt) node\[anchor=south\] [$N$]{};
in [1,...,6]{} (-1-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(-2.2,-1.5) circle (0pt) node\[anchor=south\] [$F_{k+2}-N$]{}; (-2.2,-1.5) – (-2.2,-2);
in [1,...,5]{} (-2.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,16]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,5]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
(-1,0.5) circle (0pt) node\[anchor=south\] [$F_{k+1}-N$]{}; (-1,0.5) – (-1,0);
in [1,...,11]{} (-1-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
Let $S=\{1,...,F_{k-1}-1\}$. Then $S$ is a set of vertices of degree at least 3 together with the vertex $\frac{1}{2}F_k$, which is of degree 2.
The vertex $\frac{1}{2}F_k$ is the only vertex of degree 2 adjacent to a degree 4 vertex (namely, $\frac{1}{2}F_{k-3}$) and a degree 1 vertex (namely, $F_k+\frac{1}{2}F_{k-3}$). All other degree 2 vertices have either a degree 2 neighbour or a degree 3 neighbour. Therefore $\phi(\frac{1}{2}F_k)=\frac{1}{2}F_k$
The vertices of degree at least 3 are mapped to the vertices of degree at least 3. So $\phi(S)=S$ and $\psi=\phi \mid_S$ is an automorphism of $G_{F_{k-1}-1}$. By the induction hypothesis, $\psi$ is the identity on $S$.
The vertex $F_{k-1}$ is the only degree 2 vertex adjacent to both a degree 3 and a degree 1 vertex, and so $\phi(F_{k-1})=F_{k-1}$. So, for any $x \leq F_{k-1}$, $\phi(x)=x$.
Now consider any $y \geq F_k$ with degree 1, and its only neighbour $F_{k+1}-y$. Since $F_{k+1}-y\leq F_{k-1}$, by the preceding paragraphs, $\phi(F_{k+1}-y)=F_{k+1}-y$. Since $y$ and $F_{k+1}-y$ are adjacent, $\phi(y)$ and $F_{k+1}-y$ are adjacent, but the only degree 1 vertex adjacent to $F_{k+1}-y$ is $y$. Hence, $\phi(y)=y$. So $\phi$ fixes those $y\geq F_k$ with degree 1.
Define two types of degree 2 vertices: type one are those in the interval $\left(F_{k-1}, F_{k}\right)$ and type two are in the interval $\left[F_{k+2}-N, N\right]$.
First, type one vertices are not adjacent to type two vertices.
Let $x \in \left(F_{k-1},F_{k}\right)$ be a type one vertex. Then $x$ is adjacent to $F_{k+1}-x$—another type one vertex. Vertex $x$ is also adjacent to $F_{k}-x \leq F_{k-1}$. Vertex $F_{k+1}-x$ is adjacent to $x-F_{k-1} \leq F_{k-1}$. The sum of indexes of the vertices $F_{k+1}-x$ and $x-F_{k-1}$ is $F_k$.
Let $y \in \left[F_{k+2}-N,N\right]$ be a type two vertex. Then $y$ is adjacent to $F_{k+2}-y$—another type two vertex. Vertex $y$ is also adjacent to $F_{k+1}-y \leq F_{k-1}$. Vertex $F_{k+2}-x$ is adjacent to $y-F_{k} \leq F_{k-1}$ since $(F_{k+2}-y)+(y-F_{k})=F_{k+1}$.
Type one vertices are mapped to type one vertices, since every vertex adjacent to two type one vertices has neighbours that sum to $F_k$. Type two vertices are mapped to type two vertices, since every vertex adjacent to two type two vertices has neighbours that sum to $F_{k+1}$. This means that $\phi$ preserves the type of degree 2 vertices.
For $x \in \left(F_{k-1},F_{k}\right)$, the vertex $F_k-x$ is adjacent to $x$. Therefore, $\phi(x)$ is adjacent to $\phi(F_k-x)=F_k-x$. However, $x$ is the only type one vertex adjacent to $F_k-x$, and so $\phi(x)=x$.
For $y \in \left[F_{k+2}-N,N \right]$, the vertex $F_{k+1}-y$ is adjacent to $y$. Therefore, $\phi(y)$ is adjacent to $\phi(F_{k+1}-y)=F_{k+1}-y$. However, $y$ is the only type two vertex that is adjacent to $F_{k+1}-y$, and so $\phi(y)=y$.
Therefore, the only automorphism of $G_N$ is the identity.
[Case 2:]{} $N\in \left[\frac{3}{2}F_{k}, \frac{1}{2}F_{k+3} \right)$. There are two subcases to consider.
[Case]{} 2a: $N \in \left[\frac{3}{2}F_{k},F_{k+1}\right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(-0.4,-2) circle (1pt) node\[anchor=north\] [2]{}; (-0.4,-1.2) circle (0pt) node\[anchor=south\] [$N$]{}; (-0.4,-1.2)–(-0.4,-2);
(-0.8,-2) circle (1pt) node\[anchor=north\] [2]{}; (-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{}; in [1,...,11]{} (-0.4-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(-2.6,-1.2) circle (0pt) node\[anchor=south\] [$F_{k+2}-N$]{}; (-2.6,-1.2) – (-2.6,-2);
(-2.2,-2) circle (1pt) node\[anchor=north\] [2]{}; (-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,3]{} (-2.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,16]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,4]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{}; (-5.6,0.8) circle (0pt) node\[anchor=south\] [$F_{k+1}-N$]{}; (-5.6,0.8) – (-5.6,0);
in [1,...,3]{} (-5.6-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
Let $S=\{1,...,F_{k-2}-1\}$ and let $\phi\in \mbox{Aut}(G_N)$. Then $S$ is a set of vertices of degree at least 4 and since any automorphism maps vertices of degree at least 4 to vertices of degree at least 4, $\phi(S)=S$. The map $\psi=\phi \mid_S $ is then an automorphism of $G_{F_{k-2}-1}$. By the induction hypothesis, $\psi$ is either the identity or interchanges $\frac{3}{2}F_{k-3}$ with $\frac{1}{2}F_{k-3}$ and fixes all other vertices of $S$. However, in $G_N$, the degrees of $\frac{3}{2}F_{k-3}$ and $\frac{1}{2}F_{k-3}$ are 4 and 5 respectively, so no automorphism carries one to the other. Thus, $\psi$ is the identity map on $S$.
Vertex $F_{k-2}$ is the only degree 3 vertex that does not have a neighbour of degree 3, and so, $\phi(F_{k-2})=F_{k-2}$. Each remaining degree 3 vertex has a unique neighbour less than $F_{k-3}$. Therefore $\phi$ fixes degree 3 vertices. So $\phi$ fixes every vertex in $[1,F_{k-1})\backslash\{\frac{1}{2}F_k\}$. (The last point $\frac{1}{2}F_k$ in this interval is discussed below.)
Vertex $F_k$ is the only degree 1 vertex that has a neighbour of degree 2, and so $\phi(F_k)=F_k$.
Consider a degree 1 vertex $y > F_k$. Then $y$ is adjacent to $F_{k+1}-y$. Since $F_{k+1}-y\leq F_{k-1}$, by above, $\phi(F_{k+1}-y)=F_{k+1}-y$. Since $y$ is adjacent to $F_{k+1}-y$, $\phi(y)$ is adjacent to $\phi(F_{k+1}-y)=F_{k+1}-y$. Since $y$ is the only degree 1 vertex adjacent to $F_{k+1}-y$, it follows that $\phi(y)=y$.
Identify three types of degree 2 vertices: say those of type one are in the interval $\left(F_{k-1}, F_k\right)$, those of type two are in the interval $\left[F_{k+2}-N, N\right]\backslash \{F_k+F_{k-3}, \frac{3}{2}F_k\}$, and call the remaining three $\frac{1}{2}F_k$, $F_k+\frac{1}{2}F_{k-3}$, $\frac{3}{2}F_k$, type three.
Note that type one vertices are not adjacent to type two vertices.
Since $F_k+\frac{1}{2}F_{k-3}$ is the only degree 2 vertex whose neighbours are only degree 2 vertices, it follows that $\phi(F_k+\frac{1}{2}F_{k-3})=F_k+\frac{1}{2}F_{k-3}$. Vertices $\frac{1}{2}F_k$ and $\frac{3}{2}F_k$ are the only neighbours of $F_k+\frac{1}{2}F_{k-3}$, so $\phi(\{\frac{1}{2}F_k,\frac{3}{2}F_k\})
=\{\frac{1}{2}F_k,\frac{3}{2}F_k\}$. Thus, type three vertices are mapped to type three vertices.
Let $x \in \left(F_{k-1},F_{k}\right)$ be a type one vertex. Then $x$ is adjacent to $F_{k+1}-x$ (another type one vertex), and $F_{k+1}-x \leq F_{k-1}$.
Now let $y$ be a type two vertex; then $y$ is adjacent to $F_{k+2}-x$ (another type two vertex) and $y$ is also adjacent to $F_{k+1}-y \leq F_{k-1}$. Vertex $F_{k+2}-x$ is adjacent to $y-F_{k} \leq F_{k-1}$. The sum of $F_{k+2}-y$ and $y-F_{k}$ is $F_{k+1}$.
Type one vertices are mapped to type one vertices, since every two adjacent type one vertices have neighbours that sum to $F_k$.
Type two vertices are mapped to type two vertices, since two adjacent type two vertices have neighbours that sum to $F_{k+1}$. Hence, $\phi$ preserves the type of degree 2 vertices.
For $x \in \left(F_{k-1},F_{k}\right)$, vertex $F_k-x$ is a neighbour of $x$. Therefore, $\phi(x)$ is adjacent to $\phi(F_k-x)=F_k-x$. However, $x$ is the only type one vertex adjacent to $F_k-x$, and so $\phi(x)=x$.
Let $y$ be a type two vertex. Since $F_{k+1}-y$ is a neighbour of $y$, $\phi(y)$ is adjacent to $\phi(F_{k+1}-y)=F_{k+1}-y$. However, $y$ is the only type two vertex adjacent to $F_{k+1}-y$, and so $\phi(y)=y$.
So, $\phi$ fixes each vertex of $V(G)\backslash\{\frac{1}{2}F_k, \frac{3}{2}F_k\}$.
Since $\frac{1}{2}F_k$ and $\frac{3}{2}F_k$ have the same neighbours (namely, $F_{k}+\frac{1}{2}F_{k-3}$ and $\frac{1}{2}F_{k-3}$) either $\phi$ is the identity or $\phi$ interchanges $\frac{1}{2}F_k$ and $\frac{3}{2}F_k$ and fixes all other vertices.
[Case]{} 2b: $N \in \left[F_{k+1}, \frac{1}{2}F_{k+3} \right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{}; (1.2,-2) circle (1pt) node\[anchor=north\] [1]{}; (1.2,-1.2) circle (0pt) node\[anchor=south\] [$N$]{}; (1.2,-1.2)–(1.2,-2); in [1,...,6]{} (1.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
(-0.8,-2) circle (1pt) node\[anchor=north\] [2]{}; (-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{};
in [1,...,16]{} (0-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(-5.6,-1.2) circle (0pt) node\[anchor=south\] [$F_{k+2}-N$]{}; (-5.6,-1.2) – (-5.6,-2);
(-2.2,-2) circle (1pt) node\[anchor=north\] [2]{}; (-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,12]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,4]{} (-5.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
Let $S=\{1,...,F_{k+1}-1\}$. Then $S$ is a set of vertices of degree at least 2. Since any automorphism maps vertices of degree at least 2 to vertices of degree at least 2, $\phi(S)=S$ and $\psi=\phi |_S $ is an automorphism of $G_{F_{k+1}-1}$. By Case 2a, $\psi$ is either the identity isomorphism or $\psi$ interchanges $\frac{3}{2}F_{k}$ and $\frac{1}{2}F_{k}$ and fixes all other vertices of $S$. The only degree 1 vertex that has a neighbour of degree 2 is $F_{k+1}$, and so $\phi(F_{k+1})=F_{k+1}$.
Consider $y > F_{k+1}$; then the degree of $y$ is 1, with its only neighbour $F_{k+2}-y$. Then $\frac{1}{2}F_k<F_{k+2}-y\leq F_{k}$ and therefore $\phi(F_{k+2}-y)=F_{k+2}-y$. Since $y$ is adjacent to $F_{k+2}-y$, then $\phi(y)$ is adjacent to $F_{k+2}-y$. However, $y$ is the only degree 1 vertex adjacent to $F_{k+2}-y$, and so $\phi(y)=y$. So, $\phi$ fixes each vertex in $V(G_N)\backslash \{\frac{1}{2}F_k, \frac{3}{2}F_k\}$.
Since $\frac{1}{2}F_k$ and $\frac{3}{2}F_k$ have the same set of neighbours (both are adjacent to $F_{k}+\frac{1}{2}F_{k-3}$ and $\frac{1}{2}F_{k-3}$), $\phi$ either fixes both $\frac{1}{2}F_k$ and $\frac{3}{2}F_k$ (in which case $\phi$ is the identity) or $\phi$ interchanges $\frac{1}{2}F_k$ and $\frac{3}{2}F_k$ (and fixes all other vertices).
[Case 3:]{} $N \in \left[F_{k+1}+\frac{1}{2}F_k,F_{k+2}+\frac{1}{2}F_{k-3}\right)$. Note that $F_{k+1}+\frac{1}{2}F_k=\frac{1}{2}F_{k+3}$. There are two subcases to consider:
[Case]{} 3a: $N \in \left[F_{k+1}+\frac{1}{2}F_k, F_{k+2} \right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(2.4,-2) circle (1pt) node\[anchor=north\] [2]{}; (2.4,-1.2) circle (0pt) node\[anchor=south\] [$N$]{}; (2.4,-1.2)–(2.4,-2);
in [1,...,3]{} (2.4-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{}; (1.6,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,4]{} (1.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{}; (0.8,-2) circle (1pt) node\[anchor=north\] [2]{}; (0.8,-1.2) circle (0pt) node\[anchor=south\] [$F_{k+3}-N$]{};
(0.8,-1.2)–(0.8,-2);
in [1,...,4]{} (0.8-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{}; (-0.8,-2) circle (1pt) node\[anchor=north\] [2]{}; (-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{};
in [1,...,16]{} (0-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{}; (-2.2,-2) circle (1pt) node\[anchor=north\] [2]{}; (-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,16]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{}; (6.4,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{}; (4.8,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,10]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
(1.2, 0.8) circle (0pt) node\[anchor=south\] [$F_{k+2}-N$]{}; (1.2, 0.8) – (1.2,0);
in [1,...,6]{} (1.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
Let $S=\{1,...,F_{k}-1\}$. Since $S$ is a set of vertices of degree at least 3 and any automorphism maps vertices of degree at least 3 to vertices of degree at least 3, it follows that $\phi(S)=S$ and the map $\psi=\phi |_S$ is an automorphism of $G_{F_{k}-1}$. By the induction hypothesis, $\psi$ is the identity isomorphism. Therefore, for all $x<F_k$, $\phi(x)=x$.
The only degree 2 vertex that does not have a neighbour of degree 2 is $F_k$, and so $\phi(F_k)=F_k$. As in Case 1c, $\phi$ fixes all vertices of degree 2.
Consider a degree 1 vertex $y \geq F_{k+1}$. The only neighbour of $y$ is $F_{k+2}-y\leq F_{k}$ and so $\phi(F_{k+2}-y)=F_{k+2}-y$. Thus $\phi(y)$ is adjacent to $F_{k+2}-y$. However, $y$ is the only degree 1 vertex adjacent to $F_{k+2}-y$ and so $\phi(y)=y$.
Therefore, the only automorphism of $G_N$ is the identity.
[Case]{} 3b: $N \in \left[F_{k+2}, F_{k+2}+\frac{1}{2}F_{k-3}\right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(4.0,-1.9) circle (0pt) node\[anchor=south\] ; (4.0,-2) circle (1pt) ;
(3.8,-2) circle (1pt) node\[anchor=north\] [1]{}; (3.8,-1.2) circle (0pt) node\[anchor=south\] [$N$]{}; (3.8,-1.2)–(3.8,-2);
in [1,...,3]{} (3.8-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,7]{} (3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(1.6,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,8]{} (1.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(-0.6,-1.2) circle (0pt) node\[anchor=south\] [$F_{k+3}-N$]{}; (-0.6,-1.2)–(-0.6,-2); in [1,...,3]{} (0-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(-0.8,-2) circle (1pt) node\[anchor=north\] [2]{}; (-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{};
in [1,...,12]{} (-0.8-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(-2.2,-2) circle (1pt) node\[anchor=north\] [2]{}; (-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,16]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
Let $S=\{1,...,F_{k+2}-1\}$. Then $S$ is a set of vertices of degree at least 2 together with $\frac{1}{2}F_{k+3}$. Any automorphism maps vertices of degree at least 2 to vertices of degree at least 2. The remaining vertex $\frac{1}{2}F_{k+3}$ is the only degree 1 vertex with a neighbour (namely $\frac{1}{2}F_k$) of degree 3 that has neighbours of degree 1, 2 and 6. All other degree 1 vertices have at least one second neighbour of degree 7 or higher, and so $\phi(\frac{1}{2}F_{k+3})=\frac{1}{2}F_{k+3}$. So $\phi(S)=S$ and $\psi=\phi |_S $ is an automorphism of $G_{F_{k+2}-1}$. By Case 3a, $\psi$ is the identity isomorphism. Therefore, $\phi$ fixes each $x<F_{k+2}$.
Let $y \geq F_{k+2}$. The degree of $y$ is 1, with neighbour $F_{k+3}-y$. Then $F_{k+3}-y\leq F_{k+1}$ and so $\phi(F_{k+3}-y)=F_{k+3}-y$. Since $y$ is adjacent to $F_{k+3}-y$, it follows that $\phi(y)$ is adjacent to $F_{k+3}-y$. However, $y$ is the only degree 1 vertex adjacent to $F_{k+3}-y$, and so $\phi(y)=y$.
Therefore, the only automorphism of $G_N$ is the identity.
[Case]{} 4: $N \in \left[F_{k+2}+\frac{1}{2}F_{k-3}, F_{k+2}+\frac{1}{2}F_{k+1}\right)$. There are three subcases to consider.
[Case]{} 4a: $N \in \left[F_{k+2}+\frac{1}{2}F_{k-3}, F_{k+2}+F_{k-1}\right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(4.0,-1.9) circle (0pt) node\[anchor=south\] ; (4.0,-2) circle (1pt) ;
(4.4,-2) circle (1pt) node\[anchor=north\] [1]{}; (4.4,-1.2) circle (0pt) node\[anchor=south\] [$N$]{}; (4.4,-1.2)–(4.4,-2);
in [1,...,6]{} (4.4-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,7]{} (3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(1.6,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,8]{} (1.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(-2.4,-1.2) circle (0pt) node\[anchor=south\] [$F_{k+3}-N$]{}; (-2.4,-1.2)–(-2.4,-2);
in [1,...,12]{} (0-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{};
in [1,...,4]{} (-2.4-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,16]{} (-3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
Let $S=\{1,...,F_{k-2}-1\}$. Then $S$ is the set of vertices of degree at least 5 and the vertices of degree at least 5 are mapped to the vertices of degree at least 5. $\phi(S)=S$ and $\psi=\phi \mid_S $ is an automorphism of $G_{F_{k-2}-1}$. By the induction hypothesis, $\psi$ is either the identity or the map that interchanges $\frac{1}{2}F_{k-3}$ with $\frac{3}{2}F_{k-3}$ (and fixes all other vertices of $S$). However, $\frac{1}{2}F_{k-3}$ and $\frac{3}{2}F_{k-3}$ have degrees 6 and 5 respectively; so $\psi(\frac{1}{2}F_{k-3})=\frac{1}{2}F_{k-3}$ and $\psi(\frac{3}{2}F_{k-3})=\frac{3}{2}F_{k-3}$. Therefore, for each $x<F_{k-2}$, $\phi$ fixes $x$.
Let $y\in[F_{k-2},F_{k-1})$ be a vertex of degree 4. Then $y$ is adjacent to $F_{k-1}-y$. Vertex $F_{k-1}-y \leq F_{k-3}$. Then $\phi(F_{k-1}-y)=F_{k-1}-y$ is adjacent to $\phi(y)$. Vertex $y$ is the only vertex of degree 4 adjacent to $F_{k-1}-y$. Therefore, $\phi$ fixes each $y\in[F_{k-2},F_{k-1})\backslash \{\frac{1}{2}F_k\}$.
The interval $[F_{k-1},F_k)$ consists of vertices of degree 3 that do not have a neighbour of degree 1. If a vertex of degree 3 has a neighbour of degree 1, then this vertex does not belong to interval $[F_{k-1},F_k)$. Therefore, $\phi([F_{k-1},F_k))=[F_{k-1},F_k)$. Let $y\in [F_{k-1},F_k)$; then y is adjacent to $F_k-y\leq F_{k-2}$. Therefore, $\phi(y)$ is adjacent to $\phi(F_k-y)=F_k-y$. Vertex $y$ is the only degree 3 vertex without a degree 1 neighbour adjacent to $F_k-y$. Hence, $\phi$ fixes every vertex in $ [F_{k-1},F_k)$.
Vertex $F_{k+1}$ is the only degree 2 vertex that has a neighbour of degree 1, and so $\phi(F_{k+1})=F_{k+1}$. Vertex $F_{k}$ is the only degree 2 vertex that is adjacent to $F_{k+1}$, and so $\phi(F_k)=F_k$.
Proceed as in Case 1c, defining two types of degree 2 vertices, those in the interval $\left(F_{k}, F_{k+1}\right)$ and those in the interval $\left(F_{k+1}, F_{k+2}\right)$. As in Case 1c, conclude that for any vertex $y$ of degree 2, $\phi(y)=y$.
The set $S_1=\{1,...,F_{k+1}-1\}$ is the set of degree 3 or higher vertices together with vertices of degree 2 in interval $[F_k, F_{k+1})$. Degree 3 or higher vertices are mapped by any automorphism to degree 3 or higher vertices and degree 2 vertices in $[F_k, F_{k+1})$ are fixed by $\phi$. The restriction $\psi=\phi |_{S_1} $ is an automorphism of $G_{F_{k+1}-1}$. By Case 2a, $\psi$ is either the identity or the isomorphism that interchanges only $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$. Therefore, $\phi(x)=x$ for all vertices $x<F_{k+1}$, except possibly $x=\frac{1}{2}F_k$ or $x=\frac{3}{2}F_k$.
[Case]{} 4b: $N \in \left[F_{k+2}+F_{k-1}, F_{k+2}+\frac{1}{2}F_{k+1}\right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k+3}$]{}; (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(4.0,-1.9) circle (0pt) node\[anchor=south\] ; (4.0,-2) circle (1pt) ;
(4.4,-2) circle (1pt) node\[anchor=north\] [1]{}; (4.4,-1.2) circle (0pt) node\[anchor=south\] [$N$]{}; (4.4,-1.2)–(4.4,-2);
in [1,...,6]{} (4.4-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,7]{} (3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(1.6,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,8]{} (1.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,16]{} (0-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{};
(-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,2]{} (-3.2-0.2\* , -2) circle (0.5pt) node\[anchor=north\] [4]{};
(-3.6,-1.2) circle (0pt) node\[anchor=south\] [$F_{k+3}-N$]{}; (-3.6,-1.2)–(-3.6,-2);
in [1,...,14]{} (-3.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
Let $S_1=\{1,...,F_{k+1}-1\}$. Then $S_1$ is a set of vertices of degree at least 3 and the vertices of degree at least 3 are mapped to the vertices of degree at least 3. $\phi(S_1)=S$ and $\psi=\phi |_{S_1} $ is an automorphism of $G_{F_{k+1}-1}$. By Case 2a, $\psi$ is either the identity isomorphism or the one that interchanges $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$. Therefore, $\phi(x)=x$ for all vertices $x<F_{k+1}$, except of possibly $x=\frac{1}{2}F_k$ or $x=\frac{3}{2}F_k$.
Vertex $F_{k+1}$ is the only degree 2 vertex that does not have a neighbour of degree 2. Therefore, $\phi(F_{k+1})=F_{k+1}$. Let $y \geq F_{k+2}$ be a degree 1 vertex with $y \neq \frac{1}{2}F_{k+3}$, $y \neq F_{k+2}+\frac{1}{2}F_{k-3}$ . Since $F_{k+3}-y\leq F_{k+1}$ it follows that $\phi(F_{k+3}-y)=F_{k+3}-y$. Also since $y$ is adjacent to $F_{k+3}-y$, it follows that $\phi(y)$ is adjacent to $F_{k+3}-y$. Then $y$ is the only degree 1 vertex that is adjacent to $F_{k+3}-y$, and so $\phi(y)=y$.
Consider $F_{k+1}<y<F_{k+2}$, such that degree of $y$ is 2 (so $y \neq \frac{1}{2}F_{k+3}$). Then $y$ is adjacent to $F_{k+2}-y$. $F_{k+2}-y < F_{k}$ and $F_{k+2}-y \neq \frac{1}{2}F_{k}$ and therefore $\phi(F_{k+2}-y)=F_{k+2}-y$. $y$ is adjacent to $F_{k+2}-y$, therefore $\phi(y)$ is adjacent to $F_{k+2}-y$. $y$ is the only degree 2 vertex that is adjacent to $F_{k+2}-y$. Hence, $\phi(y)=y$.
Consider $x \geq F_{k+2}$, such that degree of $x$ is 1 and $x \neq F_{k+2}+\frac{1}{2}F_{k-3}$ . Then $x$ is adjacent to $F_{k+3}-y$. $F_{k+3}-x < F_{k+1}$ and $F_{k+3}-x \neq \frac{1}{2}F_{k}$, $F_{k+3}-x \neq \frac{3}{2}F_{k}$, therefore $\phi(F_{k+3}-x)=F_{k+3}-x$. Since $x$ is adjacent to $F_{k+3}-x$, it follows that $\phi(x)$ is adjacent to $F_{k+3}-x$. Since $x$ is the only degree 1 vertex adjacent to $F_{k+3}-x$, conclude that $\phi(x)=x$.
So, $\phi$ fixes all but perhaps four of the vertices of $G_N$, with the only possible exceptions being that $\phi$ might interchange $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$ and $\frac{1}{2}F_{k+3}$ with $F_{k+2}+\frac{1}{2}F_{k-3}$. However, $\frac{1}{2}F_k$ is adjacent to $\frac{1}{2}F_{k+3}$ and $\frac{3}{2}F_k$ is adjacent to $F_{k+2}+\frac{1}{2}F_{k-3}$, so $\phi$ moves $\frac{1}{2}F_k$ iff $\phi$ moves $\frac{1}{2}F_{k+3}$.
Therefore, $\mbox{Aut}(G_N)$ consists of two mappings: the identity and the automorphism that interchanges $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$, interchanges $\frac{1}{2}F_{k+3}$ with $F_{k+2}+\frac{1}{2}F_{k-3}$, and fixes all other vertices.
[Case]{} 4c: $N \in \left[F_{k+2}+\frac{1}{2}F_{k+1}, F_{k+3}-\frac{1}{2}F_{k}\right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.6,-2) circle (0pt) node\[anchor=south\] [$F_{k+3}$]{}; (6.4,-2) circle (1pt); (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(4.0,-1.9) circle (0pt) node\[anchor=south\] ; (4.0,-2) circle (1pt);
(5.6,-1.9) circle (0pt) node\[anchor=south\] ; (5.6,-2) circle (1pt) ;
in [1,...,5]{} (5.4-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(5.2,-1.2) circle (0pt) node\[anchor=south\] [$\scriptstyle N$]{}; (5.2,-1.2)–(5.2,-2);
(4.4,-1.2) circle (0pt) node\[anchor=south\] [$\scriptstyle F_{k+4}-N$]{}; (4.4,-1.2)–(4.4,-2);
in [1,...,6]{} (4.4-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,7]{} (3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(1.6,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,8]{} (1.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,16]{} (0-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{};
(-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,12]{} (-3.2-0.2\* , -2) circle (0.5pt) node\[anchor=north\] [4]{};
(-5.6,-1.2) circle (0pt) node\[anchor=south\] [$\scriptstyle F_{k+3}-N$]{}; (-5.6,-1.2)–(-5.6,-2);
in [1,...,4]{} (-5.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [3]{};
in [1,...,8]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
Note that $\frac{1}{2}F_{k+1}$ is not an integer.
Let $S=\{1,...,F_{k+1}-1\}$. Then $S$ is a set of vertices of degree at least 3 and the vertices of degree at least 3 are mapped to the vertices of degree at least 3. So $\phi(S)=S$ and $\psi=\phi \mid_S $ is an automorphism of $G_{F_{k+1}-1}$. By the Case 2a, $\psi$ is either an identity isomorphism or one that fixes all but two vertices, interchanging $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$. Therefore, $\phi$ fixes each vertex in $[1,F_{k+1}-1]\backslash \{\frac{1}{2}F_k,\frac{3}{2}F_k\}$.
Vertex $F_{k+1}$ is the only degree 2 vertex that does not have a neighbour of degree 2. Therefore, $\phi(F_{k+1})=F_{k+1}$. Let $y \geq F_{k+2}$ be a degree 1 vertex, such that $y \neq F_{k+2}+\frac{1}{2}F_{k-3}$. Vertex $y$ is adjacent to $F_{k+3}-y$, and since $F_{k+3}-y\leq F_{k+1}$, it follows that $\phi(F_{k+3}-y)=F_{k+3}-y$. Also since $y$ is adjacent to $F_{k+3}-y$, it follows that $\phi(y)$ is adjacent to $F_{k+3}-y$. Vertex $y$ is the only degree 1 vertex that is adjacent to $F_{k+3}-y$, and therefore $\phi(y)=y$.
Proceeding as in Case 1c, with two types of degree 2 vertices (type one are degree 2 vertices in the interval $\left(F_{k+1}, F_{k+2}\right)$ and type two in$\left[F_{k+4}-N, N\right]$), it follows that $\phi$ fixes all vertices of degree 2 in $G_N$.
So, $\phi$ fixes almost all vertices of $G_N$, with the possible exceptions that $\phi$ might interchange $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$ and interchange $\frac{1}{2}F_{k+3}$ with $F_{k+2}+\frac{1}{2}F_{k-3}$. Since $\frac{1}{2}F_k$ is adjacent to $\frac{1}{2}F_{k+3}$ and $\frac{3}{2}F_k$ is adjacent to $F_{k+2}+\frac{1}{2}F_{k-3}$, $\phi$ moves $\frac{1}{2}F_k$ iff $\phi$ moves $\frac{1}{2}F_{k+3}$.
Therefore, $\mbox{Aut}(G_N)$ consists of the identity and the automorphism that interchanges $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$, interchanges $\frac{1}{2}F_{k+3}$ with $F_{k+2}+\frac{1}{2}F_{k-3}$, and fixes all other vertices.
[Case]{} 5: $N \in \left[F_{k+3}-\frac{1}{2}F_k, F_{k+3}\right)$.
(-6.4,0) – (6.4,0); (-6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-5}$]{}; (-3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-4}$]{}; (0,0) circle (1pt) node\[anchor=south\] [$F_{k-3}$]{}; (3.2,0) circle (1pt) node\[anchor=south\] [$F_{k-2}$]{}; (6.4,0) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k-3}}{2}$]{}; (4.8,0) circle (1pt) node\[anchor=south\] [$\frac{F_{k}}{2}$]{};
(-6.4,-2) – (6.4,-2); (-6.4,-2) circle (1pt) node\[anchor=south\] [$F_{k-1}$]{}; (-3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k}$]{}; (0,-2) circle (1pt) node\[anchor=south\] [$F_{k+1}$]{}; (3.2,-2) circle (1pt) node\[anchor=south\] [$F_{k+2}$]{}; (6.6,-2) circle (0pt) node\[anchor=south\] [$F_{k+3}$]{}; (6.4,-2) circle (1pt); (1.6,-2) circle (1pt) node\[anchor=south\] [$\frac{F_{k+3}}{2}$]{};
(4.0,-1.9) circle (0pt) node\[anchor=south\] ; (4.0,-2) circle (1pt);
(5.6,-1.9) circle (0pt) node\[anchor=south\] ; (5.6,-2) circle (1pt) ;
in [1,...,13]{} (6.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(6.0,-1.2) circle (0pt) node\[anchor=south\] [$\scriptstyle N$]{}; (6.0,-1.2)–(6.0,-2);
(3.6,-1.2) circle (0pt) node\[anchor=south\] [$\scriptstyle F_{k+4}-N$]{}; (3.6,-1.2)–(3.6,-2);
in [1,...,2]{} (3.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,7]{} (3.2-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
(1.6,-2) circle (0.5pt) node\[anchor=north\] [1]{};
in [1,...,8]{} (1.6-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [2]{};
in [1,...,16]{} (0-0.2\*,-2) circle (0.5pt) node\[anchor=north\] [3]{};
(-0.8,-2) circle (1pt) node\[anchor=south\] [$\frac{3F_k}{2}$]{};
(-2.2,-2) circle (1pt) node\[anchor=south\] [$\frac{2F_k+F_{k-3}}{2}$]{};
in [1,...,16]{} (-3.2-0.2\* , -2) circle (0.5pt) node\[anchor=north\] [4]{};
(6.4,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,7]{} (6.4-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
(4.8,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,4]{} (4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
(4.0,0.8) circle (0pt) node\[anchor=south\] [$\scriptstyle F_{k+3}-N$]{}; (4.0,0.8)–(4.0,0);
in [1,...,4]{} (4.0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [4]{};
in [1,...,16]{} (3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [5]{};
in [1,...,16]{} (0-0.2\*,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,7]{} (-3.2-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
(-4.8,0) circle (0.5pt) node\[anchor=north\] [6]{};
in [1,...,8]{} (-4.8-0.2\*,0) circle (0.5pt) node\[anchor=north\] [7]{};
Let $S=\{1,...,F_{k+1}-1\}$. Then $S$ is a set of vertices of degree at least 3. Since any automorphism sends the vertices of degree at least 3 to the vertices of degree at least 3, $\phi(S)=S$ and $\psi=\phi |_S $ is an automorphism of $G_{F_{k+1}-1}$. By Case 2a, $\psi$ is either an identity isomorphism or one that interchanges $\frac{1}{2}F_k$ with $\frac{3}{2}F_k$. However, the degrees of $\frac{1}{2}F_k$ and $\frac{3}{2}F_k$ are 4 and 3 respectively, and so for all $x<F_{k+1}$, $\phi(x)=x$.
Vertex $F_{k+1}$ is the only degree 2 vertex that does not have a neighbour of degree 2, and so $\phi(F_{k+1})=F_{k+1}$. Vertex $\frac{1}{2}F_{k+3}$ is the only degree 1 vertex adjacent to $\frac{1}{2}F_k$, and so $\phi(\frac{1}{2}F_{k+3})=F_{k+3}$.
Let $y \geq F_{k+2}$ be a degree 1 vertex. Then $y$ is adjacent to $F_{k+3}-y$, and since $F_{k+3}-y\leq F_{k+1}$, it follows that $\phi(F_{k+3}-y)=F_{k+3}-y$. Also, since $y$ is adjacent to $F_{k+3}-y$, it follows that $\phi(y)$ is adjacent to $F_{k+3}-y$. However, $y$ is the only degree 1 vertex adjacent to $F_{k+3}-y$ and so $\phi(y)=y$. Similar to Case 1c, $\phi$ fixes all vertices of degree 2 in $G_N$.
So $\phi$ fixes all vertices of $G_N$; that is, $\mbox{Aut}(G_N)$ consists only of the identity. This completes Case 5, and hence the proof of the induction step.
By mathematical induction, the proof of the theorem is complete.
Concluding remarks {#se:conclusion}
==================
This work may be seen as a starting point for other graphs formed by sums in any set given by a linear recurrence of order 2. Some properties for such graphs are known or conjectured (see [@Cost:08] and [@FKMOP:14] for references). Many of the known proofs (including some here) for Fibonacci-sum graphs may extend accordingly.
References {#references .unnumbered}
==========
[10]{}
K. Alladi, P. Erdős, and V. E. Hoggatt Jr., On additive partitions of integers, [*Discrete Math.*]{} [**22**]{} (1978), 201–211.
B. Barwell, Problem 2732, Problems and conjectures, [*Journal of Recreational Mathematics*]{} [**34**]{} (2006), 220–223.
H. L. Bodlaender, Dynamic programming on graphs with bounded treewidth, [*Proc. 15th International Colloquium on Automata, Languages and Programming*]{}, pp. 105–118 in Lecture Notes in Computer Science [**317**]{}, Springer-Verlag, 1988.
H. L. Bodlaender, A linear time algorithm for finding tree-decompositions of small treewidth, [SIAM J. Comput.]{} [**25**]{} (1996), 1305–1317.
C. J. Colbourn and J. A. Wald, Steiner trees, partial 2-trees, and minimum IFI networks, [*Networks*]{} [**13**]{} (1983), 159–167.
G. Costain, [*On the additive graph generated by a subset of the natural numbers*]{}, Master’s thesis, McGill University, June 2008. Available at: [digitool.library.mcgill.ca/thesisfile21911.pdf](digitool.library.mcgill.ca/thesisfile21911.pdf); accessed 15 May 2017.
R. Diestel, [*Graph theory*]{}, 4th ed., Springer, Graduate texts in Mathematics [**173**]{}, 2010.
K. Fox, W. B. Kinnersley, D. McDonald, N. Orlow, and G. J. Puleo, Spanning paths in Fibonacci-Sum graphs, [*Fibonacci Quart.*]{} [**52**]{} (2014), 46–49.
R. Halin, $S$ functions for graphs, [*J. Geometry*]{} [**8**]{} (1976), 171–186.
P. D. Seymour and R. Thomas, Graph minors II: Algorithmic aspects of tree-width, [*J. Algorithms*]{} [**7**]{} (1986), 309–322.
P. D. Seymour and R. Thomas, Graph searching and a min-max theorem for tree-width, [*J. Combin. Theory Ser. B*]{} [**58**]{} (1993), 22–33.
D. Silverman, Problem 566, [*Journal of Recreational Mathematics*]{} [**9**]{} (1976–1977), 298.
E. Zeckendorf, Représentation des nombres naturels par une somme de nombres de Fibonacci ou de nombres de Lucas. [*Bull. Soc. R. Sci. Liège* ]{} [**41**]{} (1972), 179-182.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work, we assess the accuracy of dielectric-dependent hybrid density functionals and many-body perturbation theory methods for the calculation of electron affinities of small water clusters, including hydrogen-bonded water dimer and water hexamer isomers. We show that many-body perturbation theory in the [G$_0$W$_0$]{} approximation starting with the dielectric-dependent hybrid functionals predicts electron affinities of clusters within 0.1 eV of the coupled-cluster results with single, double, and perturbative triple excitations.'
author:
- 'Alex P. Gaiduk'
- Francesco Paesani
- Giulia Galli
title: 'Electron affinities of water clusters from density-functional and many-body-perturbation theory'
---
The calculation of electron affinities of aqueous systems is a difficult task, due to the high level of theory required to describe the electronic properties of anions and the need to achieve a tight convergence as a function of numerical parameters. Here we focus on the water dimer and hexamer and we present results for their electron affinity computed using density functional theory (DFT), many body perturbation theory ([G$_0$W$_0$]{}) and the CCSD(T) method. The purpose of our work is to establish the accuracy of many-body perturbation theory calculations, starting from dielectric-dependent hybrid and semi-local density functionals.
Density functional calculations were carried out using the <span style="font-variant:small-caps;">quantum espresso</span> code [@QE-2009] with a plane-wave cutoff of 85 Ry and Hamann–Schlüter–Chiang–Vanderbilt (HSCV) pseudopotentials; [@Hamann:1979/PRL/1494; @Vanderbilt:1985/PRB/8412] many-body perturbation theory calculations in the [G$_0$W$_0$]{} approximation were performed with the <span style="font-variant:small-caps;">west</span> code.[@west] All calculations were performed in unit cells with size of 21.17 Å (40 a.u.); our computed energies varied by less than 0.01 eV when the cell size was increased to 31.75 Å (60 a.u.). In order to correct for spurious interactions between periodic images in plane-wave calculations, total energies and eigenvalues were computed using the Makov–Payne correction.[@makov_periodic_1995] Using Martyna–Tuckerman long-range interaction corrections [@martyna_reciprocal_1999] instead of the Makov–Payne scheme did not change our results in any noticeable way. We checked that the electron affinities were converged with respect to the energy cutoff within 0.001 and 0.005 eV when computed as total energy differences or lowest unoccupied molecular orbital (LUMO) energy, respectively.
Quasiparticle energies were computed using the <span style="font-variant:small-caps;">west</span> code. We tested the convergence of [G$_0$W$_0$]{} quasiparticle LUMO energies with respect to the number of eigenpotentials $N_\text{PDEP}$, and extrapolated our results to the infinite eigenpotential limit.[@govoni_2017] Prior to this analysis, we verified that the quasiparticle energies were converged within 0.01 eV in the unit cell with the size of 21.17 Å, at the maximum number of eigenpotentials employed in this work (512). We then computed quasiparticle energies of the lowest unoccupied state using [G$_0$W$_0$]{}/PBE at several $N_\text{PDEP}$ values, and fit the results to the function $a+b/N_\text{PDEP}$, where $b$ represents the [G$_0$W$_0$]{} energy in the limit $N_\text{PDEP} \rightarrow \infty$. The results denoted by “$\infty$" in Table \[tab:gw\] indicate the extrapolated quasiparticle energies for the LUMO energy. All [G$_0$W$_0$]{}calculations were performed with 256 eigenpotentials and corrected by the difference between the energies obtained at $\infty$ and $N_\text{PDEP} = 256$ ($-0.03$ eV for the dimer and $-0.15$ eV for the hexamer). All the [G$_0$W$_0$]{} values reported in this work include these corrections.
[lcc]{} &\
$N_\text{PDEP}$ & Dimer & Hexamer\
150 & $0.74$ & $0.52$\
192 & $0.72$ & $0.42$\
256 & $0.71$ & $0.39$\
320 & $0.71$ & $0.36$\
512 & $0.70$ & $0.33$\
\
$\infty$ & $0.68$ & $0.24$\
[l l @ d[2.2]{} \*4[d[2.2]{}]{}]{} & & &\
& & & & & &\
& aug-cc-pVDZ & -0.22 & 1.27 & -0.33 & 0.62 & -0.499\
& d-aug-cc-pVDZ & 0.12 & 1.28 & 0.06 & 0.66 & -0.070\
& t-aug-cc-pVDZ & 0.24 & 1.28 & 0.13 & 0.66 & -0.005\
& q-aug-cc-pVDZ & 0.31 & 1.28 & 0.18 & 0.66 & 0.0040\
& 5-aug-cc-pVDZ & 0.33 & 1.28 & 0.20 & 0.66 & 0.0048\
& 6-aug-cc-pVDZ & 0.34 & 1.28 & 0.21 & 0.66 & 0.0051\
\[4pt\] & aug-cc-pVTZ & -0.14 & 1.26 & -0.24 & 0.62 & -0.385\
& d-aug-cc-pVTZ & 0.13 & 1.26 & 0.07 & 0.64 & -0.055\
& t-aug-cc-pVTZ & 0.23 & 1.26 & 0.13 & 0.64 & -0.003\
& q-aug-cc-pVTZ & 0.30 & 1.26 & 0.18 & 0.64 & 0.0046\
\[4pt\] Plane-wave & 85 Ry$^a$ & 0.32 & 1.25 & 0.19 & 0.65 &\
The reference method chosen for benchmarking our results is the coupled cluster with singles, doubles, and perturbative correction for triples \[CCSD(T)\]. Coupled cluster calculations were performed using the <span style="font-variant:small-caps;">gaussian</span> 09 program [@g09d01] with tight convergence criteria for both the Hartree–Fock and CCSD iterative procedures. The accuracy of two-electron integrals was set to the $10^{-16}$ threshold (`Acc2E=16` keyword) to improve convergence when using very diffuse basis sets. In all cases, we performed stability calculations on converged Hartree–Fock wavefunctions to ensure the algorithm determined a minimum of the total energy and not a saddle point.
When computing electron affinities as differences of the total energies of the anion and neutral species, the size of the basis set, particularly the inclusion of diffuse basis functions with low exponents, plays a crucial role. This is especially important for water, since an extra electron is significantly delocalized, and the electron affinity is close to zero. We tested the completeness of the basis set using a hydrogen-bonded water dimer. We employed augmented Dunning basis sets with two (DZ) and three (TZ) sets of polarization functions, and variable number of added diffuse shells ranging between 1 and 6. The doubly-, triply-, and quadruply-augmented basis sets were described in the literature; [@woon_gaussian_1994] we constructed basis sets with 5 and 6 sets of diffuse functions following the recipe of Ref. . The 6-aug-cc-pVDZ and q-aug-cc-pVTZ basis sets employed in this work are listed in the Appendices \[sec:6augccpVDZ\] and \[sec:qaugccpVTZ\] in the format suitable for the <span style="font-variant:small-caps;">gaussian</span> code; basis sets with fewer diffuse functions can be obtained by sequentially removing one or more sets of outer diffuse basis functions of each type.
We start by discussing the water dimer. Comparing our DFT and [G$_0$W$_0$]{} results to the reference CCSD(T) data requires comparison of the results obtained with plane waves (PW) to those obtained with localized Gaussian-type orbitals (GTO). In order to do so, we first compared the electron affinities computed with the PBE and PBE0 functionals and plane-wave basis set with the <span style="font-variant:small-caps;">quantum espresso</span> code, with those computed with the <span style="font-variant:small-caps;">gaussian</span> 09 code. The results, reported in Table \[tab:basis\], show that the $\Delta$SCF values are not particularly sensitive to the number of polarization functions but are very sensitive to the number of diffuse functions. Namely, values computed in d-aug-cc-pVDZ and d-aug-cc-pVTZ basis sets are within 0.01–0.02 eV of each other. On the other hand, it takes about 5 sets of diffuse functions added to both double-zeta and triple-zeta basis sets to converge the differences of total energies within 0.01 eV.
[l @ d[2.2]{} \*2[d[2.2]{}]{}]{} & & &\
\[2pt\] PBE & 0.32 & 1.25 & -0.68\
PBE0 & 0.19 & 0.65 & -0.45\
RSH (0.565)$^a$ & 0.26 & 0.18 & -0.08\
RSH (1.0)$^b$ & 0.38 & 0.03 & 0.05\
\[4pt\] CCSD(T) & 0.0051\^c & &\
\[-10pt\]\
[l @ \*[4]{}[d[2.2]{}]{}]{} &\
Basis set & & & &\
aug-cc-pVDZ & -0.087 & -0.175 & -0.117 & 0.070\
t-aug-cc-pVDZ & 0.087 & 0.049 & 0.077 & 0.152\
q-aug-cc-pVDZ & 0.088 & 0.050 & 0.077 & 0.172\
[l l @ d[2.2]{} \*2[d[2.2]{}]{} d[2.3]{}]{} & & & & &\
\[2pt\] & PBE & 0.32 & 1.48 & -0.24 &\
& PBE0 & 0.23 & 0.86 & -0.14 &\
& RSH (0.565) & 0.18 & 0.32 & 0.09 &\
& RSH (1.0) & 0.37 & 0.05 & 0.20 &\
& PBE & 0.29 & 1.38 & -0.33 &\
& PBE0 & 0.21 & 0.77 & -0.21 &\
& RSH (0.565) & 0.20 & 0.25 & 0.08 &\
& RSH (1.0) & 0.38 & 0.05 & 0.19 &\
& PBE & 0.32 & 1.45 & -0.27 &\
& PBE0 & 0.22 & 0.83 & -0.16 &\
& RSH (0.565) & 0.17 & 0.30 & 0.09 &\
& RSH (1.0) & 0.38 & 0.05 & 0.20 &\
& PBE & 0.50 & 1.85 & 0.01 &\
& PBE0 & 0.36 & 1.19 & 0.03 &\
& RSH (0.565) & 0.23 & 0.52 & 0.16 &\
& RSH (1.0) & 0.14 & 0.08 & 0.24 &\
Plane-wave basis set with the cutoff of 85 Ry provides an almost complete basis-set limit for the electron affinities, as shown by the comparison of the results obtained using the largest GTO and PW basis sets. This implies that (i) the protocol for computing the total energies for water clusters in plane waves, including the finite-size correction of Makov–Payne, [@makov_periodic_1995] is reliable for the systems studied here; and (ii) the basis sets chosen here will provide accurate representation of the anions for benchmarking our DFT and GW methods against the coupled-cluster method.
![ Molecular structures for the book, cage, prism, and ring isomers of the water hexamer employed in this work. The structures were relaxed for the negatively charged species at the wB97XD[@chai_long-range_2008]/aug-cc-pVTZ level of theory. The optimized geometries are listed in the Appendix \[sec:hexamers\]. []{data-label="fig:hexamers"}](fig1.pdf){width="0.6\columnwidth"}
We first checked the convergence of the CCSD(T) calculations with respect to the basis set size for the water dimer. Our results, reported in the last column of Table \[tab:basis\], confirmed that quadruply-augmented basis sets provide essentially converged CCSD(T) electron affinity. Even for triply-augmented basis sets, the error in $\Delta$SCF values is just 0.01 eV. We compared our results to those of Kim et al.,[@kim_quantum-mechanical_1999] who used TZ($2df$,$2pd$)+($3s3p$,$3s$) basis set at the CCSD(T) level of theory and obtained $\text{EA} = 0.0044$ eV. The TZ($2df$,$2pd$)+($3s3p$,$3s$) basis is essentially the same as the cc-pVTZ basis augmented with 3 sets of diffuse $s$ and $p$ functions for oxygen, and 3 sets of diffuse $s$ functions for hydrogen. Their basis set should be comparable to, but slightly smaller than the q-aug-cc-pVTZ basis used here, explaining the similarity of the value $0.0044$ eV obtained in Ref. and our value of $0.0046$ eV obtained using the q-aug-cc-pVTZ basis set. Overall, we consider the value of $0.0051$ eV as an accurate CCSD(T) reference value for the electron affinity of the dimer, which will be used for comparison with our DFT and MBPT results.
Table \[tab:methods\] reports the electron affinities computed using PBE,[@Perdew:1996/PRL/3865; @Perdew:1997/PRL/1396] PBE0, [@Adamo:1999/JCP/6158] and RSH [@skone_rsh_2016] functionals as the difference of total energies ($\Delta$SCF), and as a DFT or [G$_0$W$_0$]{} LUMO energy. We used the RSH functional with two different values of the dielectric screening: the same as used for bulk water[@gaiduk_electron_affinity_2017] and the one used for molecules (i.e. that of vacuum) in the original definition of RSH reported in Ref. . Figure \[fig:EA\_dimers\] summarizes deviations of all these quantities from the CCSD(T) value. As expected, the largest errors are found for the DFT LUMO values, due to the fact that generalized-gradient approximations such as PBE lack piecewise linearity of the total energy with respect to the number of electrons $N$, leading to an inaccurate approximation of the vertical electron affinity by the LUMO energy.[@Perdew:1982/PRL/1691] $\Delta$SCF calculations yield improved results, as total energies are piecewise linear with respect to $N$ by construction, and their accuracy is limited only by the accuracy of a given functional. Electron affinities from [G$_0$W$_0$]{} calculations are less accurate than $\Delta$SCF results for PBE and PBE0 approximations but are equally or more accurate than DFT results for RSH functional with different fractions of the exact exchange. This is likely because the dielectric-constant-dependent functionals have lower self-interaction error than the PBE and PBE0 functionals, and thus provide a more realistic starting point for [G$_0$W$_0$]{} corrections. We note that DFT approximations consistently overestimate the electron affinity of the dimer compared to the CCSD(T) value, while [G$_0$W$_0$]{} yields the wrong sign for RSH (0.565) and a slight overestimate for RSH (1.0). Overall, deviation of the [G$_0$W$_0$]{}/RSH values from the CCSD(T) reference for the water dimer is less than 0.1 eV, irrespective of the sign.
In addition to the dimer, we computed the electron affinities of larger water hexamers shown in Figure \[fig:hexamers\]. Following the convergence study reported in Table \[tab:basis\], we computed CCSD(T) electron affinities using the q-aug-cc-pVDZ basis set, which turned out to be converged within 0.02 eV with respect to the number of diffuse functions (see Table \[tab:convergence\_hexamers\]). The results of our analysis for water hexamers are reported in Table \[tab:hexamer\] and Figure \[fig:EA\_hexamers\], showing trends similar to the dimer for the relative accuracy of various approximations. We found again that [G$_0$W$_0$]{}/RSH protocol is the most accurate, with an average deviation of the electron affinities from the CCSD(T) results of only 0.01 eV for RSH (0.565) and of 0.11 eV for RSH (1.0). Overall, we showed that many-body perturbation theory calculations coupled with dielectric-constant-dependent functional RSH predicts the electron affinities of water clusters within 0.1 eV from the golden standard of quantum chemistry, CCSD(T).
The authors gratefully acknowledge helpful discussions with Marco Govoni and Nicholas Brawand. APG and GG were supported by MICCoM as part of the Computational Materials Sciences Program funded by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES), Materials Sciences and Engineering Division (5J-30161-0010A). APG was also supported by the postdoctoral fellowship from the Natural Sciences and Engineering Research Council of Canada. FP was supported by the National Science Foundation through grant CHE-1453204 and used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by the National Science Foundation through grant ACI-1053575. An award of computer time was provided by the INCITE program. This research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under contract DEAC02-06CH11357.
{width=".8\columnwidth"}
{width=".8\columnwidth"}
Water hexamers {#sec:hexamers}
==============
All molecular geometries are in in xyz format, with the coordinates given in Angstroms.
18
Book hexamer
O 0.274 1.424 1.149
H 0.301 2.281 1.584
H 1.065 1.395 0.561
O -0.182 -1.395 1.046
H -1.007 -1.426 0.528
H -0.094 -0.465 1.303
O -1.922 1.452 -0.590
H -1.725 2.060 -1.310
H -1.152 1.512 0.006
O -2.533 -1.178 -0.486
H -3.310 -1.150 0.071
H -2.345 -0.235 -0.685
O 2.359 1.101 -0.549
H 2.137 1.586 -1.351
H 2.265 0.152 -0.753
O 2.003 -1.665 -0.565
H 2.686 -1.978 0.028
H 1.187 -1.641 -0.017
18
Cage hexamer
O 0.668 -1.715 -0.324
H 0.888 -2.653 -0.324
H 1.530 -1.248 -0.246
O -0.600 0.461 -1.631
H -1.491 0.364 -1.254
H -0.184 -0.399 -1.455
O 0.704 1.769 0.353
H 0.776 2.712 0.212
H 0.238 1.400 -0.435
O -0.833 -0.335 1.654
H -0.335 0.487 1.561
H -0.312 -0.969 1.129
O 2.829 -0.072 -0.059
H 3.390 -0.383 0.659
H 2.313 0.674 0.286
O -2.928 0.016 -0.061
H -3.279 -0.858 -0.230
H -2.258 -0.115 0.647
18
Prism hexamer
O -1.409 -0.377 1.473
H -2.071 -0.903 1.937
H -0.529 -0.766 1.643
O -1.599 -0.612 -1.295
H -2.396 -1.046 -1.619
H -1.654 -0.691 -0.322
O -0.931 1.906 -0.094
H -1.269 1.374 -0.827
H -1.226 1.389 0.673
O 1.263 -1.095 1.417
H 1.596 -0.202 1.266
H 1.243 -1.449 0.511
O 1.173 -1.205 -1.425
H 1.471 -0.300 -1.275
H 0.212 -1.123 -1.545
O 1.709 1.371 -0.093
H 2.214 2.175 -0.205
H 0.760 1.640 -0.097
18
Ring hexamer
O -1.228 2.029 0.652
H -1.731 1.307 0.229
H -1.025 1.699 1.538
O -1.145 -2.077 0.650
H -0.962 -1.739 1.536
H -0.268 -2.152 0.228
O 1.159 2.229 -0.654
H 0.294 2.210 -0.185
H 0.953 1.985 -1.557
O 1.351 -2.120 -0.652
H 1.246 -1.819 -1.556
H 1.768 -1.362 -0.183
O 2.374 0.049 0.650
H 1.999 0.845 0.226
H 1.991 0.040 1.536
O -2.512 -0.110 -0.654
H -2.063 -0.850 -0.185
H -2.195 -0.166 -1.556
6-aug-cc-pVDZ basis set {#sec:6augccpVDZ}
=======================
-H 0
S 3 1.00
13.0100000 0.0196850
1.9620000 0.1379770
0.4446000 0.4781480
S 1 1.00
0.1220000 1.0000000
S 1 1.00
0.0297400 1.0000000
S 1 1.00
0.0072500 1.0000000
S 1 1.00
0.0017674 1.0000000
S 1 1.00
0.0004309 1.0000000
S 1 1.00
0.0001051 1.0000000
S 1 1.00
0.0000256 1.0000000
P 1 1.00
0.7270000 1.0000000
P 1 1.00
0.1410000 1.0000000
P 1 1.00
0.0273000 1.0000000
P 1 1.00
0.0052857 1.0000000
P 1 1.00
0.0010234 1.0000000
P 1 1.00
0.0001981 1.0000000
P 1 1.00
0.0000383 1.0000000
****
-O 0
S 8 1.00
11720.0000000 0.0007100
1759.0000000 0.0054700
400.8000000 0.0278370
113.7000000 0.1048000
37.0300000 0.2830620
13.2700000 0.4487190
5.0250000 0.2709520
1.0130000 0.0154580
S 8 1.00
11720.0000000 -0.0001600
1759.0000000 -0.0012630
400.8000000 -0.0062670
113.7000000 -0.0257160
37.0300000 -0.0709240
13.2700000 -0.1654110
5.0250000 -0.1169550
1.0130000 0.5573680
S 1 1.00
0.3023000 1.0000000
S 1 1.00
0.0789600 1.0000000
S 1 1.00
0.0206000 1.0000000
S 1 1.00
0.0053744 1.0000000
S 1 1.00
0.0014021 1.0000000
S 1 1.00
0.0003658 1.0000000
S 1 1.00
0.0000954 1.0000000
P 3 1.00
17.7000000 0.0430180
3.8540000 0.2289130
1.0460000 0.5087280
P 1 1.00
0.2753000 1.0000000
P 1 1.00
0.0685600 1.0000000
P 1 1.00
0.0171000 1.0000000
P 1 1.00
0.0042650 1.0000000
P 1 1.00
0.0010638 1.0000000
P 1 1.00
0.0002653 1.0000000
P 1 1.00
0.0000662 1.0000000
D 1 1.00
1.1850000 1.0000000
D 1 1.00
0.3320000 1.0000000
D 1 1.00
0.0930000 1.0000000
D 1 1.00
0.0260512 1.0000000
D 1 1.00
0.0072975 1.0000000
D 1 1.00
0.0020442 1.0000000
D 1 1.00
0.0000573 1.0000000
****
q-aug-cc-pVTZ basis set {#sec:qaugccpVTZ}
=======================
-H 0
S 3 1.00
33.8700000 0.0060680
5.0950000 0.0453080
1.1590000 0.2028220
S 1 1.00
0.3258000 1.0000000
S 1 1.00
0.1027000 1.0000000
S 1 1.00
0.0252600 1.0000000
S 1 1.00
0.0062100 1.0000000
S 1 1.00
0.0015267 1.0000000
S 1 1.00
0.0003753 1.0000000
P 1 1.00
1.4070000 1.0000000
P 1 1.00
0.3880000 1.0000000
P 1 1.00
0.1020000 1.0000000
P 1 1.00
0.0268000 1.0000000
P 1 1.00
0.0070416 1.0000000
P 1 1.00
0.0018502 1.0000000
D 1 1.00
1.0570000 1.0000000
D 1 1.00
0.2470000 1.0000000
D 1 1.00
0.0577000 1.0000000
D 1 1.00
0.0134789 1.0000000
D 1 1.00
0.0031487 1.0000000
****
-O 0
S 8 1.00
15330.0000000 0.0005080
2299.0000000 0.0039290
522.4000000 0.0202430
147.3000000 0.0791810
47.5500000 0.2306870
16.7600000 0.4331180
6.2070000 0.3502600
0.6882000 -0.0081540
S 8 1.00
15330.0000000 -0.0001150
2299.0000000 -0.0008950
522.4000000 -0.0046360
147.3000000 -0.0187240
47.5500000 -0.0584630
16.7600000 -0.1364630
6.2070000 -0.1757400
0.6882000 0.6034180
S 1 1.00
1.7520000 1.0000000
S 1 1.00
0.2384000 1.0000000
S 1 1.00
0.0737600 1.0000000
S 1 1.00
0.0228000 1.0000000
S 1 1.00
0.0070477 1.0000000
S 1 1.00
0.0021785 1.0000000
P 3 1.00
34.4600000 0.0159280
7.7490000 0.0997400
2.2800000 0.3104920
P 1 1.00
0.7156000 1.0000000
P 1 1.00
0.2140000 1.0000000
P 1 1.00
0.0597400 1.0000000
P 1 1.00
0.0167000 1.0000000
P 1 1.00
0.0046684 1.0000000
P 1 1.00
0.0013050 1.0000000
D 1 1.00
2.3140000 1.0000000
D 1 1.00
0.6450000 1.0000000
D 1 1.00
0.2140000 1.0000000
D 1 1.00
0.0710000 1.0000000
D 1 1.00
0.0235561 1.0000000
D 1 1.00
0.0078153 1.0000000
F 1 1.00
1.4280000 1.0000000
F 1 1.00
0.5000000 1.0000000
F 1 1.00
0.1750000 1.0000000
F 1 1.00
0.0612500 1.0000000
F 1 1.00
0.0214375 1.0000000
****
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a finitely generated bigraded $S$-module $M$ where $S$ is a standard bigraded polynomial ring we define the Cohen–Macaulay filtration $\mathcal{F}$ of $M$ with respect to the bigraded irrelevant ideal $Q$. We call $M$ to be sequentially Cohen–Macaulay with respect $Q$ if $M$ admits a Cohen–Macaulay filtration with respect to $Q$. The algebraic properties of these modules are investigated. All hypersurface rings that are sequentially Cohen–Macaulay with respect to $Q$ are classified.'
address: ' Ahad Rahimi, Department of Mathematics, Razi University, Kermanshah, Iran. School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P. O. Box: 19395-5746, Tehran, Iran. '
author:
- Ahad Rahimi
title: 'Sequentially Cohen–Macaulay modules with respect to an irrelevant ideal'
---
Introduction {#introduction .unnumbered}
============
In [@St] Stanley introduced the notion sequentially Cohen–Macaulay for graded modules. This concept was studied by several authors. We refer to ([@CN], [@F], [@HS], [@S], [@TY]). In this paper we define this notion for bigraded modules. Some new algebraic invariants make sense to study in this case. We let $S=K[x_1, \dots, x_m, y_1, \dots, y_n]$ be the standard bigraded polynomial ring over a field $K$, $P=(x_1, \dots, x_m)$ and $Q=(y_1, \dots, y_n)$ the bigraded irrelevant ideals of $S$ and $M$ a finitely generated bigraded $S$-module. In [@AR2] we call $M$ to be relative Cohen–Macaulay with respect to $Q$ if $\grade(Q, M)=\cd(Q, M)$ where $\cd(Q, M)$ denotes the cohomological dimension of $M$ with respect to $Q$. We define a finite filtration $\mathcal{F}$: $0=M_0\varsubsetneq M_1
\varsubsetneq
\dots \varsubsetneq M_r=M$ of $M$ by bigraded submodules $M$ a Cohen–Macaulay filtration with respect to $Q$ if
- Each quotient $M_i/M_{i-1}$ is relative Cohen–Macaulay with respect to $Q$;
- $0 \leq \cd(Q, M_1/M_0)<\cd(Q, M_2/M_1)< \dots< \cd(Q, M_r/M_{r-1})$.
We call $M$ to be sequentially Cohen–Macaulay with respect to $Q$ if $M$ admits a Cohen–Macaulay filtration with respect to $Q$. Sequentially Cohen–Macaulay introduced by Stanley results from our definition if we assume $P=0$. The paper is organized as follows:
In the preliminary section, we show that “a Cohen–Macaulay filtration with respect to $Q$ is unique”. Also for a Cohen–Macaulay filtration $\mathcal{F}$ of $M$ with respect to $Q$ we observe that $H^{q_i}_Q(M)\iso H^{q_i}_Q(M_i)\iso H^{q_i}_Q(M_i/M_{i-1})$ where $q_i=\cd(Q, M_i)$ for $i=1, \dots, r$ and $H^{k}_Q(M)=0$ for $k\not
\in \{ q_1, \dots, q_r\}.$ This observation is used to derive some basic algebraic properties of sequentially Cohen–Macaulay modules with respect to $Q$ in the next section. For instance, we improve [@AR2 Proposition 2.3] as follows: Consider the graded components of $H^k_{Q}(M)$ as a finitely generated graded $K[x]$-module where $K[x]=K[x_1,
\dots, x_n]$. If $M$ is sequentially Cohen–Macaulay with respect to $Q$, then the Castelnuovo–Mumford regularity of $H^k_{Q}(M)_j$ is bounded for all $k$. In Section 2 we also show that if $M$ is sequentially Cohen–Macaulay with respect to $Q$ with $\grade(Q, M)>0$ and $|K|=\infty$, then there exists a bihomogeneous $M$-regular element $y\in Q$ of degree $(0,1)$ such that $M/yM$ is sequentially Cohen–Macaulay with respect to $Q$, too. We give an example to show that the converse does not hold.
In Section 3 we consider $L\tensor_KN$ as $S$-module where $L$ and $N$ are two non-zero finitely generated graded modules over $K[x]$ and $K[y]$, respectively. We obtain a characterization for sequentially Cohen–Macaulayness $L\tensor_KN$ with respect to $Q$ as follows: $L\tensor_KN$ is sequentially Cohen–Macaulay with respect to $Q$ if and only if $N$ is Cohen–Macaulay $K[y]$-module. We also discuss two questions on the sequentially Cohen–Macaulayness of $M$ with respect to $P$, $Q$, and $P+Q$ in general.
In the last section, we let $f\in S$ be a bihomogeneous element of degree $(a, b)$ and consider the hypersurface ring $R=S/fS$. Notice that if $a, b>0$, we have $\grade(Q, R)=n-1$ and $\cd(Q, R)=n$. Hence $R$ is not relative Cohen–Macaulay with respect to $Q$. Thus it is natural to ask whether $R$ is sequentially Cohen–Macaulay with respect to $Q$. We classify all hypersurface rings that are sequentially Cohen–Macaulay with respect to $Q$ as follows: Let $R$ be the hypersurface ring. Then $R$ is sequentially Cohen–Macaulay with respect to $Q$ if and only if $f=h_1h_2$ where $\deg(h_1)=(a, 0)$ with $a\geq 0$ and $\deg(h_2)=(0, b)$ with $b\geq 0$.
Preliminaries
=============
Let $K$ be a field and $S=K[x_1, \dots, x_m, y_1, \dots, y_n]$ be the standard bigraded polynomial ring over. In other words, $\deg x_i=(1,0)$ and $\deg y_j=(0, 1)$ for all $i$ and $j$. We set the bigraded irrelevant ideals $P=(x_1, \dots, x_m)$ and $Q=(y_1, \dots, y_n)$. Let $M$ be a finitely generated bigraded $S$-module. We denote by $\cd(Q, M)$ the [*cohomological dimension of $M$ with respect to $Q$* ]{} which is the largest integer $i$ for which $H^i_ Q
(M)\neq 0$. Notice that $0\leq \cd(Q, M) \leq n.$ We recall the following [@AR2 Definition 1.3]
\[cd\] [*Let $M $ be a finitely generated bigraded $S$-module and $q \in \ZZ$. We call M to be [*relative Cohen–Macaulay with respect to $Q$*]{} if $H^i_Q(M) = 0$ for all $i \neq q$. For simplicity, from now on we say ”$M$ is Cohen–Macaulay with respect to $Q$” instead of ”$M$ is relative Cohen–Macaulay with respect to $Q$”.* ]{}
[ *By [@AR2 \[Proposition 1.2] $M$ is Cohen–Macaulay with respect to $Q$ if and only if $\grade(Q, M) = \cd(Q, M) = q$. This is also equivalent to saying that $M_k=M_{(k, *)}=\dirsum_jM_{(k, j)}$ is finitely generated Cohen–Macaulay $K[y]$-module of dimension $q$ for all $k$. We recall the following facts from [@AR2] which will be used in the sequel. $$\begin{aligned}
\cd(P, M)=\dim M/QM \; \;\; and \;\;\;\cd(Q, M)=\dim M/PM.\end{aligned}$$ If $M$ is Cohen–Macaulay, then $$\begin{aligned}
\grade(P, M)+\cd(Q, M)=\dim M,\end{aligned}$$ see, Formula (5) in [@AR2]. If $M$ is Cohen–Macaulay with respect to $Q$ with $ K=\infty$, then $$\begin{aligned}
\cd(P, M)+\cd(Q, M)=\dim M,\end{aligned}$$ see, [@AR2 Theorem, 3.6]. We also recall the following facts from [@CJR]. If $ 0
\rightarrow M' \rightarrow M \rightarrow M'' \rightarrow 0$ is an exact sequence of $S$-modules with $M$ finitely generated, then $$\begin{aligned}
\label{1} \cd(Q,M)=\max\{\cd(Q, M'), \cd(Q,M'')\},\end{aligned}$$ see [@CJR Proposition 4.4]. If $M$ is a finitely generated $S$-module, then $$\begin{aligned}
\label{2}
\cd(Q,M) = \max \{\cd (Q, S/{\pp}): \pp \in \Ass(M)\},
\end{aligned}$$ and $$\begin{aligned}
\label{3}
\cd(Q, M) = \max \{ \cd(Q, S/\pp) : \pp \in \Supp(M)\},
\end{aligned}$$ see, [@CJR Corollary 4.6].* ]{}
\[seq1\]
*Let $M$ be a finitely generated bigraded $S$-module. We call a finite filtration $\mathcal{F}$: $0=M_0\varsubsetneq M_1 \varsubsetneq
\dots \varsubsetneq M_r=M$ of $M$ by bigraded submodules $M$ a [*Cohen–Macaulay filtration with respect to $Q$*]{} if*
- Each quotient $M_i/M_{i-1}$ is Cohen–Macaulay with respect to $Q$;
- $0 \leq \cd(Q, M_1/M_0)<\cd(Q, M_2/M_1)< \dots< \cd(Q, M_r/M_{r-1})$.
A Cohen–Macaulay filtration introduced by Stanley is in our terminology a Cohen–Macaulay filtration with respect to the maximal ideal $\mm=P+Q$.
[*Let $M$ be a finitely generated bigraded $S$-module. We call $M$ to be [*sequentially Cohen–Macaulay with respect to $Q$*]{} if $M$ admits a Cohen–Macaulay filtration with respect to $Q$.*]{}
We observe that ordinary sequentially Cohen–Macaulay modules are special cases of our definition. In fact, if we assume $P=0$. Then $m=0$, and $Q=\mm$ is the unique graded maximal ideal of $S$ with $\deg y_i=1$ for $i=1, \dots, n$ and $\cd(Q,M_i/M_{i-1})=\cd(\mm,M_i/M_{i-1})=\dim M_i/M_{i-1}$ for $i=1,
\dots, r$.
[ *Cohen–Macaulay modules with respect to $Q$ are the obvious examples of sequentially Cohen–Macaulay with respect to $Q$.*]{}
[*If $M$ is finitely generated bigraded $S$-module for which $\cd(Q, M)\leq 1$, then $M$ is sequentially Cohen–Macaulay with respect to $Q$. In fact, we may assume that $M$ is not Cohen–Macaulay with respect to $Q$. Thus $\grade(Q,M)=0$ and $\cd(Q,
M)=1$. The filtration $0=M_0 \varsubsetneq M_1 \varsubsetneq
M_2=M$ where $M_1=H^0_Q(M)$ is a Cohen-Macaulay filtration with respect to $Q$. Notice that, if $n=1$ in the polynomial ring $S$, then any bigraded $S$-module is sequentially Cohen–Macaulay with respect to $Q$.*]{}
\[reg\] Let $M$ be a finitely generated bigraded $S$-module that is sequentially Cohen–Macaulay with respect to $Q$ with the Cohen–Macaulay filtration $\mathcal{F}$: $0=M_0\varsubsetneq M_1 \varsubsetneq \dots \varsubsetneq M_r=M$ with respect to $Q$. Then we have $H^{q_i}_Q(M)\iso H^{q_i}_Q(M_i)\iso H^{q_i}_Q(M_i/M_{i-1})$ where $q_i=\cd(Q, M_i)$ for $i=1, \dots, r$ and $H^{k}_Q(M)=0$ for $k\not
\in \{ q_1, \dots, q_r\}.$
We proceed by induction on the length $r$ of the filtration of $M$. Let $r=1$ and consider the Cohen–Macaulay filtration $0=M_0\varsubsetneq M_1=M$. As $M_1=M$ is relative Cohen–Macaulay with respect to $Q$, the assertion is clear in this case. The assertion is also clear for $r=2$ by considering the exact sequence $0\rightarrow M_{1} \rightarrow M_2 \rightarrow
M_2/M_{1} \rightarrow 0$ and the fact that $M_1$ is Cohen–Macaulay with respect to $Q$. Now suppose $r\geq 3$ and the statement holds for sequentially Cohen–Macaulay modules with respect to $Q$ with filtrations of length less than $r$. We want to prove it for $M$ which is sequentially Cohen–Macaulay with respect to $Q$ and has the Cohen–Macaulay filtration $\mathcal{F}$: $0=M_0\varsubsetneq M_1\varsubsetneq
\dots \varsubsetneq M_r=M$ with respect to $Q$ of length $r$. Notice that $M_{r-1}$ which appears in the filtration $\mathcal{F}$ of $M$ is also sequentially Cohen–Macaulay with respect to $Q$. Thus by the induction hypothesis we have $H^{q_i}_Q(M_{r-1})\iso
H^{q_i}_Q(M_i)\iso H^{q_i}_Q(M_i/M_{i-1})$ for $i=1, \dots, r-1$ and $H^{k}_Q(M_{r-1})=0$ for $k\not \in \{ q_1, \dots, q_{r-1}\}.$ Now the exact sequence $0\rightarrow M_{r-1} \rightarrow M \rightarrow
M/M_{r-1} \rightarrow 0$ yields $H^{q_r}_Q(M)\iso
H^{q_r}_Q(M_r/M_{r-1})$ and $H^{t}_Q(M)\iso H^{t}_Q(M_{r-1})$ for $0\leq t < q_r.$ Therefore, the desired result follows.
In the proof of Proposition \[reg\] one obtains the following observations
\[seq\] With the assumptions and the notation of Proposition \[reg\] we have
- $\cd(Q, M_i)=\cd(Q, M_i/M_{i-1})$ for $i=1, \dots, r$ and hence
$\cd(Q, M_1)<\cd(Q, M_2)< \dots < \cd(Q, M_r)=\cd(Q, M)$.
- $\grade(Q, M_i)=q_1$ for $i=1, \dots, r$.
- $M_1=H^0_Q(M)$ if and only if $q_1=0$.
[*Parts (a) and (b) say that $\grade(Q, M)=q_1$ and $\cd(Q,
M)=q_r$.*]{}
Next, we show that the filtration $\mathcal{F}$ given in Definition \[seq1\] is unique. To answer this question we need some preparation. We let $M$ be a finitely generated bigraded $S$-module and, $\sum$ be the set of all bigraded submodules $N$ of $M$ such that $\cd(Q,N)<\cd(Q, M)$. As $M$ is a Noetherian $S$-module, $\sum$ has a maximal element with respect to inclusion, say $N$. Let $L$ be an arbitrary element in $\sum$. By using (4) it follows that $\cd(Q,
L+N)<\cd(Q, M)$ and hence the maximality $N$ yields $L\subseteq N$. Thus our discussion so far shows
There is a unique largest bigraded submodule $N$ of $M$ for which $\cd(Q,N)<\cd(Q,
M)$.
Now we can make the following definition:
\[1\] [*Let $M$ be a finitely generated bigraded $S$-module. We call a filtration $\mathcal{G}$: $0=M_0\varsubsetneq M_1
\varsubsetneq
\dots \varsubsetneq M_r=M$ of bigraded submodules of $M$ the [ *dimension filtration of $M$ with respect to $Q$* ]{} if $M_{i-1}$ is the largest submodule of $M_i$ for which $\cd(Q,M_{i-1})<\cd(Q, M_i)$ for all $i=1, \dots, r$.*]{}
A dimension filtration introduced by Schenzel [@S] is a dimension filtration with respect to the maximal ideal $\mm=P+Q$. The filtration $\mathcal{G}$ given in Definition \[1\] is unique. In order to prove the uniqueness $\mathcal{F}$ given in Definition \[seq1\], we will show that $\mathcal{F}=\mathcal{G}$.
In [@JR] $M$ is called to be [*relatively unmixed*]{} with respect to $Q$ if $\cd(Q, M)=\cd(Q, S/\pp)$ for all $\pp \in \Ass(M).$
\[CM\] Let $M$ be a finitely generated bigraded $S$-module and $N$ a non-zero bigraded submodule of $M$. If $M$ is relative Cohen–Macaulay with respect to $Q$, then $\cd(Q, N)=\cd(Q, M)$.
Since $M$ is Cohen–Macaulay with respect to $Q$, it follows from [@JR Corollary 1.11] that $M$ is relatively unmixed with respect to $Q$ and hence $\cd(Q, M)=\cd(Q, S/\pp)$ for all $\pp \in \Ass(M)$. As $N$ is a non-zero submodule of $M$ we have $\Ass(N)\neq\emptyset$ and $\Ass(N)\subseteq \Ass(M)$. Thus by using (5) we have $\cd(Q,N) = \max \{\cd (Q, S/{\pp}): \pp \in
\Ass(N)\}=\cd(Q, M)$, as desired.
By an argument similar to [@CN Lemma 4.4] we have the following
\[unique\] Let $\mathcal{F}$ be the Cohen–Macaulay filtration of $M$ with respect to $Q$ given in Definition \[seq1\] and $\mathcal{G}$ be the dimension filtration of $M$ with respect to $Q$ given in Definition \[1\]. Then $\mathcal{F}=\mathcal{G}$.
Set $\mathcal{F}$: $0=M_0\varsubsetneq M_1 \varsubsetneq \dots
\varsubsetneq M_r=M$ and $\mathcal{G}$: $0=N_0\varsubsetneq N_1
\varsubsetneq \dots \varsubsetneq N_s=M$. We will show that $r=s$ and $M_i=N_i$ for all $i$. By Corollary \[seq\](a) we have $\cd(Q, M_{i-1})<\cd(Q, M_{i})$ for all $i=1, \dots, r.$ Hence, Definition \[1\] says that $M_{r-1}\subseteq N_{s-1}$. Assume $M_{r-1}\varsubsetneq N_{s-1}$. Thus $N_{s-1}/M_{r-1}$ is a non-zero submodule of $M/M_{r-1}$. Since $M/M_{r-1}$ is Cohen–Macaulay with respect to $Q$, it follows from Lemma \[CM\] that $\cd(Q, N_{s-1}/M_{r-1})=\cd(Q,M/M_{r-1})=\cd(Q, M)$. The second equality follows from Corollary \[seq\](a). Now by using (4) the exact sequence $0 \rightarrow M_{r-1} \rightarrow N_{s-1}
\rightarrow N_{s-1}/M_{r-1} \rightarrow 0$ yields $\cd(Q,
N_{s-1})=\cd(Q, M)$, a contradiction. Thus $M_{r-1}=N_{s-1}$. Continuing in this way, we get $r=s$ and $M_i=N_i$ for all $i$. Therefore, $\mathcal{F}=\mathcal{G}$.
Some algebraic properties of sequentially Cohen–Macaulay modules with respect to $Q$
=====================================================================================
In this section, we discuss some algebraic properties of sequentially Cohen–Macaulay modules with respect to $Q$. We begin with the following
\[unmixed\] Let $M$ be a finitely generated bigraded $S$-module that is sequentially Cohen–Macaulay with respect to $Q$. If $M$ is relatively unmixed with respect to $Q$, then $M$ is Cohen–Macaulay with respect to $Q$. In particular, unmixed sequentially Cohen–Macaulay modules are Cohen–Macaulay.
Let $0=M_0\varsubsetneq M_1 \varsubsetneq \dots \varsubsetneq
M_r=M$ be the Cohen–Macaulay filtration with respect to $Q$. By Corollary \[seq\] we have $\grade(Q, M)=\grade(Q, M_1)$. Since $M_1$ is Cohen–Macaulay with respect to $Q$, it follows from [@JR Corollary 1.11] that $M_1$ is relatively unmixed with respect to $Q$. Thus $\grade(Q, M)=\grade(Q, M_1)=\cd(Q, M_1)=\cd(Q, S/\pp)$ for all $\pp \in \Ass(M_1)$. As $M$ is relatively unmixed with respect to $Q$ and $\Ass(M_1)\subseteq \Ass(M)$, we have $\grade(Q, M)=\cd(Q, M)$, as desired.
As first property of sequentially Cohen–Macaulay modules we have the following
\[regular\] Let $M$ be a finitely generated bigraded $S$-module with $\grade(Q, M)>0$ and $|K|=\infty$. If $M$ is sequentially Cohen–Macaulay with respect to $Q$, then there exists a bihomogeneous $M$-regular element $y\in Q$ of degree $(0,1)$ such that $M/yM$ is sequentially Cohen–Macaulay with respect to $Q$.
We assume that $M$ is sequentially Cohen–Macaulay with respect to $Q$ and let $\mathcal{F}$: $0=M_0\varsubsetneq M_1
\varsubsetneq \dots \varsubsetneq M_r=M$ be the Cohen–Macaulay filtration with respect to $Q$. Since $\grade(Q, M)=\grade(Q, M_1)=\cd(Q, M_1)>0$, it follows from that $\grade(Q, M_i/M_{i-1})=\cd(Q, M_i/M_{i-1})>0$ for all $i$. We set $N_i=M_i/M_{i-1}$. Thus by [@AR2 Corollary, 3.5] which also valid for finitely many modules which are Cohen–Macaulay with respect to $Q$ and have positive cohomological dimension with respect to $Q$, there exists a bihomogeneous element $y\in Q$ of degree $(0, 1)$ such that $y$ is $N_i$-regular for all $i$ and $\overline{N_i}$ is Cohen–Macaulay with respect to $Q$ with $\cd(Q, \overline{N_i})=\cd(Q, N_i)-1$. Here $\overline{L}=L/yL$ for any $S$-module $L$.
Consider the exact sequence $0 \rightarrow M_{i-1} \rightarrow M_i \rightarrow N_i \rightarrow 0$ for all $i$. Since $y$ is regular on $N_i$ for all $i$, it follows that $\Tor^S_1(S/yS, N_i)=0$ for all $i$. Hence, we get the following exact sequence $0 \rightarrow \overline{M_{i-1}} \rightarrow \overline{M_{i}} \rightarrow \overline{N_i} \rightarrow 0$ for all $i$. Now the filtration $\mathcal{G}$: $0=\overline{M_0} \varsubsetneq \overline{M_1}
\varsubsetneq \dots \varsubsetneq \overline{M_r}=M/yM$ is the Cohen–Macaulay filtration for $M/yM$ with respect to $Q$. In fact, $$\begin{aligned}
\grade(Q, \overline{M_i} /\overline{M_{i-1}}) &=&\grade(Q, \overline{N_i})\\
&=& \grade(Q, N_i)-1\\
&=& \cd(Q, N_i)-1\\
&=& \cd(Q, \overline{N_i})\\
&=& \cd(Q, \overline{M_i} /\overline{M_{i-1}}),\end{aligned}$$ and, since $\cd(Q, M_i/M_{i-1})<\cd(Q, M_{i+1}/M_{i})$ for all $i$, it follows that $\cd(Q,\overline{M_i} /\overline{M_{i-1}})<\cd(Q, \overline{M_{i+1}}/\overline{M_{i}})$ for all $i$.
The following example shows that the converse of the above proposition does not hold:
[*Consider the hypersurface ring $R=K[x_1, x_2, y_1, y_2]/(f)$ where $f=x_1y_1+x_2y_2.$ One has $\grade(Q, R)=1$ and $\cd(Q, R)=2.$ By [@AR2 Lemma 3.4] there exists a bihomogeneous $R$-regular element $y\in Q$ of degree $(0, 1)$ such that $\cd(Q, R/yR)=\cd(Q, R)-1=1$ and of course $\grade(Q, R/yR)=\grade(Q, R)-1=0$. So $R/yR$ is sequentially Cohen–Macaulay with respect to $Q$. On the other hand $R$ is not sequentially Cohen–Macaulay with respect to $Q$. Indeed, $\Ass(R)=\{(f)\}$ and $\cd(Q, R)=\cd(Q, S/(f))$ says that $R$ is relatively unmixed with respect to $Q$. If $R$ is sequentially Cohen–Macaulay with respect to $Q$, then by Lemma \[unmixed\], $R$ is Cohen–Macaulay with respect to $Q$, a contradiction.* ]{}
Let $M$ be a finitely generated bigraded $S$-module. The finiteness dimension of $M$ with respect to $Q$ is defined by $$f_Q(M)=\inf \{ i \in \NN: H^i_Q(M) \; \text {is not finitely
generated} \; \}.$$
Let the notation be as in Proposition \[reg\]. Suppose that $f_Q(M)=\cd(Q,M)$. Then $M$ is sequentially Cohen–Macaulay with respect to $Q$ with $\grade(Q, M)>0$ if and only if $M$ is Cohen–Macaulay with respect to $Q$.
Let $M$ be sequentially Cohen–Macaulay with respect to $Q$. Since $f_Q(M)=\cd(Q,M)$ with $\grade(Q, M)>0$, it follows from Proposition \[reg\] that $H^k_Q(M)=0$ for $k<\cd(Q,M)$. Hence $M$ is Cohen–Macaulay with respect to $Q$. The converse is trivial.
Let $R$ be a Noetherian ring, $I$ an ideal $R$ and $N$ an $R$-module. The module $N$ is said to be [*$I$-cofinite*]{} if $\Supp(N)\subseteq V(I)$ and $\Ext^i_R(R/I, N)$ is finitely generated for all $i$.
Let the assumptions and the notation be as in Proposition \[reg\]. Then $H^{k}_Q(M)$ is not finitely generated for all $k \in \{ q_1, \dots, q_r\}$ with $q_1>0$ while $H^{k}_Q(M)$ is $Q$-cofinite for all $k$.
The first statement is known. We will show that $H^{k}_Q(M)$ is $Q$-cofinite for all $k$. The claim is clear if $k \not \in \{ q_1,
\dots, q_r\}$ by Proposition \[reg\]. We first note that $\Supp(H^{q_i}_Q(M))\subseteq V(Q)$ for $i=1, \dots, r$. For a fix $i$, we consider the Grothendieck spectral sequence $$\Ext^{s}_S(S/Q,
H^{t}_Q(M_i/M_{i-1})) \underset{s} \Longrightarrow
\Ext^{s+t}_S(S/Q, M_i/M_{i-1}).$$ As $H^{t}_Q(M_i/M_{i-1})=0$ for $t\neq q_i$, by using Proposition \[reg\] we get the following isomorphism $$\Ext^{s}_S(S/Q, H^{q_i}_Q(M)) \iso \Ext^{s}_S(S/Q,
H^{q_i}_Q(M_i/M_{i-1}))\iso \Ext^{s+q_i}_S(S/Q, M_i/M_{i-1})$$ which shows that $\Ext^{s}_S(S/Q, H^{q_i}_Q(M))$ is finitely generated. Thus $H^{q_i}_Q(M)$ is $Q$-cofinite for all $i$. Therefore, $H^{k}_Q(M)$ is $Q$-cofinite for all $k$.
In [@JR] we have shown that if $M$ is Cohen–Macaulay with respect to $Q$ with $\cd(Q, M)=q$, then $\dim_S
H^q_Q(M)=\dim M-q$ and $\dim_{K[x]} H^q_Q(M)_j=\dim M-q$ for $j\ll0$ with $q>0$.
\[tame\] Let $|K|=\infty$. With the assumptions and the notation of Proposition \[reg\] we have
- $\dim H^{q_i}_Q(M)=\dim M_i/M_{i-1}-q_i$ for $i=1, \dots, r$.
- $\dim H^{q_i}_Q(M)_j=\dim M_i/M_{i-1}-q_i$ for $i=1, \dots, r$, $j\ll 0$ with $q_1>0$.
The assertion follows from Proposition \[reg\], [@JR Proposition 2.1] and [@JR Corollary 2.4].
Let $R$ be a graded ring. Recall that a graded $R$-module $N$ is [*tame*]{} if there exists an integer $j_0$ such that $N_j=0$ for all $j\leq j_0$, or else $N_j\neq 0$ for all $j\leq j_0$.
Let $|K|=\infty$. With the assumptions and the notation of Proposition \[reg\] we have $H^k_{Q}(M)$ is tame for all $k$.
The claim is obvious if $k\not \in \{q_1, \dots, q_r \}$ by Proposition \[reg\]. If $q_1=0$, then we have $H^0_{Q}(M)_j=0$ for $j\ll 0$ and if $q_1>0$, the assertion follows from Proposition \[tame\].
Let $M$ be a finitely generated bigraded $S$-module and set $K[x]=K[x_1, \dots, x_m]$. Then the local cohomology modules $H^i_{Q}(M)$ are naturally bigraded $S$-modules and each graded component $ H^i_{Q}(M)_j= H^i_{Q}(M)_{(*, j)}=\Dirsum_k H^i_{Q}(M)_{(k, j)}$ is finitely generated graded $K[x]$-module with grading $\big(H^i_{Q}(M)_j\big)_k=H^i_{Q}(M)_{(k,j)}$. Let $N$ be a finitely generated graded $K[x]$-module with graded minimal free resolution $$\mathbb{F}: 0 \rightarrow F_{k} \longrightarrow F_{k-1} \rightarrow
\cdots \rightarrow F_1 \longrightarrow F_0 \rightarrow 0.$$ The Castelnuovo–Mumford regularity of $N$ is the invariant $$\reg(N)=\max\{b_i(\mathbb{F})-i: i\geq 0\}$$ where $b_i(\mathbb{F})$ denotes the maximal degree of the generators of $F_i$. In [@AR2 Proposision 2.3] we have shown that if $M$ is Cohen–Macaulay with respect to $Q$ with $\cd(Q, M)=q$, then the function $f_{q, M}(j)=\reg H^q_{Q}(M)_j$ is bounded. For more results in this regard see [@AR1]. In the following we improve [@AR2 Proposision 2.3] as follows:
\[bound\] Let the assumptions and the notation be as in Proposition \[reg\]. Then for all $k$ the functions $f_{k, M}(j)=\reg H^k_{Q}(M)_j$ are bounded.
The claim is clear if $k\not \in \{ q_1, \dots, q_r\}$ by Proposition \[reg\]. Since $M_i/M_{i-1}$ is Cohen–Macaulay with respect to $Q$ for $i=1, \dots, r$, it follows from [@AR2 Proposision 2.3] that the functions $ f_{q_i, Mi/M_{i-1}}(j)=\reg
H^{q_i}_Q(M_i/M_{i-1})_j$ are bounded for $i=1, \dots, r$. Thus, $f_{q_i, M}(j)=\reg H^{q_i}_Q(M)_j$ are bounded for $i=1, \dots, r$ by Proposition \[reg\]. Therefore, $f_{k, M}(j)=\reg H^k_{Q}(M)_j$ are bounded for all $k$.
For a bigraded $S$-module $M$, we recall that the $i$th $a$-invariant of $M$ is defined by $$a^i_Q(M)=\sup \{ \mu: H_{Q}^i(M)_{(*,\mu)}\neq 0 \},$$ and so $\reg(M)=\underset{i}{\max} \{a^i_Q(M)+i: i\geq 0 \} $.
With the assumptions and the notation of Proposition \[reg\] we have $$\reg(M)=\max\{a^{q_i}_Q(M_i/M_{i-1})+q_i: i=1, \dots, r\}.$$
By Proposition \[reg\] we have $a^{q_i}_Q(M)=a^{q_i}_Q(M_i/M_{i-1})$ for $i=1, \dots, r$ and $a^{i}_Q(M)=-\infty$ for $i\not \in \{q_1, \dots, q_r \}$. Thus the result follows.
Sequentially Cohen–Macaulayness with respect to $P$, $Q$ and $P+Q$
==================================================================
In this section we discuss the relationship of sequentially Cohen–Macaulayness of $M$ with respect to $P$, $Q$ and $P+Q$. First we have the following characterization of sequentially Cohen–Macaulayness of $L\tensor_KN$ as $S$-module where $L$ and $N$ are two non-zero finitely generated graded modules over $K[x]$ and $K[y]$, respectively. For the bigraded $S$-module $M$ we define the bigraded Matlis-dual of $M$ to be $M^\vee$ where the $(i, j)$th bigraded components of $M^\vee$ is given by $\Hom_K(M_{(i, j)}, K).$
\[cmd\] Let $M$ be a finitely generated bigraded $S$-module. If $M$ is Cohen–Macaulay with respect to $Q$ with $\cd(Q, M)=q$, then $\big( H^{q}_Q(M)^\vee\big)_{(k, *)}$ is finitely generated Cohen–Macaulay $K[y]$-module of dimension $q$ for all $k$.
Note that $$\begin{aligned}
\big( H^{q}_Q(M)^\vee\big)_{(k, *)} &\iso & \big( H^{q}_Q(M)_{(-k, *)}\big)^\vee\\
&\iso & \big( H^{q}_{(y_1, \dots, y_n)}(M_{(-k, *)})\big)^\vee\\
&\iso & \Ext^{n-q}_{K[y]}(M_{(-k, *)}, S(-n)).\end{aligned}$$
Since $M$ is Cohen–Macaulay with respect to $Q$ with $\cd(Q, M)=q$, it follows from [@AR2 Proposition 1.2] that $M_{(-k, *)}$ is Cohen–Macaulay $K[y]$-module of dimension $q$. Therefore the conclusion follows.
\[joint\] Let $L$ and $N$ be two non-zero finitely generated graded modules over $K[x]$ and $K[y]$, respectively. We set $M=L\tensor_KN$. Then the following statements are equivalent:
- $M$ is sequentially Cohen–Macaulay $S$-module with respect to $Q$;
- $N$ is sequentially Cohen–Macaulay $K[y]$-module.
$(a)\Rightarrow (b)$: Let $\mathcal{F}$: $0=M_0\varsubsetneq M_1 \varsubsetneq \dots \varsubsetneq M_r=M$ be the Cohen–Macaulay filtration with respect to $Q$ . By Proposition \[reg\] we have $$H^{q_i}_Q(M)\iso H^{q_i}_Q(M_i)\iso H^{q_i}_Q(M_i/M_{i-1})$$ where $q_i=\cd(Q, M_i)=\cd(Q, M_i/M_{i-1})$ for $i=1, \dots, r$ and $H^{k}_Q(M)=0$ for $k\not
\in \{ q_1, \dots, q_r\}.$ Note that $
H^{q_i}_Q(M)\iso L\tensor_KH^{q_i}_Q(N)
$ for $i=1, \dots, r$, see the proof of [@AR2 Proposition 1.5]. Hence $H^{q_i}_Q(M)^\vee \iso L^\vee \tensor_KH^{q_i}_Q(N)^\vee$ where $(-)^\vee$ is the Matlis-dual, see [@HR Lemma 1.1]. Thus $$\begin{aligned}
\big(H^{q_i}_Q(M_i/M_{i-1})^\vee\big)_{(k, *)} &\iso& \big( H^{q_i}_Q(M)^\vee\big)_{(k, *)}\\
&\iso &(L^\vee)_k\tensor_KH^{q_i}_Q(N)^\vee \\
&\iso & \Ext^{n-q_i}_{K[y]}(N, K[y])^t,\end{aligned}$$ where $t=\dim_K(L^\vee)_k.$ Since each $M_i/M_{i-1}$ is Cohen–Macaulay with respect to $Q$ with $\cd(Q, M_i/M_{i-1})=q_i$, it follows from the above isomorphisms and Lemma \[cmd\] that $\Ext^{n-q_i}_{K[y]}(N, K[y])$ is Cohen–Macaulay of dimension $q_i$ for $i=1, \dots, r$. If $k\not\in \{ q_1, \dots, q_r\}$, then $ L\tensor_KH^{k}_Q(N) \iso H^{k}_Q(M)=0$. It follows that $H^{k}_Q(N)=0$, and hence $\Ext^{n-k}_{K[y]}(N, K[y])=0$ for $k\not\in \{ q_1, \dots, q_r\}$. Therefore the result follows from [@HS Theorem 1.4].
$(b)\Rightarrow (a)$: Let $N$ be sequentially Cohen–Macaulay $K[y]$-module with the Cohen–Macaulay filtration $0=N_0\varsubsetneq N_1 \varsubsetneq \dots \varsubsetneq N_r=N$. Consider the filtration $0=L\tensor_KN_0\varsubsetneq L\tensor_KN_1
\varsubsetneq \dots \varsubsetneq L\tensor_KN_r=L\tensor_KN$. We claim this filtration is the Cohen–Macaulay filtration with respect to $Q$. First, we note that $L\tensor_KN_i \varsubsetneq
L\tensor_KN_{i+1}$ for all $i$. Otherwise, we have $\dim N_i=\dim
N_{i+1}$ by [@STY Corollary 2.3], a contradiction. For all $k$ and $i$ we have the following isomorphisms $$\begin{aligned}
H^k_Q\big( (L\tensor_KN_i)/(L\tensor_KN_{i-1}) \big)&\iso& H^k_Q(
L\tensor_K(N_i/N_{i-1}) \big )\\ &\iso& L\tensor_K
H^k_Q(N_i/N_{i-1}).\end{aligned}$$ We set $D_i= (L\tensor_KN_i)/(L\tensor_KN_{i-1})$ for all $i$. Thus we have $\cd (Q, D_i)=\dim N_i/N_{i-1}$ for all $i$. This implies that $\cd\big (Q, D_{i-1})<\cd\big (Q, D_i)$ for all $i$. Also each quotient $D_i$ is Cohen–Macaulay with respect to $Q$ because $N_i/N_{i-1}$ is Cohen–Macaulay for all $i$.
If in Theorem \[joint\], we let $L$ be a sequentially Cohen–Macaulay $K[x]$-module and $N$ a non-sequentially Cohen–Macaulay $K[y]$-module, then $L\tensor_KN$ is sequentially Cohen–Macaulay with respect to $P$ and not sequentially Cohen–Macaulay with respect to $Q$. Thus sequentially Cohen–Macaulay with respect to $P$ does not imply sequentially Cohen–Macaulay with respect to $Q$ in general.
In [@AR2 Proposition 3.1] the relative Cohen–Macaulay property is characterized for Cohen–Macaulay modules as follows: if $M$ is a finitely generated bigraded Cohen–Macaulay $S$-module, then $M$ is Cohen–Macaulay with respect to $P$ if and only if $M$ is Cohen–Macaulay with respect to $Q$. Thus we may ask a similar question as follows:
[*\[qu1\] \[mix1\] Let $M$ be a finitely generated bigraded sequentially Cohen–Macaulay $S$-module. If $M$ is sequentially Cohen–Macaulay with respect to $P$. Is $M$ sequentially Cohen–Macaulay with respect to $Q$?*]{}
We have a positive answer to this question in the following cases:
\[tensor3\] The following statements hold:
- Let $M$ be a finitely generated bigraded $S$-module with $|K|=\infty$. Suppose $\cd(P,M)=0$. Then $M$ is sequentially Cohen–Macaulay if and only if $M$ is sequentially Cohen–Macaulay with respect to $Q$.
- Let the notation be as in Proposition \[joint\]. Then, $M$ is sequentially Cohen–Macaulay $S$-module if and only if $M$ is sequentially Cohen–Macaulay with respect to $P$ and $Q$.
For the proof (a), we first let $M$ is sequentially Cohen–Macaulay and consider the Cohen–Macaulay filtration $\mathcal{F}$: $0=M_0\varsubsetneq M_1
\varsubsetneq \dots \varsubsetneq M_r=M$. We show that $\mathcal{F}$ is Cohen–Macaulay filtration with respect to $Q$. Since $\cd(P,M)=0$, it follows from the exact sequences $0\rightarrow M_{i-1} \rightarrow M_i \rightarrow M_i/M_{i-1}
\rightarrow 0$ by using (4) that $\cd(P, M_i/M_{i-1})=0$ for all $i$. Hence by (3) we have $\cd(Q, M_i/M_{i-1})=\dim M_i/M_{i-1}$ for all $i$. This implies that $\cd(Q, M_i/M_{i-1})< \cd(Q,
M_{i+1}/M_{i})$ for all $i$. On the other hand, since $M_i/M_{i-1}$ is Cohen–Macaulay $S$-module for all $i$, it follows from (2) that $\grade(Q, M_i/M_{i-1})=\dim
M_i/M_{i-1}-\cd(P, M_i/M_{i-1})=\dim M_i/M_{i-1}$. Thus $M_i/M_{i-1}$ is relative Cohen–Macaulay with respect to $Q$. Therefore, $\mathcal{F}$ is a relative Cohen–Macaulay filtration with respect to $Q$. Conversely, let $\mathcal{G}$: $0=N_0\varsubsetneq N_1
\varsubsetneq \dots \varsubsetneq N_r=M$ be the Cohen–Macaulay filtration of $M$ with respect to $Q$. We will show that $\mathcal{G}$ is the Cohen–Macaulay filtration for $M$. Since $\cd(P, M)=0$, it follows that $\cd(P, N_i/N_{i-1})=0$ for all $i$. Thus $\cd(Q, N_i/N_{i-1})=\dim N_i/N_{i-1}$ for all $i$, by (3). This implies that $\dim N_i/N_{i-1}< \dim
N_{i+1}/N_{i}$ for all $i$. On the other hand, $\depth
N_i/N_{i-1}\geq \grade(Q, N_i/N_{i-1})=\cd(Q, N_i/N_{i-1})=\dim
N_i/N_{i-1}$ for all $i$. Thus $N_i/N_{i-1}$ is Cohen–Macaulay for all $i$. Therefore, $\mathcal{G}$ is the Cohen–Macaulay filtration for $M$.
For the proof (b), $M$ is sequentially Cohen–Macaulay $S$-module if and only if $L$ and $N$ are sequentially Cohen–Macaulay over $K[x]$ and $K[y]$, respectively, see [@STY Theorem 2.11], and this is equivalent to saying that $M$ is sequentially Cohen–Macaulay with respect to $P$ and $Q$ by Proposition \[joint\].
Even for a finitely generated bigraded Cohen–Macaulay $S$-module of dimension 2 for which $\grade(P, M)=0$ and $\cd(P, M)=1$, we do not know whether $M$ is sequentially Cohen–Macaulay with respect to $Q$.
Inspired by Question \[qu1\] we may ask the following question: Let $M$ be a finitely generated bigraded $S$-module that is sequentially Cohen–Macaulay with respect to $P$ and $Q$. Is $M$ sequentially Cohen–Macaulay? The following example shows that the answer is negative. see [@JR Example 1.2]. However, we have a positive answer to this question in Proposition \[tensor3\].
\[stanley ring\] [*Let $S=K[x_1,x_2,y_1,y_2]$ be the standard bigraded ring, $I=(x_1,y_1)\cap (x_2, y_2)$ and set $R=S/I$. The ring $R$ is unmixed but not Cohen–Macaulay. In fact, $\depth R=1$ and $\dim R=2$. One has that $R$ is Cohen–Macaulay with respect to $P$ and $Q$ with $\cd(P, R)=\cd(Q, R)=1$. In particular, $R$ is sequentially Cohen–Macaulay with respect to $P$ and $Q$. On the other hand, $R$ is not sequentially Cohen–Macaulay. Because by Lemma \[unmixed\], unmixed sequentially Cohen–Macaulay modules are Cohen–Macaulay.*]{}
Hypersurface rings which are sequentially Cohen–Macaulay with respect to $Q$
=============================================================================
Let $f\in S$ be a bihomogeneous element of degree $(a, b)$ and consider the hypersurface ring $R=S/fS$. We write $$f=\sum_{{| \alpha|=a}\atop {| \beta|=b}}c_{\alpha \beta }x^\alpha y^\beta \;\; where \;\; c_{\alpha \beta} \in K.$$ Notice that $R$ is a Cohen–Macaulay module of dimension $m+n-1.$ We have the following observations:
\[hyper\] Consider the hypersurface ring $R$. Then the following statements hold:
- If $a=0$ and $ b>0$, then $R$ is Cohen–Macaulay with respect to $P$ of $\cd(P, R)=m$ and Cohen–Macaulay with respect to $Q$ of $\cd(Q, R)=n-1.$
- If $a>0$ and $ b=0$, then $R$ is Cohen–Macaulay with respect to $P$ of $\cd(P, R)=m-1$ and Cohen–Macaulay with respect to $Q$ of $\cd(Q, R)=n.$
- If $ a>0$ and $b>0$, then $\grade(P, R)=m-1$ and $\cd(P, R)=m$, and $\grade(Q, R)=n-1$ and $\cd(Q, R)=n$.
In order to proof (a), if $a=0$, then we may write $f=\sum_{| \beta|=b} c_\beta y^\beta$. Using Formula (1) we have $\cd(P, R)=\dim S/(Q+ (f))=m$ and $\cd(Q, R)=\dim S/(P+(f))=n-1.$ On the other hand, by using Formula (2), we have $\grade(P, R)= \dim R-\cd(Q, R)=m+n-1-(n-1)=m$ and $\grade(Q, R)= \dim R-\cd(P, R)=m+n-1-m=n-1$. Thus the conclusions follows. Parts (b) and (c) are proved in the same way using Formulas (1) and (2).
Notice that if $a, b>0$, then $R$ is not Cohen–Macaulay with respect to $Q$. Thus it is natural to ask whether $R$ is sequentially Cohen–Macaulay with respect to $Q$. In the following, we classify all hypersurface rings that are sequentially Cohen–Macaulay with respect to $Q$. We need the following lemma for the proof of the main theorem
\[ass\] Let $M$ be a finitely generated bigraded $S$-module that is sequentially Cohen–Macaulay with respect to $Q$ with the Cohen–Macaulay filtration $0=M_0\varsubsetneq M_1
\varsubsetneq \dots \varsubsetneq M_r=M$. Then we have $\Ass(M_i/M_{i-1})=\{\pp \in \Ass(M_i): \cd(Q, S/\pp)=\cd(Q,
M_i)\}$ for $i=1, \dots, r.$ In particular, $\Ass(M_i/M_{i-1})=\Ass(M_i)$ if and only if $M_i$ is relatively unmixed with respect to $Q$.
We first let $\pp\in \Ass(M_i/M_{i-1})$. Since $M_i/M_{i-1}$ is Cohen–Macaulay with respect to $Q$, it follows from [@JR Corollary 1.11] that $M_i/M_{i-1}$ is relatively unmixed with respect to $Q$ and hence $\cd(Q, S/\pp)=\cd(Q,
M_i/M_{i-1})=\cd(Q, M_i).$ Thus we only need to show that $\pp\in
\Ass(M_i).$ As we always have $\Ass(M_i/M_{i-1})\subseteq \Ass(M_i)\cup
\Supp(M_{i-1})$, it suffices to show that $\pp\not\in
\Supp(M_{i-1})$. Assume $\pp\in \Supp(M_{i-1})$. Then by using (6) we have $\cd(Q, S/\pp)\leq \cd(Q, M_{i-1})< \cd(Q, M_i)$, a contradiction. Thus $\pp\not\in \Supp(M_{i-1})$ and hence $\pp\in \Ass(M_i)$. Now let $\pp\in \Ass(M_i)$ such that $\cd(Q, S/\pp)=\cd(Q, M_i)$. We will show that $\pp\in \Ass(M_i/M_{i-1})$. The exact sequence $0\rightarrow M_{i-1}\rightarrow M_i \rightarrow M_i/M_{i-1}$ yields $\Ass(M_i)\subseteq \Ass(M_{i-1})\cup\Ass(M_i/M_{i-1}).$ A similar argument as above by using (5) shows that $\pp\not\in \Ass(M_{i-1})$ and hence $\pp\in \Ass(M_i/M_{i-1})$.
\[hyp3\] Let $f\in S$ be a bihomogeneous element of degree $(a,
b)$ such that $f=h_1h_2$ where $h_1= \sum_
{\left|\alpha\right|=a}c_{\alpha} x^\alpha $ with $ c_{\alpha}\in K$ and $h_2= \sum_ {\left|\beta\right|=b}c_{\beta} y^\beta $ with $
c_{\beta}\in K$. In other words, $\deg h_1=(a, 0)$ and $\deg
h_2=(0,b)$. Consider the hypersurface ring $R=S/fS$. Then $R$ is sequentially Cohen–Macaulay with respect to $P$ and $Q$.
We show that $R$ is sequentially Cohen–Macaulay with respect to $P$. The argument for $Q$ is similar. Consider the filtration $\mathcal{F}$: $0=R_0\varsubsetneq R_1
\varsubsetneq R_2=R$ where $R_1=h_2S/fS$. We claim that this filtration is the Cohen–Macaulay filtration with respect to $P$. Observe that $R_2/R_{1}\iso S/h_2S$ is Cohen–Macaulay with respect to $P$ with $\cd(P, R_2/R_{1})=m$, by Lemma \[hyper\](a) . Now consider the map $\varphi$: $S\longrightarrow h_2S/fS$ where $g\longmapsto gh_2+fS$. We get the following isomorphism $S/h_1S\iso h_2S/fS\iso R_1/R_0$. Thus $R_1/R_0$ is Cohen–Macaulay with respect to $P$ with $\cd(P, R_1/R_0)=m-1$, by Lemma \[hyper\](b). Therefore, $\mathcal{F}$ is the Cohen–Macaulay filtration of $R$ with respect to $P$.
\[hypersurface\] Let $f\in S$ be a bihomogeneous element of degree $(a, b)$ and $R=S/fS$ be the hypersurface ring. Then the following statements are equivalent:
- $R$ is sequentially Cohen–Macaulay with respect to $Q$;
- $f=h_1h_2$ where $\deg h_1=(a, 0)$ with $a\geq 0$ and $\deg h_2=(0,b)$ with $b\geq 0$.
$(a)\Rightarrow (b)$: Let $f=f_1f_2\dots f_r$ be the unique factorization of $f$ into bihomogeneous irreducible factors $f_i$ with $\deg f_i=(a_i, b_i)$ for $i=1, \dots, r$. Note that $\sum_{i=1}^r a_i=a$ and $\sum_{i=1}^r b_i=b$. Our aim is to show that for each $f_i$ we have $\deg f_i=(a_i,
0)$ with $a_i\geq 0$ or $\deg f_i=(0 ,b_i)$ with $b_i\geq 0$. Assume this is not the case and so there exists $1\leq s \leq r$ such that $\deg f_s=(a_s ,b_s)$ with $a_s, b_s>0$. We first show that in the factorization of $f=f_1f_2\dots f_r$ some $f_i$ appear with $\deg f_i=(0, b_i)$. Otherwise, we may assume $\deg f_i=(a_i, 0)$ for $i=1, \dots, s-1$ and $\deg f_i=(a_i ,b_i)$ with $a_i,
b_i >0$ for $i=s, \dots, r$. Note that $(f)=(f_1) \cap
(f_2) \cap\dots \cap (f_r)$ and hence $\Ass(R)=\{ ( f_1 ), \dots, (
f_r ) \}$. For $i=1, \dots, s$ by using (1) and Lemma \[hyper\], we have $\cd(Q, S/( f_i
))=\dim S/(P+ ( f_i ) )=n=\cd(Q, R)$. Thus $R$ is relatively unmixed with respect to $Q$. As $R$ is sequentially Cohen–Macaulay with respect to $Q$, it follows that $R$ is Cohen–Macaulay with respect to $Q$ by Lemma \[unmixed\], a contradiction. Thus we may assume that $\deg f_i=(a_i, 0)$ with $a_i\geq 0$ for $i=1, \dots, s-1$, $\deg f_i=(a_i ,b_i)$ with $a_i, b_i>0$ for $i=s,s+1, \dots, t$ and $\deg f_i=(0 ,b_i)$ with $b_i\geq 0$ for $i=t+1, \dots, r$, and $t<r$. Now let $0=R_0 \varsubsetneq R_1 \varsubsetneq R_2=R$ be the Cohen–Macaulay filtration with respect to $Q$ where $R_1=I/(f)$ for some bihomogeneous ideal $I$ of $S$. Here we note that $r=2$ by Corollary \[seq\]. By Lemma \[ass\], $\Ass(S/I)=\Ass(R_2/R_1)=\{( f_1
), \dots, (f_t ) \}$. Thus $I=\qq_1 \cap \dots \cap
\qq_t$ with $\sqrt{\qq_i}=( f_i)$ for $i=1, \dots, t$ is a minimal primary decomposition of $I$. Notice that $I\subseteq \sqrt{I}=( g )$ where $g=\prod_{i=1}^t f_i$ with $\deg g=\big(\sum_{i=1}^ta_i,\sum_{i=1}^t b_i \big)$. If $I=(g)$, then as $\sum_{i=1}^ta_i>0,\sum_{i=1}^t b_i >0$ we have $\grade(Q, S/I)=n-1$ and $\cd(Q, S/I)=n$, by Lemma \[hyper\]. This contradicts with the fact that $S/I$ is Cohen–Macaulay with respect to $Q$. Thus $( f ) \subseteq I \varsubsetneq (
g)$ and so $( g )/I$ is a non-zero submodule of $S/I$. Since $S/I$ is Cohen–Macaulay with respect to $Q$, it follows from Lemma \[CM\] that $\cd(Q, ( g)/I)=\cd(Q,
S/I)=n.$ The map $\varphi$: $S \longrightarrow ( g )/I$ where $h \longmapsto hg+I$ induces the following exact sequence $
0\rightarrow U\rightarrow S/( f_{t+1}\cdots f_r )\rightarrow ( g
)/I \rightarrow 0
$, because $ (f_{t+1}\cdots f_r)g=f\in I$. By Formula (4) we have $\cd(Q,S/( f_{t+1}\cdots f_r ))=\max\{\cd(Q, U), \cd(Q, (g)/I)\}=n $. As $\deg f_i=(0, b_i)$ for $i=t+1, \dots, r$, by Lemma \[hyper\] it follows that $\cd(Q, S/( f_{t+1}\cdots f_r ))=n-1$, a contradiction. $(b) \Rightarrow (a)$: follows from Proposition \[hyp3\].
\[hyp\] Let $R$ be the hypersurface ring. Then $R$ is sequentially Cohen–Macaulay with respect to $Q$ if and only if $R$ is sequentially Cohen–Macaulay with respect to $P$.
In view of Corollary \[bound\] we have the following
Let $R$ be the hypersurface ring. If $R$ is sequentially Cohen–Macaulay with respect to $Q$, then $\reg H^i_Q(R)_j$ is bounded for $i=n-1, n$. In particular, $\reg H^i_Q(S/
gS)_j$ is bounded for $i=n-1, n$ where $( g )$ is a principal monomial ideal in $S$.
[*In [@AR1] we have shown that the bound for the regularity of graded components of the local cohomology of the hypersurface ring under some extra assumptions is linear.*]{}
[Acknowledgment]{}
The author would like to thank Jürgen Herzog for his helpful comments.
[99]{}
W. Bruns and J. Herzog, “Cohen–Macaulay rings” (Revised edition), Cambridge Studies in Advanced Mathematics [**39**]{}, Cambridge University Press, 1998. M. Chardin, J. P. Jouanolou, A. Rahimi, The eventual stability of depth, associated primes and cohomology of a graded module, J. Comm. Alg, to appear (arXiv:1203.4342). N. T. Cuong, L. T. Nhan, Pseudo Cohen–Macaulay and pseudo generalized Cohen–Macaulay modules, J. Algebra [**267**]{} (2003), 156–177. S. Faridi, Monomial ideals via square-free monomial ideals, Lecture Notes in Pure and Applied Mathematics, [**244**]{}(2005), 85–114. J. Herzog, A. Rahimi, Local duality for bigraded modules, Ill. J. Math, [**51**]{}(2007), 137–150.
J. Herzog, E. Sbarra, Sequentially Cohen–Macaulay modules and local cohomology. In: Tata Inst. Fund. Res. Stud. Math., vol. 16. pp. 327-340. M. Jahangiri, A. Rahimi, Relative Cohen–Macaulayness and relative unmixedness of bigraded modules, J. Comm. Alg, to appear (arXiv: 1105.3091). A. Rahimi, On the regularity of local cohomology of bigraded algebras, J. Algebra [**302**]{}(2006), 313-339. A. Rahimi, Relative Cohen-Macaulayness of bigraded modules, J. Algebra [**323**]{}(2010), 1745-1757. H. Sabzrou, M. Tousi and S. Yassemi, Simplicial join via tensor product, Manuscripta math [**126**]{} (2008), 255-272. P. Schenzel, On the dimension filtration and Cohen-Macaulay filtered modules, in Proc. of the Ferrara meeting in honour of Mario Fiorentini, University of Antwerp Wilrijk, Belgium (1998), Dekker, New York, 1999, pp. 245-264. MR 1702109 (2000i:13012). R. P. Stanley, Combinatorics and commutative algebra, Second edition, Birkhäuser Boston-Basel-Berlin, 1996. M. Tousi, S. Yassemi, Sequentially Cohen–Macaulay modules under base change, Comm. Algebra, [**33**]{}(2005), 3977-3987.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the 2.4$-$45 $\mu$m ISO-SWS spectrum of the Be star $\gamma$ Cas (B0.5 IVe). The spectrum is characterised by a thermal continuum which can be well fit by a power-law S$_{\nu}$ $\propto$ $\nu^{0.99}$ over the entire SWS wavelength range. For an isothermal disc of ionized gas with constant opening angle, this correponds to a density gradient $\rho$(r) $\propto$ r$^{-2.8}$. We report the detection of the Humphreys (6-$\infty$) bound-free jump in emission at 3.4 $\mu$m. The size of the jump is sensitive to the electron temperature of the gas in the disc, and we find T $\approx$ 9000 K, i.e. much lower than the stellar effective temperature (25000-30000 K). The spectrum is dominated by numerous emission lines, mostly from , but also some lines are detected. Several spectral features cannot be identified. The line strengths of the emission lines do not follow case B recombination line theory. The line strengths and widths suggest that many lines are optically thick and come from an inner, high density region with radius 3-5 R$_{*}$ and temperature above that of the bulk of the disc material. Only the $\alpha$, $\beta$ and $\gamma$ transitions of the series lines contain a contribution from the outer regions. The level populations deviate significantly from LTE and are highly influenced by the optically thick, local (disc) continuum radiation field. The inner disc may be rotating more rapidly than the stellar photosphere.'
author:
- 'S. Hony'
- 'L.B.F.M. Waters'
- 'P.A. Zaal'
- 'A. de Koter'
- 'J. M. Marlborough'
- 'C.E. Millar'
- 'N.R. Trams'
- 'P.W. Morris'
- 'Th. de Graauw'
bibliography:
- 'articles.bib'
date: 'Received date; accepted date'
title: 'The infrared spectrum of the Be star $\gamma$ Cassiopeiae [^1]'
---
Introduction
============
Be stars are rapidly rotating main sequence or giant stars that are characterized by the presence of (variable) H$\alpha$ emission, caused by high-density circumstellar gas. Often the H$\alpha$ line profile is double-peaked and its width correlates with the projected rotational velocity (v$\sin$i) of the underlying star . At infrared and radio wavelengths, the continuum energy distribution of Be stars is dominated by free-free and bound-free emission from the high-density circumstellar gas, which also causes the H$\alpha$ emission. Direct imaging of the circumstellar material at radio wavelengths [@1992Natur.359..808D] and in the H$\alpha$ line, shows that the gas is in a disc-like geometry. This flattened geometry is also evident from the optical continuum linear polarisation [e.g. @1979AJ.....84..812P] due to electron scattering.
The physical mechanism responsible for the disc-like geometry is related to the rapid rotation of the star. Models proposed in the literature include the wind compressed disc model [@1993ApJ...409..429B], viscocity-driven outflow in the equatorial regions [@1999iau..conf..okazaki] and non-radial pulsations [@1983ApJ...269..250V]. Which of these models is correct can be tested by investigating the density and kinematical structure of the disc. We use the infrared spectral region to explore the structure of Be star discs. At these wavelengths disc emission dominates the spectrum, and a rich hydrogen and helium emission line spectrum is available. The infrared lines probe the inner regions of the disc, and their strength and width are important diagnostic tools for the density and kinematical structure of the disc. Ground-based infrared spectra of $\gamma$ Cas were previously reported by e.g. , @1985ApJ...290..325L and @1987ApJ...318..356H. The Infrared Space Observatory (ISO) with its Short Wavelength Spectrometer (SWS) is very well suited to explore the infrared part of the spectrum of Be stars.
In this study we present a preliminary analysis of the ISO-SWS spectrum of one of the brightest and best studied Be stars in the sky, $\gamma$ Cas (B0.5IVe). We will show that the emission line spectrum of $\gamma$ Cas is not well represented by Menzel case B recombination line theory. Many line fluxes correlate with the local continuum and are independent of the intrinsic line strength (Einstein A coefficient). The observed line fluxes and widths suggest that these lines are formed in an inner region with well-determined size, and that only the intrinsically strongest lines have a contribution from outer layers. This paper is organized as follows. In Sect. 2 we briefly discuss the observations and data reduction. Section 3 discusses the continuum and Sect. 4 deals with the line spectrum. Section 5 discusses some implications of our measurements for the structure of the disc of $\gamma$ Cas.
Observations and data analysis
==============================
The Be star $\gamma$ Cas was observed with the SWS on board ISO on July 22nd, 1996, as part of the SWS guaranteed time programme BESTARS. A full spectral scan (2.4-45 $\mu$m) using Astronomical Observation Template (AOT) no. 1, speed 4 was obtained, while also several AOT02 line scans were taken. The observations were reduced using the SWS Interactive Analysis (IA$^3$) software package, with calibration files equivalent to pipeline version 7.0. Further processing consisted of bad data removal and rebinning on an equidistant wavelength grid. The flux levels are accurate to within 5 per cent for the wavelengths shortward of 7 $\mu$m. The observations between 7 and 12 $\mu$m (band 2C) suffer from memory effects; this has little influence on the measured line properties but does increase the uncertainty of the continuum flux level to 15 per cent. At even longer wavelengths the signal to noise ratio decreases dramatically and only the strongest lines can be measured with reasonable accuracy. Most of the emission lines are partially resolved with a ratio of FWHM to the FWHM of the instrumental profile between 1.4-3.5. Only 6 of the emission lines are considered unresolved having this ratio below 1.4. Since the SWS instrumental profile is approximately Gaussian and the observed lines are well fitted by Gaussians, we estimate the original line width from: $${w_\mathrm{obs}}^2={w_\mathrm{org}}^2+{w_\mathrm{inst}}^2,
\label{eqn:FWHM}$$ where $w_\mathrm{obs}$ is the observed FWHM, $w_\mathrm{org}$ is the original FWHM and $w_\mathrm{inst}$ is the FWHM of the instrumental profile. The latter value varies with wavelength. To determine $w_\mathrm{inst}$, we use measured line widths of emission lines of planetary nebulae; and , observed in the same observing mode. No significant line profile variations are observed. We show the final AOT01 spectrum in Fig. \[fig:swsspec\].
The continuum energy distribution
=================================
The continuum energy distribution of $\gamma$ Cas at IR wavelengths is dominated by free-free and bound-free emission from the ionized part of the circumstellar gas . The stellar contribution to the total flux is about 20 per cent at 2.4 $\mu$m, based on extrapolation of a Kurucz model atmosphere fitted to the UV continuum . The spectrum can be well represented by a single power-law, S$_{\nu}$ $\propto$ $\nu^{\alpha}$, with $\alpha$ = 0.99 $\pm$ 0.05. This spectral slope is slightly, but significantly flatter, than that derived by , based on IRAS broad-band photometry taken in 1983. We have used the simple isothermal disc model of to estimate the radial density gradient in the disc, assuming a power-law $\rho$(r) = $\rho_{0}$(r/R$_*$)$^{-n}$, and find n = 2.8 $\pm$ 0.1. The value of $\rho_0$ depends on the assumed opening angle $\theta$ of the disc, as well as on the stellar radius and disc temperature. We use R$_*$ = 10 R$_{\sun}$ and T$_\mathrm{disc}$ = 10$^{4}$ K (see below). Analysis of the optical linear polarisation and interferometric imaging of [@1997ApJ...477..926W] suggests a half opening angle of 2.5$\degr$. We use a 1$\degr$ half opening angle. The derived density at the stellar surface is $\rho_0$ = 3.5$\times$10$^{-11}$ g cm$^{-3}$; an emission measure EM = 1.5$\times$10$^{61}$ cm$^{-3}$ was found.
There are some wavelength ranges that show a deviation from the power-law behavior of the continuum discussed above. Near 3.28 $\mu$m the merging of the emission lines of the Humphreys series, with lower quantum level $n=6$, causes a jump. This Humphreys jump (seen in emission), which is similar to the Balmer jump at optical wavelengths, can be used to derive the average electron temperature of the emitting region. The difference in flux on both sides of the Humphreys jump is caused by a discontinuity in bound-free ($\kappa_\mathrm{ff+bf}$) opacity of the gas in the disc. We write for the total continuum opacity. $$\begin{aligned}
\kappa_\mathrm{ff+bf} & \propto & \lambda^{2} \times (1-e^{-ch
/\lambda kT}) / (ch/\lambda kT) \times T^{-3/2} \times \nonumber \\
& & \{g(\lambda ,T) + b(\lambda ,T)\},
\label{eqn:totalkappa}\end{aligned}$$ where g($\lambda$,T) and b($\lambda$,T) are the free-free and bound-free gaunt factors, respectively. b($\lambda$,T) is a sensitive function of the temperature: the jump in b($\lambda$,T) (and in flux) increases towards lower electron temperature. The change in g($\lambda$ ,T) is negligible over the wavelength range of interest. The jump in b($\lambda$,T) is thus a diagnostic of the temperature in the disc. We use the following method to determine the size of this jump: We define the normalized excess flux as Z$_{\lambda}-$1 = (F$_{\lambda}$-F$_{\lambda,*}$)/F$_{\lambda,*}$, where F$_{\lambda,*}$ is the stellar photospheric flux, see Fig. \[fig:humpschema\]. Z$_{\lambda}-$1 is normalized to the source function of the gas in the disc modulo a constant since both the disc and the star radiate in the Rayleigh-Jeans limit in this wavelength regime and thus have the same wavelength dependence. On the blue side of the discontinuity b($\lambda$,T) has a certain value, with a corresponding value of $\kappa_\mathrm{ff+bf}$, of $\tau_\mathrm{ff+bf}$ for each line of sight through the disc and thus a corresponding value of Z$_{\lambda}-$1. Beyond the discontinuity there is a drop in b($\lambda$,T), $\kappa_\mathrm{ff+bf}$, $\tau_\mathrm{ff+bf}$ and Z$_{\lambda}-$1. Since there is a wide range of $\tau_\mathrm{ff+bf}$ for different lines of sight Z$_{\lambda}-$1 is [*not*]{} a simple function of b($\lambda$,T). However $\kappa_\mathrm{ff+bf}$ steadily increases with wavelength, e.g. Eqn. \[eqn:totalkappa\], and thus there is a wavelength ($\lambda^{\prime}$) where the loss in b($\lambda$,T) is compensated by the increase in $\lambda$. At $\lambda^{\prime}$ the $\kappa_\mathrm{ff+bf}$ is equal to the previous value, so $\tau_\mathrm{ff+bf}$ and Z$_{\lambda}-$1 are also equal. We can determine $\lambda^{\prime}$ directly from the observations, e.g. Fig. \[fig:humpschema\].
Using $\lambda^{\prime}~=~3.470 \pm 0.005~\mu {\mathrm m}$ and the gaunt factors calculated by we find an electron temperature in the disc of 9500 $\pm$ 1000 K. This temperature would cause a weak but measurable jump at the Hansen-Strong series limit (near 4.5 $\mu$m) but none is observed. However, we note that near this wavelength two strong lines are present that may mask an otherwise detectable jump. The disc temperature agrees well with a density-weighted temperature of 10700 K derived by @1998ApJ...494..715M from an energy balance calculation using the Poeckert & Marlborough model for the disc of $\gamma$ Cas. Note that the inner regions of the disc may have considerably higher temperatures, because this method is insensitive to contributions of those parts of the highest density parts of the disc where the continuum is optically thick. (see also Sect. 4). A second region which deviates from the power-law is near 4.3 $\mu$m. We cannot find a reasonable identification for this spectral feature. The CO$_2$ stretch mode band is at 4.27 $\mu$m. However, the interstellar extinction towards $\gamma$ Cas is very low which implies that we must rule out this possibility.
[r @- l r l l l||r @- l r l l l]{} & & & & & & & & &\
& & & & & & & & &\
&&&& & & &&& & &\
5&4 & 4.052 & 254 $\pm$ 42 & 5.45e-14 $\pm$ 6%& 61$\pm$ 6%& 10&7 & 8.761 & 249 $\pm$ 39 & 3.70e-15 $\pm$10 % & 24$\pm$10%\
6&4 & 2.626 & 226 $\pm$ 56 & 7.14e-14 $\pm$ 5 & 91$\pm$ 5 & 12&7 & 6.772 & 364 $\pm$ 40 & 8.21e-15 $\pm$ 8 & 35$\pm$ 8\
6&5 & 7.460 & 220 $\pm$ 65 & 9.51e-15 $\pm$ 8 & 28$\pm$ 8 & 13&7 & 6.292 & 414 $\pm$ 41 & 1.02e-14 $\pm$ 7 & 37$\pm$ 7\
7&5 & 4.654 & 200 $\pm$ 43 & 1.58e-14 $\pm$ 6 & 56$\pm$ 6 & 14&7 & 5.957 & 303 $\pm$ 64 & 9.11e-15 $\pm$ 7 & 40$\pm$ 7\
8&5 & 3.741 & 267 $\pm$ 48 & 2.69e-14 $\pm$ 6 & 64$\pm$ 5 & 15&7 & 5.712 & 376 $\pm$ 56 & 1.12e-14 $\pm$ 7 & 42$\pm$ 7\
9&5 & 3.297 & 284 $\pm$ 23 & 3.33e-14 $\pm$ 5 & 77$\pm$ 5 & 16&7 & 5.525 & 393 $\pm$ 13 & 1.34e-14 $\pm$ 7 & 45$\pm$ 6\
10&5 & 3.039 & 273 $\pm$ 32 & 3.73e-14 $\pm$ 5 & 80$\pm$ 5 & 17&7 & 5.380 & 367 $\pm$ 16 & 1.21e-14 $\pm$ 7 & 46$\pm$ 6\
11&5 & 2.873 & 309 $\pm$ 33 & 4.53e-14 $\pm$ 5 & 84$\pm$ 5 & 18&7 & 5.264 & 335 $\pm$ 18 & 1.13e-14 $\pm$ 7 & 47$\pm$ 6\
12&5 & 2.758 & 313 $\pm$ 36 & 4.73e-14 $\pm$ 5 & 86$\pm$ 5 & 19&7 & 5.169 & 390 $\pm$ 16 & 1.09e-14 $\pm$ 7 & 48$\pm$ 6\
13&5 & 2.675 & 324 $\pm$ 37 & 5.37e-14 $\pm$ 5 & 88$\pm$ 5 & 20&7 & 5.091 & 396 $\pm$ 17 & 1.16e-14 $\pm$ 6 & 49$\pm$ 6\
14&5 & 2.613 & 347 $\pm$ 12 & 5.20e-14 $\pm$ 5 & 95$\pm$ 5 & 21&7 & 5.026 & 457 $\pm$ 15 & 1.11e-14 $\pm$ 6 & 49$\pm$ 6\
15&5 & 2.564 & 340 $\pm$ 13 & 5.38e-14 $\pm$ 5 & 95$\pm$ 5 & 22&7 & 4.971 & 441 $\pm$ 16 & 9.04e-15 $\pm$ 7 & 50$\pm$ 6\
16&5 & 2.526 & 344 $\pm$ 14 & 5.68e-14 $\pm$ 5 & 96$\pm$ 5 & 23&7 & 4.923 & 425 $\pm$ 17 & 8.34e-15 $\pm$ 7 & 50$\pm$ 6\
17&5 & 2.495 & 379 $\pm$ 13 & 5.53e-14 $\pm$ 5 & 97$\pm$ 5 & 24&7 & 4.883 & 551 $\pm$ 13 & 8.86e-15 $\pm$ 7 & 51$\pm$ 6\
18&5 & 2.470 & 407 $\pm$ 12 & 5.66e-14 $\pm$ 5 & 98$\pm$ 5 & 25&7 & 4.847 & 441 $\pm$ 17 & 6.69e-15 $\pm$ 7 & 51$\pm$ 6\
19&5 & 2.449 & 386 $\pm$ 13 & 5.17e-14 $\pm$ 5 & 98$\pm$ 5 & 26&7 & 4.817 & 420 $\pm$ 18 & 5.18e-15 $\pm$ 8 & 52$\pm$ 6\
20&5 & 2.431 & 390 $\pm$ 13 & 4.83e-14 $\pm$ 5 & 99$\pm$ 5 & 27&7 & 4.789 & 455 $\pm$ 17 & 4.39e-15 $\pm$ 8 & 52$\pm$ 6\
21&5 & 2.416 & 447 $\pm$ 12 & 4.79e-14 $\pm$ 5 & 99$\pm$ 5 & 28&7 & 4.765 & 551 $\pm$ 14 & 5.45e-15 $\pm$ 8 & 52$\pm$ 6\
22&5 & 2.404 & 417 $\pm$ 13 & 4.24e-14 $\pm$ 5 & 100$\pm$ 5 & 29&7 & 4.744 & 706 $\pm$ 11 & 5.59e-15 $\pm$ 8 & 52$\pm$ 6\
23&5 & 2.393 & 414 $\pm$ 13 & 3.20e-14 $\pm$ 5 & 100$\pm$ 5 & 12&8 & 10.504 & 282 $\pm$ 22 & 2.88e-15 $\pm$13 & 20$\pm$12\
24&5 & 2.383 & 491 $\pm$ 11 & 4.03e-14 $\pm$ 5 & 101$\pm$ 5 & 13&8 & 9.393 & 204 $\pm$ 41 & 2.97e-15 $\pm$11 & 23$\pm$11\
7&6 & 12.371 & 313 $\pm$ 13 & 4.34e-15 $\pm$15 & 15$\pm$15 & 14&8 & 8.666 & 339 $\pm$ 29 & 4.40e-15 $\pm$10 & 24$\pm$10\
9&6 & 5.908 & 295 $\pm$ 67 & 8.81e-15 $\pm$ 7 & 41$\pm$ 7 & 15&8 & 8.156 & 345 $\pm$ 33 & 4.69e-15 $\pm$10 & 26$\pm$ 9\
10&6 & 5.129 & 325 $\pm$ 20 & 1.46e-14 $\pm$ 6 & 48$\pm$ 6 & 16&8 & 7.781 & 291 $\pm$ 44 & 4.87e-15 $\pm$ 9 & 27$\pm$ 9\
11&6 & 4.673 & 304 $\pm$ 27 & 1.57e-14 $\pm$ 6 & 56$\pm$ 6 & 18&8 & 7.272 & 252 $\pm$ 60 & 4.24e-15 $\pm$ 9 & 29$\pm$ 8\
12&6 & 4.377 & 326 $\pm$ 30 & 1.98e-14 $\pm$ 6 & 57$\pm$ 6 & 19&8 & 7.093 & 321 $\pm$ 49 & 4.83e-15 $\pm$ 9 & 30$\pm$ 8\
13&6 & 4.171 & 348 $\pm$ 32 & 2.33e-14 $\pm$ 6 & 61$\pm$ 6 & 20&8 & 6.947 & 484 $\pm$ 28 & 5.67e-15 $\pm$ 8 & 34$\pm$ 8\
14&6 & 4.021 & 304 $\pm$ 35 & 2.52e-14 $\pm$ 6 & 60$\pm$ 6 & 21&8 & 6.827 & 445 $\pm$ 32 & 4.48e-15 $\pm$ 9 & 35$\pm$ 8\
15&6 & 3.908 & 332 $\pm$ 35 & 2.73e-14 $\pm$ 6 & 62$\pm$ 5 & 22&8 & 6.724 & 376 $\pm$ 39 & 3.67e-15 $\pm$ 9 & 36$\pm$ 8\
16&6 & 3.820 & 350 $\pm$ 35 & 2.77e-14 $\pm$ 6 & 63$\pm$ 5 & 23&8 & 6.639 & 421 $\pm$ 36 & 3.47e-15 $\pm$ 9 & 36$\pm$ 7\
17&6 & 3.750 & 360 $\pm$ 35 & 2.68e-14 $\pm$ 6 & 64$\pm$ 5 & 24&8 & 6.566 & 528 $\pm$ 29 & 3.82e-15 $\pm$ 9 & 35$\pm$ 7\
18&6 & 3.693 & 372 $\pm$ 35 & 2.63e-14 $\pm$ 6 & 65$\pm$ 5 & 26&8 & 6.447 & 671 $\pm$ 24 & 4.53e-15 $\pm$ 8 & 36$\pm$ 7\
19&6 & 3.646 & 392 $\pm$ 34 & 2.54e-14 $\pm$ 6 & 65$\pm$ 5 & 15&9 & 11.539 & 275 $\pm$ 18 & 2.71e-15 $\pm$14 & 17$\pm$14\
20&6 & 3.607 & 344 $\pm$ 40 & 2.08e-14 $\pm$ 6 & 66$\pm$ 5 & 16&9 & 10.805 & 301 $\pm$ 19 & 2.26e-15 $\pm$14 & 20$\pm$13\
21&6 & 3.574 & 387 $\pm$ 36 & 2.06e-14 $\pm$ 6 & 67$\pm$ 5 & 17&9 & 10.261 & 431 $\pm$ 15 & 4.07e-15 $\pm$12 & 21$\pm$12\
22&6 & 3.546 & 380 $\pm$ 38 & 1.95e-14 $\pm$ 6 & 67$\pm$ 5 & 18&9 & 9.848 & 292 $\pm$ 25 & 2.36e-15 $\pm$12 & 22$\pm$11\
23&6 & 3.522 & 408 $\pm$ 36 & 1.69e-14 $\pm$ 6 & 67$\pm$ 5 & 19&9 & 9.522 & 337 $\pm$ 24 & 2.21e-15 $\pm$12 & 23$\pm$11\
24&6 & 3.501 & 466 $\pm$ 32 & 1.75e-14 $\pm$ 6 & 68$\pm$ 5 & 20&9 & 9.261 & 364 $\pm$ 23 & 2.18e-15 $\pm$12 & 22$\pm$10\
25&6 & 3.483 & 476 $\pm$ 31 & 1.49e-14 $\pm$ 6 & 68$\pm$ 5 & 21&9 & 9.047 & 283 $\pm$ 32 & 2.05e-15 $\pm$12 & 23$\pm$10\
26&6 & 3.467 & 438 $\pm$ 34 & 1.18e-14 $\pm$ 6 & 68$\pm$ 5 & 22&9 & 8.870 & 422 $\pm$ 22 & 1.96e-15 $\pm$12 & 24$\pm$10\
27&6 & 3.452 & 621 $\pm$ 24 & 1.40e-14 $\pm$ 6 & 69$\pm$ 5 & 23&9 & 8.721 & 416 $\pm$ 24 & 1.88e-15 $\pm$12 & 24$\pm$10\
28&6 & 3.440 & 639 $\pm$ 24 & 1.29e-14 $\pm$ 6 & 69$\pm$ 5 & 24&9 & 8.594 & 375 $\pm$ 27 & 1.82e-15 $\pm$12 & 25$\pm$10\
29&6 & 3.429 & 641 $\pm$ 24 & 1.08e-14 $\pm$ 6 & 69$\pm$ 5 & 25&9 & 8.485 & 430 $\pm$ 24 & 1.36e-15 $\pm$14 & 25$\pm$ 9\
9&7 & 11.309 & 394 $\pm$ 13 & 2.96e-15 $\pm$14 & 17$\pm$13 & 20&10 & 12.156 & 439 $\pm$ 9 & 1.91e-15 $\pm$16 & 16$\pm$15\
\[tab:hlines\]
The emission lines
==================
The entire SWS spectral region, but especially bands 1 and 2 (2.4-12 $\mu$m), is dominated by strong and partially resolved emission lines. The vast majority of these lines are recombination lines. We find the series limit of the Humphreys and Hansen-Strong series, as well as those from lower levels 8, 9 and 10. Several lines with lower levels above 10 were also identified. We find a few lines (notably at 2.4861, 2.5729, 4.2960 and 4.0367 $\mu$m). No forbidden lines could be found, although several emission features are unidentified and could perhaps be attributed to forbidden lines. The strongest unidentified line is at 2.8934 $\mu$m, close to the 2.8964 \[\] line. The lack of fine-structure lines is consistent with the optical spectrum of $\gamma$ Cas. It is likely that the high density of the ionized gas causes collisional de-excitation of the fine-structure transitions. This situation is markedly different for the hypergiant , whose infrared spectrum is also dominated by recombination lines from circumstellar gas, but which also shows prominent emission from e.g. \[\], \[\] and \[\] in its ISO-SWS spectrum .
We have measured the line fluxes with respect to both the [*local*]{} continuum and the [*stellar*]{} photospheric continuum at line centre. The measured line properties are given in Table \[tab:hlines\]. In Fig. \[fig:coglines\] we show the resulting curve of growth for the emission lines, where we plot the equivalent width EW divided by wavelength versus log X$_\mathrm{line}$. This last quantity is proportional to the line optical depth . The observed line fluxes (Fig. \[fig:caseB\]) are all much smaller (a factor 10 or more) than expected on the basis of the continuum emission measure determination and the optically thin case B recombination line predictions [@1987MNRAS.224..801H]. This shows that even for the weakest lines the emission is optically thick: the combined line and continuum opacity is much larger than 1. This is not surprising given the shape of the continuum energy distribution; show that the continuum becomes optically thick near 0.8 $\mu$m.
The shape of the empirical curve of growth is very different from that expected on the basis of LTE line formation in a rotating, partially optically thick disc. Model calculations show that under these conditions the curve of growth has a linear part (where line emission is optically thin and proportional to line strength), and a power-law part whose slope depends on the radial density gradient of the gas , see also @1952ZA.....30...96W. The observed curve of growth however shows a very steep rise of EW/$\lambda$ with X$_\mathrm{line}$ up to log X$_\mathrm{line}$ $\approx$ $-$36.5, followed by a flat part for $-$36.5 $<$ log X$_\mathrm{line}$ $<-$35.0 whose slope is close to zero, i.e. much smaller than 0.4 expected on the basis of the density gradient derived from the continuum free-free and bound-free excess. Finally, a rising part for the strongest lines in the spectrum (log X$_\mathrm{line}$ $>-$35.0) is seen. The flat part of the curve of growth is surprising and suggests that these lines have saturated and are formed in a region with a well-defined outer radius. We note that the size of this region is not the same for every series, but increases with lower quantum level.
In order to understand better the nature of the line formation in $\gamma$ Cas, we show in Fig. \[fig:vel\] the line strength $\overline{EW}$/$\lambda$, where $\overline{EW}$ is the line equivalent width measured with respect to the [*local*]{} continuum, as a function of wavelength. Also shown in Fig. \[fig:vel\] is the line FWHM versus wavelength. Both $\overline{EW}$/$\lambda$ and the FWHM show a characteristic wavelength dependence: for each series, $\overline{EW}$/$\lambda$ increases and then saturates to a value of 5-6$\times$10$^{-4}$, while the line FWHM decreases and reaches a roughly constant value of 250-300 km s$^{-1}$. In other words, the lines reach a constant line over continuum ratio of about 1.3 at a constant FWHM of 250-300 km s$^{-1}$. Only the $\alpha$ and $\beta$ lines of each series deviate from this behaviour: their line over continuum ratio is larger and their FWHM smaller.
The decrease of FWHM with increasing line strength is expected for lines formed in a rotating disc in which the rotational velocity decreases with distance; as more line flux is coming from the outer, more slowly rotating regions, the line width will decrease. This is direct proof of the rotating nature of the line emitting region. It is remarkable that the weakest lines of each series have a very high FWHM of more than 550 km s$^{-1}$. Such high velocities are not expected given the photospheric v$\sin$i of 230 km s$^{-1}$ [@1982ApJS...50...55S]. This suggests that the inner disc is rotating more rapidly than the star. However, line broadening due to electron scattering could also cause these large line widths, but we do not observe prominent electron scattering wings.
If we assume that the line and continuum source functions are equal, no line emission from the layers with $\tau_\mathrm{ff}$ $>$ 1 should be detectable. However, the large line width of the weakest lines strongly suggests that this line emission is originating from rapidly rotating parts of the disc. The fact that the line FWHM decreases with increasing line strength shows that the broad, weak lines are formed closest to the star. In the 2.5 to 7 $\mu$m wavelength region, where most of the weakest lines in Fig. \[fig:vel\] are located, the continuum is optically thick out to a radius of 2.5 to 3 R$_{*}$. Assuming Keplerian rotation in the disc, the part of the disc which is not optically thick for continuum radiation rotates at projected speeds less than about v$_{0}$$\sin$i/1.7, where v$_0$ is the Keplerian velocity of the disc at the stellar surface. Using reasonable values for the mass and radius for a B0.5 IV star of R$_{*}$=10R$_\odot$ and M$_{*}$=15M$_\odot$ , we find a maximum speed of about 220 km s$^{-1}$. Emission lines in the disc whose source function is equal to that of the continuum therefore should have *FWZI* less than 440 km s$^{-1}$, but even the FWHM of the weak lines is significantly higher than this. It is unlikely that rotational velocities are as high as 250-300 km s$^{-1}$ as far out as 2.5-3 R$_{*}$, unless the rotational velocity field deviates significantly from Keplerian. We conclude that we detect line emission from those parts of the disc that are optically thick for continuum radiation, and hence the line source function in these inner regions must exceed that of the continuum (which is the Planck function at the local electron temperature). Several effects can cause such a larger source function: NLTE level populations, or an enhanced electron temperature near the upper part of the disc. If the latter effect is important, temperatures in the line forming regions may be as much as 30 per cent higher than in the bulk of the disc, since the stronger lines are 1.3 times the continuum.
As pointed out above, the observed line fluxes of every series first increase with intrinsic line strength, but then reach a constant value with respect to the local continuum, both in $\overline{EW}$/$\lambda$ and in the line to continuum ratio. This effect occurs for transitions from upper levels roughly between 15 and about 10, to lower levels between 9 and 5. The line fluxes are no longer determined by the intrinsic line strength, but by [*the local continuum!*]{} The contribution from outer regions, with lower rotational velocities, is absent. This can be seen from the FWHM’s of the lines with constant line over continuum ratio, that no longer decrease but are constant at about 250-300 km s$^{-1}$. This could be due to a finite outer radius of the disc. However, the ISO-SWS spectrum as well as ISO-PHOT photometry at 90 $\mu$m (Trams et al., in preparation) show that there is no change in continuum slope over a wide wavelength range. Such a change in slope would be expected in the case of a finite outer radius. We conclude that the lack of line emission from the outer regions cannot be due to a finite outer disc radius. We note that the fact that the line to continuum ratio is constant for certain lines, irrespective of the strength of the local continuum and thus of the size of the continuum emitting region, also argues against a finite outer radius.
The lack of line flux from the outer regions therefore must be due to a change (decrease) in the line source function compared to the inner regions. For a certain range of intermediate energy levels in the atom, the optically thick IR free-free disc radiation field in the inner regions is able to maintain significant level populations, but these levels rapidly become de-populated when the local free-free continuum becomes optically thin. Therefore the line flux does not increase with intrinsic line strength (as do the weaker lines from high upper levels), but reaches a line over continuum ratio determined by the ratio of source functions of line and continuum.
Discussion
==========
@1987ApJ...318..356H discuss the near-IR spectrum of $\gamma$ Cas and note that the line flux ratios of the weak Pfund lines to the Br$\gamma$ line are very difficult to reconcile with those of Br$\gamma$ and Br$\alpha$ for any reasonable choice of radial density gradient in the disc. We confirm this result and show that it is a general property of the line fluxes of $\gamma$ Cas to behave in a highly non-standard way, probably caused by a combination of non-LTE and temperature effects. We do not confirm the lack of correlation between line strength and line width noted by @1987ApJ...318..356H; the ISO data clearly show such a correlation for all series in our spectrum. This may be due to the Earth atmosphere affecting the estimate of the weak high level Pfund series lines in the spectrum of Hamann & Simon. The line widths of their stronger lines (e.g. Br$\alpha$) agree well with those of our spectrum. They also agree well with the line widths derived from high-resolution spectra published by @1985ApJ...290..325L and .
The FWHM of the weakest lines (more than 500 km s$^{-1}$) significantly exceeds that of the photospheric v$\sin$i of 230 km s$^{-1}$. While electron scattering may affect the width of the strongest lines (especially at short wavelengths where free-free opacities are small compared to the electron scattering opacity) the IR lines do not show evidence for prominent electron scattering wings. This suggests that the line width is mainly due to kinematical broadening, which would imply that the disc is rotating more rapidly than the underlying star (ignoring errors in v$\sin$i). This can only occur if some mechanism adds angular momentum to the material injected into the disc. We note that found line widths of weak emission lines in the optical spectra of Be stars that also significantly exceeded the width expected on the basis of the observed v$\sin$i. Transfer of angular momentum may lead to spin-down of the star . Using the formalism given by , and assuming a full opening angle of 2 degrees and $\rho_0$ of 3$\times$10$^{-11}$ g cm$^{-3}$, the outflow velocity in the disc near the star can be of the order of 1 km s$^{-1}$ without significant spin-down of the star during its main sequence life time. Such a value is in agreement with the observed line shape.
The picture which emerges from our analysis of the infrared spectrum of $\gamma$ Cas is that of a circumstellar region of very high density, perhaps exceeding 3 10$^{-11}$ g cm$^{-3}$, which is rotating rapidly and whose rotational velocity decreases with distance. The rotational velocity in the disc near the star exceeds that of the photosphere. This region is heated by radiation from the central star, and the surface layers which directly absorb the stellar radiation field have higher temperatures than regions closer to the equatorial plane of the disc. While the weak lines originate from these dense, warm regions, only the strongest, $\alpha$ and $\beta$, lines of each series probe the outer portions of the disc.
LBFMW and AdK acknowledge financial support from an NWO [*Pionier*]{} grant. JMM acknowledges support from an NSERC grant. JMM and LBFMW acknowledge financial support from a NATO Collaborative Resarch Grant (CRG.941220). This work was supported by NWO Spinoza grant 08-0 to E.P.J. van den Heuvel. CEM acknowledges financial support from an NSERC postgraduate scholarship.
[^1]: based on observations obtained with ISO, an ESA project with instruments funded by ESA Member states (especially the PI countries: France, Germany, the Netherlands and the United Kingdom) with the participation of ISAS and NASA
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We explore the transient dynamics associated with the emergence of the classical signal in the full quantum system. We start our study from the instability which promotes the squeezing of the quantum system. This is often interpreted as the particle production though being reversible in time. We associate this state a non-dissipative classical fluctuations and study their trigger to develop the coherent state which can be classical if sufficiently developed. The Schwinger-Keldysh in-in formalism yields the classical Langevin equation including the fluctuation force which faithfully reflects the particle production property of the original quantum system. This formalism is applied to some transient process; the initiation of the spontaneous symmetry breaking, appearance of the off-diagonal long-range order in Bose-Einstein condensation, a transient process of the classicalization of the quantum fluctuations in the inflationary cosmology,... and gives some implications on the origin of the irreversibility associated with the transition from quantum to classical.'
author:
- Masahiro Morikawa
title: |
Transient Dynamics from Quantum to Classical\
- From the Developed Coherent State via Extreme Squeezing -
---
introduction
============
Quantum mechanics is no doubt a complete theory and correctly describes nature in the most fields in laboratory science. On the other hand, in some situations when the operational control doesn’t work, quantum mechanics seems to be lacking any autonomous description. This applies to the problem of the origin of inhomogeneity in the early Universe, the black hole information loss problem, and so on. All of them seem to be related to the emergence, in quantum mechanics, of any classical signal which spontaneously violates the original symmetry of the system.
We pickup three typical examples.
i\) In the early Universe, the standard cosmology indicates that the ultimate origin of the spatially non-uniform density fluctuations $\delta\rho(t,\boldsymbol{x})$ is the quantum fluctuations of the inflaton field $\delta\hat{\phi}(t,\boldsymbol{x})$ in the inflationary stage when the Universe transforms from microscopic to macroscopic scales through the exponential expansion. There is no explicit observer nor detector in the early Universe and therefore quantum mechanics cannot yield any definite outcome of the quantum fluctuations. In order to solve this problem, some authors introduce detectors [@Nambu2009][@Matsumura2018]or modify the basic laws of quantum mechanics so that the wave function spontaneously collapses [@Simonov2016]. This situation is similar to the quantum-information-theory aspect of the black hole [@Hawking1976]. If one tries to describe the black hole and the Hawking radiation as a unitary evolution, one encounters the violation of the equivalence principle[@Almheiri2013].
ii\) A presice description of the spontaneous symmetry breaking (SSB) introduces the two limiting operations, large volume and small external field limits, which are carefully ordered. However, this is an explicit vilation of the symmetry that the violation direction is artificially determined from the begining. On the other hand, in the Bose-Einstein condensation(BEC), the classical order parameter $\varphi_{0}$ describes the amount of condensation of boson gas which accumulates into the lowest energy state. This is also the indicator how extent the phase symmetry is broken spontaneously. In the ordinary argument, this classical value is simply assumed from the beginning as simply dividing the total quantum field $\hat{\Phi}(t,\boldsymbol{x})=\hat{\phi}(t,\boldsymbol{x})+\varphi_{0}$ [@Lieb2007]. However, the trasient time development from 0 to $\varphi_{0}$ is actually needed to describe the initiation of BEC process.
iii\) Furthermore, we still do not have a full successful model of the quantum measurement apparatus, in which a definite classical signal $m(t)$ emerges that has a firm correlation with the further measurement of the same quantum system[@Leggett2005]. If we adopt the standard operational prescription of quantum mechanics, the measurement can be fully described by POVM. However, from the fundamental level, the measurement apparatus should be described by quantum mechanics as well as the system. In this case, the classical signal can have many values before the completion of the measurement. However, after the measurement it must choose one value among such possible values. This process is not simply the SSB above but must further include the back reaction process to the system[@Morikawa2006].
All the above examples have three main characteristics, beside the appearance of the macroscopic classical signal.
1\) *state selection*: The system chooses one state among many other possible states. If this all states can be transformed with each other by a local symmetry operation, then this is exactly the process of the spontaneous symmetry breaking.
2\) *classical statistical probability*: At this stage, deterministic evolution is terminated but an intrinsic probability appears which governs the fate of the system.
3\) *arrow of time*: This process is intrinsically irreversible. The system can evolve into any one state among many leaving a classical signal. However, this signal cannot be canceled and the system cannot return to the original state which allows another possibility to develop.
In order to describe the transient process of the development from quantum to classical yielding the classical signal, we follow the following two steps in this paper.
a\) *squeezed state*: If the quantum state is unstable, say by the negative mass squared, the system tends to evolve into the squeezed state. This state allows classical statistical description, in the effective action method, which is well separated from the unitary dynamical evolution. However, at the free level without any interaction, this statistical fluctuation, being non-dissipative, never appears outside. At this stage, everything is still reversible and unitary whatever the squeezing is strong. Furthermore the original symmetry, if any, is not yet broken. The statistical fluctuations, all superposed, are always symmetric.
b\) *coherent state*: If the above free system couples to the other state or the non linear interaction appears, then the above statistical fluctuations do affect the full system. If this coherent state develops either by the strong statistical fluctuations or by strong instability of the potential, the classical portion of the coherent state dominates the quantum portion. Thus the classical signal appears possibly with dissipation.
In this way, the classicality thus finally obtained in b) is a quantitative notion; there is a continuous transition from quantum to classical. On the other hand, the classicality of the noise in a) is a qualitative notion; all the fluctuations are classical.
We follow these two steps in the subsequent sections. In section 2, we study the appearance of the squeezed state in the unstable system and show that this squeezing process can be identified as the particle production which can be interpreted as the classical statistical fluctuations withoiut dissipation. In section 3, we consider the non-linearity of the system in which this fluctuation triggers the development of the coherent state. In section 4, we motivate the introduction of the Schwinger-Keldysh in-in formalism starting from the popular Langevin equation. In section 5, we introduce the in-in effective action method to describe the above two processes, a) and b), at once. In section 6, we briefly describe some relevant examples which are directly related with our formalism. In the last section 7, we summarize our study.
Squeezing - non dissipative dry noise
=====================================
An instability or the time-dependent classical source yield the squeezed state in general. We first consider a simple model which yields the squeezed state, the inverted harmonic oscillator[@Albrecht1994]. The Hamiltonian is given by
$$\hat{H}=\frac{1}{2m}\hat{p}^{2}-\frac{m\omega^{2}}{2}\hat{q}^{2}=i\frac{\mathrm{\omega\hslash}}{2}\left(\hat{a}^{2}e^{-2i\phi}-h.c.\right),\label{eq:HinvertedHO}$$
where $$\begin{array}{cc}
\hat{a}^{\dagger}= & (\frac{m\omega}{2\hbar})^{1/2}(q-i\frac{p}{m\omega})\\
\hat{a}= & (\frac{m\omega}{2\hbar})^{1/2}(q+i\frac{p}{m\omega})
\end{array},$$ and $\phi=-\pi/4.$ The ordinary cross term $\hat{a}^{\dagger}\hat{a}$ disappears because the second term in the middle of Eq.(\[eq:HinvertedHO\]) is negative. Then the wave function at time $t$ becomes
$$\begin{aligned}
\left.|\Psi\left(t\right)\right\rangle & =\exp\left[\frac{\omega t}{2}\left(\hat{a}^{2}e^{-2i\phi}-h.c.\right)\right]\left.|0\right\rangle \equiv S(t)\left.|0\right\rangle \\
& =\exp\left[-\frac{\omega t}{2}\left(\hat{a}^{\dagger}\right)^{2}e^{2i\phi}-\left(\frac{\omega t}{2}\right)^{2}\right]\left.|0\right\rangle .\end{aligned}$$
Therefore, in the state $\left.|\Psi\left(t\right)\right\rangle $, the particle pair is ’condensed’. The operator $S(t)$ defines the Bogolubov transformation from the canonical pair ${a,a^{\dagger}}$ to the new pair :
$$\left\{ \begin{array}{l}
b=S^{\dagger}aS=\hat{a}\cosh\omega t-\hat{a}^{\dagger}e^{2i\phi}\sinh\omega t,\\
b^{\dagger}=S^{\dagger}a^{\dagger}S=\hat{a}^{\dagger}\cosh\omega t-\hat{a}e^{-2i\phi}\sinh\omega t,
\end{array}\right.\label{eq:Bogolubov1}$$
and $SS^{\dagger}=S^{\dagger}S=1$. This state $\left.|\Psi\left(t\right)\right\rangle $ is unlimitedly squeezed in time toward the direction $\phi=-\pi/4$ in phase space as,
$$\left\langle \Psi\left(t\right)|\right.\left(\hat{p}\cos\phi\pm\hat{q}\sin\phi\right)\left.|\Psi\left(t\right)\right\rangle =\left\{ \begin{array}{l}
4e^{-2t}\\
4e^{2t}
\end{array}\right..$$
This state is often regarded as the particle-creating state as $$N\equiv\left\langle \Psi\left(t\right)|\right.a^{\dagger}a\left.|\Psi\left(t\right)\right\rangle =\left\langle 0|\right.b^{\dagger}b\left.|0\right\rangle =(\sinh\omega t)^{2}.$$ However, $\left.|\Psi\left(t\right)\right\rangle $ is simply a quantum mechanical state and the particles claimed to be created are clearly not the classical object before any observation. Actually, the state can be reversible to the original state $\left.|0\right\rangle $ by the operation $S^{\dagger}$. This means that this inverted harmonic oscillator, if prepared in the symmetric state initially, never fall to some particular direction. The state is always in the symmetric neutral position $\left\langle \Psi\left(t\right)|\right.\hat{x}\left.|\Psi\left(t\right)\right\rangle =0$ even if the quantum fluctuations develop infinitely. On the other hand, since the kinetic energy $K$ and the negative of the potential energy $-V$ both ever increases, the action $A=K-V$ of the system would diverge. This may indicate some classical property of this system $A\gg\hbar$ [@Albrecht1994]. Since all the one point function vanishes, unique property of the state appears in the two point functions: $$\begin{array}{cc}
\left\langle \left[x\left(t\right),x\left(t'\right)\right]\right\rangle & =\frac{i}{2}\left(\frac{m\omega}{2\hbar}\right)^{-1}\sinh\left(\omega\left(t-t'\right)\right),\\
\left\langle \left\{ x\left(t\right),x\left(t'\right)\right\} \right\rangle & =\frac{1}{2}\left(\frac{m\omega}{2\hbar}\right)^{-1}\cosh\left(\omega\left(t+t'\right)\right),
\end{array}\label{eq:2ptfunctions}$$ where the former is locally ($t\approx t'$) normal but the latter abnormally diverge in time.
More generally in quantum field theory, the Hamiltonian Eq.(\[eq:HinvertedHO\]) is the infinite collection of the harmonic oscillator labeled by the three momentum $\mathbf{k}.$ Since the momentum is conserved, the pair in the squeezed state must have exactly the opposite momentum. Thus the Bogolubov transformation Eq.(\[eq:Bogolubov1\]) should now be the form, $$\left\{ \begin{array}{l}
b_{\mathbf{k}}=\alpha_{\mathbf{k}}^{*}\hat{a}_{\mathbf{k}}-\beta\hat{a}_{\mathbf{-k}}^{\dagger},\\
b_{\mathbf{-k}}^{\dagger}=\alpha_{\mathbf{k}}\hat{a}_{-\mathbf{k}}^{\dagger}-\beta_{\mathbf{k}}^{*}\hat{a}_{\mathbf{k}}
\end{array}\right..\label{eq:Bogolubov2}$$ Therefore the particle pair of momentum $\mathbf{k},\mathbf{-k}$ is entangled with each other. Many cases of the particle production, Unruh effect, accelerated mirror, Hawking radiation from the black hole,...[@Birrell1982] correspond to this type of Bogolubov transformation and the states are slightly generalized squeezed states. These are simply the pure quantum states represented by the wave functions. These states are also similar to the entangled spin pair of Silver atoms in the SternGerlach experiments [@Gerlach1922]. All the above quantum system, even under the external force, is free and therefore keeps quantum coherence, and is even reversible, until observed or disturbed by the quantum interaction.
Coherent state -development of classicality
===========================================
We have introduced the squeezed state in the previous section in relation with quantum-classical transition. We have an another popular state, the coherent state, which is worth consideration in our context. The coherent state is define to be the eigenstate of the annihilation operator, $$\hat{a}\left|\alpha\right\rangle =\alpha\left|\alpha\right\rangle$$ which has an explicit form, $$\begin{aligned}
\left|\alpha\right\rangle & =e^{\alpha\hat{a}^{\dagger}-\alpha^{*}\hat{a}}\left|0\right\rangle \\
& \equiv C\left(\alpha\right)\left|0\right\rangle \\
& =e^{-\frac{\left|\alpha\right|^{2}}{2}}e^{\alpha\hat{a}^{\dagger}}\left|0\right\rangle .\end{aligned}$$ The operator $C\left(\alpha\right)$ shifts the creation and annihilation operators by C-number $$\left\{ \begin{array}{l}
\hat{b}=C^{\dagger}\hat{a}C=\hat{a}+\alpha,\\
\hat{b}^{\dagger}=C^{\dagger}\hat{a}^{\dagger}C=\hat{a}^{\dagger}+\alpha^{*}.
\end{array}\right.\label{eq:Ceffect}$$
The particle number expectation is $$N=\left\langle \alpha\right|\hat{a}^{\dagger}\hat{a}\left|\alpha\right\rangle =\left|\alpha\right|^{2}$$ and the two developed coherent states, *i.e.* individually large $\alpha,\beta$, have almost no superposition $$\left|\left\langle \alpha\right.|\left.\beta\right\rangle \right|^{2}=e^{-\left|\alpha-\beta\right|^{2}},$$ in particular, for the infinite system, the volume factor $V$ enhances this property. $$\propto e^{-V\left|\alpha-\beta\right|^{2}}.$$
The coherent state is not special but is generally produced by the time dependent single variable interaction $\xi\left(t\right)\widehat{x}$, where $\xi\left(t\right)$ is a classical external variable. This is contrasted with the squeezed state which is generally produced by the time dependent square variable interaction $m\left(t\right)\widehat{x}{}^{2}$. For example, let us consider the quantum system is exerted by the classical field $\xi(t)$, $$\ddot{\widehat{x}}(t)=-\gamma\dot{\widehat{x}}(t)-V'(\widehat{x}(t))+\xi(t).\label{eq:Langevinquantum}$$ Then the general solution would be $$\widehat{x}(t)=\widehat{x}(0)+\int^{t}dt'\triangle\left(t-t'\right)\xi\left(t'\right)\label{eq:qu+cl}$$ where the second term on the RHS, including any appropriate green function $\triangle\left(t-t'\right)$, represents the accumulation of the classical field. Thus, in this case, $$\widehat{x}(t)=C\left(t\right)^{\dagger}\widehat{x}\left(0\right)C\left(t\right)$$ where $$C\left(t\right)=\exp\left[\int^{t}dt'\triangle\left(t-t'\right)\xi\left(t'\right)\widehat{x}\left(0\right)\right].$$ Thus the time evolution induced by the single variable interaction $\xi\left(t\right)\widehat{x}$ yields a coherent state. If the classical part gradually dominates in Eq.(\[eq:qu+cl\]) due to the accumulation of the external force $\xi\left(t\right)$, then the state gradually becomes more classical.
It would be useful to comment that the external classical field $\xi\left(t\right)$ needs not to be a systematic force, but can be random with zero-mean. Even in this case, the classical component increases in time and develop classicality.
Langevin to in-in formalism - a motivation
==========================================
In order to consider the transient process from quantum to classical, let us consider a simple example of the Langevin equation first. This gives us a natural motivation to the Schwinger-Keldysh in-in closed path time quantum theory (CTP)[@Kita2010]. Starting from the classical Langevin equation, we derive the effective partition function. The quantization of this partition function almost deduces the CTP formalism in quantum mechanics.
The Langevin equation is a dynamical equation of a particle in the environment with the potential force $-V'$ and the random force $\xi$ with friction $\gamma$ : $$\ddot{x}(t)=-\gamma\dot{x}(t)-V'(x(t))+\xi(t).\label{eq:Langevinoriginal-1}$$ The statistical average $$\left\langle ...\right\rangle _{\xi}=\int D[\xi]...P[\xi]$$ is determined by the Gaussian weight functional $P[\xi]$, $$P[\xi]=e^{-\int\xi(t)^{2}/(2\sigma^{2})}.\label{eq:Pgiven}$$ We would like to know the action which drives this Langevin equation. We first try to construct the partition function of the system. The partition function is derived by summing all the possible motions in the whole phase space.
$$\begin{aligned}
Z[J] & \equiv & \left\langle \delta[\ddot{x}(t)+\gamma\dot{x}(t)+V'(x(t))-\xi(t)]\right\rangle _{\xi}\\
& = & \int D[\xi]P[\xi]\delta[\ddot{x}(t)+\gamma\dot{x}(t)+V'(x(t))-\xi(t)]\nonumber \\
& = & \int D[\xi]D[x']P[\xi]e^{i\int dtx'(t)\{\ddot{x}(t)+\gamma\dot{x}(t)+V'(x(t))-\xi(t)\}}\nonumber \\
& = & \int D[x']e^{i\tilde{S}[x,x']}\nonumber \end{aligned}$$
where the integral form of the delta functional is utilized introducing a fictitious variable $x'\left(t\right)$, and the action becomes $$\tilde{S}[x,x']\equiv\int dt\{-\dot{x}'(t)\dot{x}(t)+\gamma x'(t)\dot{x}(t)+x'(t)V'(x(t))+\frac{i}{2}\sigma^{2}x'(t)^{2}\}\label{eq:Langevin action}$$ where the boundary term is dropped. Note that the last term, which represents classical statistical fluctuations, is pure imaginary in the action. The rest of the terms describe the deterministic dynamics though including frictional term.
We now quantize this system because the most basic theory would be the quantum mechanics, from which classical dynamics eventually appear. The partition function for the quantized system simply becomes
$$Z[J]=\int D[x]D[x']e^{i\tilde{S}[x,x']}\label{eq:PI Langevin}$$
in the path integral form. The mixed expression of the two variables $x\left(t\right)$and $x'\left(t\right)$ in Eq.(\[eq:Langevin action\]) is dowdy. It is possible to rewrite the action more resemble to the ordinary dynamics. In order to do so, we rewrite the variables as $$x_{\pm}=x\pm\frac{1}{2}x'.$$ Then, Eq. (\[eq:PI Langevin\]) becomes $$Z[J]=\int D[x_{+}]D[x_{-}]e^{i\tilde{S}[x_{+},x_{-}]}$$ where $$\tilde{S}[x_{+},x_{-}]=\int dt\left\{ \begin{array}{cc}
\left(\left(\ddot{x}_{+}\left(t\right)\right)^{2}-V(x_{+}(t))\right)-\left(\left(\ddot{x}_{-}\left(t\right)\right)^{2}-V(x_{-}(t))\right)\\
+\frac{\gamma}{2}\left(x_{+}(t)\dot{x}_{-}(t)-\dot{x}_{+}(t)x_{-}(t)\right)+\frac{i}{2}\sigma^{2}\left(x_{+}(t)-x_{-}(t)\right)^{2}
\end{array}\right\} .\label{eq:Sderived}$$ The first line of the above represents the deterministic dynamics for the variables $x_{\pm}\left(t\right)$ separately, and the second line dissipation and fluctuation terms where the variables $x_{\pm}\left(t\right)$ mix up. It is a natural extension of this expression to introduce a closed time-path $C$: which starts from $-\infty$ to $\infty$ ($+$ branch) and then comes back from $\infty$ to $-\infty$ ($-$ branch). We suppose the supports of the variables $x_{\pm}\left(t\right)$ are, respectively, the $+$ and $-$ branches. We denote the variable $\tilde{x}\left(t\right)$ on the countour $C$ unifying the variables $x_{\pm}\left(t\right)$ : $$\tilde{x}\left(t\right)=\begin{cases}
x_{+}(t) & t\in(+\mathrm{branch})\\
x_{-}(t) & t\in(-\mathrm{branch}).
\end{cases}$$ We use this unification for all the variables and denote them by tilde.
It is possible to reverse the above logic to get to the action $\tilde{S}[x_{+},x_{-}]$; starting from the action $\tilde{S}[x_{+},x_{-}]$ to get to the Langevin equation. If we find a complex action including an extra degrees of freedom like $x'$ in the above, classical random field appears and the evolution equation becomes the Langevin equation. We study a typical case in the next section.
CTP to Langevin
===============
The best way to describe the transient process from quantum to classical would be the Schwinger-Keldysh in-in formalism[@Kita2010]. In this theory, the partition function for the system with the free action
$$S[x]=\int dt(\dot{x}^{2}+\omega^{2}x^{2}),$$ is given by
$$\begin{array}{cc}
\tilde{Z}[\tilde{J}] & =\int_{C}D\tilde{x}\exp[iS[\tilde{x}]+i\int dt\tilde{J}(t)\tilde{x}(t)]\equiv\exp i\tilde{W}\\
= & \int D\tilde{x}\exp[iS[x_{+}]-iS[x_{-}]+i\int dt\tilde{J}(t)\tilde{x}(t)],
\end{array}$$ where the tilde denotes the variables on the closed-path $C$ as before. This reduces to $$\tilde{Z}[\tilde{J}]=\exp[-\frac{1}{2}\int dt\tilde{J}(t)\tilde{G}_{0}(t,t')\tilde{J}(t')].\label{eq:Zfree}$$ where $$\tilde{G}_{0}(t,t')=\left(\begin{array}{ll}
G_{F}(t,t') & G_{+}(t,t')\\
G_{-}(t,t') & G_{\overline{F}}(t,t')
\end{array}\right)\equiv\left(\begin{array}{ll}
\mathrm{Tr}\left[Tx(t)x(t')\rho\right] & \mathrm{Tr}\left[x(t')x(t)\rho\right]\\
\mathrm{Tr}\left[x(t)x(t')\rho\right] & \mathrm{Tr}\left[\overline{T}x(t)x(t')\rho\right]
\end{array}\right)$$ where $T$ denotes the ordinary time-ordering operator and $\overline{T}$ the anti time-ordering operator. If we change the representation of the matrix by $$J_{\pm}(t)=J_{c}\pm\frac{1}{2}J_{\Delta}$$ or $$\tilde{J}=\left(\begin{array}{cc}
J_{\Delta}\\
J_{C}
\end{array}\right)=\left(\begin{array}{cc}
1 & -1\\
\frac{1}{2} & \frac{1}{2}
\end{array}\right)\left(\begin{array}{cc}
J_{+}\\
J_{-}
\end{array}\right),$$ then, we have $$\tilde{G}_{0}(t,t')=\left(\begin{array}{ll}
0 & G_{R}(t,t')\\
G_{A}(t,t') & G_{C}(t,t')
\end{array}\right)\equiv\left(\begin{array}{ll}
0 & \theta(t-t')\mathrm{Tr}[x(t'),x(t)]\rho\\
\theta(t'-t)\mathrm{Tr}[x(t'),x(t)]\rho & \mathrm{Tr}\{x(t),x(t')\}\rho
\end{array}\right).$$
In our case, $$\begin{array}{cc}
G_{R}(t,t') & =\theta(t-t')\left\langle \left[x\left(t\right),x\left(t'\right)\right]\right\rangle \\
& =\frac{i}{2}\left(\frac{m\omega}{2\hbar}\right)^{-1}\sin\left(2\phi\right)\theta(t-t')\sinh\left(\omega\left(t-t'\right)\right)\\
& =\frac{i}{2}\left(\frac{m\omega}{2\hbar}\right)^{-1}\theta(t-t')\sinh\left(\omega\left(t-t'\right)\right),
\end{array}\label{eq:Gr}$$ $$\begin{array}{cc}
G_{C}(t,t') & =\left\langle \left\{ x\left(t\right),x\left(t'\right)\right\} \right\rangle \\
& =\frac{1}{2}\left(\frac{m\omega}{2\hbar}\right)^{-1}\left(\cosh\left(\omega\left(t+t'\right)\right)-\cos\left(2\phi\right)\sinh\left(\omega\left(t+t'\right)\right)\right)\\
& =\frac{1}{2}\left(\frac{m\omega}{2\hbar}\right)^{-1}\cosh\left(\omega\left(t+t'\right)\right),
\end{array}\label{eq:Gc}$$ where $\phi=-\pi/4$.
These two types of green functions sometimes show different infrared behavior. Actually, for low frequency $\omega\left(t-t'\right)\ll1$ and $\omega\left(t+t'\right)\ll1$, $$\begin{array}{cc}
G_{R}(t,t')\propto\sinh\left(\omega\left(t-t'\right)\right)\rightarrow\omega\left(t-t'\right),\\
G_{C}(t,t')\propto\cosh\left(\omega\left(t+t'\right)\right)\rightarrow1.
\end{array}$$ $G_{R}(t,t')$ has milder IR behavior than $G_{C}(t,t')$ by the factor $\omega$. However, as we will see shortly, $G_{C}(t,t')$ can be separated from the action which describes the deterministic dynamics. In particular, this situation becomes prominent at the inflationary stage in the early Universe[@Morikawa2016].
Comparing Eqs.(\[eq:Gr\],\[eq:Gc\]), the symmetric term in Eq.(\[eq:Zfree\]) $$-\frac{1}{2}\int dtJ_{\Delta}(t)G_{C}(t,t')J_{\Delta}(t')$$ is real and positive. Therefore we can factor out this part as classical statistical fluctuations introducing an auxiliary field $\xi\left(t\right),$ $$\tilde{Z}[\tilde{J}]=\int D\xi P\left(\xi\right)\exp[-\frac{1}{2}\int dt\tilde{J}(t)\tilde{G}'_{0}(t,t')\tilde{J}(t')+i\int dtJ_{\Delta}(t)\xi(t)],\label{eq:Zderived}$$ where $$P\left(\xi\right)=\exp[-\frac{1}{2}\int dt\xi(t)G_{C}(t,t')\xi(t')],\label{eq:Pderived}$$ and $\tilde{G}'_{0}(t,t')$ is thus separated green function. This separation procedure is just a reverse of the previous section where Eq.(\[eq:Sderived\]) yields Eq.(\[eq:Pgiven\]).
As in the above, we have identified the classical statistical fluctuation, *i.e.* noise, in Eqs.(\[eq:Zderived\], \[eq:Pderived\]). However, this noise only couples to the field $J_{\Delta}(t)$, in Eq.(\[eq:Zderived\]), which is an external source term but not any dynamical variable. Therefore this noise never comes out as is. Nothing happens in free state before any measurement process or interaction according to the quantum mechanics theory. Furthermore, the noise here is dry and is not accompanied by dissipation.
This situation drastically changes if we introduce the interactions. Let us introduce the effective action, in the in-in formalism, which is the Legendre transformation of the partition function $\tilde{Z}[\tilde{J}]=\exp i[\tilde{W}[\tilde{J}]]$, $$\exp[i\tilde{\Gamma}[\tilde{X}]]=\exp i[\tilde{W}[\tilde{J}]-\int dt\tilde{J}(t)\tilde{X}(t)].$$ This becomes $$\begin{aligned}
\exp[i\tilde{\Gamma}[\tilde{X}]] & = & \exp i[\tilde{W}[\tilde{J}]-\int dt\tilde{J}(t)\tilde{X}(t)]\nonumber \\
& = & \int_{C}D\tilde{x}\exp i[\tilde{S}[x]+\int d^{4}x\tilde{J}(x)(\tilde{x}(t)-\tilde{X}(t))]\nonumber \\
& = & \int_{C}D\tilde{x}\exp i[\tilde{S}[\tilde{X}+\tilde{x}]+\int dt\tilde{J}(t)\tilde{X}(t)],\end{aligned}$$ where we shifted the path-integration variable. Expanding $\tilde{S}[\tilde{X}+\tilde{x}]$ in the series of $\tilde{x}$, we have
$$\tilde{S}[\tilde{X}+\tilde{x}]=\tilde{S}\left[\tilde{X}\right]+\tilde{S}^{\prime}\left[\tilde{X}\right]\tilde{x}+\frac{1}{2}\tilde{S}''\left[\tilde{X}\right]\tilde{x}^{2}+\frac{1}{3!}\tilde{S}'''\left[\tilde{X}\right]\tilde{x}^{3}+\ldots.\label{eq:Sseries}$$
The first term represents the action for the classical field $\tilde{X}\left(t\right),$the second term and the third term yield interaction for the quantum variable, and the second term a free action for $\tilde{x}$ in the background of $\tilde{X}$.
If the interaction is quartic, $\lambda x\left(t\right)^{4},$ and the background $\tilde{X}=0$ initially, then the term $\frac{1}{3!}\tilde{S}'''\left[\tilde{X}\right]\tilde{x}^{3}$ gives the dominant interaction term $\lambda\tilde{X}\left(t\right)\tilde{x}\left(t\right)^{3}$. The lowest contribution of this term yields the two-loop quantum effect for $\tilde{X}\left(t\right)$ $$\begin{aligned}
& \int dtdt'\lambda^{2}\tilde{X}\left(t\right)\mathrm{Tr}\left[T_{C}\rho\tilde{x}\left(t\right)^{3}\tilde{x}\left(t'\right)^{3}\right]\tilde{X}\left(t'\right)\\
= & \int dtdt'\lambda^{2}\tilde{X}\left(t\right)\mathrm{Tr}\left[T_{C}\rho\tilde{x}\left(t\right)\tilde{x}\left(t'\right)\right]^{3}\tilde{X}\left(t'\right)\\
= & \int dtdt'\lambda^{2}\tilde{X}\left(t\right)G_{C}\left(t,t'\right)^{3}\tilde{X}\left(t'\right)\\
= & \int dtdt'\lambda^{2}\left(X_{C},X_{\Delta}\right)_{t}\left(\begin{array}{cc}
0 & G_{C}^{2}\text{\ensuremath{G_{A}}}\\
\text{\ensuremath{G_{R}}}\text{\ensuremath{G_{C}}}^{2} & \text{\ensuremath{G_{C}}}^{3}
\end{array}\right)_{t,t'}\left(\begin{array}{cc}
& X_{C}\\
& X_{\Delta}
\end{array}\right)_{t'}\end{aligned}$$ Therefore, the imaginary term $X_{\Delta}\left(t\right)G_{C}\left(t,t'\right)^{3}X_{\Delta}\left(t\right)$ contributes to the fluctuation as before and can be separated from the real part of the action by the introduction of the auxiliary field $\xi\left(t\right)$,
We have, in the lowest order, $$\exp[i\Gamma[X]]=\int D\xi P\left(\xi\right)\exp[iS_{\mathrm{eff}}[X]],$$ where the real action is $$\begin{array}{cc}
S_{\mathrm{eff}}[X]= & S[X]+\int\int dtdt'X_{C}(t)\left(1+\lambda^{2}G_{C}\left(t,t'\right)^{2}\right)G_{A}\left(t,t'\right)X_{\Delta}(t')\\
& +\int\int dtdt'X_{\Delta}(t)G_{R}\left(t,t'\right)\left(1+\lambda^{2}G_{C}\left(t,t'\right)^{2}\right)X_{C}(t')+\int dt\xi(t)X_{\Delta}(t)],
\end{array}$$ and the fluctuation weight is given by $$P\left(\xi\right)=\exp[-\frac{\lambda^{2}}{2}\int dt\xi(t)G_{C}\left(t,t'\right)^{3}\xi(t')].\label{eq:P3}$$ Note that the advanced term $\int\int dtdt'X_{C}(t)\left(1+\lambda^{2}G_{C}\left(t,t'\right)^{2}\right)G_{A}\left(t,t'\right)X_{\Delta}(t')$ yields the same retarded term if we exchange the variables $t\leftrightarrow t'$.
Now the application of the variational principle for $S_{\mathrm{eff}}[X]$ $$\frac{\delta S_{\mathrm{eff}}[X]}{X_{\Delta}(t)}|_{X_{\Delta}=0}=0,$$ yields the classical Langevin equation as $$\ddot{X}_{C}(t)-\omega^{2}X_{C}(t)+2\int dt'G_{R}\left(t,t'\right)\left(1+\lambda^{2}G_{C}\left(t,t'\right)^{2}\right)X_{C}(t')+\xi(t)=0.\label{eq:generalLangevin}$$
This equation describes the rapid evolution of the classical variable $X_{C}(t)$ under a) the strong fluctuations $\xi(t)$ with Eqs.(\[eq:P3\],\[eq:Gc\]) violently disturb the system, and b) the original classical instability $-\omega^{2}X_{C}(t)$ which promotes the exponential development of the system. The retarded term $2\int dt'G_{R}\left(t,t'\right)\left(1+\lambda^{2}G_{C}\left(t,t'\right)^{2}\right)X_{C}(t')$ sometimes shows the dissipative effects.
The above equation (\[eq:generalLangevin\]) describes the evolution of the classical variable $X_{C}(t)$ from 0, the symmetric state, to a finite value, the asymmetric state. Eventually in this evolution, other interaction terms in Eq.(\[eq:Sseries\]) gradually contribute. Some of them yield the new type of noise. Individual imaginary term seems to yield individual random noise. $$\exp[-X_{\Delta}G_{C}X_{\Delta}]=\int D\xi_{1}\exp[-\xi_{1}G_{C}^{-1}\xi_{1}+i\xi_{1}X_{\Delta}],$$ and $$\exp[-X_{\Delta}G_{C}^{2}X_{\Delta}]=\int D\xi_{2}\exp[-\xi_{2}G_{C}^{-2}\xi_{2}+i\xi_{2}X_{\Delta}].$$ However, all the perturbation terms have interference with each other and therefore should be treated at once: $$\exp[-\lambda^{2}X_{\Delta}\left(G_{C}+G_{C}^{2}+G_{C}^{3}\right)X_{\Delta}]=\int D\xi\exp[-\xi\left[\lambda^{2}\left(G_{C}+G_{C}^{2}+G_{C}^{3}\right)\right]^{-1}\xi+i\xi X_{\Delta}].$$ Since the fluctuation kernel $\lambda^{2}X_{\Delta}\left(G_{C}+G_{C}^{2}+G_{C}^{3}\right)X_{\Delta}$ is unique, the random field $\xi(t)$ is also unique; random fields do not appear separately.
Some Applications and Implications
==================================
Our argument is general and will have many applications and implications. Some of them are very briefly described below. Individual argument will be reported separately.
Transient dynamics of the spontaneous symmetry breaking and the Bose-Einstein condensation
-------------------------------------------------------------------------------------------
A standard method to describe the Spontaneous symmetry breaking (SSB) needs an infinitesimal explicit violation of the symmetry with the delicate order of the two limiting operations. For example in the case of ferromagnetic materials, the full order parameter is given by $$M_{\pm}\equiv\lim_{B\rightarrow0\pm}\lim_{V\rightarrow\infty}m_{V}\left(B\right),\label{eq:doublelimit}$$ where $m_{V}\left(B\right)$ is the local average of the magnetization[@Lieb2007]. In the case of Bose-Einstein condensation (BEC), the argument starts from the assumption that the boson field can be separated as [@Pitaevskii2003] $$\hat{\phi}=\left\langle \hat{\phi}\right\rangle +\hat{\delta\phi},\label{eq:assumedseparation}$$ where the classical parameter $\left\langle \hat{\phi}\right\rangle $ represents the 0-momentum condensation.
On the other hand, based on our approach, general SSB can be described as follows. We suppose the unstable potential for the complex scalar field $\phi\left(x\right)$ such as $$V\left(\phi\right)=\frac{1}{2}m^{2}\phi^{2}+\frac{1}{4!}\lambda\phi^{4}$$ with $m^{2}<0$ and $\lambda>0$. The initial tachyonic instability around $\phi\approx0$ induces the squeezed state and thus dry noise. Through the interaction, this noise autonomously violates the $U\left(1\right)$ symmetry and allow the development of the classical degrees of freedom as a coherent state. Therefore, we do not need Eq(\[eq:doublelimit\]) and SSB is genuinely spontaneous. Furthermore, the artificial separation is not needed as in Eq.(\[eq:assumedseparation\]) and the field $\left\langle \hat{\phi}\right\rangle $ does emerge as a developed coherent state.
In the case of BEC, the condensation is not only characterized by $$\frac{\left\langle \hat{a}_{0}^{*}\hat{a}_{0}\right\rangle }{V}>0,$$ but also needs the condition[@Lieb2007] $$\frac{\left|\left\langle \hat{a}_{0}\right\rangle \right|^{2}}{V}>0.$$ which guarantees the off-diagonal long-range order or the fact that the BEC as a phase transition accompanying the spontaneous symmetry breaking. In BEC, the Gross-Pitaevskii equation is generally given as Eq.(\[eq:generalLangevin\]). The fluctuation term disappears when the system moves out from the unstable transient region ($\left|\left\langle \hat{\phi}\right\rangle \right|^{2}>-2m^{2}/\lambda$).
Transient dynamics of the quantum measurement
---------------------------------------------
The appearance of the classicality is deeply related with the quantum measurement process in which a particular state is probabilistically selected among multiple possibilities. The situation is very similar to SSB above, however, there must be a back-reaction to the quantum system from the emerged classical degrees of freedom. This back-reaction guarantees the firm correlation between the quantum system and the measurement apparatus.
The prototype of the quantum measurement model has been analyzed in this line of thought introducing the external thermal bath in[@Morikawa2006]. This model describes the transient dynamics of the detector field $\hat{\phi}$ measures the spin $\hat{\mathrm{S}}.$ The Lagrangian is given by $$L=\frac{1}{2}\left(\nabla\hat{\phi}\right)^{2}-\frac{1}{2}m^{2}\hat{\phi}^{2}-\frac{1}{4!}\lambda\hat{\phi}^{4}+\mu\hat{\phi}\boldsymbol{\mathrm{\hat{S}B}}+(bath).$$ From the present approach, the detector should be described by $X_{C}(t)$ in Eq.(\[eq:generalLangevin\]). On top of this dynamics, the back reaction of it to the spin $\hat{\mathrm{S}}$ is needed. The thermal bath degrees of freedom will not be needed. The fluctuation would be provided by the dry noise associated to the initial squeezed state triggered by the unstable potential ($m^{2}<0$ and $\mu>0$).
Transient dynamics which shows macroscopic irreversibility
----------------------------------------------------------
A classical degrees of freedom as a developed coherent state has appeared after the time evolution by the Langevin equation. This process cannot be canceled and is irreversible. This is true even if the energy dissipation does not exist. Actually, the Brownian motion of the classical degrees of freedom is described by the Langevin equation with random fluctuations. The recursion probability, the system comes back to the original position, would be vanishingly small after the Brownian motion described by the Langevin equation.
In the careful argument on the appearance of the arrow of time in quantum mechanics [@Ordonez2017], the essence of the irreversibility is the appearance of the decaying and growing pair, as well as the natural boundary condition which picks up one from the pair. The first essence is similar to our tachyonic modes and the second part is automatically selected by the Schwinger-Keldysh in-in formalism. Actually we use it to put the retarded and advanced contribution together in the evolution equation(\[eq:generalLangevin\]). We have further discussed the appearance of the classical degrees of freedom.
Transient dynamics from quantum to classical fluctuations in the inflationary cosmology
---------------------------------------------------------------------------------------
In the cosmology, the hypothetical scalar field inflaton is supposed to cause the inflation in the early Universe and also to yield the ultimate seeds of density fluctuations [@Weinberg2008]. This inflaton field and the space-time metric mix together to yield the gauge invariant variables. We consider such scalar variable $v_{\mathbf{k}}\left(\eta\right)$ where $\mathbf{k}$is the three momentum and $\eta$$\left(\in\left(-\infty,0\right)\right)$ is the conformal time variable. The Hamiltonian for $v_{\mathbf{k}}\left(\eta\right)$ in the inflation (de Sitter space) becomes [@Albrecht1994], $$\begin{aligned}
H & =\frac{1}{2}\int\frac{d^{3}k}{(2\pi)^{3/2}}\left[\hat{p}_{-\mathbf{k}}\hat{p}_{\mathbf{k}}+k^{2}\hat{v}_{-\mathbf{k}}\hat{v}_{\mathbf{k}}+\frac{1}{\eta}\left(\hat{p}_{-\mathbf{k}}\hat{v}_{\mathbf{k}}+\hat{v}_{-\mathbf{k}}\hat{p}_{\mathbf{k}}\right)\right]\\
& =\frac{1}{2}\int\frac{d^{3}k}{(2\pi)^{3/2}}\left[k\left(\hat{a}_{\mathbf{k}}^{\dagger}\hat{a}_{\mathbf{k}}+\hat{a}_{-\mathbf{k}}^{\dagger}\hat{a}_{-\mathbf{k}}+1\right)+\frac{i}{\eta}\left(e^{-2i\vartheta}\hat{a}_{-\mathbf{k}}\hat{a}_{\mathbf{k}}-h.c.\right)\right]\label{eq:AlbrechtH}\\
\nonumber \end{aligned}$$ where $\vartheta=-\pi/2$ and the factor $1/\left|\eta\right|$ in this Eq.(\[eq:AlbrechtH\]) infinitely increases and yields strong squeezed state. However this is a special quantum state and does not directly mean the appearance of the classical fluctuations as we have already studied. Introducing the non-linear interactions of $v$ and constructing the effective action, we can decompose the action into the deterministic part and the stochastic part[@Morikawa2016], $$\begin{aligned}
\exp[i\widetilde{\Gamma}[\tilde{\varphi}]] & =\int{\cal D\xi}P(\xi)\exp[iS_{0}[\tilde{\varphi;\xi}]],\\
\exp[iS_{0}[\tilde{\varphi;\xi}]] & =\exp[iS_{0}[\overset{\text{\~{ }}}{\varphi}]]\int D\phi\exp i[(\lambda\varphi(x)^{3})_{\triangle}G_{R}(x-y)(\lambda\varphi(y)^{3})_{C}+\\
& (\lambda\varphi(x)^{3})_{C}G_{A}(x-y)(\lambda\varphi(y)^{3})_{\triangle}+i(\lambda\varphi(x)^{3})_{\triangle}G_{C}(x-y)(\lambda\varphi(y)^{3})_{\triangle}]\\
P(\xi) & =\exp[-\frac{1}{4}\int d^{3}k\xi(\overrightarrow{k})G_{C}(\overrightarrow{k})^{-1}\xi(\overrightarrow{k})].\end{aligned}$$ reflecting the dry noise generated by the squeezed state from Eq.(\[eq:AlbrechtH\]). The fluctuation kernel is given by
$$G_{C}(\overset{\rightarrow}{k})=\frac{H^{2}}{k^{3}}\left((1+k^{2}\eta\eta^{\prime})\cos(k\eta)+k\eta\sin(k\eta)\right)$$
where $H$ is the Hubble constant for the inflationary de Sitter space.
The Langevin equation for the field $$3H\dot{\varphi}{}_{k}+(\lambda/2)\varphi_{0}^{2}\varphi_{k}=(\lambda/2)\varphi_{0}^{2}\xi_{k}$$ yields the classical statistical power spectrum. $$\left\langle \varphi_{k}\varphi_{k}\right\rangle _{\xi}\thickapprox\lambda^{2}\varphi_{0}^{4}\frac{H^{2}}{k^{3}}.$$ This transient process from quantum to classical has been made possible both by the sqeezed state and the interaction of the inflaton field.
Conclusions
===========
We have explored the transient emergent process of the classical degrees of freedom in the full quantum system. In this process, the existence of the squeezing state and the development of the coherent state are both essential. The former squeezed state may be triggered by an instability and yields tachyonic mode. The latter coherent state develops in the Langevin dynamics triggered by the interactions. Thus both processes are indispensable.
We first considered the generation of the squeezed state in the inverted harmonic oscillator model. The degree of squeezing increases unboundedly in time. Though the action and the expectation value of the particle number increases and therefore one may tend to think that many particles are actually produced. However, this is simply a squeezed state fully described by the quantum theory and can never be interpreted as classical nor random statistical. We have shown that some non-dissipative noise, *i.e.* dry noise, is associated to this state using the Schwinger-Keldysh in-in formalism. At this stage, the noise never destroys the original symmetry if any. Moreover, the generation process of this state is reversible and the squeezing can be canceled. This is always true for free quantum system.
Then, introducing the interaction, we considered the development of the coherent state triggered by the above dry noise. This process is described by the Langevin equation. The degree of the coherent state increases, firstly by the ever accumulating dry noise and secondly by the unstable potential. This means that the state is dominated by the classical part which determines the finite vacuum expectation values.
Thus we have considered two types of general classicality in our argument. The first one is associated to the squeezed state and is represented by the real exponent in the in-in path-integral formalism. The second one is associated to the development of the coherent state and is represented by the development of the vacuum expectation values. This latter one is the objective classicality we wanted to obtain.
The appearance of the classicality is the key feature in many important transient processes in Physics. They are the dynamics of the general spontaneous symmetry breaking, the initiation process of the Bose-Einstein condensation, the transient process of the quantum measurement and the appearance of the classical signal in the apparatus, the transient process that the arrow of time appears in the macroscopic system, the transient process of the inflation in the early Universe where the quantum fluctuations form the classical seeds of galaxies, and so on.
The author would like to thank all the members of the Astrophysics group in Ochanomizu University for useful and critical comments as well as for inspiring thorough discussions.
[10]{} Y. Nambu and Y. Ohsumi, Phys.Rev.D**80**,124031 (2009); Phys.Rev.D**84**, 044028 (2011).
Akira Matsumura and Yasusada Nambu, Phys.Rev.D**98**, 025004 (2018).
K. Simonov, B.C. Hiesmayr, Phys.Rev.A**94**, 052128 (2016) .
S. W. Hawking, Phys. Rev. D**14**, 2460 (1976).
A. Almheiri, D. Marolf, J. Polchinski and J. Sully, JHEP**1302**, 062 (2013).
Elliott H. Lieb, Robert Seiringer, Jakob Yngvason, Rep. Math. Phys.**59**, 389-399 (2007).
A. J. Leggett, Science **307**-5711, 871-872 (2005).
Masahiro Morikawa and Akika Nakamichi, Prog. Theor. Phys.**116**, 2006, 679698. (2006).
Andreas Albrecht, Pedro Ferreira, Michael Joyce, Tomislav Prokopec, Phys.Rev.D**50**, 4807-4820 (1994).
N.D.Birrell and P.C.W.Davies, *Quantum Fields in Curved Space*, Cambridge Monographs on Mathematical Physics, Cambridge University Press (1982).
Gerlach, W.and Stern, O., Zeitschrift für Physik. **9**, 349352 (1922).
T. Kita, Prog. Theor. Phys. **123**, 581 (2010).
Masahiro Morikawa, arXiv:1604.01015 [\[]{}hep-th[\]]{} (2016).
L. P. Pitaevskii and S. Stringari, *Bose-Einstein Condensation* International Series of Monographs on Physics, Clarendon Pr (2003).
Gonzalo Ordonez, Naomichi Hatano, J. Phys. A: Math. Theor. **50,** 405304 (2017).
Steven Weinberg, *Cosmology*, Oxford Univ Pr (2008).
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.